src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecGB_VdwNone_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_VdwNone_GeomP1P1_VF_avx_128_fma_double
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            None
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwNone_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     __m128i          vfitab;
  78     __m128i          ifour       = _mm_set1_epi32(4);
  79     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
  80     real             *vftab;
  81     __m128d          dummy_mask,cutoff_mask;
  82     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  83     __m128d          one     = _mm_set1_pd(1.0);
  84     __m128d          two     = _mm_set1_pd(2.0);
  85     x                = xx[0];
  86     f                = ff[0];
  87
  88     nri              = nlist->nri;
  89     iinr             = nlist->iinr;
  90     jindex           = nlist->jindex;
  91     jjnr             = nlist->jjnr;
  92     shiftidx         = nlist->shift;
  93     gid              = nlist->gid;
  94     shiftvec         = fr->shift_vec[0];
  95     fshift           = fr->fshift[0];
  96     facel            = _mm_set1_pd(fr->epsfac);
  97     charge           = mdatoms->chargeA;
  98
  99     invsqrta         = fr->invsqrta;
 100     dvda             = fr->dvda;
 101     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 102     gbtab            = fr->gbtab.data;
 103     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 104
 105     /* Avoid stupid compiler warnings */
 106     jnrA = jnrB = 0;
 107     j_coord_offsetA = 0;
 108     j_coord_offsetB = 0;
 109
 110     outeriter        = 0;
 111     inneriter        = 0;
 112
 113     /* Start outer loop over neighborlists */
 114     for(iidx=0; iidx<nri; iidx++)
 115     {
 116         /* Load shift vector for this list */
 117         i_shift_offset   = DIM*shiftidx[iidx];
 118
 119         /* Load limits for loop over neighbors */
 120         j_index_start    = jindex[iidx];
 121         j_index_end      = jindex[iidx+1];
 122
 123         /* Get outer coordinate index */
 124         inr              = iinr[iidx];
 125         i_coord_offset   = DIM*inr;
 126
 127         /* Load i particle coords and add shift vector */
 128         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 129
 130         fix0             = _mm_setzero_pd();
 131         fiy0             = _mm_setzero_pd();
 132         fiz0             = _mm_setzero_pd();
 133
 134         /* Load parameters for i particles */
 135         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 136         isai0            = _mm_load1_pd(invsqrta+inr+0);
 137
 138         /* Reset potential sums */
 139         velecsum         = _mm_setzero_pd();
 140         vgbsum           = _mm_setzero_pd();
 141         dvdasum          = _mm_setzero_pd();
 142
 143         /* Start inner kernel loop */
 144         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 145         {
 146
 147             /* Get j neighbor index, and coordinate index */
 148             jnrA             = jjnr[jidx];
 149             jnrB             = jjnr[jidx+1];
 150             j_coord_offsetA  = DIM*jnrA;
 151             j_coord_offsetB  = DIM*jnrB;
 152
 153             /* load j atom coordinates */
 154             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 155                                               &jx0,&jy0,&jz0);
 156
 157             /* Calculate displacement vector */
 158             dx00             = _mm_sub_pd(ix0,jx0);
 159             dy00             = _mm_sub_pd(iy0,jy0);
 160             dz00             = _mm_sub_pd(iz0,jz0);
 161
 162             /* Calculate squared distance and things based on it */
 163             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 164
 165             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 166
 167             /* Load parameters for j particles */
 168             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 169             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 170
 171             /**************************
 172              * CALCULATE INTERACTIONS *
 173              **************************/
 174
 175             r00              = _mm_mul_pd(rsq00,rinv00);
 176
 177             /* Compute parameters for interactions between i and j atoms */
 178             qq00             = _mm_mul_pd(iq0,jq0);
 179
 180             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 181             isaprod          = _mm_mul_pd(isai0,isaj0);
 182             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 183             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 184
 185             /* Calculate generalized born table index - this is a separate table from the normal one,
 186              * but we use the same procedure by multiplying r with scale and truncating to integer.
