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

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomW3W3_VF_avx_128_fma_single
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            None
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEwSh_VdwNone_GeomW3W3_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  90     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  92     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  93     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  94     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  95     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  96     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  97     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  98     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  99     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 102     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 105     real             *charge;
 106     __m128i          ewitab;
 107     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 108     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 109     real             *ewtab;
 110     __m128           dummy_mask,cutoff_mask;
 111     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 112     __m128           one     = _mm_set1_ps(1.0);
 113     __m128           two     = _mm_set1_ps(2.0);
 114     x                = xx[0];
 115     f                = ff[0];
 116
 117     nri              = nlist->nri;
 118     iinr             = nlist->iinr;
 119     jindex           = nlist->jindex;
 120     jjnr             = nlist->jjnr;
 121     shiftidx         = nlist->shift;
 122     gid              = nlist->gid;
 123     shiftvec         = fr->shift_vec[0];
 124     fshift           = fr->fshift[0];
 125     facel            = _mm_set1_ps(fr->epsfac);
 126     charge           = mdatoms->chargeA;
 127
 128     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 129     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 130     beta2            = _mm_mul_ps(beta,beta);
 131     beta3            = _mm_mul_ps(beta,beta2);
 132     ewtab            = fr->ic->tabq_coul_FDV0;
 133     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 134     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 135
 136     /* Setup water-specific parameters */
 137     inr              = nlist->iinr[0];
 138     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 139     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 140     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 141
 142     jq0              = _mm_set1_ps(charge[inr+0]);
 143     jq1              = _mm_set1_ps(charge[inr+1]);
 144     jq2              = _mm_set1_ps(charge[inr+2]);
 145     qq00             = _mm_mul_ps(iq0,jq0);
 146     qq01             = _mm_mul_ps(iq0,jq1);
 147     qq02             = _mm_mul_ps(iq0,jq2);
 148     qq10             = _mm_mul_ps(iq1,jq0);
 149     qq11             = _mm_mul_ps(iq1,jq1);
 150     qq12             = _mm_mul_ps(iq1,jq2);
 151     qq20             = _mm_mul_ps(iq2,jq0);
 152     qq21             = _mm_mul_ps(iq2,jq1);
 153     qq22             = _mm_mul_ps(iq2,jq2);
 154
 155     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 156     rcutoff_scalar   = fr->rcoulomb;
 157     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 158     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 159
 160     /* Avoid stupid compiler warnings */
 161     jnrA = jnrB = jnrC = jnrD = 0;
 162     j_coord_offsetA = 0;
 163     j_coord_offsetB = 0;
 164     j_coord_offsetC = 0;
 165     j_coord_offsetD = 0;
 166
 167     outeriter        = 0;
 168     inneriter        = 0;
 169
 170     for(iidx=0;iidx<4*DIM;iidx++)
 171     {
 172         scratch[iidx] = 0.0;
 173     }
 174
 175     /* Start outer loop over neighborlists */
 176     for(iidx=0; iidx<nri; iidx++)
 177     {
 178         /* Load shift vector for this list */
 179         i_shift_offset   = DIM*shiftidx[iidx];
 180
 181         /* Load limits for loop over neighbors */
 182         j_index_start    = jindex[iidx];
 183         j_index_end      = jindex[iidx+1];
 184
 185         /* Get outer coordinate index */
 186         inr              = iinr[iidx];
 187         i_coord_offset   = DIM*inr;
 188
 189         /* Load i particle coords and add shift vector */
 190         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 191                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 192
 193         fix0             = _mm_setzero_ps();
 194         fiy0             = _mm_setzero_ps();
 195         fiz0             = _mm_setzero_ps();
 196         fix1             = _mm_setzero_ps();
 197         fiy1             = _mm_setzero_ps();
 198         fiz1             = _mm_setzero_ps();
 199         fix2             = _mm_setzero_ps();
 200         fiy2             = _mm_setzero_ps();
 201         fiz2             = _mm_setzero_ps();
 202
 203         /* Reset potential sums */
 204         velecsum         = _mm_setzero_ps();
 205
 206         /* Start inner kernel loop */
 207         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 208         {
 209
 210             /* Get j neighbor index, and coordinate index */
 211             jnrA             = jjnr[jidx];
 212             jnrB             = jjnr[jidx+1];
 213             jnrC             = jjnr[jidx+2];
 214             jnrD             = jjnr[jidx+3];
 215             j_coord_offsetA  = DIM*jnrA;
 216             j_coord_offsetB  = DIM*jnrB;
 217             j_coord_offsetC  = DIM*jnrC;
 218             j_coord_offsetD  = DIM*jnrD;
 219
 220             /* load j atom coordinates */
 221             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 222                                               x+j_coord_offsetC,x+j_coord_offsetD,
 223                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 224
 225             /* Calculate displacement vector */
 226             dx00             = _mm_sub_ps(ix0,jx0);
 227             dy00             = _mm_sub_ps(iy0,jy0);
 228             dz00             = _mm_sub_ps(iz0,jz0);
 229             dx01             = _mm_sub_ps(ix0,jx1);
 230             dy01             = _mm_sub_ps(iy0,jy1);
 231             dz01             = _mm_sub_ps(iz0,jz1);
 232             dx02             = _mm_sub_ps(ix0,jx2);
 233             dy02             = _mm_sub_ps(iy0,jy2);
 234             dz02             = _mm_sub_ps(iz0,jz2);
 235             dx10             = _mm_sub_ps(ix1,jx0);
 236             dy10             = _mm_sub_ps(iy1,jy0);
 237             dz10             = _mm_sub_ps(iz1,jz0);
 238             dx11             = _mm_sub_ps(ix1,jx1);
 239             dy11             = _mm_sub_ps(iy1,jy1);
 240             dz11             = _mm_sub_ps(iz1,jz1);
 241             dx12             = _mm_sub_ps(ix1,jx2);
 242             dy12             = _mm_sub_ps(iy1,jy2);
 243             dz12             = _mm_sub_ps(iz1,jz2);
 244             dx20             = _mm_sub_ps(ix2,jx0);
 245             dy20             = _mm_sub_ps(iy2,jy0);
 246             dz20             = _mm_sub_ps(iz2,jz0);
 247             dx21             = _mm_sub_ps(ix2,jx1);
 248             dy21             = _mm_sub_ps(iy2,jy1);
 249             dz21             = _mm_sub_ps(iz2,jz1);
 250             dx22             = _mm_sub_ps(ix2,jx2);
 251             dy22             = _mm_sub_ps(iy2,jy2);
 252             dz22             = _mm_sub_ps(iz2,jz2);
 253
 254             /* Calculate squared distance and things based on it */
 255             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 256             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 257             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 258             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 259             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 260             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 261             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 262             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 263             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 264
 265             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 266             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 267             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 268             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 269             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 270             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 271             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 272             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 273             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 274
 275             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 276             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 277             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 278             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 279             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 280             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 281             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 282             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 283             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 284
 285             fjx0             = _mm_setzero_ps();
 286             fjy0             = _mm_setzero_ps();
 287             fjz0             = _mm_setzero_ps();
 288             fjx1             = _mm_setzero_ps();
 289             fjy1             = _mm_setzero_ps();
 290             fjz1             = _mm_setzero_ps();
 291             fjx2             = _mm_setzero_ps();
 292             fjy2             = _mm_setzero_ps();
 293             fjz2             = _mm_setzero_ps();
 294
 295             /**************************
 296              * CALCULATE INTERACTIONS *
 297              **************************/
 298
 299             if (gmx_mm_any_lt(rsq00,rcutoff2))
 300             {
 301
 302             r00              = _mm_mul_ps(rsq00,rinv00);
 303
 304             /* EWALD ELECTROSTATICS */
 305
 306             /* Analytical PME correction */
 307             zeta2            = _mm_mul_ps(beta2,rsq00);
 308             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 309             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 310             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 311             felec            = _mm_mul_ps(qq00,felec);
 312             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 313             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 314             velec            = _mm_mul_ps(qq00,velec);
 315
 316             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 317
 318             /* Update potential sum for this i atom from the interaction with this j atom. */
 319             velec            = _mm_and_ps(velec,cutoff_mask);
 320             velecsum         = _mm_add_ps(velecsum,velec);
 321
 322             fscal            = felec;
 323
 324             fscal            = _mm_and_ps(fscal,cutoff_mask);
 325
 326              /* Update vectorial force */
 327             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 328             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 329             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 330
 331             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 332             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 333             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 334
 335             }
 336
 337             /**************************
 338              * CALCULATE INTERACTIONS *
 339              **************************/
 340
 341             if (gmx_mm_any_lt(rsq01,rcutoff2))
 342             {
 343
 344             r01              = _mm_mul_ps(rsq01,rinv01);
 345
 346             /* EWALD ELECTROSTATICS */
 347
 348             /* Analytical PME correction */
 349             zeta2            = _mm_mul_ps(beta2,rsq01);
 350             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 351             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 352             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 353             felec            = _mm_mul_ps(qq01,felec);
 354             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 355             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 356             velec            = _mm_mul_ps(qq01,velec);
 357
 358             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 359
 360             /* Update potential sum for this i atom from the interaction with this j atom. */
 361             velec            = _mm_and_ps(velec,cutoff_mask);
 362             velecsum         = _mm_add_ps(velecsum,velec);
 363
 364             fscal            = felec;
 365
 366             fscal            = _mm_and_ps(fscal,cutoff_mask);
 367
 368              /* Update vectorial force */
 369             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 370             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 371             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 372
 373             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 374             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 375             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 376
 377             }
