src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecGB_VdwLJ_GeomP1P1_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_ElecGB_VdwLJ_GeomP1P1_VF_avx_128_fma_double
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwLJ_GeomP1P1_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              vdwjidx0A,vdwjidx0B;
  69     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  70     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  71     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  72     real             *charge;
  73     __m128i          gbitab;
  74     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
  75     __m128d          minushalf = _mm_set1_pd(-0.5);
  76     real             *invsqrta,*dvda,*gbtab;
  77     int              nvdwtype;
  78     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  79     int              *vdwtype;
  80     real             *vdwparam;
  81     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  82     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  83     __m128i          vfitab;
  84     __m128i          ifour       = _mm_set1_epi32(4);
  85     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
  86     real             *vftab;
  87     __m128d          dummy_mask,cutoff_mask;
  88     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  89     __m128d          one     = _mm_set1_pd(1.0);
  90     __m128d          two     = _mm_set1_pd(2.0);
  91     x                = xx[0];
  92     f                = ff[0];
  93
  94     nri              = nlist->nri;
  95     iinr             = nlist->iinr;
  96     jindex           = nlist->jindex;
  97     jjnr             = nlist->jjnr;
  98     shiftidx         = nlist->shift;
  99     gid              = nlist->gid;
 100     shiftvec         = fr->shift_vec[0];
 101     fshift           = fr->fshift[0];
 102     facel            = _mm_set1_pd(fr->epsfac);
 103     charge           = mdatoms->chargeA;
 104     nvdwtype         = fr->ntype;
 105     vdwparam         = fr->nbfp;
 106     vdwtype          = mdatoms->typeA;
 107
 108     invsqrta         = fr->invsqrta;
 109     dvda             = fr->dvda;
 110     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 111     gbtab            = fr->gbtab.data;
 112     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 113
 114     /* Avoid stupid compiler warnings */
 115     jnrA = jnrB = 0;
 116     j_coord_offsetA = 0;
 117     j_coord_offsetB = 0;
 118
 119     outeriter        = 0;
 120     inneriter        = 0;
 121
 122     /* Start outer loop over neighborlists */
 123     for(iidx=0; iidx<nri; iidx++)
 124     {
 125         /* Load shift vector for this list */
 126         i_shift_offset   = DIM*shiftidx[iidx];
 127
 128         /* Load limits for loop over neighbors */
 129         j_index_start    = jindex[iidx];
 130         j_index_end      = jindex[iidx+1];
 131
 132         /* Get outer coordinate index */
 133         inr              = iinr[iidx];
 134         i_coord_offset   = DIM*inr;
 135
 136         /* Load i particle coords and add shift vector */
 137         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 138
 139         fix0             = _mm_setzero_pd();
 140         fiy0             = _mm_setzero_pd();
 141         fiz0             = _mm_setzero_pd();
 142
 143         /* Load parameters for i particles */
 144         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 145         isai0            = _mm_load1_pd(invsqrta+inr+0);
 146         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 147
 148         /* Reset potential sums */
 149         velecsum         = _mm_setzero_pd();
 150         vgbsum           = _mm_setzero_pd();
 151         vvdwsum          = _mm_setzero_pd();
 152         dvdasum          = _mm_setzero_pd();
 153
 154         /* Start inner kernel loop */
 155         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 156         {
 157
 158             /* Get j neighbor index, and coordinate index */
 159             jnrA             = jjnr[jidx];
 160             jnrB             = jjnr[jidx+1];
 161             j_coord_offsetA  = DIM*jnrA;
 162             j_coord_offsetB  = DIM*jnrB;
 163
 164             /* load j atom coordinates */
 165             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 166                                               &jx0,&jy0,&jz0);
 167
 168             /* Calculate displacement vector */
 169             dx00             = _mm_sub_pd(ix0,jx0);
 170             dy00             = _mm_sub_pd(iy0,jy0);
 171             dz00             = _mm_sub_pd(iz0,jz0);
 172
 173             /* Calculate squared distance and things based on it */
 174             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 175
 176             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 177
 178             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 179
 180             /* Load parameters for j particles */
 181             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 182             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 183             vdwjidx0A        = 2*vdwtype[jnrA+0];
 184             vdwjidx0B        = 2*vdwtype[jnrB+0];
 185
 186             /**************************
 187              * CALCULATE INTERACTIONS *
 188              **************************/
 189
 190             r00              = _mm_mul_pd(rsq00,rinv00);
 191
 192             /* Compute parameters for interactions between i and j atoms */
 193             qq00             = _mm_mul_pd(iq0,jq0);
 194             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 195                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 196
 197             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 198             isaprod          = _mm_mul_pd(isai0,isaj0);
 199             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 200             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 201
 202             /* Calculate generalized born table index - this is a separate table from the normal one,
 203              * but we use the same procedure by multiplying r with scale and truncating to integer.
