src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRF_VdwNone_GeomW4P1_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_double 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_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_avx_128_fma_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            None
  40  * Geometry:                   Water4-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwNone_GeomW4P1_VF_avx_128_fma_double
  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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset1;
  67     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  68     int              vdwioffset2;
  69     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  70     int              vdwioffset3;
  71     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  75     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  76     __m128d          dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
  77     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     __m128d          dummy_mask,cutoff_mask;
  80     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  81     __m128d          one     = _mm_set1_pd(1.0);
  82     __m128d          two     = _mm_set1_pd(2.0);
  83     x                = xx[0];
  84     f                = ff[0];
  85
  86     nri              = nlist->nri;
  87     iinr             = nlist->iinr;
  88     jindex           = nlist->jindex;
  89     jjnr             = nlist->jjnr;
  90     shiftidx         = nlist->shift;
  91     gid              = nlist->gid;
  92     shiftvec         = fr->shift_vec[0];
  93     fshift           = fr->fshift[0];
  94     facel            = _mm_set1_pd(fr->epsfac);
  95     charge           = mdatoms->chargeA;
  96     krf              = _mm_set1_pd(fr->ic->k_rf);
  97     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
  98     crf              = _mm_set1_pd(fr->ic->c_rf);
  99
 100     /* Setup water-specific parameters */
 101     inr              = nlist->iinr[0];
 102     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 103     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 104     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 105
 106     /* Avoid stupid compiler warnings */
 107     jnrA = jnrB = 0;
 108     j_coord_offsetA = 0;
 109     j_coord_offsetB = 0;
 110
 111     outeriter        = 0;
 112     inneriter        = 0;
 113
 114     /* Start outer loop over neighborlists */
 115     for(iidx=0; iidx<nri; iidx++)
 116     {
 117         /* Load shift vector for this list */
 118         i_shift_offset   = DIM*shiftidx[iidx];
 119
 120         /* Load limits for loop over neighbors */
 121         j_index_start    = jindex[iidx];
 122         j_index_end      = jindex[iidx+1];
 123
 124         /* Get outer coordinate index */
 125         inr              = iinr[iidx];
 126         i_coord_offset   = DIM*inr;
 127
 128         /* Load i particle coords and add shift vector */
 129         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 130                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 131
 132         fix1             = _mm_setzero_pd();
 133         fiy1             = _mm_setzero_pd();
 134         fiz1             = _mm_setzero_pd();
 135         fix2             = _mm_setzero_pd();
 136         fiy2             = _mm_setzero_pd();
 137         fiz2             = _mm_setzero_pd();
 138         fix3             = _mm_setzero_pd();
 139         fiy3             = _mm_setzero_pd();
 140         fiz3             = _mm_setzero_pd();
 141
 142         /* Reset potential sums */
 143         velecsum         = _mm_setzero_pd();
 144
 145         /* Start inner kernel loop */
 146         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 147         {
 148
 149             /* Get j neighbor index, and coordinate index */
 150             jnrA             = jjnr[jidx];
 151             jnrB             = jjnr[jidx+1];
 152             j_coord_offsetA  = DIM*jnrA;
 153             j_coord_offsetB  = DIM*jnrB;
 154
 155             /* load j atom coordinates */
 156             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 157                                               &jx0,&jy0,&jz0);
 158
 159             /* Calculate displacement vector */
 160             dx10             = _mm_sub_pd(ix1,jx0);
 161             dy10             = _mm_sub_pd(iy1,jy0);
 162             dz10             = _mm_sub_pd(iz1,jz0);
 163             dx20             = _mm_sub_pd(ix2,jx0);
 164             dy20             = _mm_sub_pd(iy2,jy0);
 165             dz20             = _mm_sub_pd(iz2,jz0);
 166             dx30             = _mm_sub_pd(ix3,jx0);
 167             dy30             = _mm_sub_pd(iy3,jy0);
 168             dz30             = _mm_sub_pd(iz3,jz0);
 169
 170             /* Calculate squared distance and things based on it */
 171             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 172             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 173             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 174
 175             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 176             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 177             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 178
 179             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 180             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 181             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 182
 183             /* Load parameters for j particles */
 184             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 185
 186             fjx0             = _mm_setzero_pd();
 187             fjy0             = _mm_setzero_pd();
 188             fjz0             = _mm_setzero_pd();
 189
 190             /**************************
 191              * CALCULATE INTERACTIONS *
 192              **************************/
 193
 194             /* Compute parameters for interactions between i and j atoms */
 195             qq10             = _mm_mul_pd(iq1,jq0);
 196
 197             /* REACTION-FIELD ELECTROSTATICS */
 198             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 199             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 200
 201             /* Update potential sum for this i atom from the interaction with this j atom. */
