src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRF_VdwLJ_GeomW3P1_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_VdwLJ_GeomW3P1_VF_avx_128_fma_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwLJ_GeomW3P1_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              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  75     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  76     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  77     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  78     real             *charge;
  79     int              nvdwtype;
  80     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  81     int              *vdwtype;
  82     real             *vdwparam;
  83     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  84     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  85     __m128d          dummy_mask,cutoff_mask;
  86     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  87     __m128d          one     = _mm_set1_pd(1.0);
  88     __m128d          two     = _mm_set1_pd(2.0);
  89     x                = xx[0];
  90     f                = ff[0];
  91
  92     nri              = nlist->nri;
  93     iinr             = nlist->iinr;
  94     jindex           = nlist->jindex;
  95     jjnr             = nlist->jjnr;
  96     shiftidx         = nlist->shift;
  97     gid              = nlist->gid;
  98     shiftvec         = fr->shift_vec[0];
  99     fshift           = fr->fshift[0];
 100     facel            = _mm_set1_pd(fr->epsfac);
 101     charge           = mdatoms->chargeA;
 102     krf              = _mm_set1_pd(fr->ic->k_rf);
 103     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 104     crf              = _mm_set1_pd(fr->ic->c_rf);
 105     nvdwtype         = fr->ntype;
 106     vdwparam         = fr->nbfp;
 107     vdwtype          = mdatoms->typeA;
 108
 109     /* Setup water-specific parameters */
 110     inr              = nlist->iinr[0];
 111     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 112     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 113     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 114     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 115
 116     /* Avoid stupid compiler warnings */
 117     jnrA = jnrB = 0;
 118     j_coord_offsetA = 0;
 119     j_coord_offsetB = 0;
 120
 121     outeriter        = 0;
 122     inneriter        = 0;
 123
 124     /* Start outer loop over neighborlists */
 125     for(iidx=0; iidx<nri; iidx++)
 126     {
 127         /* Load shift vector for this list */
 128         i_shift_offset   = DIM*shiftidx[iidx];
 129
 130         /* Load limits for loop over neighbors */
 131         j_index_start    = jindex[iidx];
 132         j_index_end      = jindex[iidx+1];
 133
 134         /* Get outer coordinate index */
 135         inr              = iinr[iidx];
 136         i_coord_offset   = DIM*inr;
 137
 138         /* Load i particle coords and add shift vector */
 139         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 140                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 141
 142         fix0             = _mm_setzero_pd();
 143         fiy0             = _mm_setzero_pd();
 144         fiz0             = _mm_setzero_pd();
 145         fix1             = _mm_setzero_pd();
 146         fiy1             = _mm_setzero_pd();
 147         fiz1             = _mm_setzero_pd();
 148         fix2             = _mm_setzero_pd();
 149         fiy2             = _mm_setzero_pd();
 150         fiz2             = _mm_setzero_pd();
 151
 152         /* Reset potential sums */
 153         velecsum         = _mm_setzero_pd();
 154         vvdwsum          = _mm_setzero_pd();
 155
 156         /* Start inner kernel loop */
 157         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 158         {
 159
 160             /* Get j neighbor index, and coordinate index */
 161             jnrA             = jjnr[jidx];
 162             jnrB             = jjnr[jidx+1];
 163             j_coord_offsetA  = DIM*jnrA;
 164             j_coord_offsetB  = DIM*jnrB;
 165
 166             /* load j atom coordinates */
 167             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 168                                               &jx0,&jy0,&jz0);
 169
 170             /* Calculate displacement vector */
 171             dx00             = _mm_sub_pd(ix0,jx0);
 172             dy00             = _mm_sub_pd(iy0,jy0);
 173             dz00             = _mm_sub_pd(iz0,jz0);
 174             dx10             = _mm_sub_pd(ix1,jx0);
 175             dy10             = _mm_sub_pd(iy1,jy0);
 176             dz10             = _mm_sub_pd(iz1,jz0);
 177             dx20             = _mm_sub_pd(ix2,jx0);
 178             dy20             = _mm_sub_pd(iy2,jy0);
 179             dz20             = _mm_sub_pd(iz2,jz0);
 180
 181             /* Calculate squared distance and things based on it */
 182             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 183             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 184             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 185
 186             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 187             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 188             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 189
 190             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 191             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 192             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 193
 194             /* Load parameters for j particles */
 195             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 196             vdwjidx0A        = 2*vdwtype[jnrA+0];
 197             vdwjidx0B        = 2*vdwtype[jnrB+0];
 198
 199             fjx0             = _mm_setzero_pd();
 200             fjy0             = _mm_setzero_pd();
 201             fjz0             = _mm_setzero_pd();
 202
 203             /**************************
 204              * CALCULATE INTERACTIONS *
 205              **************************/
 206
 207             /* Compute parameters for interactions between i and j atoms */
 208             qq00             = _mm_mul_pd(iq0,jq0);
