src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCoul_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_ElecCoul_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_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     nvdwtype         = fr->ntype;
 104     vdwparam         = fr->nbfp;
 105     vdwtype          = mdatoms->typeA;
 106
 107     vftab            = kernel_data->table_vdw->data;
 108     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 109
 110     /* Avoid stupid compiler warnings */
 111     jnrA = jnrB = jnrC = jnrD = 0;
 112     j_coord_offsetA = 0;
 113     j_coord_offsetB = 0;
 114     j_coord_offsetC = 0;
 115     j_coord_offsetD = 0;
 116
 117     outeriter        = 0;
 118     inneriter        = 0;
 119
 120     for(iidx=0;iidx<4*DIM;iidx++)
 121     {
 122         scratch[iidx] = 0.0;
 123     }
 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_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 141
 142         fix0             = _mm_setzero_ps();
 143         fiy0             = _mm_setzero_ps();
 144         fiz0             = _mm_setzero_ps();
 145
 146         /* Load parameters for i particles */
 147         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 148         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 149
 150         /* Reset potential sums */
 151         velecsum         = _mm_setzero_ps();
 152         vvdwsum          = _mm_setzero_ps();
 153
 154         /* Start inner kernel loop */
 155         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 156         {
 157
 158             /* Get j neighbor index, and coordinate index */
 159             jnrA             = jjnr[jidx];
 160             jnrB             = jjnr[jidx+1];
 161             jnrC             = jjnr[jidx+2];
 162             jnrD             = jjnr[jidx+3];
 163             j_coord_offsetA  = DIM*jnrA;
 164             j_coord_offsetB  = DIM*jnrB;
 165             j_coord_offsetC  = DIM*jnrC;
 166             j_coord_offsetD  = DIM*jnrD;
 167
 168             /* load j atom coordinates */
 169             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 170                                               x+j_coord_offsetC,x+j_coord_offsetD,
 171                                               &jx0,&jy0,&jz0);
 172
 173             /* Calculate displacement vector */
 174             dx00             = _mm_sub_ps(ix0,jx0);
 175             dy00             = _mm_sub_ps(iy0,jy0);
 176             dz00             = _mm_sub_ps(iz0,jz0);
 177
 178             /* Calculate squared distance and things based on it */
 179             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 180
 181             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 182
 183             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 184
 185             /* Load parameters for j particles */
 186             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 187                                                               charge+jnrC+0,charge+jnrD+0);
 188             vdwjidx0A        = 2*vdwtype[jnrA+0];
 189             vdwjidx0B        = 2*vdwtype[jnrB+0];
 190             vdwjidx0C        = 2*vdwtype[jnrC+0];
 191             vdwjidx0D        = 2*vdwtype[jnrD+0];
 192
 193             /**************************
 194              * CALCULATE INTERACTIONS *
 195              **************************/
 196
 197             r00              = _mm_mul_ps(rsq00,rinv00);
 198
 199             /* Compute parameters for interactions between i and j atoms */
 200             qq00             = _mm_mul_ps(iq0,jq0);
 201             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 202                                          vdwparam+vdwioffset0+vdwjidx0B,
 203                                          vdwparam+vdwioffset0+vdwjidx0C,
 204                                          vdwparam+vdwioffset0+vdwjidx0D,
 205                                          &c6_00,&c12_00);
 206
 207             /* Calculate table index by multiplying r with table scale and truncate to integer */
 208             rt               = _mm_mul_ps(r00,vftabscale);
 209             vfitab           = _mm_cvttps_epi32(rt);
 210 #ifdef __XOP__
 211             vfeps            = _mm_frcz_ps(rt);
 212 #else
 213             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 214 #endif
 215             twovfeps         = _mm_add_ps(vfeps,vfeps);
 216             vfitab           = _mm_slli_epi32(vfitab,3);
 217
 218             /* COULOMB ELECTROSTATICS */
 219             velec            = _mm_mul_ps(qq00,rinv00);
 220             felec            = _mm_mul_ps(velec,rinvsq00);
 221
 222             /* CUBIC SPLINE TABLE DISPERSION */
 223             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 224             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 225             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 226             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 227             _MM_TRANSPOSE4_PS(Y,F,G,H);
 228             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 229             VV               = _mm_macc_ps(vfeps,Fp,Y);
 230             vvdw6            = _mm_mul_ps(c6_00,VV);
 231             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 232             fvdw6            = _mm_mul_ps(c6_00,FF);
 233