 378
 379             /**************************
 380              * CALCULATE INTERACTIONS *
 381              **************************/
 382
 383             if (gmx_mm_any_lt(rsq02,rcutoff2))
 384             {
 385
 386             r02              = _mm_mul_ps(rsq02,rinv02);
 387
 388             /* EWALD ELECTROSTATICS */
 389
 390             /* Analytical PME correction */
 391             zeta2            = _mm_mul_ps(beta2,rsq02);
 392             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 393             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 394             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 395             felec            = _mm_mul_ps(qq02,felec);
 396             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 397             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 398             velec            = _mm_mul_ps(qq02,velec);
 399
 400             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 401
 402             /* Update potential sum for this i atom from the interaction with this j atom. */
 403             velec            = _mm_and_ps(velec,cutoff_mask);
 404             velecsum         = _mm_add_ps(velecsum,velec);
 405
 406             fscal            = felec;
 407
 408             fscal            = _mm_and_ps(fscal,cutoff_mask);
 409
 410              /* Update vectorial force */
 411             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 412             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 413             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 414
 415             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 416             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 417             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 418
 419             }
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             if (gmx_mm_any_lt(rsq10,rcutoff2))
 426             {
 427
 428             r10              = _mm_mul_ps(rsq10,rinv10);
 429
 430             /* EWALD ELECTROSTATICS */
 431
 432             /* Analytical PME correction */
 433             zeta2            = _mm_mul_ps(beta2,rsq10);
 434             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 435             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 436             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 437             felec            = _mm_mul_ps(qq10,felec);
 438             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 439             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 440             velec            = _mm_mul_ps(qq10,velec);
 441
 442             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 443
 444             /* Update potential sum for this i atom from the interaction with this j atom. */
 445             velec            = _mm_and_ps(velec,cutoff_mask);
 446             velecsum         = _mm_add_ps(velecsum,velec);
 447
 448             fscal            = felec;
 449
 450             fscal            = _mm_and_ps(fscal,cutoff_mask);
 451
 452              /* Update vectorial force */
 453             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 454             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 455             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 456
 457             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 458             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 459             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 460
 461             }
 462
 463             /**************************
 464              * CALCULATE INTERACTIONS *
 465              **************************/
 466
 467             if (gmx_mm_any_lt(rsq11,rcutoff2))
 468             {
 469
 470             r11              = _mm_mul_ps(rsq11,rinv11);
 471
 472             /* EWALD ELECTROSTATICS */
 473
 474             /* Analytical PME correction */
 475             zeta2            = _mm_mul_ps(beta2,rsq11);
 476             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 477             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 478             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 479             felec            = _mm_mul_ps(qq11,felec);
 480             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 481             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
 482             velec            = _mm_mul_ps(qq11,velec);
 483
 484             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 485
 486             /* Update potential sum for this i atom from the interaction with this j atom. */
 487             velec            = _mm_and_ps(velec,cutoff_mask);
 488             velecsum         = _mm_add_ps(velecsum,velec);
 489
 490             fscal            = felec;
 491
 492             fscal            = _mm_and_ps(fscal,cutoff_mask);
 493
 494              /* Update vectorial force */
 495             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 496             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 497             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 498
 499             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 500             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 501             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 502
 503             }
 504
 505             /**************************
 506              * CALCULATE INTERACTIONS *
 507              **************************/
 508
 509             if (gmx_mm_any_lt(rsq12,rcutoff2))
 510             {
 511
 512             r12              = _mm_mul_ps(rsq12,rinv12);
 513
 514             /* EWALD ELECTROSTATICS */
 515
 516             /* Analytical PME correction */
 517             zeta2            = _mm_mul_ps(beta2,rsq12);
 518             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
 519             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 520             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 521             felec            = _mm_mul_ps(qq12,felec);
 522             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 523             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
 524             velec            = _mm_mul_ps(qq12,velec);
 525
 526             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 527
 528             /* Update potential sum for this i atom from the interaction with this j atom. */
 529             velec            = _mm_and_ps(velec,cutoff_mask);
 530             velecsum         = _mm_add_ps(velecsum,velec);
 531
 532             fscal            = felec;
 533
 534             fscal            = _mm_and_ps(fscal,cutoff_mask);
 535
 536              /* Update vectorial force */
 537             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 538             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 539             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 540
 541             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 542             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 543             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 544
 545             }
 546
 547             /**************************
 548              * CALCULATE INTERACTIONS *
 549              **************************/
 550
 551             if (gmx_mm_any_lt(rsq20,rcutoff2))
 552             {
 553
 554             r20              = _mm_mul_ps(rsq20,rinv20);
 555
 556             /* EWALD ELECTROSTATICS */
 557
 558             /* Analytical PME correction */
 559             zeta2            = _mm_mul_ps(beta2,rsq20);
 560             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
 561             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 562             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 563             felec            = _mm_mul_ps(qq20,felec);
 564             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 565             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
 566             velec            = _mm_mul_ps(qq20,velec);
 567
 568             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 569
 570             /* Update potential sum for this i atom from the interaction with this j atom. */
 571             velec            = _mm_and_ps(velec,cutoff_mask);
 572             velecsum         = _mm_add_ps(velecsum,velec);
 573
 574             fscal            = felec;
 575
 576             fscal            = _mm_and_ps(fscal,cutoff_mask);
 577
 578              /* Update vectorial force */
 579             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 580             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 581             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 582
 583             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 584             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 585             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 586
 587             }
 588
 589             /**************************
 590              * CALCULATE INTERACTIONS *
 591              **************************/
 592
 593             if (gmx_mm_any_lt(rsq21,rcutoff2))
 594             {
 595
 596             r21              = _mm_mul_ps(rsq21,rinv21);
 597
 598             /* EWALD ELECTROSTATICS */
 599
 600             /* Analytical PME correction */
 601             zeta2            = _mm_mul_ps(beta2,rsq21);
 602             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
 603             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 604             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 605             felec            = _mm_mul_ps(qq21,felec);
 606             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 607             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
 608             velec            = _mm_mul_ps(qq21,velec);
 609
 610             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 611
 612             /* Update potential sum for this i atom from the interaction with this j atom. */
 613             velec            = _mm_and_ps(velec,cutoff_mask);
 614             velecsum         = _mm_add_ps(velecsum,velec);
 615
 616             fscal            = felec;
 617
 618             fscal            = _mm_and_ps(fscal,cutoff_mask);
 619
 620              /* Update vectorial force */
 621             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 622             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 623             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 624
 625             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 626             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 627             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 628
 629             }
 630
 631             /**************************
 632              * CALCULATE INTERACTIONS *
 633              **************************/
 634
 635             if (gmx_mm_any_lt(rsq22,rcutoff2))
 636             {
 637
 638             r22              = _mm_mul_ps(rsq22,rinv22);
 639
 640             /* EWALD ELECTROSTATICS */
 641
 642             /* Analytical PME correction */
 643             zeta2            = _mm_mul_ps(beta2,rsq22);
 644             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
 645             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 646             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 647             felec            = _mm_mul_ps(qq22,felec);
 648             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 649             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
 650             velec            = _mm_mul_ps(qq22,velec);
 651
 652             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 653
 654             /* Update potential sum for this i atom from the interaction with this j atom. */
 655             velec            = _mm_and_ps(velec,cutoff_mask);
 656             velecsum         = _mm_add_ps(velecsum,velec);
 657
 658             fscal            = felec;
 659
 660             fscal            = _mm_and_ps(fscal,cutoff_mask);
 661
 662              /* Update vectorial force */
 663             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 664             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 665             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 666
 667             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 668             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 669             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 670
 671             }
 672
 673             fjptrA             = f+j_coord_offsetA;
 674             fjptrB             = f+j_coord_offsetB;
 675             fjptrC             = f+j_coord_offsetC;
 676             fjptrD             = f+j_coord_offsetD;
 677
 678             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 679                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 680
 681             /* Inner loop uses 297 flops */
 682         }
 683
 684         if(jidx<j_index_end)
 685         {
 686
 687             /* Get j neighbor index, and coordinate index */
 688             jnrlistA         = jjnr[jidx];
 689             jnrlistB         = jjnr[jidx+1];
 690             jnrlistC         = jjnr[jidx+2];
 691             jnrlistD         = jjnr[jidx+3];
 692             /* Sign of each element will be negative for non-real atoms.
