src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecGB_VdwCSTab_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_VdwCSTab_GeomP1P1_VF_avx_128_fma_double
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwCSTab_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     vftab            = kernel_data->table_vdw->data;
 109     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 110
 111     invsqrta         = fr->invsqrta;
 112     dvda             = fr->dvda;
 113     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 114     gbtab            = fr->gbtab.data;
 115     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 116
 117     /* Avoid stupid compiler warnings */
 118     jnrA = jnrB = 0;
 119     j_coord_offsetA = 0;
 120     j_coord_offsetB = 0;
 121
 122     outeriter        = 0;
 123     inneriter        = 0;
 124
 125     /* Start outer loop over neighborlists */
 126     for(iidx=0; iidx<nri; iidx++)
 127     {
 128         /* Load shift vector for this list */
 129         i_shift_offset   = DIM*shiftidx[iidx];
 130
 131         /* Load limits for loop over neighbors */
 132         j_index_start    = jindex[iidx];
 133         j_index_end      = jindex[iidx+1];
 134
 135         /* Get outer coordinate index */
 136         inr              = iinr[iidx];
 137         i_coord_offset   = DIM*inr;
 138
 139         /* Load i particle coords and add shift vector */
 140         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 141
 142         fix0             = _mm_setzero_pd();
 143         fiy0             = _mm_setzero_pd();
 144         fiz0             = _mm_setzero_pd();
 145
 146         /* Load parameters for i particles */
 147         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 148         isai0            = _mm_load1_pd(invsqrta+inr+0);
 149         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 150
 151         /* Reset potential sums */
 152         velecsum         = _mm_setzero_pd();
 153         vgbsum           = _mm_setzero_pd();
 154         vvdwsum          = _mm_setzero_pd();
 155         dvdasum          = _mm_setzero_pd();
 156
 157         /* Start inner kernel loop */
 158         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 159         {
 160
 161             /* Get j neighbor index, and coordinate index */
 162             jnrA             = jjnr[jidx];
 163             jnrB             = jjnr[jidx+1];
 164             j_coord_offsetA  = DIM*jnrA;
 165             j_coord_offsetB  = DIM*jnrB;
 166
 167             /* load j atom coordinates */
 168             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 169                                               &jx0,&jy0,&jz0);
 170
 171             /* Calculate displacement vector */
 172             dx00             = _mm_sub_pd(ix0,jx0);
 173             dy00             = _mm_sub_pd(iy0,jy0);
 174             dz00             = _mm_sub_pd(iz0,jz0);
 175
 176             /* Calculate squared distance and things based on it */
 177             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 178
 179             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 180
 181             /* Load parameters for j particles */
 182             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 183             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 184             vdwjidx0A        = 2*vdwtype[jnrA+0];
 185             vdwjidx0B        = 2*vdwtype[jnrB+0];
 186
 187             /**************************
 188              * CALCULATE INTERACTIONS *
 189              **************************/
 190
 191             r00              = _mm_mul_pd(rsq00,rinv00);
 192
 193             /* Compute parameters for interactions between i and j atoms */
 194             qq00             = _mm_mul_pd(iq0,jq0);
 195             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 196                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 197
 198             /* Calculate table index by multiplying r with table scale and truncate to integer */
 199             rt               = _mm_mul_pd(r00,vftabscale);
 200             vfitab           = _mm_cvttpd_epi32(rt);
 201 #ifdef __XOP__
 202             vfeps            = _mm_frcz_pd(rt);
 203 #else
 204             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 205 #endif
 206             twovfeps         = _mm_add_pd(vfeps,vfeps);
 207             vfitab           = _mm_slli_epi32(vfitab,3);
 208
 209             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 210             isaprod          = _mm_mul_pd(isai0,isaj0);
 211             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 212             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 213
 214             /* Calculate generalized born table index - this is a separate table from the normal one,
 215              * but we use the same procedure by multiplying r with scale and truncating to integer.
