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

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW3W3_VF_avx_128_fma_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            None
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwNone_GeomW3W3_VF_avx_128_fma_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  76     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  77     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  78     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  79     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  80     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  81     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  82     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  83     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  84     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  85     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  86     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  87     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  88     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  89     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128           dummy_mask,cutoff_mask;
  93     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  94     __m128           one     = _mm_set1_ps(1.0);
  95     __m128           two     = _mm_set1_ps(2.0);
  96     x                = xx[0];
  97     f                = ff[0];
  98
  99     nri              = nlist->nri;
 100     iinr             = nlist->iinr;
 101     jindex           = nlist->jindex;
 102     jjnr             = nlist->jjnr;
 103     shiftidx         = nlist->shift;
 104     gid              = nlist->gid;
 105     shiftvec         = fr->shift_vec[0];
 106     fshift           = fr->fshift[0];
 107     facel            = _mm_set1_ps(fr->epsfac);
 108     charge           = mdatoms->chargeA;
 109     krf              = _mm_set1_ps(fr->ic->k_rf);
 110     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 111     crf              = _mm_set1_ps(fr->ic->c_rf);
 112
 113     /* Setup water-specific parameters */
 114     inr              = nlist->iinr[0];
 115     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 116     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 117     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 118
 119     jq0              = _mm_set1_ps(charge[inr+0]);
 120     jq1              = _mm_set1_ps(charge[inr+1]);
 121     jq2              = _mm_set1_ps(charge[inr+2]);
 122     qq00             = _mm_mul_ps(iq0,jq0);
 123     qq01             = _mm_mul_ps(iq0,jq1);
 124     qq02             = _mm_mul_ps(iq0,jq2);
 125     qq10             = _mm_mul_ps(iq1,jq0);
 126     qq11             = _mm_mul_ps(iq1,jq1);
 127     qq12             = _mm_mul_ps(iq1,jq2);
 128     qq20             = _mm_mul_ps(iq2,jq0);
 129     qq21             = _mm_mul_ps(iq2,jq1);
 130     qq22             = _mm_mul_ps(iq2,jq2);
 131
 132     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 133     rcutoff_scalar   = fr->rcoulomb;
 134     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 135     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 136
 137     /* Avoid stupid compiler warnings */
 138     jnrA = jnrB = jnrC = jnrD = 0;
 139     j_coord_offsetA = 0;
 140     j_coord_offsetB = 0;
 141     j_coord_offsetC = 0;
 142     j_coord_offsetD = 0;
 143
 144     outeriter        = 0;
 145     inneriter        = 0;
 146
 147     for(iidx=0;iidx<4*DIM;iidx++)
 148     {
 149         scratch[iidx] = 0.0;
 150     }
 151
 152     /* Start outer loop over neighborlists */
 153     for(iidx=0; iidx<nri; iidx++)
 154     {
 155         /* Load shift vector for this list */
 156         i_shift_offset   = DIM*shiftidx[iidx];
 157
 158         /* Load limits for loop over neighbors */
 159         j_index_start    = jindex[iidx];
 160         j_index_end      = jindex[iidx+1];
 161
 162         /* Get outer coordinate index */
 163         inr              = iinr[iidx];
 164         i_coord_offset   = DIM*inr;
 165
 166         /* Load i particle coords and add shift vector */
 167         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 168                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 169
 170         fix0             = _mm_setzero_ps();
 171         fiy0             = _mm_setzero_ps();
 172         fiz0             = _mm_setzero_ps();
 173         fix1             = _mm_setzero_ps();
 174         fiy1             = _mm_setzero_ps();
 175         fiz1             = _mm_setzero_ps();
 176         fix2             = _mm_setzero_ps();
 177         fiy2             = _mm_setzero_ps();
 178         fiz2             = _mm_setzero_ps();
 179
 180         /* Reset potential sums */
 181         velecsum         = _mm_setzero_ps();
 182
 183         /* Start inner kernel loop */
 184         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 185         {
 186
 187             /* Get j neighbor index, and coordinate index */
 188             jnrA             = jjnr[jidx];
 189             jnrB             = jjnr[jidx+1];
 190             jnrC             = jjnr[jidx+2];
 191             jnrD             = jjnr[jidx+3];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194             j_coord_offsetC  = DIM*jnrC;
 195             j_coord_offsetD  = DIM*jnrD;
 196
 197             /* load j atom coordinates */
 198             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 199                                               x+j_coord_offsetC,x+j_coord_offsetD,
 200                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 201
 202             /* Calculate displacement vector */
 203             dx00             = _mm_sub_ps(ix0,jx0);
 204             dy00             = _mm_sub_ps(iy0,jy0);
 205             dz00             = _mm_sub_ps(iz0,jz0);
 206             dx01             = _mm_sub_ps(ix0,jx1);
 207             dy01             = _mm_sub_ps(iy0,jy1);
 208             dz01             = _mm_sub_ps(iz0,jz1);
 209             dx02             = _mm_sub_ps(ix0,jx2);
 210             dy02             = _mm_sub_ps(iy0,jy2);
 211             dz02             = _mm_sub_ps(iz0,jz2);
 212             dx10             = _mm_sub_ps(ix1,jx0);
 213             dy10             = _mm_sub_ps(iy1,jy0);
 214             dz10             = _mm_sub_ps(iz1,jz0);
 215             dx11             = _mm_sub_ps(ix1,jx1);
 216             dy11             = _mm_sub_ps(iy1,jy1);
 217             dz11             = _mm_sub_ps(iz1,jz1);
 218             dx12             = _mm_sub_ps(ix1,jx2);
 219             dy12             = _mm_sub_ps(iy1,jy2);
 220             dz12             = _mm_sub_ps(iz1,jz2);
 221             dx20             = _mm_sub_ps(ix2,jx0);
 222             dy20             = _mm_sub_ps(iy2,jy0);
 223             dz20             = _mm_sub_ps(iz2,jz0);
 224             dx21             = _mm_sub_ps(ix2,jx1);
 225             dy21             = _mm_sub_ps(iy2,jy1);
 226             dz21             = _mm_sub_ps(iz2,jz1);
 227             dx22             = _mm_sub_ps(ix2,jx2);
 228             dy22             = _mm_sub_ps(iy2,jy2);
 229             dz22             = _mm_sub_ps(iz2,jz2);
 230
 231             /* Calculate squared distance and things based on it */
 232             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 233             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 234             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 235             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 236             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 237             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 238             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 239             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 240             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 241
 242             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 243             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 244             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 245             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 246             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 247             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 248             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 249             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 250             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 251
 252             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 253             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 254             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 255             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 256             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 257             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 258             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 259             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 260             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 261
 262             fjx0             = _mm_setzero_ps();
 263             fjy0             = _mm_setzero_ps();
 264             fjz0             = _mm_setzero_ps();
 265             fjx1             = _mm_setzero_ps();
 266             fjy1             = _mm_setzero_ps();
 267             fjz1             = _mm_setzero_ps();
 268             fjx2             = _mm_setzero_ps();
 269             fjy2             = _mm_setzero_ps();
 270             fjz2             = _mm_setzero_ps();
 271
 272             /**************************
 273              * CALCULATE INTERACTIONS *
 274              **************************/
 275
 276             if (gmx_mm_any_lt(rsq00,rcutoff2))
 277             {
 278