 693              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 694              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 695              */
 696             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 697             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 698             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 699             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 700             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 701             j_coord_offsetA  = DIM*jnrA;
 702             j_coord_offsetB  = DIM*jnrB;
 703             j_coord_offsetC  = DIM*jnrC;
 704             j_coord_offsetD  = DIM*jnrD;
 705
 706             /* load j atom coordinates */
 707             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 708                                               x+j_coord_offsetC,x+j_coord_offsetD,
 709                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 710
 711             /* Calculate displacement vector */
 712             dx00             = _mm_sub_ps(ix0,jx0);
 713             dy00             = _mm_sub_ps(iy0,jy0);
 714             dz00             = _mm_sub_ps(iz0,jz0);
 715             dx01             = _mm_sub_ps(ix0,jx1);
 716             dy01             = _mm_sub_ps(iy0,jy1);
 717             dz01             = _mm_sub_ps(iz0,jz1);
 718             dx02             = _mm_sub_ps(ix0,jx2);
 719             dy02             = _mm_sub_ps(iy0,jy2);
 720             dz02             = _mm_sub_ps(iz0,jz2);
 721             dx10             = _mm_sub_ps(ix1,jx0);
 722             dy10             = _mm_sub_ps(iy1,jy0);
 723             dz10             = _mm_sub_ps(iz1,jz0);
 724             dx11             = _mm_sub_ps(ix1,jx1);
 725             dy11             = _mm_sub_ps(iy1,jy1);
 726             dz11             = _mm_sub_ps(iz1,jz1);
 727             dx12             = _mm_sub_ps(ix1,jx2);
 728             dy12             = _mm_sub_ps(iy1,jy2);
 729             dz12             = _mm_sub_ps(iz1,jz2);
 730             dx20             = _mm_sub_ps(ix2,jx0);
 731             dy20             = _mm_sub_ps(iy2,jy0);
 732             dz20             = _mm_sub_ps(iz2,jz0);
 733             dx21             = _mm_sub_ps(ix2,jx1);
 734             dy21             = _mm_sub_ps(iy2,jy1);
 735             dz21             = _mm_sub_ps(iz2,jz1);
 736             dx22             = _mm_sub_ps(ix2,jx2);
 737             dy22             = _mm_sub_ps(iy2,jy2);
 738             dz22             = _mm_sub_ps(iz2,jz2);
 739
 740             /* Calculate squared distance and things based on it */
 741             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 742             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 743             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 744             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 745             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 746             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 747             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 748             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 749             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 750
 751             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 752             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 753             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 754             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 755             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 756             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 757             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 758             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 759             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 760
 761             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 762             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 763             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 764             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 765             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 766             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 767             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 768             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 769             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 770
 771             fjx0             = _mm_setzero_ps();
 772             fjy0             = _mm_setzero_ps();
 773             fjz0             = _mm_setzero_ps();
 774             fjx1             = _mm_setzero_ps();
 775             fjy1             = _mm_setzero_ps();
 776             fjz1             = _mm_setzero_ps();
 777             fjx2             = _mm_setzero_ps();
 778             fjy2             = _mm_setzero_ps();
 779             fjz2             = _mm_setzero_ps();
 780
 781             /**************************
 782              * CALCULATE INTERACTIONS *
 783              **************************/
 784
 785             if (gmx_mm_any_lt(rsq00,rcutoff2))
 786             {
 787
 788             r00              = _mm_mul_ps(rsq00,rinv00);
 789             r00              = _mm_andnot_ps(dummy_mask,r00);
 790
 791             /* EWALD ELECTROSTATICS */
 792
 793             /* Analytical PME correction */
 794             zeta2            = _mm_mul_ps(beta2,rsq00);
 795             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 796             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 797             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 798             felec            = _mm_mul_ps(qq00,felec);
 799             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 800             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 801             velec            = _mm_mul_ps(qq00,velec);
 802
 803             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 804
 805             /* Update potential sum for this i atom from the interaction with this j atom. */
 806             velec            = _mm_and_ps(velec,cutoff_mask);
 807             velec            = _mm_andnot_ps(dummy_mask,velec);
 808             velecsum         = _mm_add_ps(velecsum,velec);
 809
 810             fscal            = felec;
 811
 812             fscal            = _mm_and_ps(fscal,cutoff_mask);
 813
 814             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 815
 816              /* Update vectorial force */
 817             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 818             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 819             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 820
 821             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 822             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 823             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 824
 825             }
 826
 827             /**************************
 828              * CALCULATE INTERACTIONS *
 829              **************************/
 830
 831             if (gmx_mm_any_lt(rsq01,rcutoff2))
 832             {
 833
 834             r01              = _mm_mul_ps(rsq01,rinv01);
 835             r01              = _mm_andnot_ps(dummy_mask,r01);
 836
 837             /* EWALD ELECTROSTATICS */
 838
 839             /* Analytical PME correction */
 840             zeta2            = _mm_mul_ps(beta2,rsq01);
 841             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
 842             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 843             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 844             felec            = _mm_mul_ps(qq01,felec);
 845             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 846             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv01,sh_ewald));
 847             velec            = _mm_mul_ps(qq01,velec);
 848
 849             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 850
 851             /* Update potential sum for this i atom from the interaction with this j atom. */
 852             velec            = _mm_and_ps(velec,cutoff_mask);
 853             velec            = _mm_andnot_ps(dummy_mask,velec);
 854             velecsum         = _mm_add_ps(velecsum,velec);
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_and_ps(fscal,cutoff_mask);
 859
 860             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 861
 862              /* Update vectorial force */
 863             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 864             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 865             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 866
 867             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 868             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 869             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 870
 871             }
 872
 873             /**************************
 874              * CALCULATE INTERACTIONS *
 875              **************************/
 876
 877             if (gmx_mm_any_lt(rsq02,rcutoff2))
 878             {
 879
 880             r02              = _mm_mul_ps(rsq02,rinv02);
 881             r02              = _mm_andnot_ps(dummy_mask,r02);
 882
 883             /* EWALD ELECTROSTATICS */
 884
 885             /* Analytical PME correction */
 886             zeta2            = _mm_mul_ps(beta2,rsq02);
 887             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
 888             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 889             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 890             felec            = _mm_mul_ps(qq02,felec);
 891             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 892             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv02,sh_ewald));
 893             velec            = _mm_mul_ps(qq02,velec);
 894
 895             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 896
 897             /* Update potential sum for this i atom from the interaction with this j atom. */
 898             velec            = _mm_and_ps(velec,cutoff_mask);
 899             velec            = _mm_andnot_ps(dummy_mask,velec);
 900             velecsum         = _mm_add_ps(velecsum,velec);
 901
 902             fscal            = felec;
 903
 904             fscal            = _mm_and_ps(fscal,cutoff_mask);
 905
 906             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 907
 908              /* Update vectorial force */
 909             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 910             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 911             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 912
 913             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 914             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 915             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 916
 917             }
 918
 919             /**************************
 920              * CALCULATE INTERACTIONS *
 921              **************************/
 922
 923             if (gmx_mm_any_lt(rsq10,rcutoff2))
 924             {
 925
 926             r10              = _mm_mul_ps(rsq10,rinv10);
 927             r10              = _mm_andnot_ps(dummy_mask,r10);
 928
 929             /* EWALD ELECTROSTATICS */
 930
 931             /* Analytical PME correction */
 932             zeta2            = _mm_mul_ps(beta2,rsq10);
 933             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
