src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwNone_GeomW3W3_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW3W3_VF_avx_128_fma_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            None
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwNone_GeomW3W3_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  91     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  93     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 103     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->ic->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = _mm_set1_ps(fr->ic->k_rf);
 123     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 124     crf              = _mm_set1_ps(fr->ic->c_rf);
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 129     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 130     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 131
 132     jq0              = _mm_set1_ps(charge[inr+0]);
 133     jq1              = _mm_set1_ps(charge[inr+1]);
 134     jq2              = _mm_set1_ps(charge[inr+2]);
 135     qq00             = _mm_mul_ps(iq0,jq0);
 136     qq01             = _mm_mul_ps(iq0,jq1);
 137     qq02             = _mm_mul_ps(iq0,jq2);
 138     qq10             = _mm_mul_ps(iq1,jq0);
 139     qq11             = _mm_mul_ps(iq1,jq1);
 140     qq12             = _mm_mul_ps(iq1,jq2);
 141     qq20             = _mm_mul_ps(iq2,jq0);
 142     qq21             = _mm_mul_ps(iq2,jq1);
 143     qq22             = _mm_mul_ps(iq2,jq2);
 144
 145     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 146     rcutoff_scalar   = fr->ic->rcoulomb;
 147     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 148     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 149
 150     /* Avoid stupid compiler warnings */
 151     jnrA = jnrB = jnrC = jnrD = 0;
 152     j_coord_offsetA = 0;
 153     j_coord_offsetB = 0;
 154     j_coord_offsetC = 0;
 155     j_coord_offsetD = 0;
 156
 157     outeriter        = 0;
 158     inneriter        = 0;
 159
 160     for(iidx=0;iidx<4*DIM;iidx++)
 161     {
 162         scratch[iidx] = 0.0;
 163     }
 164
 165     /* Start outer loop over neighborlists */
 166     for(iidx=0; iidx<nri; iidx++)
 167     {
 168         /* Load shift vector for this list */
 169         i_shift_offset   = DIM*shiftidx[iidx];
 170
 171         /* Load limits for loop over neighbors */
 172         j_index_start    = jindex[iidx];
 173         j_index_end      = jindex[iidx+1];
 174
 175         /* Get outer coordinate index */
 176         inr              = iinr[iidx];
 177         i_coord_offset   = DIM*inr;
 178
 179         /* Load i particle coords and add shift vector */
 180         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 181                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 182
 183         fix0             = _mm_setzero_ps();
 184         fiy0             = _mm_setzero_ps();
 185         fiz0             = _mm_setzero_ps();
 186         fix1             = _mm_setzero_ps();
 187         fiy1             = _mm_setzero_ps();
 188         fiz1             = _mm_setzero_ps();
 189         fix2             = _mm_setzero_ps();
 190         fiy2             = _mm_setzero_ps();
 191         fiz2             = _mm_setzero_ps();
 192
 193         /* Reset potential sums */
 194         velecsum         = _mm_setzero_ps();
 195
 196         /* Start inner kernel loop */
 197         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 198         {
 199
 200             /* Get j neighbor index, and coordinate index */
 201             jnrA             = jjnr[jidx];
 202             jnrB             = jjnr[jidx+1];
 203             jnrC             = jjnr[jidx+2];
 204             jnrD             = jjnr[jidx+3];
 205             j_coord_offsetA  = DIM*jnrA;
 206             j_coord_offsetB  = DIM*jnrB;
 207             j_coord_offsetC  = DIM*jnrC;
 208             j_coord_offsetD  = DIM*jnrD;
 209
 210             /* load j atom coordinates */
 211             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 212                                               x+j_coord_offsetC,x+j_coord_offsetD,
 213                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 214
 215             /* Calculate displacement vector */
 216             dx00             = _mm_sub_ps(ix0,jx0);
 217             dy00             = _mm_sub_ps(iy0,jy0);
 218             dz00             = _mm_sub_ps(iz0,jz0);
 219             dx01             = _mm_sub_ps(ix0,jx1);
 220             dy01             = _mm_sub_ps(iy0,jy1);
 221             dz01             = _mm_sub_ps(iz0,jz1);
 222             dx02             = _mm_sub_ps(ix0,jx2);
 223             dy02             = _mm_sub_ps(iy0,jy2);
 224             dz02             = _mm_sub_ps(iz0,jz2);
 225             dx10             = _mm_sub_ps(ix1,jx0);
 226             dy10             = _mm_sub_ps(iy1,jy0);
 227             dz10             = _mm_sub_ps(iz1,jz0);
 228             dx11             = _mm_sub_ps(ix1,jx1);
 229             dy11             = _mm_sub_ps(iy1,jy1);
 230             dz11             = _mm_sub_ps(iz1,jz1);
 231             dx12             = _mm_sub_ps(ix1,jx2);
 232             dy12             = _mm_sub_ps(iy1,jy2);
 233             dz12             = _mm_sub_ps(iz1,jz2);
 234             dx20             = _mm_sub_ps(ix2,jx0);
 235             dy20             = _mm_sub_ps(iy2,jy0);
 236             dz20             = _mm_sub_ps(iz2,jz0);
 237             dx21             = _mm_sub_ps(ix2,jx1);
 238             dy21             = _mm_sub_ps(iy2,jy1);
 239             dz21             = _mm_sub_ps(iz2,jz1);
 240             dx22             = _mm_sub_ps(ix2,jx2);
 241             dy22             = _mm_sub_ps(iy2,jy2);
 242             dz22             = _mm_sub_ps(iz2,jz2);
 243
 244             /* Calculate squared distance and things based on it */
 245             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 246             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 247             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 248             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 249             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 250             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 251             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 252             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 253             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 254
 255             rinv00           = avx128fma_invsqrt_f(rsq00);
 256             rinv01           = avx128fma_invsqrt_f(rsq01);
 257             rinv02           = avx128fma_invsqrt_f(rsq02);
 258             rinv10           = avx128fma_invsqrt_f(rsq10);
 259             rinv11           = avx128fma_invsqrt_f(rsq11);
 260             rinv12           = avx128fma_invsqrt_f(rsq12);
 261             rinv20           = avx128fma_invsqrt_f(rsq20);
 262             rinv21           = avx128fma_invsqrt_f(rsq21);
 263             rinv22           = avx128fma_invsqrt_f(rsq22);
 264
 265             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 266             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 267             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 268             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 269             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 270             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 271             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 272             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 273             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 274
 275             fjx0             = _mm_setzero_ps();
 276             fjy0             = _mm_setzero_ps();
 277             fjz0             = _mm_setzero_ps();
 278             fjx1             = _mm_setzero_ps();
 279             fjy1             = _mm_setzero_ps();
 280             fjz1             = _mm_setzero_ps();
 281             fjx2             = _mm_setzero_ps();
 282             fjy2             = _mm_setzero_ps();
 283             fjz2             = _mm_setzero_ps();
 284
 285             /**************************
 286              * CALCULATE INTERACTIONS *
 287              **************************/
 288
 289             if (gmx_mm_any_lt(rsq00,rcutoff2))
 290             {
 291
 292             /* REACTION-FIELD ELECTROSTATICS */
 293             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 294             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 295
 296             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 297
 298             /* Update potential sum for this i atom from the interaction with this j atom. */
 299             velec            = _mm_and_ps(velec,cutoff_mask);
 300             velecsum         = _mm_add_ps(velecsum,velec);
 301
 302             fscal            = felec;
 303
 304             fscal            = _mm_and_ps(fscal,cutoff_mask);
 305
 306              /* Update vectorial force */
 307             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 308             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 309             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 310
 311             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 312             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 313             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 314
 315             }
 316
 317             /**************************
 318              * CALCULATE INTERACTIONS *
 319              **************************/
 320
 321             if (gmx_mm_any_lt(rsq01,rcutoff2))
 322             {
 323
 324             /* REACTION-FIELD ELECTROSTATICS */
 325             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 326             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 327
 328             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 329
 330             /* Update potential sum for this i atom from the interaction with this j atom. */
 331             velec            = _mm_and_ps(velec,cutoff_mask);
 332             velecsum         = _mm_add_ps(velecsum,velec);
 333
 334             fscal            = felec;
 335
 336             fscal            = _mm_and_ps(fscal,cutoff_mask);
 337
 338              /* Update vectorial force */
