src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_single 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_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  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,C,D refer to j loop unrolling done with AVX_128, e.g. for the four 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,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  72     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  73     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  74     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  75     real             *charge;
  76     int              nvdwtype;
  77     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  78     int              *vdwtype;
  79     real             *vdwparam;
  80     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  81     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  82     __m128i          vfitab;
  83     __m128i          ifour       = _mm_set1_epi32(4);
  84     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  85     real             *vftab;
  86     __m128           dummy_mask,cutoff_mask;
  87     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  88     __m128           one     = _mm_set1_ps(1.0);
  89     __m128           two     = _mm_set1_ps(2.0);
  90     x                = xx[0];
  91     f                = ff[0];
  92
  93     nri              = nlist->nri;
  94     iinr             = nlist->iinr;
  95     jindex           = nlist->jindex;
  96     jjnr             = nlist->jjnr;
  97     shiftidx         = nlist->shift;
  98     gid              = nlist->gid;
  99     shiftvec         = fr->shift_vec[0];
 100     fshift           = fr->fshift[0];
 101     facel            = _mm_set1_ps(fr->epsfac);
 102     charge           = mdatoms->chargeA;
 103     krf              = _mm_set1_ps(fr->ic->k_rf);
 104     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 105     crf              = _mm_set1_ps(fr->ic->c_rf);
 106     nvdwtype         = fr->ntype;
 107     vdwparam         = fr->nbfp;
 108     vdwtype          = mdatoms->typeA;
 109
 110     vftab            = kernel_data->table_vdw->data;
 111     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 112
 113     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 114     rcutoff_scalar   = fr->rcoulomb;
 115     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 116     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 117
 118     /* Avoid stupid compiler warnings */
 119     jnrA = jnrB = jnrC = jnrD = 0;
 120     j_coord_offsetA = 0;
 121     j_coord_offsetB = 0;
 122     j_coord_offsetC = 0;
 123     j_coord_offsetD = 0;
 124
 125     outeriter        = 0;
 126     inneriter        = 0;
 127
 128     for(iidx=0;iidx<4*DIM;iidx++)
 129     {
 130         scratch[iidx] = 0.0;
 131     }
 132
 133     /* Start outer loop over neighborlists */
 134     for(iidx=0; iidx<nri; iidx++)
 135     {
 136         /* Load shift vector for this list */
 137         i_shift_offset   = DIM*shiftidx[iidx];
 138
 139         /* Load limits for loop over neighbors */
 140         j_index_start    = jindex[iidx];
 141         j_index_end      = jindex[iidx+1];
 142
 143         /* Get outer coordinate index */
 144         inr              = iinr[iidx];
 145         i_coord_offset   = DIM*inr;
 146
 147         /* Load i particle coords and add shift vector */
 148         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 149
 150         fix0             = _mm_setzero_ps();
 151         fiy0             = _mm_setzero_ps();
 152         fiz0             = _mm_setzero_ps();
 153
 154         /* Load parameters for i particles */
 155         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 156         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 157
 158         /* Reset potential sums */
 159         velecsum         = _mm_setzero_ps();
 160         vvdwsum          = _mm_setzero_ps();
 161
 162         /* Start inner kernel loop */
 163         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 164         {
 165
 166             /* Get j neighbor index, and coordinate index */
 167             jnrA             = jjnr[jidx];
 168             jnrB             = jjnr[jidx+1];
 169             jnrC             = jjnr[jidx+2];
 170             jnrD             = jjnr[jidx+3];
 171             j_coord_offsetA  = DIM*jnrA;
 172             j_coord_offsetB  = DIM*jnrB;
 173             j_coord_offsetC  = DIM*jnrC;
 174             j_coord_offsetD  = DIM*jnrD;
 175
 176             /* load j atom coordinates */
 177             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 178                                               x+j_coord_offsetC,x+j_coord_offsetD,
 179                                               &jx0,&jy0,&jz0);
 180
 181             /* Calculate displacement vector */
 182             dx00             = _mm_sub_ps(ix0,jx0);
 183             dy00             = _mm_sub_ps(iy0,jy0);
 184             dz00             = _mm_sub_ps(iz0,jz0);
 185
 186             /* Calculate squared distance and things based on it */
 187             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 188
 189             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 190
 191             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 192
 193             /* Load parameters for j particles */