 209             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 210                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 211
 212             /* REACTION-FIELD ELECTROSTATICS */
 213             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 214             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 215
 216             /* LENNARD-JONES DISPERSION/REPULSION */
 217
 218             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 219             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 220             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 221             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 222             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 223
 224             /* Update potential sum for this i atom from the interaction with this j atom. */
 225             velecsum         = _mm_add_pd(velecsum,velec);
 226             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 227
 228             fscal            = _mm_add_pd(felec,fvdw);
 229
 230             /* Update vectorial force */
 231             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 232             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 233             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 234
 235             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 236             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 237             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 238
 239             /**************************
 240              * CALCULATE INTERACTIONS *
 241              **************************/
 242
 243             /* Compute parameters for interactions between i and j atoms */
 244             qq10             = _mm_mul_pd(iq1,jq0);
 245
 246             /* REACTION-FIELD ELECTROSTATICS */
 247             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 248             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 249
 250             /* Update potential sum for this i atom from the interaction with this j atom. */
 251             velecsum         = _mm_add_pd(velecsum,velec);
 252
 253             fscal            = felec;
 254
 255             /* Update vectorial force */
 256             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 257             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 258             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 259
 260             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 261             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 262             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 263
 264             /**************************
 265              * CALCULATE INTERACTIONS *
 266              **************************/
 267
 268             /* Compute parameters for interactions between i and j atoms */
 269             qq20             = _mm_mul_pd(iq2,jq0);
 270
 271             /* REACTION-FIELD ELECTROSTATICS */
 272             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 273             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 274
 275             /* Update potential sum for this i atom from the interaction with this j atom. */
 276             velecsum         = _mm_add_pd(velecsum,velec);
 277
 278             fscal            = felec;
 279
 280             /* Update vectorial force */
 281             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 282             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 283             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 284
 285             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 286             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 287             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 288
 289             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 290
 291             /* Inner loop uses 120 flops */
 292         }
 293
 294         if(jidx<j_index_end)
 295         {
 296
 297             jnrA             = jjnr[jidx];
 298             j_coord_offsetA  = DIM*jnrA;
 299
 300             /* load j atom coordinates */
 301             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 302                                               &jx0,&jy0,&jz0);
 303
 304             /* Calculate displacement vector */
 305             dx00             = _mm_sub_pd(ix0,jx0);
 306             dy00             = _mm_sub_pd(iy0,jy0);
 307             dz00             = _mm_sub_pd(iz0,jz0);
 308             dx10             = _mm_sub_pd(ix1,jx0);
 309             dy10             = _mm_sub_pd(iy1,jy0);
 310             dz10             = _mm_sub_pd(iz1,jz0);
 311             dx20             = _mm_sub_pd(ix2,jx0);
 312             dy20             = _mm_sub_pd(iy2,jy0);
 313             dz20             = _mm_sub_pd(iz2,jz0);
 314
 315             /* Calculate squared distance and things based on it */
 316             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 317             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 318             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 319
 320             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 321             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 322             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 323
 324             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 325             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 326             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 327
 328             /* Load parameters for j particles */
 329             jq0              = _mm_load_sd(charge+jnrA+0);
 330             vdwjidx0A        = 2*vdwtype[jnrA+0];
 331
 332             fjx0             = _mm_setzero_pd();
 333             fjy0             = _mm_setzero_pd();
 334             fjz0             = _mm_setzero_pd();
 335
 336             /**************************
 337              * CALCULATE INTERACTIONS *
 338              **************************/
 339
 340             /* Compute parameters for interactions between i and j atoms */
 341             qq00             = _mm_mul_pd(iq0,jq0);
 342             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 343
 344             /* REACTION-FIELD ELECTROSTATICS */