 202             velecsum         = _mm_add_pd(velecsum,velec);
 203
 204             fscal            = felec;
 205
 206             /* Update vectorial force */
 207             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 208             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 209             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 210
 211             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 212             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 213             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 214
 215             /**************************
 216              * CALCULATE INTERACTIONS *
 217              **************************/
 218
 219             /* Compute parameters for interactions between i and j atoms */
 220             qq20             = _mm_mul_pd(iq2,jq0);
 221
 222             /* REACTION-FIELD ELECTROSTATICS */
 223             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 224             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 225
 226             /* Update potential sum for this i atom from the interaction with this j atom. */
 227             velecsum         = _mm_add_pd(velecsum,velec);
 228
 229             fscal            = felec;
 230
 231             /* Update vectorial force */
 232             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 233             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 234             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 235
 236             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 237             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 238             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 239
 240             /**************************
 241              * CALCULATE INTERACTIONS *
 242              **************************/
 243
 244             /* Compute parameters for interactions between i and j atoms */
 245             qq30             = _mm_mul_pd(iq3,jq0);
 246
 247             /* REACTION-FIELD ELECTROSTATICS */
 248             velec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_macc_pd(krf,rsq30,rinv30),crf));
 249             felec            = _mm_mul_pd(qq30,_mm_msub_pd(rinv30,rinvsq30,krf2));
 250
 251             /* Update potential sum for this i atom from the interaction with this j atom. */
 252             velecsum         = _mm_add_pd(velecsum,velec);
 253
 254             fscal            = felec;
 255
 256             /* Update vectorial force */
 257             fix3             = _mm_macc_pd(dx30,fscal,fix3);
 258             fiy3             = _mm_macc_pd(dy30,fscal,fiy3);
 259             fiz3             = _mm_macc_pd(dz30,fscal,fiz3);
 260
 261             fjx0             = _mm_macc_pd(dx30,fscal,fjx0);
 262             fjy0             = _mm_macc_pd(dy30,fscal,fjy0);
 263             fjz0             = _mm_macc_pd(dz30,fscal,fjz0);
 264
 265             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 266
 267             /* Inner loop uses 108 flops */
 268         }
 269
 270         if(jidx<j_index_end)
 271         {
 272
 273             jnrA             = jjnr[jidx];
 274             j_coord_offsetA  = DIM*jnrA;
 275
 276             /* load j atom coordinates */
 277             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 278                                               &jx0,&jy0,&jz0);
 279
 280             /* Calculate displacement vector */
 281             dx10             = _mm_sub_pd(ix1,jx0);
 282             dy10             = _mm_sub_pd(iy1,jy0);
 283             dz10             = _mm_sub_pd(iz1,jz0);
 284             dx20             = _mm_sub_pd(ix2,jx0);
 285             dy20             = _mm_sub_pd(iy2,jy0);
 286             dz20             = _mm_sub_pd(iz2,jz0);
 287             dx30             = _mm_sub_pd(ix3,jx0);
 288             dy30             = _mm_sub_pd(iy3,jy0);
 289             dz30             = _mm_sub_pd(iz3,jz0);
 290
 291             /* Calculate squared distance and things based on it */
 292             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 293             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 294             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 295
 296             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 297             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 298             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 299
 300             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 301             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 302             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 303
 304             /* Load parameters for j particles */
 305             jq0              = _mm_load_sd(charge+jnrA+0);
 306
 307             fjx0             = _mm_setzero_pd();
 308             fjy0             = _mm_setzero_pd();
 309             fjz0             = _mm_setzero_pd();
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             /* Compute parameters for interactions between i and j atoms */
 316             qq10             = _mm_mul_pd(iq1,jq0);
 317
 318             /* REACTION-FIELD ELECTROSTATICS */
 319             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 320             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 321
 322             /* Update potential sum for this i atom from the interaction with this j atom. */
 323             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 324             velecsum         = _mm_add_pd(velecsum,velec);
 325
 326             fscal            = felec;
 327
 328             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 329
 330             /* Update vectorial force */
 331             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 332             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 333             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 334
 335             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 336             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 337             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 338
 339             /**************************
 340              * CALCULATE INTERACTIONS *