 204              */
 205             rt               = _mm_mul_pd(r00,gbscale);
 206             gbitab           = _mm_cvttpd_epi32(rt);
 207 #ifdef __XOP__
 208             gbeps            = _mm_frcz_pd(rt);
 209 #else
 210             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 211 #endif
 212             gbitab           = _mm_slli_epi32(gbitab,2);
 213
 214             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 215             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 216             GMX_MM_TRANSPOSE2_PD(Y,F);
 217             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 218             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 219             GMX_MM_TRANSPOSE2_PD(G,H);
 220             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 221             VV               = _mm_macc_pd(gbeps,Fp,Y);
 222             vgb              = _mm_mul_pd(gbqqfactor,VV);
 223
 224             twogbeps         = _mm_add_pd(gbeps,gbeps);
 225             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 226             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 227             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 228             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 229             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 230             velec            = _mm_mul_pd(qq00,rinv00);
 231             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 232
 233             /* LENNARD-JONES DISPERSION/REPULSION */
 234
 235             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 236             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 237             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 238             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 239             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 240
 241             /* Update potential sum for this i atom from the interaction with this j atom. */
 242             velecsum         = _mm_add_pd(velecsum,velec);
 243             vgbsum           = _mm_add_pd(vgbsum,vgb);
 244             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 245
 246             fscal            = _mm_add_pd(felec,fvdw);
 247
 248             /* Update vectorial force */
 249             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 250             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 251             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 252
 253             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 254                                                    _mm_mul_pd(dx00,fscal),
 255                                                    _mm_mul_pd(dy00,fscal),
 256                                                    _mm_mul_pd(dz00,fscal));
 257
 258             /* Inner loop uses 74 flops */
 259         }
 260
 261         if(jidx<j_index_end)
 262         {
 263
 264             jnrA             = jjnr[jidx];
 265             j_coord_offsetA  = DIM*jnrA;
 266
 267             /* load j atom coordinates */
 268             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 269                                               &jx0,&jy0,&jz0);
 270
 271             /* Calculate displacement vector */
 272             dx00             = _mm_sub_pd(ix0,jx0);
 273             dy00             = _mm_sub_pd(iy0,jy0);
 274             dz00             = _mm_sub_pd(iz0,jz0);
 275
 276             /* Calculate squared distance and things based on it */
 277             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 278
 279             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 280
 281             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 282
 283             /* Load parameters for j particles */
 284             jq0              = _mm_load_sd(charge+jnrA+0);
 285             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 286             vdwjidx0A        = 2*vdwtype[jnrA+0];
 287
 288             /**************************
 289              * CALCULATE INTERACTIONS *
 290              **************************/
 291
 292             r00              = _mm_mul_pd(rsq00,rinv00);
 293
 294             /* Compute parameters for interactions between i and j atoms */
 295             qq00             = _mm_mul_pd(iq0,jq0);
 296             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 297
 298             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 299             isaprod          = _mm_mul_pd(isai0,isaj0);
 300             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 301             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 302
 303             /* Calculate generalized born table index - this is a separate table from the normal one,
 304              * but we use the same procedure by multiplying r with scale and truncating to integer.