 345             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 346             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 347
 348             /* LENNARD-JONES DISPERSION/REPULSION */
 349
 350             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 351             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 352             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 353             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 354             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 355
 356             /* Update potential sum for this i atom from the interaction with this j atom. */
 357             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 358             velecsum         = _mm_add_pd(velecsum,velec);
 359             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 360             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 361
 362             fscal            = _mm_add_pd(felec,fvdw);
 363
 364             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 365
 366             /* Update vectorial force */
 367             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 368             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 369             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 370
 371             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 372             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 373             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 374
 375             /**************************
 376              * CALCULATE INTERACTIONS *
 377              **************************/
 378
 379             /* Compute parameters for interactions between i and j atoms */
 380             qq10             = _mm_mul_pd(iq1,jq0);
 381
 382             /* REACTION-FIELD ELECTROSTATICS */
 383             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 384             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 385
 386             /* Update potential sum for this i atom from the interaction with this j atom. */
 387             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 388             velecsum         = _mm_add_pd(velecsum,velec);
 389
 390             fscal            = felec;
 391
 392             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 393
 394             /* Update vectorial force */
 395             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 396             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 397             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 398
 399             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 400             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 401             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 402
 403             /**************************
 404              * CALCULATE INTERACTIONS *
 405              **************************/
 406
 407             /* Compute parameters for interactions between i and j atoms */
 408             qq20             = _mm_mul_pd(iq2,jq0);
 409
 410             /* REACTION-FIELD ELECTROSTATICS */
 411             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 412             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 413
 414             /* Update potential sum for this i atom from the interaction with this j atom. */
 415             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 416             velecsum         = _mm_add_pd(velecsum,velec);
 417
 418             fscal            = felec;
 419
 420             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 421
 422             /* Update vectorial force */
 423             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 424             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 425             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 426
 427             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 428             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 429             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 430
 431             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 432
 433             /* Inner loop uses 120 flops */
 434         }
 435
 436         /* End of innermost loop */
 437
 438         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 439                                               f+i_coord_offset,fshift+i_shift_offset);
 440
 441         ggid                        = gid[iidx];
 442         /* Update potential energies */
 443         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 444         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 445
 446         /* Increment number of inner iterations */
 447         inneriter                  += j_index_end - j_index_start;
 448
 449         /* Outer loop uses 20 flops */
 450     }
 451
 452     /* Increment number of outer iterations */
 453     outeriter        += nri;
 454
 455     /* Update outer/inner flops */
 456
 457     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*120);
 458 }
 459 /*
 460  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_avx_128_fma_double
 461  * Electrostatics interaction: ReactionField
 462  * VdW interaction:            LennardJones
 463  * Geometry:                   Water3-Particle
 464  * Calculate force/pot:        Force
 465  */
 466 void
 467 nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_avx_128_fma_double
 468                     (t_nblist * gmx_restrict                nlist,
 469                      rvec * gmx_restrict                    xx,
 470                      rvec * gmx_restrict                    ff,
 471                      t_forcerec * gmx_restrict              fr,
 472                      t_mdatoms * gmx_restrict               mdatoms,
 473                      nb_kernel_data_t * gmx_restrict        kernel_data,
 474                      t_nrnb * gmx_restrict                  nrnb)
 475 {
 476     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 477      * just 0 for non-waters.
 478      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 479      * jnr indices corresponding to data put in the four positions in the SIMD register.