 934             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 935             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 936             felec            = _mm_mul_ps(qq10,felec);
 937             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 938             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv10,sh_ewald));
 939             velec            = _mm_mul_ps(qq10,velec);
 940
 941             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 942
 943             /* Update potential sum for this i atom from the interaction with this j atom. */
 944             velec            = _mm_and_ps(velec,cutoff_mask);
 945             velec            = _mm_andnot_ps(dummy_mask,velec);
 946             velecsum         = _mm_add_ps(velecsum,velec);
 947
 948             fscal            = felec;
 949
 950             fscal            = _mm_and_ps(fscal,cutoff_mask);
 951
 952             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 953
 954              /* Update vectorial force */
 955             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 956             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 957             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 958
 959             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 960             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 961             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 962
 963             }
 964
 965             /**************************
 966              * CALCULATE INTERACTIONS *
 967              **************************/
 968
 969             if (gmx_mm_any_lt(rsq11,rcutoff2))
 970             {
 971
 972             r11              = _mm_mul_ps(rsq11,rinv11);
 973             r11              = _mm_andnot_ps(dummy_mask,r11);
 974
 975             /* EWALD ELECTROSTATICS */
 976
 977             /* Analytical PME correction */
 978             zeta2            = _mm_mul_ps(beta2,rsq11);
 979             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
 980             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 981             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 982             felec            = _mm_mul_ps(qq11,felec);
 983             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 984             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv11,sh_ewald));
 985             velec            = _mm_mul_ps(qq11,velec);
 986
 987             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 988
 989             /* Update potential sum for this i atom from the interaction with this j atom. */
 990             velec            = _mm_and_ps(velec,cutoff_mask);
 991             velec            = _mm_andnot_ps(dummy_mask,velec);
 992             velecsum         = _mm_add_ps(velecsum,velec);
 993
 994             fscal            = felec;
 995
 996             fscal            = _mm_and_ps(fscal,cutoff_mask);
 997
 998             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 999
1000              /* Update vectorial force */
1001             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1002             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1003             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1004
1005             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1006             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1007             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1008
1009             }
1010
1011             /**************************
1012              * CALCULATE INTERACTIONS *
1013              **************************/
1014
1015             if (gmx_mm_any_lt(rsq12,rcutoff2))
1016             {
1017
1018             r12              = _mm_mul_ps(rsq12,rinv12);
1019             r12              = _mm_andnot_ps(dummy_mask,r12);
1020
1021             /* EWALD ELECTROSTATICS */
1022
1023             /* Analytical PME correction */
1024             zeta2            = _mm_mul_ps(beta2,rsq12);
1025             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1026             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1027             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1028             felec            = _mm_mul_ps(qq12,felec);
1029             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1030             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv12,sh_ewald));
1031             velec            = _mm_mul_ps(qq12,velec);
1032
1033             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1034
1035             /* Update potential sum for this i atom from the interaction with this j atom. */
1036             velec            = _mm_and_ps(velec,cutoff_mask);
1037             velec            = _mm_andnot_ps(dummy_mask,velec);
1038             velecsum         = _mm_add_ps(velecsum,velec);
1039
1040             fscal            = felec;
1041
1042             fscal            = _mm_and_ps(fscal,cutoff_mask);
1043
1044             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1045
1046              /* Update vectorial force */
1047             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1048             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1049             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1050
1051             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1052             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1053             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1054
1055             }
1056
1057             /**************************
1058              * CALCULATE INTERACTIONS *
1059              **************************/
1060
1061             if (gmx_mm_any_lt(rsq20,rcutoff2))
1062             {
1063
1064             r20              = _mm_mul_ps(rsq20,rinv20);
1065             r20              = _mm_andnot_ps(dummy_mask,r20);
1066
1067             /* EWALD ELECTROSTATICS */
1068
1069             /* Analytical PME correction */
1070             zeta2            = _mm_mul_ps(beta2,rsq20);
1071             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1072             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1073             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1074             felec            = _mm_mul_ps(qq20,felec);
1075             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1076             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv20,sh_ewald));
1077             velec            = _mm_mul_ps(qq20,velec);
1078
1079             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1080
1081             /* Update potential sum for this i atom from the interaction with this j atom. */
1082             velec            = _mm_and_ps(velec,cutoff_mask);
1083             velec            = _mm_andnot_ps(dummy_mask,velec);
1084             velecsum         = _mm_add_ps(velecsum,velec);
1085
1086             fscal            = felec;
1087
1088             fscal            = _mm_and_ps(fscal,cutoff_mask);
1089
1090             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1091
1092              /* Update vectorial force */
1093             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1094             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1095             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1096
1097             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1098             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1099             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1100
1101             }
1102
1103             /**************************
1104              * CALCULATE INTERACTIONS *
1105              **************************/
1106
1107             if (gmx_mm_any_lt(rsq21,rcutoff2))
1108             {
1109
1110             r21              = _mm_mul_ps(rsq21,rinv21);
1111             r21              = _mm_andnot_ps(dummy_mask,r21);
1112
1113             /* EWALD ELECTROSTATICS */
1114
1115             /* Analytical PME correction */
1116             zeta2            = _mm_mul_ps(beta2,rsq21);
1117             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1118             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1119             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1120             felec            = _mm_mul_ps(qq21,felec);
1121             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1122             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv21,sh_ewald));
1123             velec            = _mm_mul_ps(qq21,velec);
1124
1125             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1126
1127             /* Update potential sum for this i atom from the interaction with this j atom. */
1128             velec            = _mm_and_ps(velec,cutoff_mask);
1129             velec            = _mm_andnot_ps(dummy_mask,velec);
1130             velecsum         = _mm_add_ps(velecsum,velec);
1131
1132             fscal            = felec;
1133
1134             fscal            = _mm_and_ps(fscal,cutoff_mask);
1135
1136             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1137
1138              /* Update vectorial force */
1139             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1140             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1141             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1142
1143             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1144             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1145             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1146
1147             }
1148
1149             /**************************
1150              * CALCULATE INTERACTIONS *
1151              **************************/
1152
1153             if (gmx_mm_any_lt(rsq22,rcutoff2))
1154             {
1155
1156             r22              = _mm_mul_ps(rsq22,rinv22);
1157             r22              = _mm_andnot_ps(dummy_mask,r22);
1158
1159             /* EWALD ELECTROSTATICS */
1160
1161             /* Analytical PME correction */
1162             zeta2            = _mm_mul_ps(beta2,rsq22);
1163             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1164             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1165             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1166             felec            = _mm_mul_ps(qq22,felec);
1167             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
1168             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv22,sh_ewald));
1169             velec            = _mm_mul_ps(qq22,velec);
1170
1171             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1172
1173             /* Update potential sum for this i atom from the interaction with this j atom. */
1174             velec            = _mm_and_ps(velec,cutoff_mask);
1175             velec            = _mm_andnot_ps(dummy_mask,velec);
1176             velecsum         = _mm_add_ps(velecsum,velec);
1177
1178             fscal            = felec;
1179
1180             fscal            = _mm_and_ps(fscal,cutoff_mask);
1181
1182             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1183
1184              /* Update vectorial force */
1185             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1186             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1187             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1188
1189             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1190             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1191             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1192