 216              */
 217             rt               = _mm_mul_pd(r00,gbscale);
 218             gbitab           = _mm_cvttpd_epi32(rt);
 219 #ifdef __XOP__
 220             gbeps            = _mm_frcz_pd(rt);
 221 #else
 222             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 223 #endif
 224             gbitab           = _mm_slli_epi32(gbitab,2);
 225
 226             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 227             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 228             GMX_MM_TRANSPOSE2_PD(Y,F);
 229             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 230             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 231             GMX_MM_TRANSPOSE2_PD(G,H);
 232             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 233             VV               = _mm_macc_pd(gbeps,Fp,Y);
 234             vgb              = _mm_mul_pd(gbqqfactor,VV);
 235
 236             twogbeps         = _mm_add_pd(gbeps,gbeps);
 237             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 238             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 239             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 240             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 241             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 242             velec            = _mm_mul_pd(qq00,rinv00);
 243             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 244
 245             /* CUBIC SPLINE TABLE DISPERSION */
 246             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 247             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 248             GMX_MM_TRANSPOSE2_PD(Y,F);
 249             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 250             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 251             GMX_MM_TRANSPOSE2_PD(G,H);
 252             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 253             VV               = _mm_macc_pd(vfeps,Fp,Y);
 254             vvdw6            = _mm_mul_pd(c6_00,VV);
 255             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 256             fvdw6            = _mm_mul_pd(c6_00,FF);
 257
 258             /* CUBIC SPLINE TABLE REPULSION */
 259             vfitab           = _mm_add_epi32(vfitab,ifour);
 260             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 261             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 262             GMX_MM_TRANSPOSE2_PD(Y,F);
 263             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 264             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 265             GMX_MM_TRANSPOSE2_PD(G,H);
 266             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 267             VV               = _mm_macc_pd(vfeps,Fp,Y);
 268             vvdw12           = _mm_mul_pd(c12_00,VV);
 269             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 270             fvdw12           = _mm_mul_pd(c12_00,FF);
 271             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 272             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 273
 274             /* Update potential sum for this i atom from the interaction with this j atom. */
 275             velecsum         = _mm_add_pd(velecsum,velec);
 276             vgbsum           = _mm_add_pd(vgbsum,vgb);
 277             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 278
 279             fscal            = _mm_add_pd(felec,fvdw);
 280
 281             /* Update vectorial force */
 282             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 283             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 284             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 285
 286             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 287                                                    _mm_mul_pd(dx00,fscal),
 288                                                    _mm_mul_pd(dy00,fscal),
 289                                                    _mm_mul_pd(dz00,fscal));
 290
 291             /* Inner loop uses 95 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
 309             /* Calculate squared distance and things based on it */
 310             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 311
 312             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 313
 314             /* Load parameters for j particles */
 315             jq0              = _mm_load_sd(charge+jnrA+0);
 316             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 317             vdwjidx0A        = 2*vdwtype[jnrA+0];
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             r00              = _mm_mul_pd(rsq00,rinv00);
 324
 325             /* Compute parameters for interactions between i and j atoms */
 326             qq00             = _mm_mul_pd(iq0,jq0);
 327             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 328
 329             /* Calculate table index by multiplying r with table scale and truncate to integer */
 330             rt               = _mm_mul_pd(r00,vftabscale);
 331             vfitab           = _mm_cvttpd_epi32(rt);
 332 #ifdef __XOP__
 333             vfeps            = _mm_frcz_pd(rt);
 334 #else
 335             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 336 #endif
 337             twovfeps         = _mm_add_pd(vfeps,vfeps);
 338             vfitab           = _mm_slli_epi32(vfitab,3);
 339
 340             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 341             isaprod          = _mm_mul_pd(isai0,isaj0);
 342             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 343             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 344
 345             /* Calculate generalized born table index - this is a separate table from the normal one,
 346              * but we use the same procedure by multiplying r with scale and truncating to integer.