 279             /* REACTION-FIELD ELECTROSTATICS */
 280             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 281             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 282
 283             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 284
 285             /* Update potential sum for this i atom from the interaction with this j atom. */
 286             velec            = _mm_and_ps(velec,cutoff_mask);
 287             velecsum         = _mm_add_ps(velecsum,velec);
 288
 289             fscal            = felec;
 290
 291             fscal            = _mm_and_ps(fscal,cutoff_mask);
 292
 293              /* Update vectorial force */
 294             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 295             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 296             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 297
 298             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 299             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 300             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 301
 302             }
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             if (gmx_mm_any_lt(rsq01,rcutoff2))
 309             {
 310
 311             /* REACTION-FIELD ELECTROSTATICS */
 312             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 313             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 314
 315             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 316
 317             /* Update potential sum for this i atom from the interaction with this j atom. */
 318             velec            = _mm_and_ps(velec,cutoff_mask);
 319             velecsum         = _mm_add_ps(velecsum,velec);
 320
 321             fscal            = felec;
 322
 323             fscal            = _mm_and_ps(fscal,cutoff_mask);
 324
 325              /* Update vectorial force */
 326             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 327             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 328             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 329
 330             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 331             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 332             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 333
 334             }
 335
 336             /**************************
 337              * CALCULATE INTERACTIONS *
 338              **************************/
 339
 340             if (gmx_mm_any_lt(rsq02,rcutoff2))
 341             {
 342
 343             /* REACTION-FIELD ELECTROSTATICS */
 344             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 345             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 346
 347             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 348
 349             /* Update potential sum for this i atom from the interaction with this j atom. */
 350             velec            = _mm_and_ps(velec,cutoff_mask);
 351             velecsum         = _mm_add_ps(velecsum,velec);
 352
 353             fscal            = felec;
 354
 355             fscal            = _mm_and_ps(fscal,cutoff_mask);
 356
 357              /* Update vectorial force */
 358             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 359             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 360             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 361
 362             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 363             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 364             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 365
 366             }
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             if (gmx_mm_any_lt(rsq10,rcutoff2))
 373             {
 374
 375             /* REACTION-FIELD ELECTROSTATICS */
 376             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 377             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 378
 379             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velec            = _mm_and_ps(velec,cutoff_mask);
 383             velecsum         = _mm_add_ps(velecsum,velec);
 384
 385             fscal            = felec;
 386
 387             fscal            = _mm_and_ps(fscal,cutoff_mask);
 388
 389              /* Update vectorial force */
 390             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 391             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 392             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 393
 394             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 395             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 396             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 397
 398             }
 399
 400             /**************************
 401              * CALCULATE INTERACTIONS *
 402              **************************/
 403
 404             if (gmx_mm_any_lt(rsq11,rcutoff2))
 405             {
 406
 407             /* REACTION-FIELD ELECTROSTATICS */
 408             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 409             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 410
 411             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 412
 413             /* Update potential sum for this i atom from the interaction with this j atom. */
 414             velec            = _mm_and_ps(velec,cutoff_mask);
 415             velecsum         = _mm_add_ps(velecsum,velec);
 416
 417             fscal            = felec;
 418
 419             fscal            = _mm_and_ps(fscal,cutoff_mask);
 420
 421              /* Update vectorial force */
 422             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 423             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 424             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 425
 426             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 427             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 428             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 429
 430             }
 431
 432             /**************************
 433              * CALCULATE INTERACTIONS *
 434              **************************/
 435
 436             if (gmx_mm_any_lt(rsq12,rcutoff2))
 437             {
 438
 439             /* REACTION-FIELD ELECTROSTATICS */
 440             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 441             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 442
 443             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 444
 445             /* Update potential sum for this i atom from the interaction with this j atom. */
 446             velec            = _mm_and_ps(velec,cutoff_mask);
 447             velecsum         = _mm_add_ps(velecsum,velec);
 448
 449             fscal            = felec;
 450
 451             fscal            = _mm_and_ps(fscal,cutoff_mask);
 452
 453              /* Update vectorial force */
 454             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 455             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 456             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 457
 458             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 459             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 460             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 461
 462             }
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             if (gmx_mm_any_lt(rsq20,rcutoff2))
 469             {
 470
 471             /* REACTION-FIELD ELECTROSTATICS */
 472             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 473             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 474
 475             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 476
 477             /* Update potential sum for this i atom from the interaction with this j atom. */
 478             velec            = _mm_and_ps(velec,cutoff_mask);
 479             velecsum         = _mm_add_ps(velecsum,velec);
 480
 481             fscal            = felec;
 482
 483             fscal            = _mm_and_ps(fscal,cutoff_mask);
 484
 485              /* Update vectorial force */
 486             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 487             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 488             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 489
 490             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 491             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 492             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 493
 494             }
 495
 496             /**************************
 497              * CALCULATE INTERACTIONS *
 498              **************************/
 499
 500             if (gmx_mm_any_lt(rsq21,rcutoff2))
 501             {
 502
 503             /* REACTION-FIELD ELECTROSTATICS */
 504             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 505             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 506
 507             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 508
 509             /* Update potential sum for this i atom from the interaction with this j atom. */
 510             velec            = _mm_and_ps(velec,cutoff_mask);
 511             velecsum         = _mm_add_ps(velecsum,velec);
 512
 513             fscal            = felec;
 514
 515             fscal            = _mm_and_ps(fscal,cutoff_mask);
 516
 517              /* Update vectorial force */
 518             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 519             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 520             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 521