 194             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 195                                                               charge+jnrC+0,charge+jnrD+0);
 196             vdwjidx0A        = 2*vdwtype[jnrA+0];
 197             vdwjidx0B        = 2*vdwtype[jnrB+0];
 198             vdwjidx0C        = 2*vdwtype[jnrC+0];
 199             vdwjidx0D        = 2*vdwtype[jnrD+0];
 200
 201             /**************************
 202              * CALCULATE INTERACTIONS *
 203              **************************/
 204
 205             if (gmx_mm_any_lt(rsq00,rcutoff2))
 206             {
 207
 208             r00              = _mm_mul_ps(rsq00,rinv00);
 209
 210             /* Compute parameters for interactions between i and j atoms */
 211             qq00             = _mm_mul_ps(iq0,jq0);
 212             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 213                                          vdwparam+vdwioffset0+vdwjidx0B,
 214                                          vdwparam+vdwioffset0+vdwjidx0C,
 215                                          vdwparam+vdwioffset0+vdwjidx0D,
 216                                          &c6_00,&c12_00);
 217
 218             /* Calculate table index by multiplying r with table scale and truncate to integer */
 219             rt               = _mm_mul_ps(r00,vftabscale);
 220             vfitab           = _mm_cvttps_epi32(rt);
 221 #ifdef __XOP__
 222             vfeps            = _mm_frcz_ps(rt);
 223 #else
 224             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 225 #endif
 226             twovfeps         = _mm_add_ps(vfeps,vfeps);
 227             vfitab           = _mm_slli_epi32(vfitab,3);
 228
 229             /* REACTION-FIELD ELECTROSTATICS */
 230             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 231             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 232
 233             /* CUBIC SPLINE TABLE DISPERSION */
 234             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 235             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 236             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 237             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 238             _MM_TRANSPOSE4_PS(Y,F,G,H);
 239             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 240             VV               = _mm_macc_ps(vfeps,Fp,Y);
 241             vvdw6            = _mm_mul_ps(c6_00,VV);
 242             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 243             fvdw6            = _mm_mul_ps(c6_00,FF);
 244
 245             /* CUBIC SPLINE TABLE REPULSION */
 246             vfitab           = _mm_add_epi32(vfitab,ifour);
 247             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 248             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 249             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 250             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 251             _MM_TRANSPOSE4_PS(Y,F,G,H);
 252             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 253             VV               = _mm_macc_ps(vfeps,Fp,Y);
 254             vvdw12           = _mm_mul_ps(c12_00,VV);
 255             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 256             fvdw12           = _mm_mul_ps(c12_00,FF);
 257             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 258             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 259
 260             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 261
 262             /* Update potential sum for this i atom from the interaction with this j atom. */
 263             velec            = _mm_and_ps(velec,cutoff_mask);
 264             velecsum         = _mm_add_ps(velecsum,velec);
 265             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 266             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 267
 268             fscal            = _mm_add_ps(felec,fvdw);
 269
 270             fscal            = _mm_and_ps(fscal,cutoff_mask);
 271
 272              /* Update vectorial force */
 273             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 274             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 275             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 276
 277             fjptrA             = f+j_coord_offsetA;
 278             fjptrB             = f+j_coord_offsetB;
 279             fjptrC             = f+j_coord_offsetC;
 280             fjptrD             = f+j_coord_offsetD;
 281             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 282                                                    _mm_mul_ps(dx00,fscal),
 283                                                    _mm_mul_ps(dy00,fscal),
 284                                                    _mm_mul_ps(dz00,fscal));
 285
 286             }
 287
 288             /* Inner loop uses 75 flops */
 289         }
 290
 291         if(jidx<j_index_end)
 292         {
 293
 294             /* Get j neighbor index, and coordinate index */
 295             jnrlistA         = jjnr[jidx];
 296             jnrlistB         = jjnr[jidx+1];
 297             jnrlistC         = jjnr[jidx+2];
 298             jnrlistD         = jjnr[jidx+3];
 299             /* Sign of each element will be negative for non-real atoms.