 234             /* CUBIC SPLINE TABLE REPULSION */
 235             vfitab           = _mm_add_epi32(vfitab,ifour);
 236             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 237             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 238             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 239             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 240             _MM_TRANSPOSE4_PS(Y,F,G,H);
 241             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 242             VV               = _mm_macc_ps(vfeps,Fp,Y);
 243             vvdw12           = _mm_mul_ps(c12_00,VV);
 244             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 245             fvdw12           = _mm_mul_ps(c12_00,FF);
 246             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 247             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 248
 249             /* Update potential sum for this i atom from the interaction with this j atom. */
 250             velecsum         = _mm_add_ps(velecsum,velec);
 251             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 252
 253             fscal            = _mm_add_ps(felec,fvdw);
 254
 255              /* Update vectorial force */
 256             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 257             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 258             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 259
 260             fjptrA             = f+j_coord_offsetA;
 261             fjptrB             = f+j_coord_offsetB;
 262             fjptrC             = f+j_coord_offsetC;
 263             fjptrD             = f+j_coord_offsetD;
 264             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 265                                                    _mm_mul_ps(dx00,fscal),
 266                                                    _mm_mul_ps(dy00,fscal),
 267                                                    _mm_mul_ps(dz00,fscal));
 268
 269             /* Inner loop uses 66 flops */
 270         }
 271
 272         if(jidx<j_index_end)
 273         {
 274
 275             /* Get j neighbor index, and coordinate index */
 276             jnrlistA         = jjnr[jidx];
 277             jnrlistB         = jjnr[jidx+1];
 278             jnrlistC         = jjnr[jidx+2];
 279             jnrlistD         = jjnr[jidx+3];
 280             /* Sign of each element will be negative for non-real atoms.
 281              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 282              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 283              */
 284             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 285             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 286             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 287             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 288             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 289             j_coord_offsetA  = DIM*jnrA;
 290             j_coord_offsetB  = DIM*jnrB;
 291             j_coord_offsetC  = DIM*jnrC;
 292             j_coord_offsetD  = DIM*jnrD;
 293
 294             /* load j atom coordinates */
 295             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 296                                               x+j_coord_offsetC,x+j_coord_offsetD,
 297                                               &jx0,&jy0,&jz0);
 298
 299             /* Calculate displacement vector */
 300             dx00             = _mm_sub_ps(ix0,jx0);
 301             dy00             = _mm_sub_ps(iy0,jy0);
 302             dz00             = _mm_sub_ps(iz0,jz0);
 303
 304             /* Calculate squared distance and things based on it */
 305             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 306
 307             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 308
 309             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 310
 311             /* Load parameters for j particles */
 312             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 313                                                               charge+jnrC+0,charge+jnrD+0);
 314             vdwjidx0A        = 2*vdwtype[jnrA+0];
 315             vdwjidx0B        = 2*vdwtype[jnrB+0];
 316             vdwjidx0C        = 2*vdwtype[jnrC+0];
 317             vdwjidx0D        = 2*vdwtype[jnrD+0];
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             r00              = _mm_mul_ps(rsq00,rinv00);
 324             r00              = _mm_andnot_ps(dummy_mask,r00);
 325
 326             /* Compute parameters for interactions between i and j atoms */
 327             qq00             = _mm_mul_ps(iq0,jq0);
 328             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 329                                          vdwparam+vdwioffset0+vdwjidx0B,
 330                                          vdwparam+vdwioffset0+vdwjidx0C,
 331                                          vdwparam+vdwioffset0+vdwjidx0D,
 332                                          &c6_00,&c12_00);
 333
 334             /* Calculate table index by multiplying r with table scale and truncate to integer */
 335             rt               = _mm_mul_ps(r00,vftabscale);
 336             vfitab           = _mm_cvttps_epi32(rt);
 337 #ifdef __XOP__
 338             vfeps            = _mm_frcz_ps(rt);