1193             }
1194
1195             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1196             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1197             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1198             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1199
1200             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1201                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1202
1203             /* Inner loop uses 306 flops */
1204         }
1205
1206         /* End of innermost loop */
1207
1208         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1209                                               f+i_coord_offset,fshift+i_shift_offset);
1210
1211         ggid                        = gid[iidx];
1212         /* Update potential energies */
1213         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1214
1215         /* Increment number of inner iterations */
1216         inneriter                  += j_index_end - j_index_start;
1217
1218         /* Outer loop uses 19 flops */
1219     }
1220
1221     /* Increment number of outer iterations */
1222     outeriter        += nri;
1223
1224     /* Update outer/inner flops */
1225
1226     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*306);
1227 }
1228 /*
1229  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomW3W3_F_avx_128_fma_single
1230  * Electrostatics interaction: Ewald
1231  * VdW interaction:            None
1232  * Geometry:                   Water3-Water3
1233  * Calculate force/pot:        Force
1234  */
1235 void
1236 nb_kernel_ElecEwSh_VdwNone_GeomW3W3_F_avx_128_fma_single
1237                     (t_nblist                    * gmx_restrict       nlist,
1238                      rvec                        * gmx_restrict          xx,
1239                      rvec                        * gmx_restrict          ff,
1240                      t_forcerec                  * gmx_restrict          fr,
1241                      t_mdatoms                   * gmx_restrict     mdatoms,
1242                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1243                      t_nrnb                      * gmx_restrict        nrnb)
1244 {
1245     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1246      * just 0 for non-waters.
1247      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1248      * jnr indices corresponding to data put in the four positions in the SIMD register.
1249      */
1250     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1251     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1252     int              jnrA,jnrB,jnrC,jnrD;
1253     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1254     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1255     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1256     real             rcutoff_scalar;
1257     real             *shiftvec,*fshift,*x,*f;
1258     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1259     real             scratch[4*DIM];
1260     __m128           fscal,rcutoff,rcutoff2,jidxall;
1261     int              vdwioffset0;
1262     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1263     int              vdwioffset1;
1264     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1265     int              vdwioffset2;
1266     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1267     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1268     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1269     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1270     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1271     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1272     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1273     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1274     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1275     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1276     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1277     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1278     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1279     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1280     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1281     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1282     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1283     real             *charge;
1284     __m128i          ewitab;
1285     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1286     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
1287     real             *ewtab;
1288     __m128           dummy_mask,cutoff_mask;
1289     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1290     __m128           one     = _mm_set1_ps(1.0);
1291     __m128           two     = _mm_set1_ps(2.0);
1292     x                = xx[0];
1293     f                = ff[0];
1294
1295     nri              = nlist->nri;
1296     iinr             = nlist->iinr;
1297     jindex           = nlist->jindex;
1298     jjnr             = nlist->jjnr;
1299     shiftidx         = nlist->shift;
1300     gid              = nlist->gid;
1301     shiftvec         = fr->shift_vec[0];
1302     fshift           = fr->fshift[0];
1303     facel            = _mm_set1_ps(fr->epsfac);
1304     charge           = mdatoms->chargeA;
1305
1306     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
1307     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
1308     beta2            = _mm_mul_ps(beta,beta);
1309     beta3            = _mm_mul_ps(beta,beta2);
1310     ewtab            = fr->ic->tabq_coul_F;
1311     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
1312     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
1313
1314     /* Setup water-specific parameters */
1315     inr              = nlist->iinr[0];
1316     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1317     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1318     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1319
1320     jq0              = _mm_set1_ps(charge[inr+0]);
1321     jq1              = _mm_set1_ps(charge[inr+1]);
1322     jq2              = _mm_set1_ps(charge[inr+2]);
1323     qq00             = _mm_mul_ps(iq0,jq0);
1324     qq01             = _mm_mul_ps(iq0,jq1);
1325     qq02             = _mm_mul_ps(iq0,jq2);
1326     qq10             = _mm_mul_ps(iq1,jq0);
1327     qq11             = _mm_mul_ps(iq1,jq1);
1328     qq12             = _mm_mul_ps(iq1,jq2);
1329     qq20             = _mm_mul_ps(iq2,jq0);
1330     qq21             = _mm_mul_ps(iq2,jq1);
1331     qq22             = _mm_mul_ps(iq2,jq2);
1332
1333     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1334     rcutoff_scalar   = fr->rcoulomb;
1335     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1336     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1337
1338     /* Avoid stupid compiler warnings */
1339     jnrA = jnrB = jnrC = jnrD = 0;
1340     j_coord_offsetA = 0;
1341     j_coord_offsetB = 0;
1342     j_coord_offsetC = 0;
1343     j_coord_offsetD = 0;
1344
1345     outeriter        = 0;
1346     inneriter        = 0;
1347
1348     for(iidx=0;iidx<4*DIM;iidx++)
1349     {
1350         scratch[iidx] = 0.0;
1351     }
1352
1353     /* Start outer loop over neighborlists */
1354     for(iidx=0; iidx<nri; iidx++)
1355     {
1356         /* Load shift vector for this list */
1357         i_shift_offset   = DIM*shiftidx[iidx];
1358
1359         /* Load limits for loop over neighbors */
1360         j_index_start    = jindex[iidx];
1361         j_index_end      = jindex[iidx+1];
1362
1363         /* Get outer coordinate index */
1364         inr              = iinr[iidx];
1365         i_coord_offset   = DIM*inr;
1366
1367         /* Load i particle coords and add shift vector */
1368         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1369                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1370
1371         fix0             = _mm_setzero_ps();
1372         fiy0             = _mm_setzero_ps();
1373         fiz0             = _mm_setzero_ps();
1374         fix1             = _mm_setzero_ps();
1375         fiy1             = _mm_setzero_ps();
1376         fiz1             = _mm_setzero_ps();
1377         fix2             = _mm_setzero_ps();
1378         fiy2             = _mm_setzero_ps();
1379         fiz2             = _mm_setzero_ps();
1380
1381         /* Start inner kernel loop */
1382         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1383         {
1384
1385             /* Get j neighbor index, and coordinate index */
1386             jnrA             = jjnr[jidx];
1387             jnrB             = jjnr[jidx+1];
1388             jnrC             = jjnr[jidx+2];
1389             jnrD             = jjnr[jidx+3];
1390             j_coord_offsetA  = DIM*jnrA;
1391             j_coord_offsetB  = DIM*jnrB;
1392             j_coord_offsetC  = DIM*jnrC;
1393             j_coord_offsetD  = DIM*jnrD;
1394
1395             /* load j atom coordinates */
1396             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1397                                               x+j_coord_offsetC,x+j_coord_offsetD,
1398                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1399
1400             /* Calculate displacement vector */
1401             dx00             = _mm_sub_ps(ix0,jx0);
1402             dy00             = _mm_sub_ps(iy0,jy0);
1403             dz00             = _mm_sub_ps(iz0,jz0);
1404             dx01             = _mm_sub_ps(ix0,jx1);
1405             dy01             = _mm_sub_ps(iy0,jy1);
1406             dz01             = _mm_sub_ps(iz0,jz1);
1407             dx02             = _mm_sub_ps(ix0,jx2);
1408             dy02             = _mm_sub_ps(iy0,jy2);
1409             dz02             = _mm_sub_ps(iz0,jz2);
1410             dx10             = _mm_sub_ps(ix1,jx0);
1411             dy10             = _mm_sub_ps(iy1,jy0);
1412             dz10             = _mm_sub_ps(iz1,jz0);
1413             dx11             = _mm_sub_ps(ix1,jx1);
1414             dy11             = _mm_sub_ps(iy1,jy1);
1415             dz11             = _mm_sub_ps(iz1,jz1);
1416             dx12             = _mm_sub_ps(ix1,jx2);
1417             dy12             = _mm_sub_ps(iy1,jy2);
1418             dz12             = _mm_sub_ps(iz1,jz2);
1419             dx20             = _mm_sub_ps(ix2,jx0);
1420             dy20             = _mm_sub_ps(iy2,jy0);
1421             dz20             = _mm_sub_ps(iz2,jz0);
1422             dx21             = _mm_sub_ps(ix2,jx1);
1423             dy21             = _mm_sub_ps(iy2,jy1);
1424             dz21             = _mm_sub_ps(iz2,jz1);
1425             dx22             = _mm_sub_ps(ix2,jx2);
1426             dy22             = _mm_sub_ps(iy2,jy2);
1427             dz22             = _mm_sub_ps(iz2,jz2);
1428
1429             /* Calculate squared distance and things based on it */
1430             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1431             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1432             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1433             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1434             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1435             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1436             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1437             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1438             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1439