 300              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 301              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 302              */
 303             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 304             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 305             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 306             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 307             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 308             j_coord_offsetA  = DIM*jnrA;
 309             j_coord_offsetB  = DIM*jnrB;
 310             j_coord_offsetC  = DIM*jnrC;
 311             j_coord_offsetD  = DIM*jnrD;
 312
 313             /* load j atom coordinates */
 314             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 315                                               x+j_coord_offsetC,x+j_coord_offsetD,
 316                                               &jx0,&jy0,&jz0);
 317
 318             /* Calculate displacement vector */
 319             dx00             = _mm_sub_ps(ix0,jx0);
 320             dy00             = _mm_sub_ps(iy0,jy0);
 321             dz00             = _mm_sub_ps(iz0,jz0);
 322
 323             /* Calculate squared distance and things based on it */
 324             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 325
 326             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 327
 328             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 329
 330             /* Load parameters for j particles */
 331             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 332                                                               charge+jnrC+0,charge+jnrD+0);
 333             vdwjidx0A        = 2*vdwtype[jnrA+0];
 334             vdwjidx0B        = 2*vdwtype[jnrB+0];
 335             vdwjidx0C        = 2*vdwtype[jnrC+0];
 336             vdwjidx0D        = 2*vdwtype[jnrD+0];
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             if (gmx_mm_any_lt(rsq00,rcutoff2))
 343             {
 344
 345             r00              = _mm_mul_ps(rsq00,rinv00);
 346             r00              = _mm_andnot_ps(dummy_mask,r00);
 347
 348             /* Compute parameters for interactions between i and j atoms */
 349             qq00             = _mm_mul_ps(iq0,jq0);
 350             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 351                                          vdwparam+vdwioffset0+vdwjidx0B,
 352                                          vdwparam+vdwioffset0+vdwjidx0C,
 353                                          vdwparam+vdwioffset0+vdwjidx0D,
 354                                          &c6_00,&c12_00);
 355
 356             /* Calculate table index by multiplying r with table scale and truncate to integer */
 357             rt               = _mm_mul_ps(r00,vftabscale);
 358             vfitab           = _mm_cvttps_epi32(rt);
 359 #ifdef __XOP__
 360             vfeps            = _mm_frcz_ps(rt);
 361 #else
 362             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 363 #endif
 364             twovfeps         = _mm_add_ps(vfeps,vfeps);
 365             vfitab           = _mm_slli_epi32(vfitab,3);
 366
 367             /* REACTION-FIELD ELECTROSTATICS */
 368             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 369             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 370
 371             /* CUBIC SPLINE TABLE DISPERSION */
 372             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 373             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 374             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 375             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 376             _MM_TRANSPOSE4_PS(Y,F,G,H);
 377             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 378             VV               = _mm_macc_ps(vfeps,Fp,Y);
 379             vvdw6            = _mm_mul_ps(c6_00,VV);
 380             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 381             fvdw6            = _mm_mul_ps(c6_00,FF);
 382
 383             /* CUBIC SPLINE TABLE REPULSION */
 384             vfitab           = _mm_add_epi32(vfitab,ifour);
 385             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 386             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 387             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 388             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 389             _MM_TRANSPOSE4_PS(Y,F,G,H);
 390             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 391             VV               = _mm_macc_ps(vfeps,Fp,Y);
 392             vvdw12           = _mm_mul_ps(c12_00,VV);
 393             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 394             fvdw12           = _mm_mul_ps(c12_00,FF);
 395             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 396             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 397
 398             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 399
 400             /* Update potential sum for this i atom from the interaction with this j atom. */
 401             velec            = _mm_and_ps(velec,cutoff_mask);
 402             velec            = _mm_andnot_ps(dummy_mask,velec);
 403             velecsum         = _mm_add_ps(velecsum,velec);
 404             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 405             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 406             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 407
 408             fscal            = _mm_add_ps(felec,fvdw);
 409
 410             fscal            = _mm_and_ps(fscal,cutoff_mask);
 411
 412             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 413
 414              /* Update vectorial force */
 415             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 416             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 417             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 418
 419             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 420             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 421             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 422             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 423             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 424                                                    _mm_mul_ps(dx00,fscal),
 425                                                    _mm_mul_ps(dy00,fscal),
 426                                                    _mm_mul_ps(dz00,fscal));
 427
 428             }
 429
 430             /* Inner loop uses 76 flops */
 431         }
 432
 433         /* End of innermost loop */
 434
 435         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 436                                               f+i_coord_offset,fshift+i_shift_offset);
 437
 438         ggid                        = gid[iidx];
 439         /* Update potential energies */
 440         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 441         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 442
 443         /* Increment number of inner iterations */
 444         inneriter                  += j_index_end - j_index_start;
 445
 446         /* Outer loop uses 9 flops */
 447     }
 448
 449     /* Increment number of outer iterations */
 450     outeriter        += nri;
 451
 452     /* Update outer/inner flops */
 453
 454     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*76);
 455 }
 456 /*
 457  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 458  * Electrostatics interaction: ReactionField
 459  * VdW interaction:            CubicSplineTable
 460  * Geometry:                   Particle-Particle
 461  * Calculate force/pot:        Force
 462  */
 463 void
 464 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 465                     (t_nblist * gmx_restrict                nlist,
 466                      rvec * gmx_restrict                    xx,
 467                      rvec * gmx_restrict                    ff,
 468                      t_forcerec * gmx_restrict              fr,
 469                      t_mdatoms * gmx_restrict               mdatoms,
 470                      nb_kernel_data_t * gmx_restrict        kernel_data,
 471                      t_nrnb * gmx_restrict                  nrnb)
 472 {
 473     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 474      * just 0 for non-waters.