 339             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 340             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 341             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 342
 343             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 344             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 345             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 346
 347             }
 348
 349             /**************************
 350              * CALCULATE INTERACTIONS *
 351              **************************/
 352
 353             if (gmx_mm_any_lt(rsq02,rcutoff2))
 354             {
 355
 356             /* REACTION-FIELD ELECTROSTATICS */
 357             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 358             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 359
 360             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 361
 362             /* Update potential sum for this i atom from the interaction with this j atom. */
 363             velec            = _mm_and_ps(velec,cutoff_mask);
 364             velecsum         = _mm_add_ps(velecsum,velec);
 365
 366             fscal            = felec;
 367
 368             fscal            = _mm_and_ps(fscal,cutoff_mask);
 369
 370              /* Update vectorial force */
 371             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 372             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 373             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 374
 375             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 376             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 377             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 378
 379             }
 380
 381             /**************************
 382              * CALCULATE INTERACTIONS *
 383              **************************/
 384
 385             if (gmx_mm_any_lt(rsq10,rcutoff2))
 386             {
 387
 388             /* REACTION-FIELD ELECTROSTATICS */
 389             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 390             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 391
 392             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 393
 394             /* Update potential sum for this i atom from the interaction with this j atom. */
 395             velec            = _mm_and_ps(velec,cutoff_mask);
 396             velecsum         = _mm_add_ps(velecsum,velec);
 397
 398             fscal            = felec;
 399
 400             fscal            = _mm_and_ps(fscal,cutoff_mask);
 401
 402              /* Update vectorial force */
 403             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 404             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 405             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 406
 407             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 408             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 409             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 410
 411             }
 412
 413             /**************************
 414              * CALCULATE INTERACTIONS *
 415              **************************/
 416
 417             if (gmx_mm_any_lt(rsq11,rcutoff2))
 418             {
 419
 420             /* REACTION-FIELD ELECTROSTATICS */
 421             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 422             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 423
 424             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 425
 426             /* Update potential sum for this i atom from the interaction with this j atom. */
 427             velec            = _mm_and_ps(velec,cutoff_mask);
 428             velecsum         = _mm_add_ps(velecsum,velec);
 429
 430             fscal            = felec;
 431
 432             fscal            = _mm_and_ps(fscal,cutoff_mask);
 433
 434              /* Update vectorial force */
 435             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 436             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 437             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 438
 439             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 440             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 441             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 442
 443             }
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             if (gmx_mm_any_lt(rsq12,rcutoff2))
 450             {
 451
 452             /* REACTION-FIELD ELECTROSTATICS */
 453             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 454             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 455
 456             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 457
 458             /* Update potential sum for this i atom from the interaction with this j atom. */
 459             velec            = _mm_and_ps(velec,cutoff_mask);
 460             velecsum         = _mm_add_ps(velecsum,velec);
 461
 462             fscal            = felec;
 463
 464             fscal            = _mm_and_ps(fscal,cutoff_mask);
 465
 466              /* Update vectorial force */
 467             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 468             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 469             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 470
 471             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 472             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 473             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 474
 475             }
 476
 477             /**************************
 478              * CALCULATE INTERACTIONS *
 479              **************************/
 480
 481             if (gmx_mm_any_lt(rsq20,rcutoff2))
 482             {
 483
 484             /* REACTION-FIELD ELECTROSTATICS */
 485             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 486             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 487
 488             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 489
 490             /* Update potential sum for this i atom from the interaction with this j atom. */
 491             velec            = _mm_and_ps(velec,cutoff_mask);
 492             velecsum         = _mm_add_ps(velecsum,velec);
 493
 494             fscal            = felec;
 495
 496             fscal            = _mm_and_ps(fscal,cutoff_mask);
 497
 498              /* Update vectorial force */
 499             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 500             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 501             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 502
 503             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 504             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 505             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 506
 507             }
 508
 509             /**************************
 510              * CALCULATE INTERACTIONS *
 511              **************************/
 512
 513             if (gmx_mm_any_lt(rsq21,rcutoff2))
 514             {
 515
 516             /* REACTION-FIELD ELECTROSTATICS */
 517             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 518             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 519
 520             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 521
 522             /* Update potential sum for this i atom from the interaction with this j atom. */
 523             velec            = _mm_and_ps(velec,cutoff_mask);
 524             velecsum         = _mm_add_ps(velecsum,velec);
 525
 526             fscal            = felec;
 527
 528             fscal            = _mm_and_ps(fscal,cutoff_mask);
 529
 530              /* Update vectorial force */
 531             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 532             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 533             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 534
 535             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 536             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 537             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 538
 539             }
 540
 541             /**************************
 542              * CALCULATE INTERACTIONS *
 543              **************************/
 544
 545             if (gmx_mm_any_lt(rsq22,rcutoff2))
 546             {
 547
 548             /* REACTION-FIELD ELECTROSTATICS */
 549             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 550             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 551
 552             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 553
 554             /* Update potential sum for this i atom from the interaction with this j atom. */
 555             velec            = _mm_and_ps(velec,cutoff_mask);
 556             velecsum         = _mm_add_ps(velecsum,velec);
 557
 558             fscal            = felec;
 559
 560             fscal            = _mm_and_ps(fscal,cutoff_mask);
 561
 562              /* Update vectorial force */
 563             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 564             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 565             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 566
 567             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 568             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 569             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 570
 571             }
 572
 573             fjptrA             = f+j_coord_offsetA;
 574             fjptrB             = f+j_coord_offsetB;
 575             fjptrC             = f+j_coord_offsetC;
 576             fjptrD             = f+j_coord_offsetD;
 577
 578             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 579                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 580
 581             /* Inner loop uses 351 flops */
 582         }
 583
 584         if(jidx<j_index_end)
 585         {
 586
 587             /* Get j neighbor index, and coordinate index */
 588             jnrlistA         = jjnr[jidx];
 589             jnrlistB         = jjnr[jidx+1];
 590             jnrlistC         = jjnr[jidx+2];
 591             jnrlistD         = jjnr[jidx+3];
 592             /* Sign of each element will be negative for non-real atoms.