1440             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1441             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1442             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1443             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1444             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1445             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1446             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1447             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1448             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1449
1450             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1451             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1452             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1453             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1454             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1455             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1456             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1457             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1458             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1459
1460             fjx0             = _mm_setzero_ps();
1461             fjy0             = _mm_setzero_ps();
1462             fjz0             = _mm_setzero_ps();
1463             fjx1             = _mm_setzero_ps();
1464             fjy1             = _mm_setzero_ps();
1465             fjz1             = _mm_setzero_ps();
1466             fjx2             = _mm_setzero_ps();
1467             fjy2             = _mm_setzero_ps();
1468             fjz2             = _mm_setzero_ps();
1469
1470             /**************************
1471              * CALCULATE INTERACTIONS *
1472              **************************/
1473
1474             if (gmx_mm_any_lt(rsq00,rcutoff2))
1475             {
1476
1477             r00              = _mm_mul_ps(rsq00,rinv00);
1478
1479             /* EWALD ELECTROSTATICS */
1480
1481             /* Analytical PME correction */
1482             zeta2            = _mm_mul_ps(beta2,rsq00);
1483             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1484             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1485             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1486             felec            = _mm_mul_ps(qq00,felec);
1487
1488             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1489
1490             fscal            = felec;
1491
1492             fscal            = _mm_and_ps(fscal,cutoff_mask);
1493
1494              /* Update vectorial force */
1495             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1496             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1497             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1498
1499             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1500             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1501             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1502
1503             }
1504
1505             /**************************
1506              * CALCULATE INTERACTIONS *
1507              **************************/
1508
1509             if (gmx_mm_any_lt(rsq01,rcutoff2))
1510             {
1511
1512             r01              = _mm_mul_ps(rsq01,rinv01);
1513
1514             /* EWALD ELECTROSTATICS */
1515
1516             /* Analytical PME correction */
1517             zeta2            = _mm_mul_ps(beta2,rsq01);
1518             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1519             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1520             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1521             felec            = _mm_mul_ps(qq01,felec);
1522
1523             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1524
1525             fscal            = felec;
1526
1527             fscal            = _mm_and_ps(fscal,cutoff_mask);
1528
1529              /* Update vectorial force */
1530             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1531             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1532             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1533
1534             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1535             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1536             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1537
1538             }
1539
1540             /**************************
1541              * CALCULATE INTERACTIONS *
1542              **************************/
1543
1544             if (gmx_mm_any_lt(rsq02,rcutoff2))
1545             {
1546
1547             r02              = _mm_mul_ps(rsq02,rinv02);
1548
1549             /* EWALD ELECTROSTATICS */
1550
1551             /* Analytical PME correction */
1552             zeta2            = _mm_mul_ps(beta2,rsq02);
1553             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1554             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1555             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1556             felec            = _mm_mul_ps(qq02,felec);
1557
1558             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1559
1560             fscal            = felec;
1561
1562             fscal            = _mm_and_ps(fscal,cutoff_mask);
1563
1564              /* Update vectorial force */
1565             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1566             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1567             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1568
1569             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1570             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1571             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1572
1573             }
1574
1575             /**************************
1576              * CALCULATE INTERACTIONS *
1577              **************************/
1578
1579             if (gmx_mm_any_lt(rsq10,rcutoff2))
1580             {
1581
1582             r10              = _mm_mul_ps(rsq10,rinv10);
1583
1584             /* EWALD ELECTROSTATICS */
1585
1586             /* Analytical PME correction */
1587             zeta2            = _mm_mul_ps(beta2,rsq10);
1588             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
1589             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1590             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1591             felec            = _mm_mul_ps(qq10,felec);
1592
1593             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1594
1595             fscal            = felec;
1596
1597             fscal            = _mm_and_ps(fscal,cutoff_mask);
1598
1599              /* Update vectorial force */
1600             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1601             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1602             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1603
1604             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1605             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1606             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1607
1608             }
1609
1610             /**************************
1611              * CALCULATE INTERACTIONS *
1612              **************************/
1613
1614             if (gmx_mm_any_lt(rsq11,rcutoff2))
1615             {
1616
1617             r11              = _mm_mul_ps(rsq11,rinv11);
1618
1619             /* EWALD ELECTROSTATICS */
1620
1621             /* Analytical PME correction */
1622             zeta2            = _mm_mul_ps(beta2,rsq11);
1623             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
1624             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1625             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1626             felec            = _mm_mul_ps(qq11,felec);
1627
1628             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1629
1630             fscal            = felec;
1631
1632             fscal            = _mm_and_ps(fscal,cutoff_mask);
1633
1634              /* Update vectorial force */
1635             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1636             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1637             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1638
1639             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1640             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1641             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1642
1643             }
1644
1645             /**************************
1646              * CALCULATE INTERACTIONS *
1647              **************************/
1648
1649             if (gmx_mm_any_lt(rsq12,rcutoff2))
1650             {
1651
1652             r12              = _mm_mul_ps(rsq12,rinv12);
1653
1654             /* EWALD ELECTROSTATICS */
1655
1656             /* Analytical PME correction */
1657             zeta2            = _mm_mul_ps(beta2,rsq12);
1658             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
1659             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1660             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1661             felec            = _mm_mul_ps(qq12,felec);
1662
1663             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1664
1665             fscal            = felec;
1666
1667             fscal            = _mm_and_ps(fscal,cutoff_mask);
1668
1669              /* Update vectorial force */
1670             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1671             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1672             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1673
1674             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1675             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1676             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1677
1678             }
1679
1680             /**************************
1681              * CALCULATE INTERACTIONS *
1682              **************************/
1683
1684             if (gmx_mm_any_lt(rsq20,rcutoff2))
1685             {
1686
1687             r20              = _mm_mul_ps(rsq20,rinv20);
1688
1689             /* EWALD ELECTROSTATICS */
1690
1691             /* Analytical PME correction */
1692             zeta2            = _mm_mul_ps(beta2,rsq20);
1693             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
1694             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1695             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1696             felec            = _mm_mul_ps(qq20,felec);
1697
1698             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1699
1700             fscal            = felec;
1701
1702             fscal            = _mm_and_ps(fscal,cutoff_mask);
1703
1704              /* Update vectorial force */
1705             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1706             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1707             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1708
1709             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1710             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1711             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1712
1713             }
1714
1715             /**************************
1716              * CALCULATE INTERACTIONS *
1717              **************************/
1718
1719             if (gmx_mm_any_lt(rsq21,rcutoff2))
1720             {
1721