 187              */
 188             rt               = _mm_mul_pd(r00,gbscale);
 189             gbitab           = _mm_cvttpd_epi32(rt);
 190 #ifdef __XOP__
 191             gbeps            = _mm_frcz_pd(rt);
 192 #else
 193             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 194 #endif
 195             gbitab           = _mm_slli_epi32(gbitab,2);
 196
 197             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 198             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 199             GMX_MM_TRANSPOSE2_PD(Y,F);
 200             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 201             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 202             GMX_MM_TRANSPOSE2_PD(G,H);
 203             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 204             VV               = _mm_macc_pd(gbeps,Fp,Y);
 205             vgb              = _mm_mul_pd(gbqqfactor,VV);
 206
 207             twogbeps         = _mm_add_pd(gbeps,gbeps);
 208             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 209             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 210             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 211             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 212             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 213             velec            = _mm_mul_pd(qq00,rinv00);
 214             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 215
 216             /* Update potential sum for this i atom from the interaction with this j atom. */
 217             velecsum         = _mm_add_pd(velecsum,velec);
 218             vgbsum           = _mm_add_pd(vgbsum,vgb);
 219
 220             fscal            = felec;
 221
 222             /* Update vectorial force */
 223             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 224             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 225             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 226
 227             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 228                                                    _mm_mul_pd(dx00,fscal),
 229                                                    _mm_mul_pd(dy00,fscal),
 230                                                    _mm_mul_pd(dz00,fscal));
 231
 232             /* Inner loop uses 61 flops */
 233         }
 234
 235         if(jidx<j_index_end)
 236         {
 237
 238             jnrA             = jjnr[jidx];
 239             j_coord_offsetA  = DIM*jnrA;
 240
 241             /* load j atom coordinates */
 242             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 243                                               &jx0,&jy0,&jz0);
 244
 245             /* Calculate displacement vector */
 246             dx00             = _mm_sub_pd(ix0,jx0);
 247             dy00             = _mm_sub_pd(iy0,jy0);
 248             dz00             = _mm_sub_pd(iz0,jz0);
 249
 250             /* Calculate squared distance and things based on it */
 251             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 252
 253             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 254
 255             /* Load parameters for j particles */
 256             jq0              = _mm_load_sd(charge+jnrA+0);
 257             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 258
 259             /**************************
 260              * CALCULATE INTERACTIONS *
 261              **************************/
 262
 263             r00              = _mm_mul_pd(rsq00,rinv00);
 264
 265             /* Compute parameters for interactions between i and j atoms */
 266             qq00             = _mm_mul_pd(iq0,jq0);
 267
 268             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 269             isaprod          = _mm_mul_pd(isai0,isaj0);
 270             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 271             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 272
 273             /* Calculate generalized born table index - this is a separate table from the normal one,
 274              * but we use the same procedure by multiplying r with scale and truncating to integer.
 275              */
 276             rt               = _mm_mul_pd(r00,gbscale);
 277             gbitab           = _mm_cvttpd_epi32(rt);
 278 #ifdef __XOP__
 279             gbeps            = _mm_frcz_pd(rt);
 280 #else
 281             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 282 #endif
 283             gbitab           = _mm_slli_epi32(gbitab,2);
 284
 285             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 286             F                = _mm_setzero_pd();
 287             GMX_MM_TRANSPOSE2_PD(Y,F);
 288             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 289             H                = _mm_setzero_pd();
 290             GMX_MM_TRANSPOSE2_PD(G,H);
 291             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 292             VV               = _mm_macc_pd(gbeps,Fp,Y);
 293             vgb              = _mm_mul_pd(gbqqfactor,VV);
 294
 295             twogbeps         = _mm_add_pd(gbeps,gbeps);
 296             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 297             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 298             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 299             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 300             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 301             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 302             velec            = _mm_mul_pd(qq00,rinv00);
 303             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 304
 305             /* Update potential sum for this i atom from the interaction with this j atom. */
 306             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 307             velecsum         = _mm_add_pd(velecsum,velec);
 308             vgb              = _mm_unpacklo_pd(vgb,_mm_setzero_pd());
 309             vgbsum           = _mm_add_pd(vgbsum,vgb);
 310
 311             fscal            = felec;
 312
 313             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 314
 315             /* Update vectorial force */
 316             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 317             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 318             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 319
 320             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 321                                                    _mm_mul_pd(dx00,fscal),
 322                                                    _mm_mul_pd(dy00,fscal),
 323                                                    _mm_mul_pd(dz00,fscal));
 324
 325             /* Inner loop uses 61 flops */
 326         }
 327
 328         /* End of innermost loop */
 329
 330         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 331                                               f+i_coord_offset,fshift+i_shift_offset);
 332
 333         ggid                        = gid[iidx];
 334         /* Update potential energies */
 335         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 336         gmx_mm_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 337         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 338         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 339
 340         /* Increment number of inner iterations */
 341         inneriter                  += j_index_end - j_index_start;
 342
 343         /* Outer loop uses 9 flops */
 344     }
 345
 346     /* Increment number of outer iterations */
 347     outeriter        += nri;
 348
 349     /* Update outer/inner flops */
 350
 351     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*61);
 352 }
 353 /*
 354  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_double
 355  * Electrostatics interaction: GeneralizedBorn
 356  * VdW interaction:            None
 357  * Geometry:                   Particle-Particle
 358  * Calculate force/pot:        Force
 359  */
 360 void
 361 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_double
 362                     (t_nblist * gmx_restrict                nlist,
 363                      rvec * gmx_restrict                    xx,
 364                      rvec * gmx_restrict                    ff,
 365                      t_forcerec * gmx_restrict              fr,
 366                      t_mdatoms * gmx_restrict               mdatoms,
 367                      nb_kernel_data_t * gmx_restrict        kernel_data,
 368                      t_nrnb * gmx_restrict                  nrnb)
 369 {
 370     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 371      * just 0 for non-waters.