 522             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 523             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 524             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 525
 526             }
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             if (gmx_mm_any_lt(rsq22,rcutoff2))
 533             {
 534
 535             /* REACTION-FIELD ELECTROSTATICS */
 536             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 537             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 538
 539             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 540
 541             /* Update potential sum for this i atom from the interaction with this j atom. */
 542             velec            = _mm_and_ps(velec,cutoff_mask);
 543             velecsum         = _mm_add_ps(velecsum,velec);
 544
 545             fscal            = felec;
 546
 547             fscal            = _mm_and_ps(fscal,cutoff_mask);
 548
 549              /* Update vectorial force */
 550             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 551             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 552             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 553
 554             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 555             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 556             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 557
 558             }
 559
 560             fjptrA             = f+j_coord_offsetA;
 561             fjptrB             = f+j_coord_offsetB;
 562             fjptrC             = f+j_coord_offsetC;
 563             fjptrD             = f+j_coord_offsetD;
 564
 565             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 566                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 567
 568             /* Inner loop uses 351 flops */
 569         }
 570
 571         if(jidx<j_index_end)
 572         {
 573
 574             /* Get j neighbor index, and coordinate index */
 575             jnrlistA         = jjnr[jidx];
 576             jnrlistB         = jjnr[jidx+1];
 577             jnrlistC         = jjnr[jidx+2];
 578             jnrlistD         = jjnr[jidx+3];
 579             /* Sign of each element will be negative for non-real atoms.
 580              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 581              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 582              */
 583             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 584             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 585             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 586             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 587             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 588             j_coord_offsetA  = DIM*jnrA;
 589             j_coord_offsetB  = DIM*jnrB;
 590             j_coord_offsetC  = DIM*jnrC;
 591             j_coord_offsetD  = DIM*jnrD;
 592
 593             /* load j atom coordinates */
 594             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 595                                               x+j_coord_offsetC,x+j_coord_offsetD,
 596                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 597
 598             /* Calculate displacement vector */
 599             dx00             = _mm_sub_ps(ix0,jx0);
 600             dy00             = _mm_sub_ps(iy0,jy0);
 601             dz00             = _mm_sub_ps(iz0,jz0);
 602             dx01             = _mm_sub_ps(ix0,jx1);
 603             dy01             = _mm_sub_ps(iy0,jy1);
 604             dz01             = _mm_sub_ps(iz0,jz1);
 605             dx02             = _mm_sub_ps(ix0,jx2);
 606             dy02             = _mm_sub_ps(iy0,jy2);
 607             dz02             = _mm_sub_ps(iz0,jz2);
 608             dx10             = _mm_sub_ps(ix1,jx0);
 609             dy10             = _mm_sub_ps(iy1,jy0);
 610             dz10             = _mm_sub_ps(iz1,jz0);
 611             dx11             = _mm_sub_ps(ix1,jx1);
 612             dy11             = _mm_sub_ps(iy1,jy1);
 613             dz11             = _mm_sub_ps(iz1,jz1);
 614             dx12             = _mm_sub_ps(ix1,jx2);
 615             dy12             = _mm_sub_ps(iy1,jy2);
 616             dz12             = _mm_sub_ps(iz1,jz2);
 617             dx20             = _mm_sub_ps(ix2,jx0);
 618             dy20             = _mm_sub_ps(iy2,jy0);
 619             dz20             = _mm_sub_ps(iz2,jz0);
 620             dx21             = _mm_sub_ps(ix2,jx1);
 621             dy21             = _mm_sub_ps(iy2,jy1);
 622             dz21             = _mm_sub_ps(iz2,jz1);
 623             dx22             = _mm_sub_ps(ix2,jx2);
 624             dy22             = _mm_sub_ps(iy2,jy2);
 625             dz22             = _mm_sub_ps(iz2,jz2);
 626
 627             /* Calculate squared distance and things based on it */
 628             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 629             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 630             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 631             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 632             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 633             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 634             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 635             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 636             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 637
 638             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 639             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 640             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 641             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 642             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 643             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 644             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 645             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 646             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 647
 648             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 649             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 650             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 651             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 652             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 653             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 654             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 655             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 656             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 657
 658             fjx0             = _mm_setzero_ps();
 659             fjy0             = _mm_setzero_ps();
 660             fjz0             = _mm_setzero_ps();
 661             fjx1             = _mm_setzero_ps();
 662             fjy1             = _mm_setzero_ps();
 663             fjz1             = _mm_setzero_ps();
 664             fjx2             = _mm_setzero_ps();
 665             fjy2             = _mm_setzero_ps();
 666             fjz2             = _mm_setzero_ps();
 667
 668             /**************************
 669              * CALCULATE INTERACTIONS *
 670              **************************/
 671
 672             if (gmx_mm_any_lt(rsq00,rcutoff2))
 673             {
 674
 675             /* REACTION-FIELD ELECTROSTATICS */
 676             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 677             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 678
 679             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 680
 681             /* Update potential sum for this i atom from the interaction with this j atom. */
 682             velec            = _mm_and_ps(velec,cutoff_mask);
 683             velec            = _mm_andnot_ps(dummy_mask,velec);
 684             velecsum         = _mm_add_ps(velecsum,velec);
 685
 686             fscal            = felec;
 687
 688             fscal            = _mm_and_ps(fscal,cutoff_mask);
 689
 690             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 691
 692              /* Update vectorial force */
 693             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 694             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 695             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 696
 697             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 698             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 699             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 700
 701             }
 702
 703             /**************************
 704              * CALCULATE INTERACTIONS *
 705              **************************/
 706
 707             if (gmx_mm_any_lt(rsq01,rcutoff2))
 708             {
 709
 710             /* REACTION-FIELD ELECTROSTATICS */
 711             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 712             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 713
 714             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 715
 716             /* Update potential sum for this i atom from the interaction with this j atom. */
 717             velec            = _mm_and_ps(velec,cutoff_mask);
 718             velec            = _mm_andnot_ps(dummy_mask,velec);
 719             velecsum         = _mm_add_ps(velecsum,velec);
 720
 721             fscal            = felec;
 722
 723             fscal            = _mm_and_ps(fscal,cutoff_mask);
 724
 725             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 726
 727              /* Update vectorial force */
 728             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 729             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 730             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 731