 339 #else
 340             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 341 #endif
 342             twovfeps         = _mm_add_ps(vfeps,vfeps);
 343             vfitab           = _mm_slli_epi32(vfitab,3);
 344
 345             /* COULOMB ELECTROSTATICS */
 346             velec            = _mm_mul_ps(qq00,rinv00);
 347             felec            = _mm_mul_ps(velec,rinvsq00);
 348
 349             /* CUBIC SPLINE TABLE DISPERSION */
 350             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 351             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 352             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 353             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 354             _MM_TRANSPOSE4_PS(Y,F,G,H);
 355             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 356             VV               = _mm_macc_ps(vfeps,Fp,Y);
 357             vvdw6            = _mm_mul_ps(c6_00,VV);
 358             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 359             fvdw6            = _mm_mul_ps(c6_00,FF);
 360
 361             /* CUBIC SPLINE TABLE REPULSION */
 362             vfitab           = _mm_add_epi32(vfitab,ifour);
 363             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 364             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 365             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 366             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 367             _MM_TRANSPOSE4_PS(Y,F,G,H);
 368             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 369             VV               = _mm_macc_ps(vfeps,Fp,Y);
 370             vvdw12           = _mm_mul_ps(c12_00,VV);
 371             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 372             fvdw12           = _mm_mul_ps(c12_00,FF);
 373             vvdw             = _mm_add_ps(vvdw12,vvdw6);
 374             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 375
 376             /* Update potential sum for this i atom from the interaction with this j atom. */
 377             velec            = _mm_andnot_ps(dummy_mask,velec);
 378             velecsum         = _mm_add_ps(velecsum,velec);
 379             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 380             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 381
 382             fscal            = _mm_add_ps(felec,fvdw);
 383
 384             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 385
 386              /* Update vectorial force */
 387             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 388             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 389             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 390
 391             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 392             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 393             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 394             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 395             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 396                                                    _mm_mul_ps(dx00,fscal),
 397                                                    _mm_mul_ps(dy00,fscal),
 398                                                    _mm_mul_ps(dz00,fscal));
 399
 400             /* Inner loop uses 67 flops */
 401         }
 402
 403         /* End of innermost loop */
 404
 405         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 406                                               f+i_coord_offset,fshift+i_shift_offset);
 407
 408         ggid                        = gid[iidx];
 409         /* Update potential energies */
 410         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 411         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 412
 413         /* Increment number of inner iterations */
 414         inneriter                  += j_index_end - j_index_start;
 415
 416         /* Outer loop uses 9 flops */
 417     }
 418
 419     /* Increment number of outer iterations */
 420     outeriter        += nri;
 421
 422     /* Update outer/inner flops */
 423
 424     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*67);
 425 }
 426 /*
 427  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 428  * Electrostatics interaction: Coulomb
 429  * VdW interaction:            CubicSplineTable
 430  * Geometry:                   Particle-Particle
 431  * Calculate force/pot:        Force
 432  */
 433 void
 434 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_128_fma_single
 435                     (t_nblist * gmx_restrict                nlist,
 436                      rvec * gmx_restrict                    xx,
 437                      rvec * gmx_restrict                    ff,
 438                      t_forcerec * gmx_restrict              fr,
 439                      t_mdatoms * gmx_restrict               mdatoms,
 440                      nb_kernel_data_t * gmx_restrict        kernel_data,
 441                      t_nrnb * gmx_restrict                  nrnb)
 442 {
 443     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 444      * just 0 for non-waters.
 445      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 446      * jnr indices corresponding to data put in the four positions in the SIMD register.