 347              */
 348             rt               = _mm_mul_pd(r00,gbscale);
 349             gbitab           = _mm_cvttpd_epi32(rt);
 350 #ifdef __XOP__
 351             gbeps            = _mm_frcz_pd(rt);
 352 #else
 353             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 354 #endif
 355             gbitab           = _mm_slli_epi32(gbitab,2);
 356
 357             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 358             F                = _mm_setzero_pd();
 359             GMX_MM_TRANSPOSE2_PD(Y,F);
 360             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 361             H                = _mm_setzero_pd();
 362             GMX_MM_TRANSPOSE2_PD(G,H);
 363             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 364             VV               = _mm_macc_pd(gbeps,Fp,Y);
 365             vgb              = _mm_mul_pd(gbqqfactor,VV);
 366
 367             twogbeps         = _mm_add_pd(gbeps,gbeps);
 368             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 369             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 370             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 371             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 372             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 373             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 374             velec            = _mm_mul_pd(qq00,rinv00);
 375             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 376
 377             /* CUBIC SPLINE TABLE DISPERSION */
 378             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 379             F                = _mm_setzero_pd();
 380             GMX_MM_TRANSPOSE2_PD(Y,F);
 381             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 382             H                = _mm_setzero_pd();
 383             GMX_MM_TRANSPOSE2_PD(G,H);
 384             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 385             VV               = _mm_macc_pd(vfeps,Fp,Y);
 386             vvdw6            = _mm_mul_pd(c6_00,VV);
 387             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 388             fvdw6            = _mm_mul_pd(c6_00,FF);
 389
 390             /* CUBIC SPLINE TABLE REPULSION */
 391             vfitab           = _mm_add_epi32(vfitab,ifour);
 392             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 393             F                = _mm_setzero_pd();
 394             GMX_MM_TRANSPOSE2_PD(Y,F);
 395             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 396             H                = _mm_setzero_pd();
 397             GMX_MM_TRANSPOSE2_PD(G,H);
 398             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 399             VV               = _mm_macc_pd(vfeps,Fp,Y);
 400             vvdw12           = _mm_mul_pd(c12_00,VV);
 401             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 402             fvdw12           = _mm_mul_pd(c12_00,FF);
 403             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 404             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 408             velecsum         = _mm_add_pd(velecsum,velec);
 409             vgb              = _mm_unpacklo_pd(vgb,_mm_setzero_pd());
 410             vgbsum           = _mm_add_pd(vgbsum,vgb);
 411             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 412             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 413
 414             fscal            = _mm_add_pd(felec,fvdw);
 415
 416             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 417
 418             /* Update vectorial force */
 419             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 420             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 421             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 422
 423             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 424                                                    _mm_mul_pd(dx00,fscal),
 425                                                    _mm_mul_pd(dy00,fscal),
 426                                                    _mm_mul_pd(dz00,fscal));
 427
 428             /* Inner loop uses 95 flops */
 429         }
 430
 431         /* End of innermost loop */
 432
 433         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 434                                               f+i_coord_offset,fshift+i_shift_offset);
 435
 436         ggid                        = gid[iidx];
 437         /* Update potential energies */
 438         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 439         gmx_mm_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
 440         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 441         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 442         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 443
 444         /* Increment number of inner iterations */
 445         inneriter                  += j_index_end - j_index_start;
 446
 447         /* Outer loop uses 10 flops */
 448     }
 449
 450     /* Increment number of outer iterations */
 451     outeriter        += nri;
 452
 453     /* Update outer/inner flops */
 454
 455     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*95);
 456 }
 457 /*
 458  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_double
 459  * Electrostatics interaction: GeneralizedBorn
 460  * VdW interaction:            CubicSplineTable
 461  * Geometry:                   Particle-Particle
 462  * Calculate force/pot:        Force
 463  */
 464 void
 465 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_double
 466                     (t_nblist * gmx_restrict                nlist,
 467                      rvec * gmx_restrict                    xx,
 468                      rvec * gmx_restrict                    ff,
 469                      t_forcerec * gmx_restrict              fr,
 470                      t_mdatoms * gmx_restrict               mdatoms,
 471                      nb_kernel_data_t * gmx_restrict        kernel_data,
 472                      t_nrnb * gmx_restrict                  nrnb)
 473 {
 474     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 475      * just 0 for non-waters.