 732             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 733             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 734             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 735
 736             }
 737
 738             /**************************
 739              * CALCULATE INTERACTIONS *
 740              **************************/
 741
 742             if (gmx_mm_any_lt(rsq02,rcutoff2))
 743             {
 744
 745             /* REACTION-FIELD ELECTROSTATICS */
 746             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 747             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 748
 749             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 750
 751             /* Update potential sum for this i atom from the interaction with this j atom. */
 752             velec            = _mm_and_ps(velec,cutoff_mask);
 753             velec            = _mm_andnot_ps(dummy_mask,velec);
 754             velecsum         = _mm_add_ps(velecsum,velec);
 755
 756             fscal            = felec;
 757
 758             fscal            = _mm_and_ps(fscal,cutoff_mask);
 759
 760             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 761
 762              /* Update vectorial force */
 763             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 764             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 765             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 766
 767             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 768             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 769             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 770
 771             }
 772
 773             /**************************
 774              * CALCULATE INTERACTIONS *
 775              **************************/
 776
 777             if (gmx_mm_any_lt(rsq10,rcutoff2))
 778             {
 779
 780             /* REACTION-FIELD ELECTROSTATICS */
 781             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 782             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 783
 784             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 785
 786             /* Update potential sum for this i atom from the interaction with this j atom. */
 787             velec            = _mm_and_ps(velec,cutoff_mask);
 788             velec            = _mm_andnot_ps(dummy_mask,velec);
 789             velecsum         = _mm_add_ps(velecsum,velec);
 790
 791             fscal            = felec;
 792
 793             fscal            = _mm_and_ps(fscal,cutoff_mask);
 794
 795             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 796
 797              /* Update vectorial force */
 798             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 799             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 800             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 801
 802             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 803             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 804             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 805
 806             }
 807
 808             /**************************
 809              * CALCULATE INTERACTIONS *
 810              **************************/
 811
 812             if (gmx_mm_any_lt(rsq11,rcutoff2))
 813             {
 814
 815             /* REACTION-FIELD ELECTROSTATICS */
 816             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 817             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 818
 819             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 820
 821             /* Update potential sum for this i atom from the interaction with this j atom. */
 822             velec            = _mm_and_ps(velec,cutoff_mask);
 823             velec            = _mm_andnot_ps(dummy_mask,velec);
 824             velecsum         = _mm_add_ps(velecsum,velec);
 825
 826             fscal            = felec;
 827
 828             fscal            = _mm_and_ps(fscal,cutoff_mask);
 829
 830             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 831
 832              /* Update vectorial force */
 833             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 834             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 835             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 836
 837             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 838             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 839             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 840
 841             }
 842
 843             /**************************
 844              * CALCULATE INTERACTIONS *
 845              **************************/
 846
 847             if (gmx_mm_any_lt(rsq12,rcutoff2))
 848             {
 849
 850             /* REACTION-FIELD ELECTROSTATICS */
 851             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 852             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 853
 854             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 855
 856             /* Update potential sum for this i atom from the interaction with this j atom. */
 857             velec            = _mm_and_ps(velec,cutoff_mask);
 858             velec            = _mm_andnot_ps(dummy_mask,velec);
 859             velecsum         = _mm_add_ps(velecsum,velec);
 860
 861             fscal            = felec;
 862
 863             fscal            = _mm_and_ps(fscal,cutoff_mask);
 864
 865             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 866
 867              /* Update vectorial force */
 868             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 869             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 870             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 871
 872             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 873             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 874             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 875
 876             }
 877
 878             /**************************
 879              * CALCULATE INTERACTIONS *
 880              **************************/
 881
 882             if (gmx_mm_any_lt(rsq20,rcutoff2))
 883             {
 884
 885             /* REACTION-FIELD ELECTROSTATICS */
 886             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 887             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 888
 889             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 890
 891             /* Update potential sum for this i atom from the interaction with this j atom. */
 892             velec            = _mm_and_ps(velec,cutoff_mask);
 893             velec            = _mm_andnot_ps(dummy_mask,velec);
 894             velecsum         = _mm_add_ps(velecsum,velec);
 895
 896             fscal            = felec;
 897
 898             fscal            = _mm_and_ps(fscal,cutoff_mask);
 899
 900             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 901
 902              /* Update vectorial force */
 903             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 904             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 905             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 906
 907             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 908             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 909             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 910
 911             }
 912
 913             /**************************
 914              * CALCULATE INTERACTIONS *
 915              **************************/
 916
 917             if (gmx_mm_any_lt(rsq21,rcutoff2))
 918             {
 919
 920             /* REACTION-FIELD ELECTROSTATICS */
 921             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 922             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 923
 924             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 925
 926             /* Update potential sum for this i atom from the interaction with this j atom. */
 927             velec            = _mm_and_ps(velec,cutoff_mask);
 928             velec            = _mm_andnot_ps(dummy_mask,velec);
 929             velecsum         = _mm_add_ps(velecsum,velec);
 930
 931             fscal            = felec;
 932
 933             fscal            = _mm_and_ps(fscal,cutoff_mask);
 934
 935             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 936
 937              /* Update vectorial force */
 938             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 939             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 940             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 941
 942             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 943             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 944             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 945
 946             }
 947
 948             /**************************
 949              * CALCULATE INTERACTIONS *
 950              **************************/
 951
 952             if (gmx_mm_any_lt(rsq22,rcutoff2))
 953             {
 954
 955             /* REACTION-FIELD ELECTROSTATICS */
 956             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 957             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 958
 959             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 960
 961             /* Update potential sum for this i atom from the interaction with this j atom. */
 962             velec            = _mm_and_ps(velec,cutoff_mask);
 963             velec            = _mm_andnot_ps(dummy_mask,velec);