1722             r21              = _mm_mul_ps(rsq21,rinv21);
1723
1724             /* EWALD ELECTROSTATICS */
1725
1726             /* Analytical PME correction */
1727             zeta2            = _mm_mul_ps(beta2,rsq21);
1728             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
1729             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1730             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1731             felec            = _mm_mul_ps(qq21,felec);
1732
1733             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1734
1735             fscal            = felec;
1736
1737             fscal            = _mm_and_ps(fscal,cutoff_mask);
1738
1739              /* Update vectorial force */
1740             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1741             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1742             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1743
1744             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1745             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1746             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1747
1748             }
1749
1750             /**************************
1751              * CALCULATE INTERACTIONS *
1752              **************************/
1753
1754             if (gmx_mm_any_lt(rsq22,rcutoff2))
1755             {
1756
1757             r22              = _mm_mul_ps(rsq22,rinv22);
1758
1759             /* EWALD ELECTROSTATICS */
1760
1761             /* Analytical PME correction */
1762             zeta2            = _mm_mul_ps(beta2,rsq22);
1763             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
1764             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1765             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1766             felec            = _mm_mul_ps(qq22,felec);
1767
1768             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1769
1770             fscal            = felec;
1771
1772             fscal            = _mm_and_ps(fscal,cutoff_mask);
1773
1774              /* Update vectorial force */
1775             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1776             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1777             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1778
1779             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1780             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1781             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1782
1783             }
1784
1785             fjptrA             = f+j_coord_offsetA;
1786             fjptrB             = f+j_coord_offsetB;
1787             fjptrC             = f+j_coord_offsetC;
1788             fjptrD             = f+j_coord_offsetD;
1789
1790             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1791                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1792
1793             /* Inner loop uses 279 flops */
1794         }
1795
1796         if(jidx<j_index_end)
1797         {
1798
1799             /* Get j neighbor index, and coordinate index */
1800             jnrlistA         = jjnr[jidx];
1801             jnrlistB         = jjnr[jidx+1];
1802             jnrlistC         = jjnr[jidx+2];
1803             jnrlistD         = jjnr[jidx+3];
1804             /* Sign of each element will be negative for non-real atoms.
1805              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1806              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1807              */
1808             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1809             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1810             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1811             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1812             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1813             j_coord_offsetA  = DIM*jnrA;
1814             j_coord_offsetB  = DIM*jnrB;
1815             j_coord_offsetC  = DIM*jnrC;
1816             j_coord_offsetD  = DIM*jnrD;
1817
1818             /* load j atom coordinates */
1819             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1820                                               x+j_coord_offsetC,x+j_coord_offsetD,
1821                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1822
1823             /* Calculate displacement vector */
1824             dx00             = _mm_sub_ps(ix0,jx0);
1825             dy00             = _mm_sub_ps(iy0,jy0);
1826             dz00             = _mm_sub_ps(iz0,jz0);
1827             dx01             = _mm_sub_ps(ix0,jx1);
1828             dy01             = _mm_sub_ps(iy0,jy1);
1829             dz01             = _mm_sub_ps(iz0,jz1);
1830             dx02             = _mm_sub_ps(ix0,jx2);
1831             dy02             = _mm_sub_ps(iy0,jy2);
1832             dz02             = _mm_sub_ps(iz0,jz2);
1833             dx10             = _mm_sub_ps(ix1,jx0);
1834             dy10             = _mm_sub_ps(iy1,jy0);
1835             dz10             = _mm_sub_ps(iz1,jz0);
1836             dx11             = _mm_sub_ps(ix1,jx1);
1837             dy11             = _mm_sub_ps(iy1,jy1);
1838             dz11             = _mm_sub_ps(iz1,jz1);
1839             dx12             = _mm_sub_ps(ix1,jx2);
1840             dy12             = _mm_sub_ps(iy1,jy2);
1841             dz12             = _mm_sub_ps(iz1,jz2);
1842             dx20             = _mm_sub_ps(ix2,jx0);
1843             dy20             = _mm_sub_ps(iy2,jy0);
1844             dz20             = _mm_sub_ps(iz2,jz0);
1845             dx21             = _mm_sub_ps(ix2,jx1);
1846             dy21             = _mm_sub_ps(iy2,jy1);
1847             dz21             = _mm_sub_ps(iz2,jz1);
1848             dx22             = _mm_sub_ps(ix2,jx2);
1849             dy22             = _mm_sub_ps(iy2,jy2);
1850             dz22             = _mm_sub_ps(iz2,jz2);
1851
1852             /* Calculate squared distance and things based on it */
1853             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1854             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1855             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1856             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1857             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1858             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1859             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1860             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1861             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1862
1863             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1864             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1865             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1866             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1867             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1868             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1869             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1870             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1871             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1872
1873             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1874             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1875             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1876             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1877             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1878             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1879             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1880             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1881             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1882
1883             fjx0             = _mm_setzero_ps();
1884             fjy0             = _mm_setzero_ps();
1885             fjz0             = _mm_setzero_ps();
1886             fjx1             = _mm_setzero_ps();
1887             fjy1             = _mm_setzero_ps();
1888             fjz1             = _mm_setzero_ps();
1889             fjx2             = _mm_setzero_ps();
1890             fjy2             = _mm_setzero_ps();
1891             fjz2             = _mm_setzero_ps();
1892
1893             /**************************
1894              * CALCULATE INTERACTIONS *
1895              **************************/
1896
1897             if (gmx_mm_any_lt(rsq00,rcutoff2))
1898             {
1899
1900             r00              = _mm_mul_ps(rsq00,rinv00);
1901             r00              = _mm_andnot_ps(dummy_mask,r00);
1902
1903             /* EWALD ELECTROSTATICS */
1904
1905             /* Analytical PME correction */
1906             zeta2            = _mm_mul_ps(beta2,rsq00);
1907             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
1908             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1909             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1910             felec            = _mm_mul_ps(qq00,felec);
1911
1912             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1913
1914             fscal            = felec;
1915
1916             fscal            = _mm_and_ps(fscal,cutoff_mask);
1917
1918             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1919
1920              /* Update vectorial force */
1921             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1922             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1923             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1924
1925             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1926             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1927             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1928
1929             }
1930
1931             /**************************
1932              * CALCULATE INTERACTIONS *
1933              **************************/
1934
1935             if (gmx_mm_any_lt(rsq01,rcutoff2))
1936             {
1937
1938             r01              = _mm_mul_ps(rsq01,rinv01);
1939             r01              = _mm_andnot_ps(dummy_mask,r01);
1940
1941             /* EWALD ELECTROSTATICS */
1942
1943             /* Analytical PME correction */
1944             zeta2            = _mm_mul_ps(beta2,rsq01);
1945             rinv3            = _mm_mul_ps(rinvsq01,rinv01);
1946             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1947             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1948             felec            = _mm_mul_ps(qq01,felec);
1949
1950             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1951
1952             fscal            = felec;
1953
1954             fscal            = _mm_and_ps(fscal,cutoff_mask);
1955
1956             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1957
1958              /* Update vectorial force */
1959             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1960             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1961             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1962
1963             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1964             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1965             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1966
1967             }
1968
1969             /**************************
1970              * CALCULATE INTERACTIONS *
1971              **************************/
1972
1973             if (gmx_mm_any_lt(rsq02,rcutoff2))