 476      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 477      * jnr indices corresponding to data put in the four positions in the SIMD register.
 478      */
 479     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 480     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 481     int              jnrA,jnrB;
 482     int              j_coord_offsetA,j_coord_offsetB;
 483     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 484     real             rcutoff_scalar;
 485     real             *shiftvec,*fshift,*x,*f;
 486     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 487     int              vdwioffset0;
 488     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 489     int              vdwjidx0A,vdwjidx0B;
 490     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 491     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 492     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 493     real             *charge;
 494     __m128i          gbitab;
 495     __m128d          vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
 496     __m128d          minushalf = _mm_set1_pd(-0.5);
 497     real             *invsqrta,*dvda,*gbtab;
 498     int              nvdwtype;
 499     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 500     int              *vdwtype;
 501     real             *vdwparam;
 502     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 503     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 504     __m128i          vfitab;
 505     __m128i          ifour       = _mm_set1_epi32(4);
 506     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 507     real             *vftab;
 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     nvdwtype         = fr->ntype;
 526     vdwparam         = fr->nbfp;
 527     vdwtype          = mdatoms->typeA;
 528
 529     vftab            = kernel_data->table_vdw->data;
 530     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 531
 532     invsqrta         = fr->invsqrta;
 533     dvda             = fr->dvda;
 534     gbtabscale       = _mm_set1_pd(fr->gbtab.scale);
 535     gbtab            = fr->gbtab.data;
 536     gbinvepsdiff     = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 537
 538     /* Avoid stupid compiler warnings */
 539     jnrA = jnrB = 0;
 540     j_coord_offsetA = 0;
 541     j_coord_offsetB = 0;
 542
 543     outeriter        = 0;
 544     inneriter        = 0;
 545
 546     /* Start outer loop over neighborlists */
 547     for(iidx=0; iidx<nri; iidx++)
 548     {
 549         /* Load shift vector for this list */
 550         i_shift_offset   = DIM*shiftidx[iidx];
 551
 552         /* Load limits for loop over neighbors */
 553         j_index_start    = jindex[iidx];
 554         j_index_end      = jindex[iidx+1];
 555
 556         /* Get outer coordinate index */
 557         inr              = iinr[iidx];
 558         i_coord_offset   = DIM*inr;
 559
 560         /* Load i particle coords and add shift vector */
 561         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 562
 563         fix0             = _mm_setzero_pd();
 564         fiy0             = _mm_setzero_pd();
 565         fiz0             = _mm_setzero_pd();
 566
 567         /* Load parameters for i particles */
 568         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 569         isai0            = _mm_load1_pd(invsqrta+inr+0);
 570         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 571
 572         dvdasum          = _mm_setzero_pd();
 573
 574         /* Start inner kernel loop */
 575         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 576         {
 577
 578             /* Get j neighbor index, and coordinate index */
 579             jnrA             = jjnr[jidx];
 580             jnrB             = jjnr[jidx+1];
 581             j_coord_offsetA  = DIM*jnrA;
 582             j_coord_offsetB  = DIM*jnrB;
 583
 584             /* load j atom coordinates */
 585             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 586                                               &jx0,&jy0,&jz0);
 587
 588             /* Calculate displacement vector */
 589             dx00             = _mm_sub_pd(ix0,jx0);
 590             dy00             = _mm_sub_pd(iy0,jy0);
 591             dz00             = _mm_sub_pd(iz0,jz0);
 592
 593             /* Calculate squared distance and things based on it */
 594             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 595
 596             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 597
 598             /* Load parameters for j particles */
 599             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 600             isaj0            = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
 601             vdwjidx0A        = 2*vdwtype[jnrA+0];
 602             vdwjidx0B        = 2*vdwtype[jnrB+0];
 603
 604             /**************************
 605              * CALCULATE INTERACTIONS *
 606              **************************/
 607
 608             r00              = _mm_mul_pd(rsq00,rinv00);
 609
 610             /* Compute parameters for interactions between i and j atoms */
 611             qq00             = _mm_mul_pd(iq0,jq0);
 612             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 613                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 614
 615             /* Calculate table index by multiplying r with table scale and truncate to integer */
 616             rt               = _mm_mul_pd(r00,vftabscale);
 617             vfitab           = _mm_cvttpd_epi32(rt);
 618 #ifdef __XOP__
 619             vfeps            = _mm_frcz_pd(rt);
 620 #else
 621             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 622 #endif
 623             twovfeps         = _mm_add_pd(vfeps,vfeps);
 624             vfitab           = _mm_slli_epi32(vfitab,3);
 625
 626             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 627             isaprod          = _mm_mul_pd(isai0,isaj0);
 628             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 629             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 630
 631             /* Calculate generalized born table index - this is a separate table from the normal one,
 632              * but we use the same procedure by multiplying r with scale and truncating to integer.