 593              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 594              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 595              */
 596             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 597             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 598             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 599             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 600             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 601             j_coord_offsetA  = DIM*jnrA;
 602             j_coord_offsetB  = DIM*jnrB;
 603             j_coord_offsetC  = DIM*jnrC;
 604             j_coord_offsetD  = DIM*jnrD;
 605
 606             /* load j atom coordinates */
 607             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 608                                               x+j_coord_offsetC,x+j_coord_offsetD,
 609                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 610
 611             /* Calculate displacement vector */
 612             dx00             = _mm_sub_ps(ix0,jx0);
 613             dy00             = _mm_sub_ps(iy0,jy0);
 614             dz00             = _mm_sub_ps(iz0,jz0);
 615             dx01             = _mm_sub_ps(ix0,jx1);
 616             dy01             = _mm_sub_ps(iy0,jy1);
 617             dz01             = _mm_sub_ps(iz0,jz1);
 618             dx02             = _mm_sub_ps(ix0,jx2);
 619             dy02             = _mm_sub_ps(iy0,jy2);
 620             dz02             = _mm_sub_ps(iz0,jz2);
 621             dx10             = _mm_sub_ps(ix1,jx0);
 622             dy10             = _mm_sub_ps(iy1,jy0);
 623             dz10             = _mm_sub_ps(iz1,jz0);
 624             dx11             = _mm_sub_ps(ix1,jx1);
 625             dy11             = _mm_sub_ps(iy1,jy1);
 626             dz11             = _mm_sub_ps(iz1,jz1);
 627             dx12             = _mm_sub_ps(ix1,jx2);
 628             dy12             = _mm_sub_ps(iy1,jy2);
 629             dz12             = _mm_sub_ps(iz1,jz2);
 630             dx20             = _mm_sub_ps(ix2,jx0);
 631             dy20             = _mm_sub_ps(iy2,jy0);
 632             dz20             = _mm_sub_ps(iz2,jz0);
 633             dx21             = _mm_sub_ps(ix2,jx1);
 634             dy21             = _mm_sub_ps(iy2,jy1);
 635             dz21             = _mm_sub_ps(iz2,jz1);
 636             dx22             = _mm_sub_ps(ix2,jx2);
 637             dy22             = _mm_sub_ps(iy2,jy2);
 638             dz22             = _mm_sub_ps(iz2,jz2);
 639
 640             /* Calculate squared distance and things based on it */
 641             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 642             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 643             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 644             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 645             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 646             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 647             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 648             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 649             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 650
 651             rinv00           = avx128fma_invsqrt_f(rsq00);
 652             rinv01           = avx128fma_invsqrt_f(rsq01);
 653             rinv02           = avx128fma_invsqrt_f(rsq02);
 654             rinv10           = avx128fma_invsqrt_f(rsq10);
 655             rinv11           = avx128fma_invsqrt_f(rsq11);
 656             rinv12           = avx128fma_invsqrt_f(rsq12);
 657             rinv20           = avx128fma_invsqrt_f(rsq20);
 658             rinv21           = avx128fma_invsqrt_f(rsq21);
 659             rinv22           = avx128fma_invsqrt_f(rsq22);
 660
 661             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 662             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 663             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 664             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 665             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 666             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 667             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 668             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 669             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 670
 671             fjx0             = _mm_setzero_ps();
 672             fjy0             = _mm_setzero_ps();
 673             fjz0             = _mm_setzero_ps();
 674             fjx1             = _mm_setzero_ps();
 675             fjy1             = _mm_setzero_ps();
 676             fjz1             = _mm_setzero_ps();
 677             fjx2             = _mm_setzero_ps();
 678             fjy2             = _mm_setzero_ps();
 679             fjz2             = _mm_setzero_ps();
 680
 681             /**************************
 682              * CALCULATE INTERACTIONS *
 683              **************************/
 684
 685             if (gmx_mm_any_lt(rsq00,rcutoff2))
 686             {
 687
 688             /* REACTION-FIELD ELECTROSTATICS */
 689             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 690             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 691
 692             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 693
 694             /* Update potential sum for this i atom from the interaction with this j atom. */
 695             velec            = _mm_and_ps(velec,cutoff_mask);
 696             velec            = _mm_andnot_ps(dummy_mask,velec);
 697             velecsum         = _mm_add_ps(velecsum,velec);
 698
 699             fscal            = felec;
 700
 701             fscal            = _mm_and_ps(fscal,cutoff_mask);
 702
 703             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 704
 705              /* Update vectorial force */
 706             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 707             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 708             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 709
 710             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 711             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 712             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 713
 714             }
 715
 716             /**************************
 717              * CALCULATE INTERACTIONS *
 718              **************************/
 719
 720             if (gmx_mm_any_lt(rsq01,rcutoff2))
 721             {
 722
 723             /* REACTION-FIELD ELECTROSTATICS */
 724             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 725             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 726
 727             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 728
 729             /* Update potential sum for this i atom from the interaction with this j atom. */
 730             velec            = _mm_and_ps(velec,cutoff_mask);
 731             velec            = _mm_andnot_ps(dummy_mask,velec);
 732             velecsum         = _mm_add_ps(velecsum,velec);
 733
 734             fscal            = felec;
 735
 736             fscal            = _mm_and_ps(fscal,cutoff_mask);
 737
 738             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 739
 740              /* Update vectorial force */
 741             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 742             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 743             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 744
 745             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 746             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 747             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 748
 749             }
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             if (gmx_mm_any_lt(rsq02,rcutoff2))
 756             {
 757
 758             /* REACTION-FIELD ELECTROSTATICS */
 759             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 760             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 761
 762             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 763
 764             /* Update potential sum for this i atom from the interaction with this j atom. */
 765             velec            = _mm_and_ps(velec,cutoff_mask);
 766             velec            = _mm_andnot_ps(dummy_mask,velec);
 767             velecsum         = _mm_add_ps(velecsum,velec);
 768
 769             fscal            = felec;
 770
 771             fscal            = _mm_and_ps(fscal,cutoff_mask);
 772
 773             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 774
 775              /* Update vectorial force */
 776             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 777             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 778             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 779
 780             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 781             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 782             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 783
 784             }
 785
 786             /**************************
 787              * CALCULATE INTERACTIONS *
 788              **************************/
 789
 790             if (gmx_mm_any_lt(rsq10,rcutoff2))
 791             {
 792