1974             {
1975
1976             r02              = _mm_mul_ps(rsq02,rinv02);
1977             r02              = _mm_andnot_ps(dummy_mask,r02);
1978
1979             /* EWALD ELECTROSTATICS */
1980
1981             /* Analytical PME correction */
1982             zeta2            = _mm_mul_ps(beta2,rsq02);
1983             rinv3            = _mm_mul_ps(rinvsq02,rinv02);
1984             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
1985             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
1986             felec            = _mm_mul_ps(qq02,felec);
1987
1988             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1989
1990             fscal            = felec;
1991
1992             fscal            = _mm_and_ps(fscal,cutoff_mask);
1993
1994             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1995
1996              /* Update vectorial force */
1997             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1998             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1999             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
2000
2001             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
2002             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
2003             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
2004
2005             }
2006
2007             /**************************
2008              * CALCULATE INTERACTIONS *
2009              **************************/
2010
2011             if (gmx_mm_any_lt(rsq10,rcutoff2))
2012             {
2013
2014             r10              = _mm_mul_ps(rsq10,rinv10);
2015             r10              = _mm_andnot_ps(dummy_mask,r10);
2016
2017             /* EWALD ELECTROSTATICS */
2018
2019             /* Analytical PME correction */
2020             zeta2            = _mm_mul_ps(beta2,rsq10);
2021             rinv3            = _mm_mul_ps(rinvsq10,rinv10);
2022             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2023             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2024             felec            = _mm_mul_ps(qq10,felec);
2025
2026             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
2027
2028             fscal            = felec;
2029
2030             fscal            = _mm_and_ps(fscal,cutoff_mask);
2031
2032             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2033
2034              /* Update vectorial force */
2035             fix1             = _mm_macc_ps(dx10,fscal,fix1);
2036             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
2037             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
2038
2039             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
2040             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
2041             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
2042
2043             }
2044
2045             /**************************
2046              * CALCULATE INTERACTIONS *
2047              **************************/
2048
2049             if (gmx_mm_any_lt(rsq11,rcutoff2))
2050             {
2051
2052             r11              = _mm_mul_ps(rsq11,rinv11);
2053             r11              = _mm_andnot_ps(dummy_mask,r11);
2054
2055             /* EWALD ELECTROSTATICS */
2056
2057             /* Analytical PME correction */
2058             zeta2            = _mm_mul_ps(beta2,rsq11);
2059             rinv3            = _mm_mul_ps(rinvsq11,rinv11);
2060             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2061             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2062             felec            = _mm_mul_ps(qq11,felec);
2063
2064             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
2065
2066             fscal            = felec;
2067
2068             fscal            = _mm_and_ps(fscal,cutoff_mask);
2069
2070             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2071
2072              /* Update vectorial force */
2073             fix1             = _mm_macc_ps(dx11,fscal,fix1);
2074             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
2075             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
2076
2077             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
2078             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
2079             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
2080
2081             }
2082
2083             /**************************
2084              * CALCULATE INTERACTIONS *
2085              **************************/
2086
2087             if (gmx_mm_any_lt(rsq12,rcutoff2))
2088             {
2089
2090             r12              = _mm_mul_ps(rsq12,rinv12);
2091             r12              = _mm_andnot_ps(dummy_mask,r12);
2092
2093             /* EWALD ELECTROSTATICS */
2094
2095             /* Analytical PME correction */
2096             zeta2            = _mm_mul_ps(beta2,rsq12);
2097             rinv3            = _mm_mul_ps(rinvsq12,rinv12);
2098             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2099             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2100             felec            = _mm_mul_ps(qq12,felec);
2101
2102             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
2103
2104             fscal            = felec;
2105
2106             fscal            = _mm_and_ps(fscal,cutoff_mask);
2107
2108             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2109
2110              /* Update vectorial force */
2111             fix1             = _mm_macc_ps(dx12,fscal,fix1);
2112             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
2113             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
2114
2115             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
2116             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
2117             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
2118
2119             }
2120
2121             /**************************
2122              * CALCULATE INTERACTIONS *
2123              **************************/
2124
2125             if (gmx_mm_any_lt(rsq20,rcutoff2))
2126             {
2127
2128             r20              = _mm_mul_ps(rsq20,rinv20);
2129             r20              = _mm_andnot_ps(dummy_mask,r20);
2130
2131             /* EWALD ELECTROSTATICS */
2132
2133             /* Analytical PME correction */
2134             zeta2            = _mm_mul_ps(beta2,rsq20);
2135             rinv3            = _mm_mul_ps(rinvsq20,rinv20);
2136             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2137             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2138             felec            = _mm_mul_ps(qq20,felec);
2139
2140             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
2141
2142             fscal            = felec;
2143
2144             fscal            = _mm_and_ps(fscal,cutoff_mask);
2145
2146             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2147
2148              /* Update vectorial force */
2149             fix2             = _mm_macc_ps(dx20,fscal,fix2);
2150             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
2151             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
2152
2153             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
2154             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
2155             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
2156
2157             }
2158
2159             /**************************
2160              * CALCULATE INTERACTIONS *
2161              **************************/
2162
2163             if (gmx_mm_any_lt(rsq21,rcutoff2))
2164             {
2165
2166             r21              = _mm_mul_ps(rsq21,rinv21);
2167             r21              = _mm_andnot_ps(dummy_mask,r21);
2168
2169             /* EWALD ELECTROSTATICS */
2170
2171             /* Analytical PME correction */
2172             zeta2            = _mm_mul_ps(beta2,rsq21);
2173             rinv3            = _mm_mul_ps(rinvsq21,rinv21);
2174             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2175             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2176             felec            = _mm_mul_ps(qq21,felec);
2177
2178             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
2179
2180             fscal            = felec;
2181
2182             fscal            = _mm_and_ps(fscal,cutoff_mask);
2183
2184             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2185
2186              /* Update vectorial force */
2187             fix2             = _mm_macc_ps(dx21,fscal,fix2);
2188             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
2189             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
2190
2191             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
2192             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
2193             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
2194
2195             }
2196
2197             /**************************
2198              * CALCULATE INTERACTIONS *
2199              **************************/
2200
2201             if (gmx_mm_any_lt(rsq22,rcutoff2))
2202             {
2203
2204             r22              = _mm_mul_ps(rsq22,rinv22);
2205             r22              = _mm_andnot_ps(dummy_mask,r22);
2206
2207             /* EWALD ELECTROSTATICS */
2208
2209             /* Analytical PME correction */
2210             zeta2            = _mm_mul_ps(beta2,rsq22);
2211             rinv3            = _mm_mul_ps(rinvsq22,rinv22);
2212             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
2213             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
2214             felec            = _mm_mul_ps(qq22,felec);
2215
2216             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2217
2218             fscal            = felec;
2219
2220             fscal            = _mm_and_ps(fscal,cutoff_mask);
2221
2222             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2223
2224              /* Update vectorial force */
2225             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2226             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2227             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2228
2229             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2230             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2231             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2232
2233             }
2234
2235             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2236             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2237             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2238             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2239
2240             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2241                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2242
2243             /* Inner loop uses 288 flops */
2244         }
2245
2246         /* End of innermost loop */
2247
2248         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2249                                               f+i_coord_offset,fshift+i_shift_offset);
2250
2251         /* Increment number of inner iterations */
2252         inneriter                  += j_index_end - j_index_start;
2253
2254         /* Outer loop uses 18 flops */
2255     }
2256
2257     /* Increment number of outer iterations */
2258     outeriter        += nri;
2259
2260     /* Update outer/inner flops */
2261
2262     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*288);
2263 }