 475      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 476      * jnr indices corresponding to data put in the four positions in the SIMD register.
 477      */
 478     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 479     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 480     int              jnrA,jnrB,jnrC,jnrD;
 481     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 482     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 483     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 484     real             rcutoff_scalar;
 485     real             *shiftvec,*fshift,*x,*f;
 486     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 487     real             scratch[4*DIM];
 488     __m128           fscal,rcutoff,rcutoff2,jidxall;
 489     int              vdwioffset0;
 490     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 491     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 492     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 493     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 494     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 495     real             *charge;
 496     int              nvdwtype;
 497     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 498     int              *vdwtype;
 499     real             *vdwparam;
 500     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 501     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 502     __m128i          vfitab;
 503     __m128i          ifour       = _mm_set1_epi32(4);
 504     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 505     real             *vftab;
 506     __m128           dummy_mask,cutoff_mask;
 507     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 508     __m128           one     = _mm_set1_ps(1.0);
 509     __m128           two     = _mm_set1_ps(2.0);
 510     x                = xx[0];
 511     f                = ff[0];
 512
 513     nri              = nlist->nri;
 514     iinr             = nlist->iinr;
 515     jindex           = nlist->jindex;
 516     jjnr             = nlist->jjnr;
 517     shiftidx         = nlist->shift;
 518     gid              = nlist->gid;
 519     shiftvec         = fr->shift_vec[0];
 520     fshift           = fr->fshift[0];
 521     facel            = _mm_set1_ps(fr->epsfac);
 522     charge           = mdatoms->chargeA;
 523     krf              = _mm_set1_ps(fr->ic->k_rf);
 524     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 525     crf              = _mm_set1_ps(fr->ic->c_rf);
 526     nvdwtype         = fr->ntype;
 527     vdwparam         = fr->nbfp;
 528     vdwtype          = mdatoms->typeA;
 529
 530     vftab            = kernel_data->table_vdw->data;
 531     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 532
 533     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 534     rcutoff_scalar   = fr->rcoulomb;
 535     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 536     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 537
 538     /* Avoid stupid compiler warnings */
 539     jnrA = jnrB = jnrC = jnrD = 0;
 540     j_coord_offsetA = 0;
 541     j_coord_offsetB = 0;
 542     j_coord_offsetC = 0;
 543     j_coord_offsetD = 0;
 544
 545     outeriter        = 0;
 546     inneriter        = 0;
 547
 548     for(iidx=0;iidx<4*DIM;iidx++)
 549     {
 550         scratch[iidx] = 0.0;
 551     }
 552
 553     /* Start outer loop over neighborlists */
 554     for(iidx=0; iidx<nri; iidx++)
 555     {
 556         /* Load shift vector for this list */
 557         i_shift_offset   = DIM*shiftidx[iidx];
 558
 559         /* Load limits for loop over neighbors */
 560         j_index_start    = jindex[iidx];
 561         j_index_end      = jindex[iidx+1];
 562
 563         /* Get outer coordinate index */
 564         inr              = iinr[iidx];
 565         i_coord_offset   = DIM*inr;
 566
 567         /* Load i particle coords and add shift vector */
 568         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 569
 570         fix0             = _mm_setzero_ps();
 571         fiy0             = _mm_setzero_ps();
 572         fiz0             = _mm_setzero_ps();
 573
 574         /* Load parameters for i particles */
 575         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 576         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 577
 578         /* Start inner kernel loop */
 579         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 580         {
 581
 582             /* Get j neighbor index, and coordinate index */
 583             jnrA             = jjnr[jidx];
 584             jnrB             = jjnr[jidx+1];
 585             jnrC             = jjnr[jidx+2];
 586             jnrD             = jjnr[jidx+3];
 587             j_coord_offsetA  = DIM*jnrA;
 588             j_coord_offsetB  = DIM*jnrB;
 589             j_coord_offsetC  = DIM*jnrC;
 590             j_coord_offsetD  = DIM*jnrD;
 591
 592             /* load j atom coordinates */
 593             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 594                                               x+j_coord_offsetC,x+j_coord_offsetD,
 595                                               &jx0,&jy0,&jz0);