 305              */
 306             rt               = _mm_mul_pd(r00,gbscale);
 307             gbitab           = _mm_cvttpd_epi32(rt);
 308 #ifdef __XOP__
 309             gbeps            = _mm_frcz_pd(rt);
 310 #else
 311             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 312 #endif
 313             gbitab           = _mm_slli_epi32(gbitab,2);
 314
 315             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 316             F                = _mm_setzero_pd();
 317             GMX_MM_TRANSPOSE2_PD(Y,F);
 318             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 319             H                = _mm_setzero_pd();
 320             GMX_MM_TRANSPOSE2_PD(G,H);
 321             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 322             VV               = _mm_macc_pd(gbeps,Fp,Y);
 323             vgb              = _mm_mul_pd(gbqqfactor,VV);
 324
 325             twogbeps         = _mm_add_pd(gbeps,gbeps);
 326             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 327             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 328             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 329             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 330             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 331             velec            = _mm_mul_pd(qq00,rinv00);
 332             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 333
 334             /* LENNARD-JONES DISPERSION/REPULSION */
 335
 336             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 337             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 338             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 339             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 340             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 341
 342             /* Update potential sum for this i atom from the interaction with this j atom. */
 343             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 344             velecsum         = _mm_add_pd(velecsum,velec);
 345             vgb              = _mm_unpacklo_pd(vgb,_mm_setzero_pd());
 346             vgbsum           = _mm_add_pd(vgbsum,vgb);
 347             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 348             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 349
 350             fscal            = _mm_add_pd(felec,fvdw);
 351
 352             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 353
 354             /* Update vectorial force */
 355             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 356             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 357             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 358
 359             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 360                                                    _mm_mul_pd(dx00,fscal),
 361                                                    _mm_mul_pd(dy00,fscal),
 362                                                    _mm_mul_pd(dz00,fscal));
 363
 364             /* Inner loop uses 74 flops */
 365         }
 366
 367         /* End of innermost loop */
 368
 369         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 370                                               f+i_coord_offset,fshift+i_shift_offset);
 371
 372         ggid                        = gid[iidx];
 373         /* Update potential energies */
 374         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 375         gmx_mm_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 376         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 377         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 378         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 379
 380         /* Increment number of inner iterations */
 381         inneriter                  += j_index_end - j_index_start;
 382
 383         /* Outer loop uses 10 flops */
 384     }
 385
 386     /* Increment number of outer iterations */
 387     outeriter        += nri;
 388
 389     /* Update outer/inner flops */
 390
 391     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*74);
 392 }
 393 /*
 394  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_double
 395  * Electrostatics interaction: GeneralizedBorn
 396  * VdW interaction:            LennardJones
 397  * Geometry:                   Particle-Particle
 398  * Calculate force/pot:        Force
 399  */
 400 void
 401 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_double
 402                     (t_nblist * gmx_restrict                nlist,
 403                      rvec * gmx_restrict                    xx,
 404                      rvec * gmx_restrict                    ff,
 405                      t_forcerec * gmx_restrict              fr,
 406                      t_mdatoms * gmx_restrict               mdatoms,
 407                      nb_kernel_data_t * gmx_restrict        kernel_data,
 408                      t_nrnb * gmx_restrict                  nrnb)
 409 {
 410     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 411      * just 0 for non-waters.
 412      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 413      * jnr indices corresponding to data put in the four positions in the SIMD register.