 447      */
 448     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 449     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 450     int              jnrA,jnrB,jnrC,jnrD;
 451     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 452     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 453     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 454     real             rcutoff_scalar;
 455     real             *shiftvec,*fshift,*x,*f;
 456     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 457     real             scratch[4*DIM];
 458     __m128           fscal,rcutoff,rcutoff2,jidxall;
 459     int              vdwioffset0;
 460     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 461     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 462     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 463     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 464     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 465     real             *charge;
 466     int              nvdwtype;
 467     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 468     int              *vdwtype;
 469     real             *vdwparam;
 470     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 471     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 472     __m128i          vfitab;
 473     __m128i          ifour       = _mm_set1_epi32(4);
 474     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 475     real             *vftab;
 476     __m128           dummy_mask,cutoff_mask;
 477     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 478     __m128           one     = _mm_set1_ps(1.0);
 479     __m128           two     = _mm_set1_ps(2.0);
 480     x                = xx[0];
 481     f                = ff[0];
 482
 483     nri              = nlist->nri;
 484     iinr             = nlist->iinr;
 485     jindex           = nlist->jindex;
 486     jjnr             = nlist->jjnr;
 487     shiftidx         = nlist->shift;
 488     gid              = nlist->gid;
 489     shiftvec         = fr->shift_vec[0];
 490     fshift           = fr->fshift[0];
 491     facel            = _mm_set1_ps(fr->epsfac);
 492     charge           = mdatoms->chargeA;
 493     nvdwtype         = fr->ntype;
 494     vdwparam         = fr->nbfp;
 495     vdwtype          = mdatoms->typeA;
 496
 497     vftab            = kernel_data->table_vdw->data;
 498     vftabscale       = _mm_set1_ps(kernel_data->table_vdw->scale);
 499
 500     /* Avoid stupid compiler warnings */
 501     jnrA = jnrB = jnrC = jnrD = 0;
 502     j_coord_offsetA = 0;
 503     j_coord_offsetB = 0;
 504     j_coord_offsetC = 0;
 505     j_coord_offsetD = 0;
 506
 507     outeriter        = 0;
 508     inneriter        = 0;
 509
 510     for(iidx=0;iidx<4*DIM;iidx++)
 511     {
 512         scratch[iidx] = 0.0;
 513     }
 514
 515     /* Start outer loop over neighborlists */
 516     for(iidx=0; iidx<nri; iidx++)
 517     {
 518         /* Load shift vector for this list */
 519         i_shift_offset   = DIM*shiftidx[iidx];
 520
 521         /* Load limits for loop over neighbors */
 522         j_index_start    = jindex[iidx];
 523         j_index_end      = jindex[iidx+1];
 524
 525         /* Get outer coordinate index */
 526         inr              = iinr[iidx];
 527         i_coord_offset   = DIM*inr;
 528
 529         /* Load i particle coords and add shift vector */
 530         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 531
 532         fix0             = _mm_setzero_ps();
 533         fiy0             = _mm_setzero_ps();
 534         fiz0             = _mm_setzero_ps();
 535
 536         /* Load parameters for i particles */
 537         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 538         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 539
 540         /* Start inner kernel loop */
 541         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 542         {
 543
 544             /* Get j neighbor index, and coordinate index */
 545             jnrA             = jjnr[jidx];
 546             jnrB             = jjnr[jidx+1];
 547             jnrC             = jjnr[jidx+2];
 548             jnrD             = jjnr[jidx+3];
 549             j_coord_offsetA  = DIM*jnrA;
 550             j_coord_offsetB  = DIM*jnrB;
 551             j_coord_offsetC  = DIM*jnrC;
 552             j_coord_offsetD  = DIM*jnrD;
 553
 554             /* load j atom coordinates */
 555             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 556                                               x+j_coord_offsetC,x+j_coord_offsetD,
 557                                               &jx0,&jy0,&jz0);
 558
 559             /* Calculate displacement vector */
 560             dx00             = _mm_sub_ps(ix0,jx0);
 561             dy00             = _mm_sub_ps(iy0,jy0);