 964             velecsum         = _mm_add_ps(velecsum,velec);
 965
 966             fscal            = felec;
 967
 968             fscal            = _mm_and_ps(fscal,cutoff_mask);
 969
 970             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 971
 972              /* Update vectorial force */
 973             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 974             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 975             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 976
 977             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 978             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 979             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 980
 981             }
 982
 983             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 984             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 985             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 986             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 987
 988             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 989                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 990
 991             /* Inner loop uses 351 flops */
 992         }
 993
 994         /* End of innermost loop */
 995
 996         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 997                                               f+i_coord_offset,fshift+i_shift_offset);
 998
 999         ggid                        = gid[iidx];
1000         /* Update potential energies */
1001         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1002
1003         /* Increment number of inner iterations */
1004         inneriter                  += j_index_end - j_index_start;
1005
1006         /* Outer loop uses 19 flops */
1007     }
1008
1009     /* Increment number of outer iterations */
1010     outeriter        += nri;
1011
1012     /* Update outer/inner flops */
1013
1014     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*351);
1015 }
1016 /*
1017  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_avx_128_fma_single
1018  * Electrostatics interaction: ReactionField
1019  * VdW interaction:            None
1020  * Geometry:                   Water3-Water3
1021  * Calculate force/pot:        Force
1022  */
1023 void
1024 nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_avx_128_fma_single
1025                     (t_nblist * gmx_restrict                nlist,
1026                      rvec * gmx_restrict                    xx,
1027                      rvec * gmx_restrict                    ff,
1028                      t_forcerec * gmx_restrict              fr,
1029                      t_mdatoms * gmx_restrict               mdatoms,
1030                      nb_kernel_data_t * gmx_restrict        kernel_data,
1031                      t_nrnb * gmx_restrict                  nrnb)
1032 {
1033     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1034      * just 0 for non-waters.
1035      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1036      * jnr indices corresponding to data put in the four positions in the SIMD register.
1037      */
1038     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1039     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1040     int              jnrA,jnrB,jnrC,jnrD;
1041     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1042     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1043     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1044     real             rcutoff_scalar;
1045     real             *shiftvec,*fshift,*x,*f;
1046     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1047     real             scratch[4*DIM];
1048     __m128           fscal,rcutoff,rcutoff2,jidxall;
1049     int              vdwioffset0;
1050     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1051     int              vdwioffset1;
1052     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1053     int              vdwioffset2;
1054     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1055     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1056     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1057     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1058     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1059     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1060     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1061     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1062     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1063     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1064     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1065     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1066     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1067     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1068     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1069     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1070     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1071     real             *charge;
1072     __m128           dummy_mask,cutoff_mask;
1073     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1074     __m128           one     = _mm_set1_ps(1.0);
1075     __m128           two     = _mm_set1_ps(2.0);
1076     x                = xx[0];
1077     f                = ff[0];
1078
1079     nri              = nlist->nri;
1080     iinr             = nlist->iinr;
1081     jindex           = nlist->jindex;
1082     jjnr             = nlist->jjnr;
1083     shiftidx         = nlist->shift;
1084     gid              = nlist->gid;
1085     shiftvec         = fr->shift_vec[0];
1086     fshift           = fr->fshift[0];
1087     facel            = _mm_set1_ps(fr->epsfac);
1088     charge           = mdatoms->chargeA;
1089     krf              = _mm_set1_ps(fr->ic->k_rf);
1090     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1091     crf              = _mm_set1_ps(fr->ic->c_rf);
1092
1093     /* Setup water-specific parameters */
1094     inr              = nlist->iinr[0];
1095     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1096     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1097     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1098
1099     jq0              = _mm_set1_ps(charge[inr+0]);
1100     jq1              = _mm_set1_ps(charge[inr+1]);
1101     jq2              = _mm_set1_ps(charge[inr+2]);
1102     qq00             = _mm_mul_ps(iq0,jq0);
1103     qq01             = _mm_mul_ps(iq0,jq1);
1104     qq02             = _mm_mul_ps(iq0,jq2);
1105     qq10             = _mm_mul_ps(iq1,jq0);
1106     qq11             = _mm_mul_ps(iq1,jq1);
1107     qq12             = _mm_mul_ps(iq1,jq2);
1108     qq20             = _mm_mul_ps(iq2,jq0);
1109     qq21             = _mm_mul_ps(iq2,jq1);
1110     qq22             = _mm_mul_ps(iq2,jq2);
1111
1112     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1113     rcutoff_scalar   = fr->rcoulomb;
1114     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1115     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1116
1117     /* Avoid stupid compiler warnings */
1118     jnrA = jnrB = jnrC = jnrD = 0;
1119     j_coord_offsetA = 0;
1120     j_coord_offsetB = 0;
1121     j_coord_offsetC = 0;
1122     j_coord_offsetD = 0;
1123
1124     outeriter        = 0;
1125     inneriter        = 0;
1126
1127     for(iidx=0;iidx<4*DIM;iidx++)
1128     {
1129         scratch[iidx] = 0.0;
1130     }
1131
1132     /* Start outer loop over neighborlists */
1133     for(iidx=0; iidx<nri; iidx++)
1134     {
1135         /* Load shift vector for this list */
1136         i_shift_offset   = DIM*shiftidx[iidx];
1137
1138         /* Load limits for loop over neighbors */
1139         j_index_start    = jindex[iidx];
1140         j_index_end      = jindex[iidx+1];
1141
1142         /* Get outer coordinate index */
1143         inr              = iinr[iidx];
1144         i_coord_offset   = DIM*inr;
1145
1146         /* Load i particle coords and add shift vector */
1147         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1148                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1149
1150         fix0             = _mm_setzero_ps();
1151         fiy0             = _mm_setzero_ps();
1152         fiz0             = _mm_setzero_ps();
1153         fix1             = _mm_setzero_ps();
1154         fiy1             = _mm_setzero_ps();
1155         fiz1             = _mm_setzero_ps();
1156         fix2             = _mm_setzero_ps();
1157         fiy2             = _mm_setzero_ps();
1158         fiz2             = _mm_setzero_ps();
1159
1160         /* Start inner kernel loop */
1161         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1162         {
1163
1164             /* Get j neighbor index, and coordinate index */
1165             jnrA             = jjnr[jidx];
1166             jnrB             = jjnr[jidx+1];
1167             jnrC             = jjnr[jidx+2];
1168             jnrD             = jjnr[jidx+3];
1169             j_coord_offsetA  = DIM*jnrA;
1170             j_coord_offsetB  = DIM*jnrB;
1171             j_coord_offsetC  = DIM*jnrC;
1172             j_coord_offsetD  = DIM*jnrD;
1173
1174             /* load j atom coordinates */
1175             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1176                                               x+j_coord_offsetC,x+j_coord_offsetD,
1177                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1178
1179             /* Calculate displacement vector */
1180             dx00             = _mm_sub_ps(ix0,jx0);
1181             dy00             = _mm_sub_ps(iy0,jy0);
1182             dz00             = _mm_sub_ps(iz0,jz0);
1183             dx01             = _mm_sub_ps(ix0,jx1);
1184             dy01             = _mm_sub_ps(iy0,jy1);
1185             dz01             = _mm_sub_ps(iz0,jz1);