 414      */
 415     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 416     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 417     int              jnrA,jnrB;
 418     int              j_coord_offsetA,j_coord_offsetB;
 419     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 420     real             rcutoff_scalar;
 421     real             *shiftvec,*fshift,*x,*f;
 422     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 423     int              vdwioffset0;
 424     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 425     int              vdwjidx0A,vdwjidx0B;
 426     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 427     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 428     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 429     real             *charge;
 430     __m128i          gbitab;
 431     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
 432     __m128d          minushalf = _mm_set1_pd(-0.5);
 433     real             *invsqrta,*dvda,*gbtab;
 434     int              nvdwtype;
 435     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 436     int              *vdwtype;
 437     real             *vdwparam;
 438     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 439     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 440     __m128i          vfitab;
 441     __m128i          ifour       = _mm_set1_epi32(4);
 442     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 443     real             *vftab;
 444     __m128d          dummy_mask,cutoff_mask;
 445     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 446     __m128d          one     = _mm_set1_pd(1.0);
 447     __m128d          two     = _mm_set1_pd(2.0);
 448     x                = xx[0];
 449     f                = ff[0];
 450
 451     nri              = nlist->nri;
 452     iinr             = nlist->iinr;
 453     jindex           = nlist->jindex;
 454     jjnr             = nlist->jjnr;
 455     shiftidx         = nlist->shift;
 456     gid              = nlist->gid;
 457     shiftvec         = fr->shift_vec[0];
 458     fshift           = fr->fshift[0];
 459     facel            = _mm_set1_pd(fr->epsfac);
 460     charge           = mdatoms->chargeA;
 461     nvdwtype         = fr->ntype;
 462     vdwparam         = fr->nbfp;
 463     vdwtype          = mdatoms->typeA;
 464
 465     invsqrta         = fr->invsqrta;
 466     dvda             = fr->dvda;
 467     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 468     gbtab            = fr->gbtab.data;
 469     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 470
 471     /* Avoid stupid compiler warnings */
 472     jnrA = jnrB = 0;
 473     j_coord_offsetA = 0;
 474     j_coord_offsetB = 0;
 475
 476     outeriter        = 0;
 477     inneriter        = 0;
 478
 479     /* Start outer loop over neighborlists */
 480     for(iidx=0; iidx<nri; iidx++)
 481     {
 482         /* Load shift vector for this list */
 483         i_shift_offset   = DIM*shiftidx[iidx];
 484
 485         /* Load limits for loop over neighbors */
 486         j_index_start    = jindex[iidx];
 487         j_index_end      = jindex[iidx+1];
 488
 489         /* Get outer coordinate index */
 490         inr              = iinr[iidx];
 491         i_coord_offset   = DIM*inr;
 492
 493         /* Load i particle coords and add shift vector */
 494         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 495
 496         fix0             = _mm_setzero_pd();
 497         fiy0             = _mm_setzero_pd();
 498         fiz0             = _mm_setzero_pd();
 499
 500         /* Load parameters for i particles */
 501         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 502         isai0            = _mm_load1_pd(invsqrta+inr+0);
 503         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 504
 505         dvdasum          = _mm_setzero_pd();
 506
 507         /* Start inner kernel loop */
 508         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 509         {
 510
 511             /* Get j neighbor index, and coordinate index */
 512             jnrA             = jjnr[jidx];
 513             jnrB             = jjnr[jidx+1];
 514             j_coord_offsetA  = DIM*jnrA;
 515             j_coord_offsetB  = DIM*jnrB;
 516
 517             /* load j atom coordinates */
 518             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 519                                               &jx0,&jy0,&jz0);
 520
 521             /* Calculate displacement vector */
 522             dx00             = _mm_sub_pd(ix0,jx0);
 523             dy00             = _mm_sub_pd(iy0,jy0);
 524             dz00             = _mm_sub_pd(iz0,jz0);
 525
 526             /* Calculate squared distance and things based on it */
 527             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 528
 529             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 530
 531             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 532
 533             /* Load parameters for j particles */
 534             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 535             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 536             vdwjidx0A        = 2*vdwtype[jnrA+0];
 537             vdwjidx0B        = 2*vdwtype[jnrB+0];
 538
 539             /**************************
 540              * CALCULATE INTERACTIONS *
 541              **************************/
 542
 543             r00              = _mm_mul_pd(rsq00,rinv00);
 544
 545             /* Compute parameters for interactions between i and j atoms */
 546             qq00             = _mm_mul_pd(iq0,jq0);
 547             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 548                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 549
 550             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 551             isaprod          = _mm_mul_pd(isai0,isaj0);
 552             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 553             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 554
 555             /* Calculate generalized born table index - this is a separate table from the normal one,
 556              * but we use the same procedure by multiplying r with scale and truncating to integer.