 633              */
 634             rt               = _mm_mul_pd(r00,gbscale);
 635             gbitab           = _mm_cvttpd_epi32(rt);
 636 #ifdef __XOP__
 637             gbeps            = _mm_frcz_pd(rt);
 638 #else
 639             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 640 #endif
 641             gbitab           = _mm_slli_epi32(gbitab,2);
 642
 643             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 644             F                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
 645             GMX_MM_TRANSPOSE2_PD(Y,F);
 646             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 647             H                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
 648             GMX_MM_TRANSPOSE2_PD(G,H);
 649             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 650             VV               = _mm_macc_pd(gbeps,Fp,Y);
 651             vgb              = _mm_mul_pd(gbqqfactor,VV);
 652
 653             twogbeps         = _mm_add_pd(gbeps,gbeps);
 654             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 655             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 656             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 657             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 658             gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 659             velec            = _mm_mul_pd(qq00,rinv00);
 660             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 661
 662             /* CUBIC SPLINE TABLE DISPERSION */
 663             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 664             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 665             GMX_MM_TRANSPOSE2_PD(Y,F);
 666             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 667             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 668             GMX_MM_TRANSPOSE2_PD(G,H);
 669             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 670             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 671             fvdw6            = _mm_mul_pd(c6_00,FF);
 672
 673             /* CUBIC SPLINE TABLE REPULSION */
 674             vfitab           = _mm_add_epi32(vfitab,ifour);
 675             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 676             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 677             GMX_MM_TRANSPOSE2_PD(Y,F);
 678             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 679             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 680             GMX_MM_TRANSPOSE2_PD(G,H);
 681             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 682             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 683             fvdw12           = _mm_mul_pd(c12_00,FF);
 684             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 685
 686             fscal            = _mm_add_pd(felec,fvdw);
 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_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 694                                                    _mm_mul_pd(dx00,fscal),
 695                                                    _mm_mul_pd(dy00,fscal),
 696                                                    _mm_mul_pd(dz00,fscal));
 697
 698             /* Inner loop uses 85 flops */
 699         }
 700
 701         if(jidx<j_index_end)
 702         {
 703
 704             jnrA             = jjnr[jidx];
 705             j_coord_offsetA  = DIM*jnrA;
 706
 707             /* load j atom coordinates */
 708             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 709                                               &jx0,&jy0,&jz0);
 710
 711             /* Calculate displacement vector */
 712             dx00             = _mm_sub_pd(ix0,jx0);
 713             dy00             = _mm_sub_pd(iy0,jy0);
 714             dz00             = _mm_sub_pd(iz0,jz0);
 715
 716             /* Calculate squared distance and things based on it */
 717             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 718
 719             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 720
 721             /* Load parameters for j particles */
 722             jq0              = _mm_load_sd(charge+jnrA+0);
 723             isaj0            = _mm_load_sd(invsqrta+jnrA+0);
 724             vdwjidx0A        = 2*vdwtype[jnrA+0];
 725
 726             /**************************
 727              * CALCULATE INTERACTIONS *
 728              **************************/
 729
 730             r00              = _mm_mul_pd(rsq00,rinv00);
 731
 732             /* Compute parameters for interactions between i and j atoms */
 733             qq00             = _mm_mul_pd(iq0,jq0);
 734             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 735
 736             /* Calculate table index by multiplying r with table scale and truncate to integer */
 737             rt               = _mm_mul_pd(r00,vftabscale);
 738             vfitab           = _mm_cvttpd_epi32(rt);
 739 #ifdef __XOP__
 740             vfeps            = _mm_frcz_pd(rt);
 741 #else
 742             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 743 #endif
 744             twovfeps         = _mm_add_pd(vfeps,vfeps);
 745             vfitab           = _mm_slli_epi32(vfitab,3);
 746
 747             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 748             isaprod          = _mm_mul_pd(isai0,isaj0);
 749             gbqqfactor       = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
 750             gbscale          = _mm_mul_pd(isaprod,gbtabscale);
 751
 752             /* Calculate generalized born table index - this is a separate table from the normal one,
 753              * but we use the same procedure by multiplying r with scale and truncating to integer.