1186             dx02             = _mm_sub_ps(ix0,jx2);
1187             dy02             = _mm_sub_ps(iy0,jy2);
1188             dz02             = _mm_sub_ps(iz0,jz2);
1189             dx10             = _mm_sub_ps(ix1,jx0);
1190             dy10             = _mm_sub_ps(iy1,jy0);
1191             dz10             = _mm_sub_ps(iz1,jz0);
1192             dx11             = _mm_sub_ps(ix1,jx1);
1193             dy11             = _mm_sub_ps(iy1,jy1);
1194             dz11             = _mm_sub_ps(iz1,jz1);
1195             dx12             = _mm_sub_ps(ix1,jx2);
1196             dy12             = _mm_sub_ps(iy1,jy2);
1197             dz12             = _mm_sub_ps(iz1,jz2);
1198             dx20             = _mm_sub_ps(ix2,jx0);
1199             dy20             = _mm_sub_ps(iy2,jy0);
1200             dz20             = _mm_sub_ps(iz2,jz0);
1201             dx21             = _mm_sub_ps(ix2,jx1);
1202             dy21             = _mm_sub_ps(iy2,jy1);
1203             dz21             = _mm_sub_ps(iz2,jz1);
1204             dx22             = _mm_sub_ps(ix2,jx2);
1205             dy22             = _mm_sub_ps(iy2,jy2);
1206             dz22             = _mm_sub_ps(iz2,jz2);
1207
1208             /* Calculate squared distance and things based on it */
1209             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1210             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1211             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1212             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1213             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1214             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1215             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1216             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1217             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1218
1219             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1220             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1221             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1222             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1223             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1224             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1225             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1226             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1227             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1228
1229             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1230             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1231             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1232             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1233             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1234             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1235             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1236             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1237             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1238
1239             fjx0             = _mm_setzero_ps();
1240             fjy0             = _mm_setzero_ps();
1241             fjz0             = _mm_setzero_ps();
1242             fjx1             = _mm_setzero_ps();
1243             fjy1             = _mm_setzero_ps();
1244             fjz1             = _mm_setzero_ps();
1245             fjx2             = _mm_setzero_ps();
1246             fjy2             = _mm_setzero_ps();
1247             fjz2             = _mm_setzero_ps();
1248
1249             /**************************
1250              * CALCULATE INTERACTIONS *
1251              **************************/
1252
1253             if (gmx_mm_any_lt(rsq00,rcutoff2))
1254             {
1255
1256             /* REACTION-FIELD ELECTROSTATICS */
1257             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1258
1259             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1260
1261             fscal            = felec;
1262
1263             fscal            = _mm_and_ps(fscal,cutoff_mask);
1264
1265              /* Update vectorial force */
1266             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1267             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1268             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1269
1270             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1271             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1272             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1273
1274             }
1275
1276             /**************************
1277              * CALCULATE INTERACTIONS *
1278              **************************/
1279
1280             if (gmx_mm_any_lt(rsq01,rcutoff2))
1281             {
1282
1283             /* REACTION-FIELD ELECTROSTATICS */
1284             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1285
1286             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1287
1288             fscal            = felec;
1289
1290             fscal            = _mm_and_ps(fscal,cutoff_mask);
1291
1292              /* Update vectorial force */
1293             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1294             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1295             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1296
1297             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1298             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1299             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1300
1301             }
1302
1303             /**************************
1304              * CALCULATE INTERACTIONS *
1305              **************************/
1306
1307             if (gmx_mm_any_lt(rsq02,rcutoff2))
1308             {
1309
1310             /* REACTION-FIELD ELECTROSTATICS */
1311             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1312
1313             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1314
1315             fscal            = felec;
1316
1317             fscal            = _mm_and_ps(fscal,cutoff_mask);
1318
1319              /* Update vectorial force */
1320             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1321             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1322             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1323
1324             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1325             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1326             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1327
1328             }
1329
1330             /**************************
1331              * CALCULATE INTERACTIONS *
1332              **************************/
1333
1334             if (gmx_mm_any_lt(rsq10,rcutoff2))
1335             {
1336
1337             /* REACTION-FIELD ELECTROSTATICS */
1338             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1339
1340             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1341
1342             fscal            = felec;
1343
1344             fscal            = _mm_and_ps(fscal,cutoff_mask);
1345
1346              /* Update vectorial force */
1347             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1348             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1349             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1350
1351             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1352             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1353             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1354
1355             }
1356
1357             /**************************
1358              * CALCULATE INTERACTIONS *
1359              **************************/
1360
1361             if (gmx_mm_any_lt(rsq11,rcutoff2))
1362             {
1363
1364             /* REACTION-FIELD ELECTROSTATICS */
1365             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1366
1367             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1368
1369             fscal            = felec;
1370
1371             fscal            = _mm_and_ps(fscal,cutoff_mask);
1372
1373              /* Update vectorial force */
1374             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1375             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1376             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1377
1378             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1379             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1380             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1381
1382             }
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             if (gmx_mm_any_lt(rsq12,rcutoff2))
1389             {
1390
1391             /* REACTION-FIELD ELECTROSTATICS */
1392             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1393
1394             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1395
1396             fscal            = felec;
1397
1398             fscal            = _mm_and_ps(fscal,cutoff_mask);
1399
1400              /* Update vectorial force */
1401             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1402             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1403             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1404
1405             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1406             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1407             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1408
1409             }
1410
1411             /**************************
1412              * CALCULATE INTERACTIONS *
1413              **************************/
1414