 793             /* REACTION-FIELD ELECTROSTATICS */
 794             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 795             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 796
 797             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 798
 799             /* Update potential sum for this i atom from the interaction with this j atom. */
 800             velec            = _mm_and_ps(velec,cutoff_mask);
 801             velec            = _mm_andnot_ps(dummy_mask,velec);
 802             velecsum         = _mm_add_ps(velecsum,velec);
 803
 804             fscal            = felec;
 805
 806             fscal            = _mm_and_ps(fscal,cutoff_mask);
 807
 808             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 809
 810              /* Update vectorial force */
 811             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 812             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 813             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 814
 815             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 816             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 817             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 818
 819             }
 820
 821             /**************************
 822              * CALCULATE INTERACTIONS *
 823              **************************/
 824
 825             if (gmx_mm_any_lt(rsq11,rcutoff2))
 826             {
 827
 828             /* REACTION-FIELD ELECTROSTATICS */
 829             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 830             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 831
 832             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 833
 834             /* Update potential sum for this i atom from the interaction with this j atom. */
 835             velec            = _mm_and_ps(velec,cutoff_mask);
 836             velec            = _mm_andnot_ps(dummy_mask,velec);
 837             velecsum         = _mm_add_ps(velecsum,velec);
 838
 839             fscal            = felec;
 840
 841             fscal            = _mm_and_ps(fscal,cutoff_mask);
 842
 843             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 844
 845              /* Update vectorial force */
 846             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 847             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 848             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 849
 850             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 851             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 852             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 853
 854             }
 855
 856             /**************************
 857              * CALCULATE INTERACTIONS *
 858              **************************/
 859
 860             if (gmx_mm_any_lt(rsq12,rcutoff2))
 861             {
 862
 863             /* REACTION-FIELD ELECTROSTATICS */
 864             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 865             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 866
 867             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 868
 869             /* Update potential sum for this i atom from the interaction with this j atom. */
 870             velec            = _mm_and_ps(velec,cutoff_mask);
 871             velec            = _mm_andnot_ps(dummy_mask,velec);
 872             velecsum         = _mm_add_ps(velecsum,velec);
 873
 874             fscal            = felec;
 875
 876             fscal            = _mm_and_ps(fscal,cutoff_mask);
 877
 878             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 879
 880              /* Update vectorial force */
 881             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 882             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 883             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 884
 885             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 886             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 887             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 888
 889             }
 890
 891             /**************************
 892              * CALCULATE INTERACTIONS *
 893              **************************/
 894
 895             if (gmx_mm_any_lt(rsq20,rcutoff2))
 896             {
 897
 898             /* REACTION-FIELD ELECTROSTATICS */
 899             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 900             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 901
 902             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 903
 904             /* Update potential sum for this i atom from the interaction with this j atom. */
 905             velec            = _mm_and_ps(velec,cutoff_mask);
 906             velec            = _mm_andnot_ps(dummy_mask,velec);
 907             velecsum         = _mm_add_ps(velecsum,velec);
 908
 909             fscal            = felec;
 910
 911             fscal            = _mm_and_ps(fscal,cutoff_mask);
 912
 913             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 914
 915              /* Update vectorial force */
 916             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 917             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 918             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 919
 920             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 921             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 922             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 923
 924             }
 925
 926             /**************************
 927              * CALCULATE INTERACTIONS *
 928              **************************/
 929
 930             if (gmx_mm_any_lt(rsq21,rcutoff2))
 931             {
 932
 933             /* REACTION-FIELD ELECTROSTATICS */
 934             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 935             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 936
 937             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 938
 939             /* Update potential sum for this i atom from the interaction with this j atom. */
 940             velec            = _mm_and_ps(velec,cutoff_mask);
 941             velec            = _mm_andnot_ps(dummy_mask,velec);
 942             velecsum         = _mm_add_ps(velecsum,velec);
 943
 944             fscal            = felec;
 945
 946             fscal            = _mm_and_ps(fscal,cutoff_mask);
 947
 948             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 949
 950              /* Update vectorial force */
 951             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 952             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 953             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 954
 955             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 956             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 957             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 958
 959             }
 960
 961             /**************************
 962              * CALCULATE INTERACTIONS *
 963              **************************/
 964
 965             if (gmx_mm_any_lt(rsq22,rcutoff2))
 966             {
 967
 968             /* REACTION-FIELD ELECTROSTATICS */
 969             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 970             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 971
 972             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 973
 974             /* Update potential sum for this i atom from the interaction with this j atom. */
 975             velec            = _mm_and_ps(velec,cutoff_mask);
 976             velec            = _mm_andnot_ps(dummy_mask,velec);
 977             velecsum         = _mm_add_ps(velecsum,velec);
 978
 979             fscal            = felec;
 980
 981             fscal            = _mm_and_ps(fscal,cutoff_mask);
 982
 983             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 984
 985              /* Update vectorial force */
 986             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 987             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 988             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 989
 990             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 991             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 992             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 993
 994             }
 995
 996             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 997             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 998             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 999             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1000
1001             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1002                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1003
1004             /* Inner loop uses 351 flops */
1005         }
1006
1007         /* End of innermost loop */
1008
1009         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1010                                               f+i_coord_offset,fshift+i_shift_offset);
1011
1012         ggid                        = gid[iidx];
1013         /* Update potential energies */
1014         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1015