 480      */
 481     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 482     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 483     int              jnrA,jnrB;
 484     int              j_coord_offsetA,j_coord_offsetB;
 485     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 486     real             rcutoff_scalar;
 487     real             *shiftvec,*fshift,*x,*f;
 488     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 489     int              vdwioffset0;
 490     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 491     int              vdwioffset1;
 492     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 493     int              vdwioffset2;
 494     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 495     int              vdwjidx0A,vdwjidx0B;
 496     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 497     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 498     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 499     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 500     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 501     real             *charge;
 502     int              nvdwtype;
 503     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 504     int              *vdwtype;
 505     real             *vdwparam;
 506     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 507     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 508     __m128d          dummy_mask,cutoff_mask;
 509     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 510     __m128d          one     = _mm_set1_pd(1.0);
 511     __m128d          two     = _mm_set1_pd(2.0);
 512     x                = xx[0];
 513     f                = ff[0];
 514
 515     nri              = nlist->nri;
 516     iinr             = nlist->iinr;
 517     jindex           = nlist->jindex;
 518     jjnr             = nlist->jjnr;
 519     shiftidx         = nlist->shift;
 520     gid              = nlist->gid;
 521     shiftvec         = fr->shift_vec[0];
 522     fshift           = fr->fshift[0];
 523     facel            = _mm_set1_pd(fr->epsfac);
 524     charge           = mdatoms->chargeA;
 525     krf              = _mm_set1_pd(fr->ic->k_rf);
 526     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 527     crf              = _mm_set1_pd(fr->ic->c_rf);
 528     nvdwtype         = fr->ntype;
 529     vdwparam         = fr->nbfp;
 530     vdwtype          = mdatoms->typeA;
 531
 532     /* Setup water-specific parameters */
 533     inr              = nlist->iinr[0];
 534     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 535     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 536     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 537     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 538
 539     /* Avoid stupid compiler warnings */
 540     jnrA = jnrB = 0;
 541     j_coord_offsetA = 0;
 542     j_coord_offsetB = 0;
 543
 544     outeriter        = 0;
 545     inneriter        = 0;
 546
 547     /* Start outer loop over neighborlists */
 548     for(iidx=0; iidx<nri; iidx++)
 549     {
 550         /* Load shift vector for this list */
 551         i_shift_offset   = DIM*shiftidx[iidx];
 552
 553         /* Load limits for loop over neighbors */
 554         j_index_start    = jindex[iidx];
 555         j_index_end      = jindex[iidx+1];
 556
 557         /* Get outer coordinate index */
 558         inr              = iinr[iidx];
 559         i_coord_offset   = DIM*inr;
 560
 561         /* Load i particle coords and add shift vector */
 562         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 563                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 564
 565         fix0             = _mm_setzero_pd();
 566         fiy0             = _mm_setzero_pd();
 567         fiz0             = _mm_setzero_pd();
 568         fix1             = _mm_setzero_pd();
 569         fiy1             = _mm_setzero_pd();
 570         fiz1             = _mm_setzero_pd();
 571         fix2             = _mm_setzero_pd();
 572         fiy2             = _mm_setzero_pd();
 573         fiz2             = _mm_setzero_pd();
 574
 575         /* Start inner kernel loop */
 576         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 577         {
 578
 579             /* Get j neighbor index, and coordinate index */
 580             jnrA             = jjnr[jidx];
 581             jnrB             = jjnr[jidx+1];
 582             j_coord_offsetA  = DIM*jnrA;
 583             j_coord_offsetB  = DIM*jnrB;
 584
 585             /* load j atom coordinates */
 586             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 587                                               &jx0,&jy0,&jz0);
 588
 589             /* Calculate displacement vector */
 590             dx00             = _mm_sub_pd(ix0,jx0);
 591             dy00             = _mm_sub_pd(iy0,jy0);
 592             dz00             = _mm_sub_pd(iz0,jz0);
 593             dx10             = _mm_sub_pd(ix1,jx0);
 594             dy10             = _mm_sub_pd(iy1,jy0);
 595             dz10             = _mm_sub_pd(iz1,jz0);
 596             dx20             = _mm_sub_pd(ix2,jx0);
 597             dy20             = _mm_sub_pd(iy2,jy0);
 598             dz20             = _mm_sub_pd(iz2,jz0);
 599
 600             /* Calculate squared distance and things based on it */
 601             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 602             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 603             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 604
 605             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 606             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 607             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 608
 609             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 610             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 611             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 612
 613             /* Load parameters for j particles */
 614             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 615             vdwjidx0A        = 2*vdwtype[jnrA+0];
 616             vdwjidx0B        = 2*vdwtype[jnrB+0];
 617
 618             fjx0             = _mm_setzero_pd();
 619             fjy0             = _mm_setzero_pd();
 620             fjz0             = _mm_setzero_pd();
 621
 622             /**************************
 623              * CALCULATE INTERACTIONS *
 624              **************************/
 625
 626             /* Compute parameters for interactions between i and j atoms */
 627             qq00             = _mm_mul_pd(iq0,jq0);
 628             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 629                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 630
 631             /* REACTION-FIELD ELECTROSTATICS */
 632             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 633
 634             /* LENNARD-JONES DISPERSION/REPULSION */
 635
 636             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 637             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 638
 639             fscal            = _mm_add_pd(felec,fvdw);
 640
 641             /* Update vectorial force */
 642             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 643             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 644             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 645
 646             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 647             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 648             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             /* Compute parameters for interactions between i and j atoms */