 372      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 373      * jnr indices corresponding to data put in the four positions in the SIMD register.
 374      */
 375     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 376     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 377     int              jnrA,jnrB;
 378     int              j_coord_offsetA,j_coord_offsetB;
 379     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 380     real             rcutoff_scalar;
 381     real             *shiftvec,*fshift,*x,*f;
 382     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 383     int              vdwioffset0;
 384     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 385     int              vdwjidx0A,vdwjidx0B;
 386     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 387     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 388     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 389     real             *charge;
 390     __m128i          gbitab;
 391     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
 392     __m128d          minushalf = _mm_set1_pd(-0.5);
 393     real             *invsqrta,*dvda,*gbtab;
 394     __m128i          vfitab;
 395     __m128i          ifour       = _mm_set1_epi32(4);
 396     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 397     real             *vftab;
 398     __m128d          dummy_mask,cutoff_mask;
 399     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 400     __m128d          one     = _mm_set1_pd(1.0);
 401     __m128d          two     = _mm_set1_pd(2.0);
 402     x                = xx[0];
 403     f                = ff[0];
 404
 405     nri              = nlist->nri;
 406     iinr             = nlist->iinr;
 407     jindex           = nlist->jindex;
 408     jjnr             = nlist->jjnr;
 409     shiftidx         = nlist->shift;
 410     gid              = nlist->gid;
 411     shiftvec         = fr->shift_vec[0];
 412     fshift           = fr->fshift[0];
 413     facel            = _mm_set1_pd(fr->epsfac);
 414     charge           = mdatoms->chargeA;
 415
 416     invsqrta         = fr->invsqrta;
 417     dvda             = fr->dvda;
 418     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 419     gbtab            = fr->gbtab.data;
 420     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 421
 422     /* Avoid stupid compiler warnings */
 423     jnrA = jnrB = 0;
 424     j_coord_offsetA = 0;
 425     j_coord_offsetB = 0;
 426
 427     outeriter        = 0;
 428     inneriter        = 0;
 429
 430     /* Start outer loop over neighborlists */
 431     for(iidx=0; iidx<nri; iidx++)
 432     {
 433         /* Load shift vector for this list */
 434         i_shift_offset   = DIM*shiftidx[iidx];
 435
 436         /* Load limits for loop over neighbors */
 437         j_index_start    = jindex[iidx];
 438         j_index_end      = jindex[iidx+1];
 439
 440         /* Get outer coordinate index */
 441         inr              = iinr[iidx];
 442         i_coord_offset   = DIM*inr;
 443
 444         /* Load i particle coords and add shift vector */
 445         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 446
 447         fix0             = _mm_setzero_pd();
 448         fiy0             = _mm_setzero_pd();
 449         fiz0             = _mm_setzero_pd();
 450
 451         /* Load parameters for i particles */
 452         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 453         isai0            = _mm_load1_pd(invsqrta+inr+0);
 454
 455         dvdasum          = _mm_setzero_pd();
 456
 457         /* Start inner kernel loop */
 458         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 459         {
 460
 461             /* Get j neighbor index, and coordinate index */
 462             jnrA             = jjnr[jidx];
 463             jnrB             = jjnr[jidx+1];
 464             j_coord_offsetA  = DIM*jnrA;
 465             j_coord_offsetB  = DIM*jnrB;
 466
 467             /* load j atom coordinates */
 468             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 469                                               &jx0,&jy0,&jz0);
 470
 471             /* Calculate displacement vector */
 472             dx00             = _mm_sub_pd(ix0,jx0);
 473             dy00             = _mm_sub_pd(iy0,jy0);
 474             dz00             = _mm_sub_pd(iz0,jz0);
 475
 476             /* Calculate squared distance and things based on it */
 477             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 478
 479             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 480
 481             /* Load parameters for j particles */
 482             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 483             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 484
 485             /**************************
 486              * CALCULATE INTERACTIONS *
 487              **************************/
 488
 489             r00              = _mm_mul_pd(rsq00,rinv00);
 490
 491             /* Compute parameters for interactions between i and j atoms */
 492             qq00             = _mm_mul_pd(iq0,jq0);
 493
 494             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 495             isaprod          = _mm_mul_pd(isai0,isaj0);
 496             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 497             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 498
 499             /* Calculate generalized born table index - this is a separate table from the normal one,
 500              * but we use the same procedure by multiplying r with scale and truncating to integer.