1016         /* Increment number of inner iterations */
1017         inneriter                  += j_index_end - j_index_start;
1018
1019         /* Outer loop uses 19 flops */
1020     }
1021
1022     /* Increment number of outer iterations */
1023     outeriter        += nri;
1024
1025     /* Update outer/inner flops */
1026
1027     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*351);
1028 }
1029 /*
1030  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_avx_128_fma_single
1031  * Electrostatics interaction: ReactionField
1032  * VdW interaction:            None
1033  * Geometry:                   Water3-Water3
1034  * Calculate force/pot:        Force
1035  */
1036 void
1037 nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_avx_128_fma_single
1038                     (t_nblist                    * gmx_restrict       nlist,
1039                      rvec                        * gmx_restrict          xx,
1040                      rvec                        * gmx_restrict          ff,
1041                      struct t_forcerec           * gmx_restrict          fr,
1042                      t_mdatoms                   * gmx_restrict     mdatoms,
1043                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1044                      t_nrnb                      * gmx_restrict        nrnb)
1045 {
1046     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1047      * just 0 for non-waters.
1048      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1049      * jnr indices corresponding to data put in the four positions in the SIMD register.
1050      */
1051     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1052     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1053     int              jnrA,jnrB,jnrC,jnrD;
1054     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1055     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1056     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1057     real             rcutoff_scalar;
1058     real             *shiftvec,*fshift,*x,*f;
1059     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1060     real             scratch[4*DIM];
1061     __m128           fscal,rcutoff,rcutoff2,jidxall;
1062     int              vdwioffset0;
1063     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1064     int              vdwioffset1;
1065     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1066     int              vdwioffset2;
1067     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1068     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1069     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1070     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1071     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1072     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1073     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1074     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1075     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1076     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1077     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1078     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1079     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1080     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1081     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1082     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1083     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1084     real             *charge;
1085     __m128           dummy_mask,cutoff_mask;
1086     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1087     __m128           one     = _mm_set1_ps(1.0);
1088     __m128           two     = _mm_set1_ps(2.0);
1089     x                = xx[0];
1090     f                = ff[0];
1091
1092     nri              = nlist->nri;
1093     iinr             = nlist->iinr;
1094     jindex           = nlist->jindex;
1095     jjnr             = nlist->jjnr;
1096     shiftidx         = nlist->shift;
1097     gid              = nlist->gid;
1098     shiftvec         = fr->shift_vec[0];
1099     fshift           = fr->fshift[0];
1100     facel            = _mm_set1_ps(fr->ic->epsfac);
1101     charge           = mdatoms->chargeA;
1102     krf              = _mm_set1_ps(fr->ic->k_rf);
1103     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1104     crf              = _mm_set1_ps(fr->ic->c_rf);
1105
1106     /* Setup water-specific parameters */
1107     inr              = nlist->iinr[0];
1108     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1109     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1110     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1111
1112     jq0              = _mm_set1_ps(charge[inr+0]);
1113     jq1              = _mm_set1_ps(charge[inr+1]);
1114     jq2              = _mm_set1_ps(charge[inr+2]);
1115     qq00             = _mm_mul_ps(iq0,jq0);
1116     qq01             = _mm_mul_ps(iq0,jq1);
1117     qq02             = _mm_mul_ps(iq0,jq2);
1118     qq10             = _mm_mul_ps(iq1,jq0);
1119     qq11             = _mm_mul_ps(iq1,jq1);
1120     qq12             = _mm_mul_ps(iq1,jq2);
1121     qq20             = _mm_mul_ps(iq2,jq0);
1122     qq21             = _mm_mul_ps(iq2,jq1);
1123     qq22             = _mm_mul_ps(iq2,jq2);
1124
1125     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1126     rcutoff_scalar   = fr->ic->rcoulomb;
1127     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1128     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1129
1130     /* Avoid stupid compiler warnings */
1131     jnrA = jnrB = jnrC = jnrD = 0;
1132     j_coord_offsetA = 0;
1133     j_coord_offsetB = 0;
1134     j_coord_offsetC = 0;
1135     j_coord_offsetD = 0;
1136
1137     outeriter        = 0;
1138     inneriter        = 0;
1139
1140     for(iidx=0;iidx<4*DIM;iidx++)
1141     {
1142         scratch[iidx] = 0.0;
1143     }
1144
1145     /* Start outer loop over neighborlists */
1146     for(iidx=0; iidx<nri; iidx++)
1147     {
1148         /* Load shift vector for this list */
1149         i_shift_offset   = DIM*shiftidx[iidx];
1150
1151         /* Load limits for loop over neighbors */
1152         j_index_start    = jindex[iidx];
1153         j_index_end      = jindex[iidx+1];
1154
1155         /* Get outer coordinate index */
1156         inr              = iinr[iidx];
1157         i_coord_offset   = DIM*inr;
1158
1159         /* Load i particle coords and add shift vector */
1160         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1161                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1162
1163         fix0             = _mm_setzero_ps();
1164         fiy0             = _mm_setzero_ps();
1165         fiz0             = _mm_setzero_ps();
1166         fix1             = _mm_setzero_ps();
1167         fiy1             = _mm_setzero_ps();
1168         fiz1             = _mm_setzero_ps();
1169         fix2             = _mm_setzero_ps();
1170         fiy2             = _mm_setzero_ps();
1171         fiz2             = _mm_setzero_ps();
1172
1173         /* Start inner kernel loop */
1174         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1175         {
1176
1177             /* Get j neighbor index, and coordinate index */
1178             jnrA             = jjnr[jidx];
1179             jnrB             = jjnr[jidx+1];
1180             jnrC             = jjnr[jidx+2];
1181             jnrD             = jjnr[jidx+3];
1182             j_coord_offsetA  = DIM*jnrA;
1183             j_coord_offsetB  = DIM*jnrB;
1184             j_coord_offsetC  = DIM*jnrC;
1185             j_coord_offsetD  = DIM*jnrD;
1186
1187             /* load j atom coordinates */
1188             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1189                                               x+j_coord_offsetC,x+j_coord_offsetD,
1190                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1191
1192             /* Calculate displacement vector */
1193             dx00             = _mm_sub_ps(ix0,jx0);
1194             dy00             = _mm_sub_ps(iy0,jy0);
1195             dz00             = _mm_sub_ps(iz0,jz0);
1196             dx01             = _mm_sub_ps(ix0,jx1);
1197             dy01             = _mm_sub_ps(iy0,jy1);
1198             dz01             = _mm_sub_ps(iz0,jz1);
1199             dx02             = _mm_sub_ps(ix0,jx2);
1200             dy02             = _mm_sub_ps(iy0,jy2);
1201             dz02             = _mm_sub_ps(iz0,jz2);
1202             dx10             = _mm_sub_ps(ix1,jx0);
1203             dy10             = _mm_sub_ps(iy1,jy0);
1204             dz10             = _mm_sub_ps(iz1,jz0);
1205             dx11             = _mm_sub_ps(ix1,jx1);
1206             dy11             = _mm_sub_ps(iy1,jy1);
1207             dz11             = _mm_sub_ps(iz1,jz1);
1208             dx12             = _mm_sub_ps(ix1,jx2);
1209             dy12             = _mm_sub_ps(iy1,jy2);
1210             dz12             = _mm_sub_ps(iz1,jz2);
1211             dx20             = _mm_sub_ps(ix2,jx0);
1212             dy20             = _mm_sub_ps(iy2,jy0);
1213             dz20             = _mm_sub_ps(iz2,jz0);
1214             dx21             = _mm_sub_ps(ix2,jx1);
1215             dy21             = _mm_sub_ps(iy2,jy1);
1216             dz21             = _mm_sub_ps(iz2,jz1);
1217             dx22             = _mm_sub_ps(ix2,jx2);
1218             dy22             = _mm_sub_ps(iy2,jy2);
1219             dz22             = _mm_sub_ps(iz2,jz2);
1220