 341              **************************/
 342
 343             /* Compute parameters for interactions between i and j atoms */
 344             qq20             = _mm_mul_pd(iq2,jq0);
 345
 346             /* REACTION-FIELD ELECTROSTATICS */
 347             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 348             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 349
 350             /* Update potential sum for this i atom from the interaction with this j atom. */
 351             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 352             velecsum         = _mm_add_pd(velecsum,velec);
 353
 354             fscal            = felec;
 355
 356             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 357
 358             /* Update vectorial force */
 359             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 360             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 361             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 362
 363             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 364             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 365             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 366
 367             /**************************
 368              * CALCULATE INTERACTIONS *
 369              **************************/
 370
 371             /* Compute parameters for interactions between i and j atoms */
 372             qq30             = _mm_mul_pd(iq3,jq0);
 373
 374             /* REACTION-FIELD ELECTROSTATICS */
 375             velec            = _mm_mul_pd(qq30,_mm_sub_pd(_mm_macc_pd(krf,rsq30,rinv30),crf));
 376             felec            = _mm_mul_pd(qq30,_mm_msub_pd(rinv30,rinvsq30,krf2));
 377
 378             /* Update potential sum for this i atom from the interaction with this j atom. */
 379             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 380             velecsum         = _mm_add_pd(velecsum,velec);
 381
 382             fscal            = felec;
 383
 384             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 385
 386             /* Update vectorial force */
 387             fix3             = _mm_macc_pd(dx30,fscal,fix3);
 388             fiy3             = _mm_macc_pd(dy30,fscal,fiy3);
 389             fiz3             = _mm_macc_pd(dz30,fscal,fiz3);
 390
 391             fjx0             = _mm_macc_pd(dx30,fscal,fjx0);
 392             fjy0             = _mm_macc_pd(dy30,fscal,fjy0);
 393             fjz0             = _mm_macc_pd(dz30,fscal,fjz0);
 394
 395             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 396
 397             /* Inner loop uses 108 flops */
 398         }
 399
 400         /* End of innermost loop */
 401
 402         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 403                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 404
 405         ggid                        = gid[iidx];
 406         /* Update potential energies */
 407         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 408
 409         /* Increment number of inner iterations */
 410         inneriter                  += j_index_end - j_index_start;
 411
 412         /* Outer loop uses 19 flops */
 413     }
 414
 415     /* Increment number of outer iterations */
 416     outeriter        += nri;
 417
 418     /* Update outer/inner flops */
 419
 420     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4_VF,outeriter*19 + inneriter*108);
 421 }
 422 /*
 423  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4P1_F_avx_128_fma_double
 424  * Electrostatics interaction: ReactionField
 425  * VdW interaction:            None
 426  * Geometry:                   Water4-Particle
 427  * Calculate force/pot:        Force
 428  */
 429 void
 430 nb_kernel_ElecRF_VdwNone_GeomW4P1_F_avx_128_fma_double
 431                     (t_nblist * gmx_restrict                nlist,
 432                      rvec * gmx_restrict                    xx,
 433                      rvec * gmx_restrict                    ff,
 434                      t_forcerec * gmx_restrict              fr,
 435                      t_mdatoms * gmx_restrict               mdatoms,
 436                      nb_kernel_data_t * gmx_restrict        kernel_data,
 437                      t_nrnb * gmx_restrict                  nrnb)
 438 {
 439     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 440      * just 0 for non-waters.
 441      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 442      * jnr indices corresponding to data put in the four positions in the SIMD register.
 443      */
 444     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 445     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 446     int              jnrA,jnrB;
 447     int              j_coord_offsetA,j_coord_offsetB;
 448     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 449     real             rcutoff_scalar;
 450     real             *shiftvec,*fshift,*x,*f;
 451     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 452     int              vdwioffset1;
 453     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 454     int              vdwioffset2;
 455     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 456     int              vdwioffset3;
 457     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 458     int              vdwjidx0A,vdwjidx0B;
 459     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 460     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 461     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 462     __m128d          dx30,dy30,dz30,rsq30,rinv30,rinvsq30,r30,qq30,c6_30,c12_30;
 463     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 464     real             *charge;
 465     __m128d          dummy_mask,cutoff_mask;
 466     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 467     __m128d          one     = _mm_set1_pd(1.0);
 468     __m128d          two     = _mm_set1_pd(2.0);
 469     x                = xx[0];
 470     f                = ff[0];
 471
 472     nri              = nlist->nri;
 473     iinr             = nlist->iinr;
 474     jindex           = nlist->jindex;
 475     jjnr             = nlist->jjnr;
 476     shiftidx         = nlist->shift;
 477     gid              = nlist->gid;
 478     shiftvec         = fr->shift_vec[0];