 655             qq10             = _mm_mul_pd(iq1,jq0);
 656
 657             /* REACTION-FIELD ELECTROSTATICS */
 658             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 659
 660             fscal            = felec;
 661
 662             /* Update vectorial force */
 663             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 664             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 665             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 666
 667             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 668             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 669             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 670
 671             /**************************
 672              * CALCULATE INTERACTIONS *
 673              **************************/
 674
 675             /* Compute parameters for interactions between i and j atoms */
 676             qq20             = _mm_mul_pd(iq2,jq0);
 677
 678             /* REACTION-FIELD ELECTROSTATICS */
 679             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 680
 681             fscal            = felec;
 682
 683             /* Update vectorial force */
 684             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 685             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 686             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 687
 688             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 689             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 690             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 691
 692             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
 693
 694             /* Inner loop uses 100 flops */
 695         }
 696
 697         if(jidx<j_index_end)
 698         {
 699
 700             jnrA             = jjnr[jidx];
 701             j_coord_offsetA  = DIM*jnrA;
 702
 703             /* load j atom coordinates */
 704             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 705                                               &jx0,&jy0,&jz0);
 706
 707             /* Calculate displacement vector */
 708             dx00             = _mm_sub_pd(ix0,jx0);
 709             dy00             = _mm_sub_pd(iy0,jy0);
 710             dz00             = _mm_sub_pd(iz0,jz0);
 711             dx10             = _mm_sub_pd(ix1,jx0);
 712             dy10             = _mm_sub_pd(iy1,jy0);
 713             dz10             = _mm_sub_pd(iz1,jz0);
 714             dx20             = _mm_sub_pd(ix2,jx0);
 715             dy20             = _mm_sub_pd(iy2,jy0);
 716             dz20             = _mm_sub_pd(iz2,jz0);
 717
 718             /* Calculate squared distance and things based on it */
 719             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 720             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 721             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 722
 723             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 724             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 725             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 726
 727             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 728             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 729             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 730
 731             /* Load parameters for j particles */
 732             jq0              = _mm_load_sd(charge+jnrA+0);
 733             vdwjidx0A        = 2*vdwtype[jnrA+0];
 734
 735             fjx0             = _mm_setzero_pd();
 736             fjy0             = _mm_setzero_pd();
 737             fjz0             = _mm_setzero_pd();
 738
 739             /**************************
 740              * CALCULATE INTERACTIONS *
 741              **************************/
 742
 743             /* Compute parameters for interactions between i and j atoms */
 744             qq00             = _mm_mul_pd(iq0,jq0);
 745             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 746
 747             /* REACTION-FIELD ELECTROSTATICS */
 748             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 749
 750             /* LENNARD-JONES DISPERSION/REPULSION */
 751
 752             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 753             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 754
 755             fscal            = _mm_add_pd(felec,fvdw);
 756
 757             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 758
 759             /* Update vectorial force */
 760             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 761             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 762             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 763
 764             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 765             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 766             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 767
 768             /**************************
 769              * CALCULATE INTERACTIONS *
 770              **************************/
 771
 772             /* Compute parameters for interactions between i and j atoms */
 773             qq10             = _mm_mul_pd(iq1,jq0);
 774
 775             /* REACTION-FIELD ELECTROSTATICS */
 776             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 777
 778             fscal            = felec;
 779
 780             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 781
 782             /* Update vectorial force */
 783             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 784             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 785             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 786
 787             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 788             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 789             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 790
 791             /**************************
 792              * CALCULATE INTERACTIONS *
 793              **************************/
 794
 795             /* Compute parameters for interactions between i and j atoms */
 796             qq20             = _mm_mul_pd(iq2,jq0);
 797
 798             /* REACTION-FIELD ELECTROSTATICS */
 799             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 800
 801             fscal            = felec;
 802
 803             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 804
 805             /* Update vectorial force */
 806             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 807             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 808             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 809
 810             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 811             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 812             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 813
 814             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
 815
 816             /* Inner loop uses 100 flops */
 817         }
 818
 819         /* End of innermost loop */
 820
 821         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 822                                               f+i_coord_offset,fshift+i_shift_offset);
 823
 824         /* Increment number of inner iterations */
 825         inneriter                  += j_index_end - j_index_start;
 826
 827         /* Outer loop uses 18 flops */
 828     }
 829
 830     /* Increment number of outer iterations */
 831     outeriter        += nri;
 832
 833     /* Update outer/inner flops */
 834
 835     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*100);
 836 }