1221             /* Calculate squared distance and things based on it */
1222             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1223             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1224             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1225             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1226             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1227             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1228             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1229             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1230             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1231
1232             rinv00           = avx128fma_invsqrt_f(rsq00);
1233             rinv01           = avx128fma_invsqrt_f(rsq01);
1234             rinv02           = avx128fma_invsqrt_f(rsq02);
1235             rinv10           = avx128fma_invsqrt_f(rsq10);
1236             rinv11           = avx128fma_invsqrt_f(rsq11);
1237             rinv12           = avx128fma_invsqrt_f(rsq12);
1238             rinv20           = avx128fma_invsqrt_f(rsq20);
1239             rinv21           = avx128fma_invsqrt_f(rsq21);
1240             rinv22           = avx128fma_invsqrt_f(rsq22);
1241
1242             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1243             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1244             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1245             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1246             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1247             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1248             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1249             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1250             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1251
1252             fjx0             = _mm_setzero_ps();
1253             fjy0             = _mm_setzero_ps();
1254             fjz0             = _mm_setzero_ps();
1255             fjx1             = _mm_setzero_ps();
1256             fjy1             = _mm_setzero_ps();
1257             fjz1             = _mm_setzero_ps();
1258             fjx2             = _mm_setzero_ps();
1259             fjy2             = _mm_setzero_ps();
1260             fjz2             = _mm_setzero_ps();
1261
1262             /**************************
1263              * CALCULATE INTERACTIONS *
1264              **************************/
1265
1266             if (gmx_mm_any_lt(rsq00,rcutoff2))
1267             {
1268
1269             /* REACTION-FIELD ELECTROSTATICS */
1270             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1271
1272             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1273
1274             fscal            = felec;
1275
1276             fscal            = _mm_and_ps(fscal,cutoff_mask);
1277
1278              /* Update vectorial force */
1279             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1280             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1281             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1282
1283             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1284             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1285             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1286
1287             }
1288
1289             /**************************
1290              * CALCULATE INTERACTIONS *
1291              **************************/
1292
1293             if (gmx_mm_any_lt(rsq01,rcutoff2))
1294             {
1295
1296             /* REACTION-FIELD ELECTROSTATICS */
1297             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1298
1299             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1300
1301             fscal            = felec;
1302
1303             fscal            = _mm_and_ps(fscal,cutoff_mask);
1304
1305              /* Update vectorial force */
1306             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1307             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1308             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1309
1310             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1311             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1312             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1313
1314             }
1315
1316             /**************************
1317              * CALCULATE INTERACTIONS *
1318              **************************/
1319
1320             if (gmx_mm_any_lt(rsq02,rcutoff2))
1321             {
1322
1323             /* REACTION-FIELD ELECTROSTATICS */
1324             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1325
1326             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1327
1328             fscal            = felec;
1329
1330             fscal            = _mm_and_ps(fscal,cutoff_mask);
1331
1332              /* Update vectorial force */
1333             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1334             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1335             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1336
1337             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1338             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1339             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1340
1341             }
1342
1343             /**************************
1344              * CALCULATE INTERACTIONS *
1345              **************************/
1346
1347             if (gmx_mm_any_lt(rsq10,rcutoff2))
1348             {
1349
1350             /* REACTION-FIELD ELECTROSTATICS */
1351             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1352
1353             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1354
1355             fscal            = felec;
1356
1357             fscal            = _mm_and_ps(fscal,cutoff_mask);
1358
1359              /* Update vectorial force */
1360             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1361             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1362             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1363
1364             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1365             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1366             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1367
1368             }
1369
1370             /**************************
1371              * CALCULATE INTERACTIONS *
1372              **************************/
1373
1374             if (gmx_mm_any_lt(rsq11,rcutoff2))
1375             {
1376
1377             /* REACTION-FIELD ELECTROSTATICS */
1378             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1379
1380             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1381
1382             fscal            = felec;
1383
1384             fscal            = _mm_and_ps(fscal,cutoff_mask);
1385
1386              /* Update vectorial force */
1387             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1388             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1389             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1390
1391             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1392             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1393             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1394
1395             }
1396
1397             /**************************
1398              * CALCULATE INTERACTIONS *
1399              **************************/
1400
1401             if (gmx_mm_any_lt(rsq12,rcutoff2))
1402             {
1403
1404             /* REACTION-FIELD ELECTROSTATICS */
1405             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1406
1407             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1408
1409             fscal            = felec;
1410
1411             fscal            = _mm_and_ps(fscal,cutoff_mask);
1412
1413              /* Update vectorial force */
1414             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1415             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1416             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1417
1418             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1419             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1420             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1421
1422             }
1423
1424             /**************************
1425              * CALCULATE INTERACTIONS *
1426              **************************/
1427
1428             if (gmx_mm_any_lt(rsq20,rcutoff2))
1429             {
1430
1431             /* REACTION-FIELD ELECTROSTATICS */
1432             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1433
1434             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1435
1436             fscal            = felec;
1437
1438             fscal            = _mm_and_ps(fscal,cutoff_mask);
1439
1440              /* Update vectorial force */
1441             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1442             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1443             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1444
1445             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1446             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1447             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1448