 596
 597             /* Calculate displacement vector */
 598             dx00             = _mm_sub_ps(ix0,jx0);
 599             dy00             = _mm_sub_ps(iy0,jy0);
 600             dz00             = _mm_sub_ps(iz0,jz0);
 601
 602             /* Calculate squared distance and things based on it */
 603             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 604
 605             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 606
 607             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 608
 609             /* Load parameters for j particles */
 610             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 611                                                               charge+jnrC+0,charge+jnrD+0);
 612             vdwjidx0A        = 2*vdwtype[jnrA+0];
 613             vdwjidx0B        = 2*vdwtype[jnrB+0];
 614             vdwjidx0C        = 2*vdwtype[jnrC+0];
 615             vdwjidx0D        = 2*vdwtype[jnrD+0];
 616
 617             /**************************
 618              * CALCULATE INTERACTIONS *
 619              **************************/
 620
 621             if (gmx_mm_any_lt(rsq00,rcutoff2))
 622             {
 623
 624             r00              = _mm_mul_ps(rsq00,rinv00);
 625
 626             /* Compute parameters for interactions between i and j atoms */
 627             qq00             = _mm_mul_ps(iq0,jq0);
 628             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 629                                          vdwparam+vdwioffset0+vdwjidx0B,
 630                                          vdwparam+vdwioffset0+vdwjidx0C,
 631                                          vdwparam+vdwioffset0+vdwjidx0D,
 632                                          &c6_00,&c12_00);
 633
 634             /* Calculate table index by multiplying r with table scale and truncate to integer */
 635             rt               = _mm_mul_ps(r00,vftabscale);
 636             vfitab           = _mm_cvttps_epi32(rt);
 637 #ifdef __XOP__
 638             vfeps            = _mm_frcz_ps(rt);
 639 #else
 640             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 641 #endif
 642             twovfeps         = _mm_add_ps(vfeps,vfeps);
 643             vfitab           = _mm_slli_epi32(vfitab,3);
 644
 645             /* REACTION-FIELD ELECTROSTATICS */
 646             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 647
 648             /* CUBIC SPLINE TABLE DISPERSION */
 649             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 650             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 651             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 652             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 653             _MM_TRANSPOSE4_PS(Y,F,G,H);
 654             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 655             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 656             fvdw6            = _mm_mul_ps(c6_00,FF);
 657
 658             /* CUBIC SPLINE TABLE REPULSION */
 659             vfitab           = _mm_add_epi32(vfitab,ifour);
 660             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 661             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 662             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 663             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 664             _MM_TRANSPOSE4_PS(Y,F,G,H);
 665             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 666             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 667             fvdw12           = _mm_mul_ps(c12_00,FF);
 668             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 669
 670             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 671
 672             fscal            = _mm_add_ps(felec,fvdw);
 673
 674             fscal            = _mm_and_ps(fscal,cutoff_mask);
 675
 676              /* Update vectorial force */
 677             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 678             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 679             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 680
 681             fjptrA             = f+j_coord_offsetA;
 682             fjptrB             = f+j_coord_offsetB;
 683             fjptrC             = f+j_coord_offsetC;
 684             fjptrD             = f+j_coord_offsetD;
 685             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 686                                                    _mm_mul_ps(dx00,fscal),
 687                                                    _mm_mul_ps(dy00,fscal),
 688                                                    _mm_mul_ps(dz00,fscal));
 689
 690             }
 691
 692             /* Inner loop uses 60 flops */
 693         }
 694
 695         if(jidx<j_index_end)
 696         {
 697
 698             /* Get j neighbor index, and coordinate index */
 699             jnrlistA         = jjnr[jidx];
 700             jnrlistB         = jjnr[jidx+1];
 701             jnrlistC         = jjnr[jidx+2];
 702             jnrlistD         = jjnr[jidx+3];
 703             /* Sign of each element will be negative for non-real atoms.