 479     fshift           = fr->fshift[0];
 480     facel            = _mm_set1_pd(fr->epsfac);
 481     charge           = mdatoms->chargeA;
 482     krf              = _mm_set1_pd(fr->ic->k_rf);
 483     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 484     crf              = _mm_set1_pd(fr->ic->c_rf);
 485
 486     /* Setup water-specific parameters */
 487     inr              = nlist->iinr[0];
 488     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 489     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 490     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 491
 492     /* Avoid stupid compiler warnings */
 493     jnrA = jnrB = 0;
 494     j_coord_offsetA = 0;
 495     j_coord_offsetB = 0;
 496
 497     outeriter        = 0;
 498     inneriter        = 0;
 499
 500     /* Start outer loop over neighborlists */
 501     for(iidx=0; iidx<nri; iidx++)
 502     {
 503         /* Load shift vector for this list */
 504         i_shift_offset   = DIM*shiftidx[iidx];
 505
 506         /* Load limits for loop over neighbors */
 507         j_index_start    = jindex[iidx];
 508         j_index_end      = jindex[iidx+1];
 509
 510         /* Get outer coordinate index */
 511         inr              = iinr[iidx];
 512         i_coord_offset   = DIM*inr;
 513
 514         /* Load i particle coords and add shift vector */
 515         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 516                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 517
 518         fix1             = _mm_setzero_pd();
 519         fiy1             = _mm_setzero_pd();
 520         fiz1             = _mm_setzero_pd();
 521         fix2             = _mm_setzero_pd();
 522         fiy2             = _mm_setzero_pd();
 523         fiz2             = _mm_setzero_pd();
 524         fix3             = _mm_setzero_pd();
 525         fiy3             = _mm_setzero_pd();
 526         fiz3             = _mm_setzero_pd();
 527
 528         /* Start inner kernel loop */
 529         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 530         {
 531
 532             /* Get j neighbor index, and coordinate index */
 533             jnrA             = jjnr[jidx];
 534             jnrB             = jjnr[jidx+1];
 535             j_coord_offsetA  = DIM*jnrA;
 536             j_coord_offsetB  = DIM*jnrB;
 537
 538             /* load j atom coordinates */
 539             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 540                                               &jx0,&jy0,&jz0);
 541
 542             /* Calculate displacement vector */
 543             dx10             = _mm_sub_pd(ix1,jx0);
 544             dy10             = _mm_sub_pd(iy1,jy0);
 545             dz10             = _mm_sub_pd(iz1,jz0);
 546             dx20             = _mm_sub_pd(ix2,jx0);
 547             dy20             = _mm_sub_pd(iy2,jy0);
 548             dz20             = _mm_sub_pd(iz2,jz0);
 549             dx30             = _mm_sub_pd(ix3,jx0);
 550             dy30             = _mm_sub_pd(iy3,jy0);
 551             dz30             = _mm_sub_pd(iz3,jz0);
 552
 553             /* Calculate squared distance and things based on it */
 554             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 555             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 556             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 557
 558             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 559             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 560             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 561
 562             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 563             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 564             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 565
 566             /* Load parameters for j particles */
 567             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 568
 569             fjx0             = _mm_setzero_pd();
 570             fjy0             = _mm_setzero_pd();
 571             fjz0             = _mm_setzero_pd();
 572
 573             /**************************
 574              * CALCULATE INTERACTIONS *
 575              **************************/
 576
 577             /* Compute parameters for interactions between i and j atoms */
 578             qq10             = _mm_mul_pd(iq1,jq0);
 579
 580             /* REACTION-FIELD ELECTROSTATICS */
 581             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 582
 583             fscal            = felec;
 584
 585             /* Update vectorial force */
 586             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 587             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 588             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 589
 590             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 591             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 592             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 593
 594             /**************************
 595              * CALCULATE INTERACTIONS *
 596              **************************/
 597
 598             /* Compute parameters for interactions between i and j atoms */
 599             qq20             = _mm_mul_pd(iq2,jq0);
 600
 601             /* REACTION-FIELD ELECTROSTATICS */
 602             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 603
 604             fscal            = felec;
 605
 606             /* Update vectorial force */
 607             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 608             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 609             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 610
 611             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 612             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 613             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 614
 615             /**************************
 616              * CALCULATE INTERACTIONS *
 617              **************************/
 618
 619             /* Compute parameters for interactions between i and j atoms */
 620             qq30             = _mm_mul_pd(iq3,jq0);
 621
 622             /* REACTION-FIELD ELECTROSTATICS */
 623             felec            = _mm_mul_pd(qq30,_mm_msub_pd(rinv30,rinvsq30,krf2));