1449             }
1450
1451             /**************************
1452              * CALCULATE INTERACTIONS *
1453              **************************/
1454
1455             if (gmx_mm_any_lt(rsq21,rcutoff2))
1456             {
1457
1458             /* REACTION-FIELD ELECTROSTATICS */
1459             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1460
1461             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1462
1463             fscal            = felec;
1464
1465             fscal            = _mm_and_ps(fscal,cutoff_mask);
1466
1467              /* Update vectorial force */
1468             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1469             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1470             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1471
1472             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1473             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1474             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1475
1476             }
1477
1478             /**************************
1479              * CALCULATE INTERACTIONS *
1480              **************************/
1481
1482             if (gmx_mm_any_lt(rsq22,rcutoff2))
1483             {
1484
1485             /* REACTION-FIELD ELECTROSTATICS */
1486             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1487
1488             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1489
1490             fscal            = felec;
1491
1492             fscal            = _mm_and_ps(fscal,cutoff_mask);
1493
1494              /* Update vectorial force */
1495             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1496             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1497             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1498
1499             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1500             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1501             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1502
1503             }
1504
1505             fjptrA             = f+j_coord_offsetA;
1506             fjptrB             = f+j_coord_offsetB;
1507             fjptrC             = f+j_coord_offsetC;
1508             fjptrD             = f+j_coord_offsetD;
1509
1510             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1511                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1512
1513             /* Inner loop uses 297 flops */
1514         }
1515
1516         if(jidx<j_index_end)
1517         {
1518
1519             /* Get j neighbor index, and coordinate index */
1520             jnrlistA         = jjnr[jidx];
1521             jnrlistB         = jjnr[jidx+1];
1522             jnrlistC         = jjnr[jidx+2];
1523             jnrlistD         = jjnr[jidx+3];
1524             /* Sign of each element will be negative for non-real atoms.
1525              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1526              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1527              */
1528             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1529             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1530             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1531             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1532             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1533             j_coord_offsetA  = DIM*jnrA;
1534             j_coord_offsetB  = DIM*jnrB;
1535             j_coord_offsetC  = DIM*jnrC;
1536             j_coord_offsetD  = DIM*jnrD;
1537
1538             /* load j atom coordinates */
1539             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1540                                               x+j_coord_offsetC,x+j_coord_offsetD,
1541                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1542
1543             /* Calculate displacement vector */
1544             dx00             = _mm_sub_ps(ix0,jx0);
1545             dy00             = _mm_sub_ps(iy0,jy0);
1546             dz00             = _mm_sub_ps(iz0,jz0);
1547             dx01             = _mm_sub_ps(ix0,jx1);
1548             dy01             = _mm_sub_ps(iy0,jy1);
1549             dz01             = _mm_sub_ps(iz0,jz1);
1550             dx02             = _mm_sub_ps(ix0,jx2);
1551             dy02             = _mm_sub_ps(iy0,jy2);
1552             dz02             = _mm_sub_ps(iz0,jz2);
1553             dx10             = _mm_sub_ps(ix1,jx0);
1554             dy10             = _mm_sub_ps(iy1,jy0);
1555             dz10             = _mm_sub_ps(iz1,jz0);
1556             dx11             = _mm_sub_ps(ix1,jx1);
1557             dy11             = _mm_sub_ps(iy1,jy1);
1558             dz11             = _mm_sub_ps(iz1,jz1);
1559             dx12             = _mm_sub_ps(ix1,jx2);
1560             dy12             = _mm_sub_ps(iy1,jy2);
1561             dz12             = _mm_sub_ps(iz1,jz2);
1562             dx20             = _mm_sub_ps(ix2,jx0);
1563             dy20             = _mm_sub_ps(iy2,jy0);
1564             dz20             = _mm_sub_ps(iz2,jz0);
1565             dx21             = _mm_sub_ps(ix2,jx1);
1566             dy21             = _mm_sub_ps(iy2,jy1);
1567             dz21             = _mm_sub_ps(iz2,jz1);
1568             dx22             = _mm_sub_ps(ix2,jx2);
1569             dy22             = _mm_sub_ps(iy2,jy2);
1570             dz22             = _mm_sub_ps(iz2,jz2);
1571
1572             /* Calculate squared distance and things based on it */
1573             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1574             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1575             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1576             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1577             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1578             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1579             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1580             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1581             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1582
1583             rinv00           = avx128fma_invsqrt_f(rsq00);
1584             rinv01           = avx128fma_invsqrt_f(rsq01);
1585             rinv02           = avx128fma_invsqrt_f(rsq02);
1586             rinv10           = avx128fma_invsqrt_f(rsq10);
1587             rinv11           = avx128fma_invsqrt_f(rsq11);
1588             rinv12           = avx128fma_invsqrt_f(rsq12);
1589             rinv20           = avx128fma_invsqrt_f(rsq20);
1590             rinv21           = avx128fma_invsqrt_f(rsq21);
1591             rinv22           = avx128fma_invsqrt_f(rsq22);
1592
1593             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1594             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1595             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1596             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1597             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1598             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1599             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1600             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1601             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1602
1603             fjx0             = _mm_setzero_ps();
1604             fjy0             = _mm_setzero_ps();
1605             fjz0             = _mm_setzero_ps();
1606             fjx1             = _mm_setzero_ps();
1607             fjy1             = _mm_setzero_ps();
1608             fjz1             = _mm_setzero_ps();
1609             fjx2             = _mm_setzero_ps();
1610             fjy2             = _mm_setzero_ps();
1611             fjz2             = _mm_setzero_ps();
1612
1613             /**************************
1614              * CALCULATE INTERACTIONS *
1615              **************************/
1616
1617             if (gmx_mm_any_lt(rsq00,rcutoff2))
1618             {
1619
1620             /* REACTION-FIELD ELECTROSTATICS */
1621             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1622
1623             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1624
1625             fscal            = felec;
1626
1627             fscal            = _mm_and_ps(fscal,cutoff_mask);
1628
1629             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1630
1631              /* Update vectorial force */
1632             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1633             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1634             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1635
1636             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1637             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1638             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1639
1640             }
1641
1642             /**************************
1643              * CALCULATE INTERACTIONS *
1644              **************************/
1645
1646             if (gmx_mm_any_lt(rsq01,rcutoff2))
1647             {
1648
1649             /* REACTION-FIELD ELECTROSTATICS */
1650             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1651
1652             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1653
1654             fscal            = felec;