 501              */
 502             rt               = _mm_mul_pd(r00,gbscale);
 503             gbitab           = _mm_cvttpd_epi32(rt);
 504 #ifdef __XOP__
 505             gbeps            = _mm_frcz_pd(rt);
 506 #else
 507             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 508 #endif
 509             gbitab           = _mm_slli_epi32(gbitab,2);
 510
 511             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 512             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 513             GMX_MM_TRANSPOSE2_PD(Y,F);
 514             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 515             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 516             GMX_MM_TRANSPOSE2_PD(G,H);
 517             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 518             VV               = _mm_macc_pd(gbeps,Fp,Y);
 519             vgb              = _mm_mul_pd(gbqqfactor,VV);
 520
 521             twogbeps         = _mm_add_pd(gbeps,gbeps);
 522             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 523             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 524             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 525             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 526             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 527             velec            = _mm_mul_pd(qq00,rinv00);
 528             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 529
 530             fscal            = felec;
 531
 532             /* Update vectorial force */
 533             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 534             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 535             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 536
 537             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 538                                                    _mm_mul_pd(dx00,fscal),
 539                                                    _mm_mul_pd(dy00,fscal),
 540                                                    _mm_mul_pd(dz00,fscal));
 541
 542             /* Inner loop uses 59 flops */
 543         }
 544
 545         if(jidx<j_index_end)
 546         {
 547
 548             jnrA             = jjnr[jidx];
 549             j_coord_offsetA  = DIM*jnrA;
 550
 551             /* load j atom coordinates */
 552             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 553                                               &jx0,&jy0,&jz0);
 554
 555             /* Calculate displacement vector */
 556             dx00             = _mm_sub_pd(ix0,jx0);
 557             dy00             = _mm_sub_pd(iy0,jy0);
 558             dz00             = _mm_sub_pd(iz0,jz0);
 559
 560             /* Calculate squared distance and things based on it */
 561             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 562
 563             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 564
 565             /* Load parameters for j particles */
 566             jq0              = _mm_load_sd(charge+jnrA+0);
 567             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 568
 569             /**************************
 570              * CALCULATE INTERACTIONS *
 571              **************************/
 572
 573             r00              = _mm_mul_pd(rsq00,rinv00);
 574
 575             /* Compute parameters for interactions between i and j atoms */
 576             qq00             = _mm_mul_pd(iq0,jq0);
 577
 578             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 579             isaprod          = _mm_mul_pd(isai0,isaj0);
 580             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 581             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 582
 583             /* Calculate generalized born table index - this is a separate table from the normal one,
 584              * but we use the same procedure by multiplying r with scale and truncating to integer.
 585              */
 586             rt               = _mm_mul_pd(r00,gbscale);
 587             gbitab           = _mm_cvttpd_epi32(rt);
 588 #ifdef __XOP__
 589             gbeps            = _mm_frcz_pd(rt);
 590 #else
 591             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 592 #endif
 593             gbitab           = _mm_slli_epi32(gbitab,2);
 594
 595             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 596             F                = _mm_setzero_pd();
 597             GMX_MM_TRANSPOSE2_PD(Y,F);
 598             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 599             H                = _mm_setzero_pd();
 600             GMX_MM_TRANSPOSE2_PD(G,H);
 601             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 602             VV               = _mm_macc_pd(gbeps,Fp,Y);
 603             vgb              = _mm_mul_pd(gbqqfactor,VV);
 604
 605             twogbeps         = _mm_add_pd(gbeps,gbeps);
 606             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 607             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 608             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 609             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 610             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 611             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 612             velec            = _mm_mul_pd(qq00,rinv00);
 613             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 614
 615             fscal            = felec;
 616
 617             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 618
 619             /* Update vectorial force */
 620             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 621             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 622             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 623
 624             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 625                                                    _mm_mul_pd(dx00,fscal),
 626                                                    _mm_mul_pd(dy00,fscal),
 627                                                    _mm_mul_pd(dz00,fscal));
 628
 629             /* Inner loop uses 59 flops */
 630         }
 631
 632         /* End of innermost loop */
 633
 634         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 635                                               f+i_coord_offset,fshift+i_shift_offset);
 636
 637         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 638         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 639
 640         /* Increment number of inner iterations */
 641         inneriter                  += j_index_end - j_index_start;
 642
 643         /* Outer loop uses 7 flops */
 644     }
 645
 646     /* Increment number of outer iterations */
 647     outeriter        += nri;
 648
 649     /* Update outer/inner flops */
 650
 651     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*59);
 652 }