 624
 625             fscal            = felec;
 626
 627             /* Update vectorial force */
 628             fix3             = _mm_macc_pd(dx30,fscal,fix3);
 629             fiy3             = _mm_macc_pd(dy30,fscal,fiy3);
 630             fiz3             = _mm_macc_pd(dz30,fscal,fiz3);
 631
 632             fjx0             = _mm_macc_pd(dx30,fscal,fjx0);
 633             fjy0             = _mm_macc_pd(dy30,fscal,fjy0);
 634             fjz0             = _mm_macc_pd(dz30,fscal,fjz0);
 635
 636             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 637
 638             /* Inner loop uses 93 flops */
 639         }
 640
 641         if(jidx<j_index_end)
 642         {
 643
 644             jnrA             = jjnr[jidx];
 645             j_coord_offsetA  = DIM*jnrA;
 646
 647             /* load j atom coordinates */
 648             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 649                                               &jx0,&jy0,&jz0);
 650
 651             /* Calculate displacement vector */
 652             dx10             = _mm_sub_pd(ix1,jx0);
 653             dy10             = _mm_sub_pd(iy1,jy0);
 654             dz10             = _mm_sub_pd(iz1,jz0);
 655             dx20             = _mm_sub_pd(ix2,jx0);
 656             dy20             = _mm_sub_pd(iy2,jy0);
 657             dz20             = _mm_sub_pd(iz2,jz0);
 658             dx30             = _mm_sub_pd(ix3,jx0);
 659             dy30             = _mm_sub_pd(iy3,jy0);
 660             dz30             = _mm_sub_pd(iz3,jz0);
 661
 662             /* Calculate squared distance and things based on it */
 663             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 664             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 665             rsq30            = gmx_mm_calc_rsq_pd(dx30,dy30,dz30);
 666
 667             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 668             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 669             rinv30           = gmx_mm_invsqrt_pd(rsq30);
 670
 671             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 672             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 673             rinvsq30         = _mm_mul_pd(rinv30,rinv30);
 674
 675             /* Load parameters for j particles */
 676             jq0              = _mm_load_sd(charge+jnrA+0);
 677
 678             fjx0             = _mm_setzero_pd();
 679             fjy0             = _mm_setzero_pd();
 680             fjz0             = _mm_setzero_pd();
 681
 682             /**************************
 683              * CALCULATE INTERACTIONS *
 684              **************************/
 685
 686             /* Compute parameters for interactions between i and j atoms */
 687             qq10             = _mm_mul_pd(iq1,jq0);
 688
 689             /* REACTION-FIELD ELECTROSTATICS */
 690             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 691
 692             fscal            = felec;
 693
 694             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 695
 696             /* Update vectorial force */
 697             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 698             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 699             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 700
 701             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 702             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 703             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 704
 705             /**************************
 706              * CALCULATE INTERACTIONS *
 707              **************************/
 708
 709             /* Compute parameters for interactions between i and j atoms */
 710             qq20             = _mm_mul_pd(iq2,jq0);
 711
 712             /* REACTION-FIELD ELECTROSTATICS */
 713             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 714
 715             fscal            = felec;
 716
 717             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 718
 719             /* Update vectorial force */
 720             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 721             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 722             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 723
 724             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 725             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 726             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 727
 728             /**************************
 729              * CALCULATE INTERACTIONS *
 730              **************************/
 731
 732             /* Compute parameters for interactions between i and j atoms */
 733             qq30             = _mm_mul_pd(iq3,jq0);
 734
 735             /* REACTION-FIELD ELECTROSTATICS */
 736             felec            = _mm_mul_pd(qq30,_mm_msub_pd(rinv30,rinvsq30,krf2));
 737
 738             fscal            = felec;
 739
 740             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 741
 742             /* Update vectorial force */
 743             fix3             = _mm_macc_pd(dx30,fscal,fix3);
 744             fiy3             = _mm_macc_pd(dy30,fscal,fiy3);
 745             fiz3             = _mm_macc_pd(dz30,fscal,fiz3);
 746
 747             fjx0             = _mm_macc_pd(dx30,fscal,fjx0);
 748             fjy0             = _mm_macc_pd(dy30,fscal,fjy0);
 749             fjz0             = _mm_macc_pd(dz30,fscal,fjz0);
 750
 751             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 752
 753             /* Inner loop uses 93 flops */
 754         }
 755
 756         /* End of innermost loop */
 757
 758         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 759                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 760
 761         /* Increment number of inner iterations */
 762         inneriter                  += j_index_end - j_index_start;
 763
 764         /* Outer loop uses 18 flops */
 765     }
 766
 767     /* Increment number of outer iterations */
 768     outeriter        += nri;
 769
 770     /* Update outer/inner flops */
 771
 772     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4_F,outeriter*18 + inneriter*93);
 773 }