1655
1656             fscal            = _mm_and_ps(fscal,cutoff_mask);
1657
1658             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1659
1660              /* Update vectorial force */
1661             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1662             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1663             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1664
1665             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1666             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1667             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1668
1669             }
1670
1671             /**************************
1672              * CALCULATE INTERACTIONS *
1673              **************************/
1674
1675             if (gmx_mm_any_lt(rsq02,rcutoff2))
1676             {
1677
1678             /* REACTION-FIELD ELECTROSTATICS */
1679             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1680
1681             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1682
1683             fscal            = felec;
1684
1685             fscal            = _mm_and_ps(fscal,cutoff_mask);
1686
1687             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1688
1689              /* Update vectorial force */
1690             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1691             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1692             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1693
1694             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1695             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1696             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1697
1698             }
1699
1700             /**************************
1701              * CALCULATE INTERACTIONS *
1702              **************************/
1703
1704             if (gmx_mm_any_lt(rsq10,rcutoff2))
1705             {
1706
1707             /* REACTION-FIELD ELECTROSTATICS */
1708             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1709
1710             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1711
1712             fscal            = felec;
1713
1714             fscal            = _mm_and_ps(fscal,cutoff_mask);
1715
1716             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1717
1718              /* Update vectorial force */
1719             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1720             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1721             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1722
1723             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1724             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1725             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1726
1727             }
1728
1729             /**************************
1730              * CALCULATE INTERACTIONS *
1731              **************************/
1732
1733             if (gmx_mm_any_lt(rsq11,rcutoff2))
1734             {
1735
1736             /* REACTION-FIELD ELECTROSTATICS */
1737             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1738
1739             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1740
1741             fscal            = felec;
1742
1743             fscal            = _mm_and_ps(fscal,cutoff_mask);
1744
1745             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1746
1747              /* Update vectorial force */
1748             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1749             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1750             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1751
1752             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1753             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1754             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1755
1756             }
1757
1758             /**************************
1759              * CALCULATE INTERACTIONS *
1760              **************************/
1761
1762             if (gmx_mm_any_lt(rsq12,rcutoff2))
1763             {
1764
1765             /* REACTION-FIELD ELECTROSTATICS */
1766             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1767
1768             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1769
1770             fscal            = felec;
1771
1772             fscal            = _mm_and_ps(fscal,cutoff_mask);
1773
1774             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1775
1776              /* Update vectorial force */
1777             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1778             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1779             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1780
1781             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1782             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1783             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1784
1785             }
1786
1787             /**************************
1788              * CALCULATE INTERACTIONS *
1789              **************************/
1790
1791             if (gmx_mm_any_lt(rsq20,rcutoff2))
1792             {
1793
1794             /* REACTION-FIELD ELECTROSTATICS */
1795             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1796
1797             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1798
1799             fscal            = felec;
1800
1801             fscal            = _mm_and_ps(fscal,cutoff_mask);
1802
1803             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1804
1805              /* Update vectorial force */
1806             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1807             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1808             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1809
1810             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1811             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1812             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1813
1814             }
1815
1816             /**************************
1817              * CALCULATE INTERACTIONS *
1818              **************************/
1819
1820             if (gmx_mm_any_lt(rsq21,rcutoff2))
1821             {
1822
1823             /* REACTION-FIELD ELECTROSTATICS */
1824             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1825
1826             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1827
1828             fscal            = felec;
1829
1830             fscal            = _mm_and_ps(fscal,cutoff_mask);
1831
1832             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1833
1834              /* Update vectorial force */
1835             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1836             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1837             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1838
1839             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1840             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1841             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1842
1843             }
1844
1845             /**************************
1846              * CALCULATE INTERACTIONS *
1847              **************************/
1848
1849             if (gmx_mm_any_lt(rsq22,rcutoff2))
1850             {
1851
1852             /* REACTION-FIELD ELECTROSTATICS */
1853             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1854
1855             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1856
1857             fscal            = felec;
1858
1859             fscal            = _mm_and_ps(fscal,cutoff_mask);
1860
1861             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1862
1863              /* Update vectorial force */
1864             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1865             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1866             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1867
1868             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1869             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1870             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1871
1872             }
1873
1874             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1875             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1876             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1877             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1878
1879             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1880                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1881
1882             /* Inner loop uses 297 flops */
1883         }
1884
1885         /* End of innermost loop */
1886
1887         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1888                                               f+i_coord_offset,fshift+i_shift_offset);
1889
1890         /* Increment number of inner iterations */
1891         inneriter                  += j_index_end - j_index_start;
1892
1893         /* Outer loop uses 18 flops */
1894     }
1895
1896     /* Increment number of outer iterations */
1897     outeriter        += nri;
1898
1899     /* Update outer/inner flops */
1900
1901     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*297);
1902 }