 562             dz00             = _mm_sub_ps(iz0,jz0);
 563
 564             /* Calculate squared distance and things based on it */
 565             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 566
 567             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 568
 569             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 570
 571             /* Load parameters for j particles */
 572             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 573                                                               charge+jnrC+0,charge+jnrD+0);
 574             vdwjidx0A        = 2*vdwtype[jnrA+0];
 575             vdwjidx0B        = 2*vdwtype[jnrB+0];
 576             vdwjidx0C        = 2*vdwtype[jnrC+0];
 577             vdwjidx0D        = 2*vdwtype[jnrD+0];
 578
 579             /**************************
 580              * CALCULATE INTERACTIONS *
 581              **************************/
 582
 583             r00              = _mm_mul_ps(rsq00,rinv00);
 584
 585             /* Compute parameters for interactions between i and j atoms */
 586             qq00             = _mm_mul_ps(iq0,jq0);
 587             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 588                                          vdwparam+vdwioffset0+vdwjidx0B,
 589                                          vdwparam+vdwioffset0+vdwjidx0C,
 590                                          vdwparam+vdwioffset0+vdwjidx0D,
 591                                          &c6_00,&c12_00);
 592
 593             /* Calculate table index by multiplying r with table scale and truncate to integer */
 594             rt               = _mm_mul_ps(r00,vftabscale);
 595             vfitab           = _mm_cvttps_epi32(rt);
 596 #ifdef __XOP__
 597             vfeps            = _mm_frcz_ps(rt);
 598 #else
 599             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 600 #endif
 601             twovfeps         = _mm_add_ps(vfeps,vfeps);
 602             vfitab           = _mm_slli_epi32(vfitab,3);
 603
 604             /* COULOMB ELECTROSTATICS */
 605             velec            = _mm_mul_ps(qq00,rinv00);
 606             felec            = _mm_mul_ps(velec,rinvsq00);
 607
 608             /* CUBIC SPLINE TABLE DISPERSION */
 609             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 610             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 611             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 612             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 613             _MM_TRANSPOSE4_PS(Y,F,G,H);
 614             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 615             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 616             fvdw6            = _mm_mul_ps(c6_00,FF);
 617
 618             /* CUBIC SPLINE TABLE REPULSION */
 619             vfitab           = _mm_add_epi32(vfitab,ifour);
 620             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 621             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 622             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 623             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 624             _MM_TRANSPOSE4_PS(Y,F,G,H);
 625             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 626             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 627             fvdw12           = _mm_mul_ps(c12_00,FF);
 628             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 629
 630             fscal            = _mm_add_ps(felec,fvdw);
 631
 632              /* Update vectorial force */
 633             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 634             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 635             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 636
 637             fjptrA             = f+j_coord_offsetA;
 638             fjptrB             = f+j_coord_offsetB;
 639             fjptrC             = f+j_coord_offsetC;
 640             fjptrD             = f+j_coord_offsetD;
 641             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 642                                                    _mm_mul_ps(dx00,fscal),
 643                                                    _mm_mul_ps(dy00,fscal),
 644                                                    _mm_mul_ps(dz00,fscal));
 645
 646             /* Inner loop uses 57 flops */
 647         }
 648
 649         if(jidx<j_index_end)
 650         {
 651
 652             /* Get j neighbor index, and coordinate index */
 653             jnrlistA         = jjnr[jidx];
 654             jnrlistB         = jjnr[jidx+1];
 655             jnrlistC         = jjnr[jidx+2];
 656             jnrlistD         = jjnr[jidx+3];
 657             /* Sign of each element will be negative for non-real atoms.