1415             if (gmx_mm_any_lt(rsq20,rcutoff2))
1416             {
1417
1418             /* REACTION-FIELD ELECTROSTATICS */
1419             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1420
1421             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1422
1423             fscal            = felec;
1424
1425             fscal            = _mm_and_ps(fscal,cutoff_mask);
1426
1427              /* Update vectorial force */
1428             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1429             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1430             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1431
1432             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1433             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1434             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1435
1436             }
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             if (gmx_mm_any_lt(rsq21,rcutoff2))
1443             {
1444
1445             /* REACTION-FIELD ELECTROSTATICS */
1446             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1447
1448             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1449
1450             fscal            = felec;
1451
1452             fscal            = _mm_and_ps(fscal,cutoff_mask);
1453
1454              /* Update vectorial force */
1455             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1456             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1457             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1458
1459             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1460             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1461             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1462
1463             }
1464
1465             /**************************
1466              * CALCULATE INTERACTIONS *
1467              **************************/
1468
1469             if (gmx_mm_any_lt(rsq22,rcutoff2))
1470             {
1471
1472             /* REACTION-FIELD ELECTROSTATICS */
1473             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1474
1475             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1476
1477             fscal            = felec;
1478
1479             fscal            = _mm_and_ps(fscal,cutoff_mask);
1480
1481              /* Update vectorial force */
1482             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1483             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1484             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1485
1486             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1487             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1488             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1489
1490             }
1491
1492             fjptrA             = f+j_coord_offsetA;
1493             fjptrB             = f+j_coord_offsetB;
1494             fjptrC             = f+j_coord_offsetC;
1495             fjptrD             = f+j_coord_offsetD;
1496
1497             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1498                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1499
1500             /* Inner loop uses 297 flops */
1501         }
1502
1503         if(jidx<j_index_end)
1504         {
1505
1506             /* Get j neighbor index, and coordinate index */
1507             jnrlistA         = jjnr[jidx];
1508             jnrlistB         = jjnr[jidx+1];
1509             jnrlistC         = jjnr[jidx+2];
1510             jnrlistD         = jjnr[jidx+3];
1511             /* Sign of each element will be negative for non-real atoms.
1512              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1513              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1514              */
1515             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1516             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1517             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1518             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1519             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1520             j_coord_offsetA  = DIM*jnrA;
1521             j_coord_offsetB  = DIM*jnrB;
1522             j_coord_offsetC  = DIM*jnrC;
1523             j_coord_offsetD  = DIM*jnrD;
1524
1525             /* load j atom coordinates */
1526             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1527                                               x+j_coord_offsetC,x+j_coord_offsetD,
1528                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1529
1530             /* Calculate displacement vector */
1531             dx00             = _mm_sub_ps(ix0,jx0);
1532             dy00             = _mm_sub_ps(iy0,jy0);
1533             dz00             = _mm_sub_ps(iz0,jz0);
1534             dx01             = _mm_sub_ps(ix0,jx1);
1535             dy01             = _mm_sub_ps(iy0,jy1);
1536             dz01             = _mm_sub_ps(iz0,jz1);
1537             dx02             = _mm_sub_ps(ix0,jx2);
1538             dy02             = _mm_sub_ps(iy0,jy2);
1539             dz02             = _mm_sub_ps(iz0,jz2);
1540             dx10             = _mm_sub_ps(ix1,jx0);
1541             dy10             = _mm_sub_ps(iy1,jy0);
1542             dz10             = _mm_sub_ps(iz1,jz0);
1543             dx11             = _mm_sub_ps(ix1,jx1);
1544             dy11             = _mm_sub_ps(iy1,jy1);
1545             dz11             = _mm_sub_ps(iz1,jz1);
1546             dx12             = _mm_sub_ps(ix1,jx2);
1547             dy12             = _mm_sub_ps(iy1,jy2);
1548             dz12             = _mm_sub_ps(iz1,jz2);
1549             dx20             = _mm_sub_ps(ix2,jx0);
1550             dy20             = _mm_sub_ps(iy2,jy0);
1551             dz20             = _mm_sub_ps(iz2,jz0);
1552             dx21             = _mm_sub_ps(ix2,jx1);
1553             dy21             = _mm_sub_ps(iy2,jy1);
1554             dz21             = _mm_sub_ps(iz2,jz1);
1555             dx22             = _mm_sub_ps(ix2,jx2);
1556             dy22             = _mm_sub_ps(iy2,jy2);
1557             dz22             = _mm_sub_ps(iz2,jz2);
1558
1559             /* Calculate squared distance and things based on it */
1560             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1561             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1562             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1563             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1564             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1565             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1566             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1567             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1568             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1569
1570             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1571             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1572             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1573             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1574             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1575             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1576             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1577             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1578             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1579
1580             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1581             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1582             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1583             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1584             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1585             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1586             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1587             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1588             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1589
1590             fjx0             = _mm_setzero_ps();
1591             fjy0             = _mm_setzero_ps();
1592             fjz0             = _mm_setzero_ps();
1593             fjx1             = _mm_setzero_ps();
1594             fjy1             = _mm_setzero_ps();
1595             fjz1             = _mm_setzero_ps();
1596             fjx2             = _mm_setzero_ps();
1597             fjy2             = _mm_setzero_ps();
1598             fjz2             = _mm_setzero_ps();
1599
1600             /**************************
1601              * CALCULATE INTERACTIONS *
1602              **************************/
1603
1604             if (gmx_mm_any_lt(rsq00,rcutoff2))
1605             {
1606
1607             /* REACTION-FIELD ELECTROSTATICS */
1608             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1609
1610             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1611
1612             fscal            = felec;
1613
1614             fscal            = _mm_and_ps(fscal,cutoff_mask);
1615
1616             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1617
1618              /* Update vectorial force */
1619             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1620             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1621             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1622
1623             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1624             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1625             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1626
1627             }
1628
1629             /**************************
1630              * CALCULATE INTERACTIONS *