 704              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 705              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 706              */
 707             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 708             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 709             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 710             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 711             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 712             j_coord_offsetA  = DIM*jnrA;
 713             j_coord_offsetB  = DIM*jnrB;
 714             j_coord_offsetC  = DIM*jnrC;
 715             j_coord_offsetD  = DIM*jnrD;
 716
 717             /* load j atom coordinates */
 718             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 719                                               x+j_coord_offsetC,x+j_coord_offsetD,
 720                                               &jx0,&jy0,&jz0);
 721
 722             /* Calculate displacement vector */
 723             dx00             = _mm_sub_ps(ix0,jx0);
 724             dy00             = _mm_sub_ps(iy0,jy0);
 725             dz00             = _mm_sub_ps(iz0,jz0);
 726
 727             /* Calculate squared distance and things based on it */
 728             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 729
 730             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 731
 732             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 733
 734             /* Load parameters for j particles */
 735             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 736                                                               charge+jnrC+0,charge+jnrD+0);
 737             vdwjidx0A        = 2*vdwtype[jnrA+0];
 738             vdwjidx0B        = 2*vdwtype[jnrB+0];
 739             vdwjidx0C        = 2*vdwtype[jnrC+0];
 740             vdwjidx0D        = 2*vdwtype[jnrD+0];
 741
 742             /**************************
 743              * CALCULATE INTERACTIONS *
 744              **************************/
 745
 746             if (gmx_mm_any_lt(rsq00,rcutoff2))
 747             {
 748
 749             r00              = _mm_mul_ps(rsq00,rinv00);
 750             r00              = _mm_andnot_ps(dummy_mask,r00);
 751
 752             /* Compute parameters for interactions between i and j atoms */
 753             qq00             = _mm_mul_ps(iq0,jq0);
 754             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 755                                          vdwparam+vdwioffset0+vdwjidx0B,
 756                                          vdwparam+vdwioffset0+vdwjidx0C,
 757                                          vdwparam+vdwioffset0+vdwjidx0D,
 758                                          &c6_00,&c12_00);
 759
 760             /* Calculate table index by multiplying r with table scale and truncate to integer */
 761             rt               = _mm_mul_ps(r00,vftabscale);
 762             vfitab           = _mm_cvttps_epi32(rt);
 763 #ifdef __XOP__
 764             vfeps            = _mm_frcz_ps(rt);
 765 #else
 766             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 767 #endif
 768             twovfeps         = _mm_add_ps(vfeps,vfeps);
 769             vfitab           = _mm_slli_epi32(vfitab,3);
 770
 771             /* REACTION-FIELD ELECTROSTATICS */
 772             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 773
 774             /* CUBIC SPLINE TABLE DISPERSION */
 775             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 776             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 777             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 778             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 779             _MM_TRANSPOSE4_PS(Y,F,G,H);
 780             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 781             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 782             fvdw6            = _mm_mul_ps(c6_00,FF);
 783
 784             /* CUBIC SPLINE TABLE REPULSION */
 785             vfitab           = _mm_add_epi32(vfitab,ifour);
 786             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 787             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 788             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 789             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 790             _MM_TRANSPOSE4_PS(Y,F,G,H);
 791             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 792             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 793             fvdw12           = _mm_mul_ps(c12_00,FF);
 794             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 795
 796             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 797
 798             fscal            = _mm_add_ps(felec,fvdw);
 799
 800             fscal            = _mm_and_ps(fscal,cutoff_mask);
 801
 802             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 803
 804              /* Update vectorial force */
 805             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 806             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 807             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 808
 809             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 810             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 811             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 812             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 813             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 814                                                    _mm_mul_ps(dx00,fscal),
 815                                                    _mm_mul_ps(dy00,fscal),
 816                                                    _mm_mul_ps(dz00,fscal));
 817
 818             }
 819
 820             /* Inner loop uses 61 flops */
 821         }
 822
 823         /* End of innermost loop */
 824
 825         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 826                                               f+i_coord_offset,fshift+i_shift_offset);
 827
 828         /* Increment number of inner iterations */
 829         inneriter                  += j_index_end - j_index_start;
 830
 831         /* Outer loop uses 7 flops */
 832     }
 833
 834     /* Increment number of outer iterations */
 835     outeriter        += nri;
 836
 837     /* Update outer/inner flops */
 838
 839     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*61);
 840 }