 658              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 659              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 660              */
 661             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 662             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 663             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 664             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 665             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 666             j_coord_offsetA  = DIM*jnrA;
 667             j_coord_offsetB  = DIM*jnrB;
 668             j_coord_offsetC  = DIM*jnrC;
 669             j_coord_offsetD  = DIM*jnrD;
 670
 671             /* load j atom coordinates */
 672             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 673                                               x+j_coord_offsetC,x+j_coord_offsetD,
 674                                               &jx0,&jy0,&jz0);
 675
 676             /* Calculate displacement vector */
 677             dx00             = _mm_sub_ps(ix0,jx0);
 678             dy00             = _mm_sub_ps(iy0,jy0);
 679             dz00             = _mm_sub_ps(iz0,jz0);
 680
 681             /* Calculate squared distance and things based on it */
 682             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 683
 684             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 685
 686             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 687
 688             /* Load parameters for j particles */
 689             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 690                                                               charge+jnrC+0,charge+jnrD+0);
 691             vdwjidx0A        = 2*vdwtype[jnrA+0];
 692             vdwjidx0B        = 2*vdwtype[jnrB+0];
 693             vdwjidx0C        = 2*vdwtype[jnrC+0];
 694             vdwjidx0D        = 2*vdwtype[jnrD+0];
 695
 696             /**************************
 697              * CALCULATE INTERACTIONS *
 698              **************************/
 699
 700             r00              = _mm_mul_ps(rsq00,rinv00);
 701             r00              = _mm_andnot_ps(dummy_mask,r00);
 702
 703             /* Compute parameters for interactions between i and j atoms */
 704             qq00             = _mm_mul_ps(iq0,jq0);
 705             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 706                                          vdwparam+vdwioffset0+vdwjidx0B,
 707                                          vdwparam+vdwioffset0+vdwjidx0C,
 708                                          vdwparam+vdwioffset0+vdwjidx0D,
 709                                          &c6_00,&c12_00);
 710
 711             /* Calculate table index by multiplying r with table scale and truncate to integer */
 712             rt               = _mm_mul_ps(r00,vftabscale);
 713             vfitab           = _mm_cvttps_epi32(rt);
 714 #ifdef __XOP__
 715             vfeps            = _mm_frcz_ps(rt);
 716 #else
 717             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 718 #endif
 719             twovfeps         = _mm_add_ps(vfeps,vfeps);
 720             vfitab           = _mm_slli_epi32(vfitab,3);
 721
 722             /* COULOMB ELECTROSTATICS */
 723             velec            = _mm_mul_ps(qq00,rinv00);
 724             felec            = _mm_mul_ps(velec,rinvsq00);
 725
 726             /* CUBIC SPLINE TABLE DISPERSION */
 727             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 728             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 729             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 730             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 731             _MM_TRANSPOSE4_PS(Y,F,G,H);
 732             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 733             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 734             fvdw6            = _mm_mul_ps(c6_00,FF);
 735
 736             /* CUBIC SPLINE TABLE REPULSION */
 737             vfitab           = _mm_add_epi32(vfitab,ifour);
 738             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 739             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 740             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 741             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 742             _MM_TRANSPOSE4_PS(Y,F,G,H);
 743             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 744             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 745             fvdw12           = _mm_mul_ps(c12_00,FF);
 746             fvdw             = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
 747
 748             fscal            = _mm_add_ps(felec,fvdw);
 749
 750             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 751
 752              /* Update vectorial force */
 753             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 754             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 755             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 756
 757             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 758             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 759             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 760             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 761             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 762                                                    _mm_mul_ps(dx00,fscal),
 763                                                    _mm_mul_ps(dy00,fscal),
 764                                                    _mm_mul_ps(dz00,fscal));
 765
 766             /* Inner loop uses 58 flops */
 767         }
 768
 769         /* End of innermost loop */
 770
 771         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 772                                               f+i_coord_offset,fshift+i_shift_offset);
 773
 774         /* Increment number of inner iterations */
 775         inneriter                  += j_index_end - j_index_start;
 776
 777         /* Outer loop uses 7 flops */
 778     }
 779
 780     /* Increment number of outer iterations */
 781     outeriter        += nri;
 782
 783     /* Update outer/inner flops */
 784
 785     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*58);
 786 }