 557              */
 558             rt               = _mm_mul_pd(r00,gbscale);
 559             gbitab           = _mm_cvttpd_epi32(rt);
 560 #ifdef __XOP__
 561             gbeps            = _mm_frcz_pd(rt);
 562 #else
 563             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 564 #endif
 565             gbitab           = _mm_slli_epi32(gbitab,2);
 566
 567             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 568             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 569             GMX_MM_TRANSPOSE2_PD(Y,F);
 570             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 571             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 572             GMX_MM_TRANSPOSE2_PD(G,H);
 573             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 574             VV               = _mm_macc_pd(gbeps,Fp,Y);
 575             vgb              = _mm_mul_pd(gbqqfactor,VV);
 576
 577             twogbeps         = _mm_add_pd(gbeps,gbeps);
 578             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 579             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 580             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 581             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 582             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 583             velec            = _mm_mul_pd(qq00,rinv00);
 584             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 585
 586             /* LENNARD-JONES DISPERSION/REPULSION */
 587
 588             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 589             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 590
 591             fscal            = _mm_add_pd(felec,fvdw);
 592
 593             /* Update vectorial force */
 594             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 595             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 596             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 597
 598             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 599                                                    _mm_mul_pd(dx00,fscal),
 600                                                    _mm_mul_pd(dy00,fscal),
 601                                                    _mm_mul_pd(dz00,fscal));
 602
 603             /* Inner loop uses 67 flops */
 604         }
 605
 606         if(jidx<j_index_end)
 607         {
 608
 609             jnrA             = jjnr[jidx];
 610             j_coord_offsetA  = DIM*jnrA;
 611
 612             /* load j atom coordinates */
 613             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 614                                               &jx0,&jy0,&jz0);
 615
 616             /* Calculate displacement vector */
 617             dx00             = _mm_sub_pd(ix0,jx0);
 618             dy00             = _mm_sub_pd(iy0,jy0);
 619             dz00             = _mm_sub_pd(iz0,jz0);
 620
 621             /* Calculate squared distance and things based on it */
 622             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 623
 624             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 625
 626             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 627
 628             /* Load parameters for j particles */
 629             jq0              = _mm_load_sd(charge+jnrA+0);
 630             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 631             vdwjidx0A        = 2*vdwtype[jnrA+0];
 632
 633             /**************************
 634              * CALCULATE INTERACTIONS *
 635              **************************/
 636
 637             r00              = _mm_mul_pd(rsq00,rinv00);
 638
 639             /* Compute parameters for interactions between i and j atoms */
 640             qq00             = _mm_mul_pd(iq0,jq0);
 641             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 642
 643             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 644             isaprod          = _mm_mul_pd(isai0,isaj0);
 645             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 646             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 647
 648             /* Calculate generalized born table index - this is a separate table from the normal one,
 649              * but we use the same procedure by multiplying r with scale and truncating to integer.
 650              */
 651             rt               = _mm_mul_pd(r00,gbscale);
 652             gbitab           = _mm_cvttpd_epi32(rt);
 653 #ifdef __XOP__
 654             gbeps            = _mm_frcz_pd(rt);
 655 #else
 656             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 657 #endif
 658             gbitab           = _mm_slli_epi32(gbitab,2);
 659
 660             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 661             F                = _mm_setzero_pd();
 662             GMX_MM_TRANSPOSE2_PD(Y,F);
 663             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 664             H                = _mm_setzero_pd();
 665             GMX_MM_TRANSPOSE2_PD(G,H);
 666             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 667             VV               = _mm_macc_pd(gbeps,Fp,Y);
 668             vgb              = _mm_mul_pd(gbqqfactor,VV);
 669
 670             twogbeps         = _mm_add_pd(gbeps,gbeps);
 671             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 672             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 673             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 674             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 675             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 676             velec            = _mm_mul_pd(qq00,rinv00);
 677             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 678
 679             /* LENNARD-JONES DISPERSION/REPULSION */
 680
 681             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 682             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
 683
 684             fscal            = _mm_add_pd(felec,fvdw);
 685
 686             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 687
 688             /* Update vectorial force */
 689             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 690             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 691             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 692
 693             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 694                                                    _mm_mul_pd(dx00,fscal),
 695                                                    _mm_mul_pd(dy00,fscal),
 696                                                    _mm_mul_pd(dz00,fscal));
 697
 698             /* Inner loop uses 67 flops */
 699         }
 700
 701         /* End of innermost loop */
 702
 703         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 704                                               f+i_coord_offset,fshift+i_shift_offset);
 705
 706         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 707         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 708
 709         /* Increment number of inner iterations */
 710         inneriter                  += j_index_end - j_index_start;
 711
 712         /* Outer loop uses 7 flops */
 713     }
 714
 715     /* Increment number of outer iterations */
 716     outeriter        += nri;
 717
 718     /* Update outer/inner flops */
 719
 720     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*67);
 721 }