1631              **************************/
1632
1633             if (gmx_mm_any_lt(rsq01,rcutoff2))
1634             {
1635
1636             /* REACTION-FIELD ELECTROSTATICS */
1637             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1638
1639             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1640
1641             fscal            = felec;
1642
1643             fscal            = _mm_and_ps(fscal,cutoff_mask);
1644
1645             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1646
1647              /* Update vectorial force */
1648             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1649             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1650             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1651
1652             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1653             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1654             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1655
1656             }
1657
1658             /**************************
1659              * CALCULATE INTERACTIONS *
1660              **************************/
1661
1662             if (gmx_mm_any_lt(rsq02,rcutoff2))
1663             {
1664
1665             /* REACTION-FIELD ELECTROSTATICS */
1666             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1667
1668             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1669
1670             fscal            = felec;
1671
1672             fscal            = _mm_and_ps(fscal,cutoff_mask);
1673
1674             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1675
1676              /* Update vectorial force */
1677             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1678             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1679             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1680
1681             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1682             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1683             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1684
1685             }
1686
1687             /**************************
1688              * CALCULATE INTERACTIONS *
1689              **************************/
1690
1691             if (gmx_mm_any_lt(rsq10,rcutoff2))
1692             {
1693
1694             /* REACTION-FIELD ELECTROSTATICS */
1695             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1696
1697             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1698
1699             fscal            = felec;
1700
1701             fscal            = _mm_and_ps(fscal,cutoff_mask);
1702
1703             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1704
1705              /* Update vectorial force */
1706             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1707             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1708             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1709
1710             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1711             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1712             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1713
1714             }
1715
1716             /**************************
1717              * CALCULATE INTERACTIONS *
1718              **************************/
1719
1720             if (gmx_mm_any_lt(rsq11,rcutoff2))
1721             {
1722
1723             /* REACTION-FIELD ELECTROSTATICS */
1724             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1725
1726             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1727
1728             fscal            = felec;
1729
1730             fscal            = _mm_and_ps(fscal,cutoff_mask);
1731
1732             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1733
1734              /* Update vectorial force */
1735             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1736             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1737             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1738
1739             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1740             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1741             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1742
1743             }
1744
1745             /**************************
1746              * CALCULATE INTERACTIONS *
1747              **************************/
1748
1749             if (gmx_mm_any_lt(rsq12,rcutoff2))
1750             {
1751
1752             /* REACTION-FIELD ELECTROSTATICS */
1753             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1754
1755             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1756
1757             fscal            = felec;
1758
1759             fscal            = _mm_and_ps(fscal,cutoff_mask);
1760
1761             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1762
1763              /* Update vectorial force */
1764             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1765             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1766             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1767
1768             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1769             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1770             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1771
1772             }
1773
1774             /**************************
1775              * CALCULATE INTERACTIONS *
1776              **************************/
1777
1778             if (gmx_mm_any_lt(rsq20,rcutoff2))
1779             {
1780
1781             /* REACTION-FIELD ELECTROSTATICS */
1782             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1783
1784             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1785
1786             fscal            = felec;
1787
1788             fscal            = _mm_and_ps(fscal,cutoff_mask);
1789
1790             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1791
1792              /* Update vectorial force */
1793             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1794             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1795             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1796
1797             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1798             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1799             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1800
1801             }
1802
1803             /**************************
1804              * CALCULATE INTERACTIONS *
1805              **************************/
1806
1807             if (gmx_mm_any_lt(rsq21,rcutoff2))
1808             {
1809
1810             /* REACTION-FIELD ELECTROSTATICS */
1811             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1812
1813             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1814
1815             fscal            = felec;
1816
1817             fscal            = _mm_and_ps(fscal,cutoff_mask);
1818
1819             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1820
1821              /* Update vectorial force */
1822             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1823             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1824             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1825
1826             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1827             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1828             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1829
1830             }
1831
1832             /**************************
1833              * CALCULATE INTERACTIONS *
1834              **************************/
1835
1836             if (gmx_mm_any_lt(rsq22,rcutoff2))
1837             {
1838
1839             /* REACTION-FIELD ELECTROSTATICS */
1840             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1841
1842             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1843
1844             fscal            = felec;
1845
1846             fscal            = _mm_and_ps(fscal,cutoff_mask);
1847
1848             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1849
1850              /* Update vectorial force */
1851             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1852             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1853             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1854
1855             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1856             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1857             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1858
1859             }
1860
1861             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1862             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1863             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1864             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1865
1866             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1867                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1868
1869             /* Inner loop uses 297 flops */
1870         }
1871
1872         /* End of innermost loop */
1873
1874         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1875                                               f+i_coord_offset,fshift+i_shift_offset);
1876
1877         /* Increment number of inner iterations */
1878         inneriter                  += j_index_end - j_index_start;
1879
1880         /* Outer loop uses 18 flops */
1881     }
1882
1883     /* Increment number of outer iterations */
1884     outeriter        += nri;
1885
1886     /* Update outer/inner flops */
1887
1888     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*297);
1889 }