 754              */
 755             rt               = _mm_mul_pd(r00,gbscale);
 756             gbitab           = _mm_cvttpd_epi32(rt);
 757 #ifdef __XOP__
 758             gbeps            = _mm_frcz_pd(rt);
 759 #else
 760             gbeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 761 #endif
 762             gbitab           = _mm_slli_epi32(gbitab,2);
 763
 764             Y                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
 765             F                = _mm_setzero_pd();
 766             GMX_MM_TRANSPOSE2_PD(Y,F);
 767             G                = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
 768             H                = _mm_setzero_pd();
 769             GMX_MM_TRANSPOSE2_PD(G,H);
 770             Fp               = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
 771             VV               = _mm_macc_pd(gbeps,Fp,Y);
 772             vgb              = _mm_mul_pd(gbqqfactor,VV);
 773
 774             twogbeps         = _mm_add_pd(gbeps,gbeps);
 775             FF               = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
 776             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
 777             dvdatmp          = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
 778             dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
 779             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
 780             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
 781             velec            = _mm_mul_pd(qq00,rinv00);
 782             felec            = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
 783
 784             /* CUBIC SPLINE TABLE DISPERSION */
 785             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 786             F                = _mm_setzero_pd();
 787             GMX_MM_TRANSPOSE2_PD(Y,F);
 788             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 789             H                = _mm_setzero_pd();
 790             GMX_MM_TRANSPOSE2_PD(G,H);
 791             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 792             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 793             fvdw6            = _mm_mul_pd(c6_00,FF);
 794
 795             /* CUBIC SPLINE TABLE REPULSION */
 796             vfitab           = _mm_add_epi32(vfitab,ifour);
 797             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 798             F                = _mm_setzero_pd();
 799             GMX_MM_TRANSPOSE2_PD(Y,F);
 800             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 801             H                = _mm_setzero_pd();
 802             GMX_MM_TRANSPOSE2_PD(G,H);
 803             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 804             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 805             fvdw12           = _mm_mul_pd(c12_00,FF);
 806             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 807
 808             fscal            = _mm_add_pd(felec,fvdw);
 809
 810             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 811
 812             /* Update vectorial force */
 813             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 814             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 815             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 816
 817             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 818                                                    _mm_mul_pd(dx00,fscal),
 819                                                    _mm_mul_pd(dy00,fscal),
 820                                                    _mm_mul_pd(dz00,fscal));
 821
 822             /* Inner loop uses 85 flops */
 823         }
 824
 825         /* End of innermost loop */
 826
 827         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 828                                               f+i_coord_offset,fshift+i_shift_offset);
 829
 830         dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
 831         gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
 832
 833         /* Increment number of inner iterations */
 834         inneriter                  += j_index_end - j_index_start;
 835
 836         /* Outer loop uses 7 flops */
 837     }
 838
 839     /* Increment number of outer iterations */
 840     outeriter        += nri;
 841
 842     /* Update outer/inner flops */
 843
 844     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*85);
 845 }