src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwLJ_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_ElecGB_VdwLJ_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_128_fma_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     __m128i          gbitab;
  77     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
  78     __m128           minushalf = _mm_set1_ps(-0.5);
  79     real             *invsqrta,*dvda,*gbtab;
  80     int              nvdwtype;
  81     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  82     int              *vdwtype;
  83     real             *vdwparam;
  84     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  85     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  86     __m128i          vfitab;
  87     __m128i          ifour       = _mm_set1_epi32(4);
  88     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  89     real             *vftab;
  90     __m128           dummy_mask,cutoff_mask;
  91     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  92     __m128           one     = _mm_set1_ps(1.0);
  93     __m128           two     = _mm_set1_ps(2.0);
  94     x                = xx[0];
  95     f                = ff[0];
  96
  97     nri              = nlist->nri;
  98     iinr             = nlist->iinr;
  99     jindex           = nlist->jindex;
 100     jjnr             = nlist->jjnr;
 101     shiftidx         = nlist->shift;
 102     gid              = nlist->gid;
 103     shiftvec         = fr->shift_vec[0];
 104     fshift           = fr->fshift[0];
 105     facel            = _mm_set1_ps(fr->epsfac);
 106     charge           = mdatoms->chargeA;
 107     nvdwtype         = fr->ntype;
 108     vdwparam         = fr->nbfp;
 109     vdwtype          = mdatoms->typeA;
 110
 111     invsqrta         = fr->invsqrta;
 112     dvda             = fr->dvda;
 113     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 114     gbtab            = fr->gbtab.data;
 115     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 116
 117     /* Avoid stupid compiler warnings */
 118     jnrA = jnrB = jnrC = jnrD = 0;
 119     j_coord_offsetA = 0;
 120     j_coord_offsetB = 0;
 121     j_coord_offsetC = 0;
 122     j_coord_offsetD = 0;
 123
 124     outeriter        = 0;
 125     inneriter        = 0;
 126
 127     for(iidx=0;iidx<4*DIM;iidx++)
 128     {
 129         scratch[iidx] = 0.0;
 130     }
 131
 132     /* Start outer loop over neighborlists */
 133     for(iidx=0; iidx<nri; iidx++)
 134     {
 135         /* Load shift vector for this list */
 136         i_shift_offset   = DIM*shiftidx[iidx];
 137
 138         /* Load limits for loop over neighbors */
 139         j_index_start    = jindex[iidx];
 140         j_index_end      = jindex[iidx+1];
 141
 142         /* Get outer coordinate index */
 143         inr              = iinr[iidx];
 144         i_coord_offset   = DIM*inr;
 145
 146         /* Load i particle coords and add shift vector */
 147         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 148
 149         fix0             = _mm_setzero_ps();
 150         fiy0             = _mm_setzero_ps();
 151         fiz0             = _mm_setzero_ps();
 152
 153         /* Load parameters for i particles */
 154         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 155         isai0            = _mm_load1_ps(invsqrta+inr+0);
 156         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 157
 158         /* Reset potential sums */
 159         velecsum         = _mm_setzero_ps();
 160         vgbsum           = _mm_setzero_ps();
 161         vvdwsum          = _mm_setzero_ps();
 162         dvdasum          = _mm_setzero_ps();
 163
 164         /* Start inner kernel loop */
 165         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 166         {
 167
 168             /* Get j neighbor index, and coordinate index */
 169             jnrA             = jjnr[jidx];
 170             jnrB             = jjnr[jidx+1];
 171             jnrC             = jjnr[jidx+2];
 172             jnrD             = jjnr[jidx+3];
 173             j_coord_offsetA  = DIM*jnrA;
 174             j_coord_offsetB  = DIM*jnrB;
 175             j_coord_offsetC  = DIM*jnrC;
 176             j_coord_offsetD  = DIM*jnrD;
 177
 178             /* load j atom coordinates */
 179             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 180                                               x+j_coord_offsetC,x+j_coord_offsetD,
 181                                               &jx0,&jy0,&jz0);
 182
 183             /* Calculate displacement vector */
 184             dx00             = _mm_sub_ps(ix0,jx0);
 185             dy00             = _mm_sub_ps(iy0,jy0);
 186             dz00             = _mm_sub_ps(iz0,jz0);
 187
 188             /* Calculate squared distance and things based on it */
 189             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 190
 191             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 192
 193             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 194
 195             /* Load parameters for j particles */
 196             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 197                                                               charge+jnrC+0,charge+jnrD+0);
 198             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 199                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 200             vdwjidx0A        = 2*vdwtype[jnrA+0];
 201             vdwjidx0B        = 2*vdwtype[jnrB+0];
 202             vdwjidx0C        = 2*vdwtype[jnrC+0];
 203             vdwjidx0D        = 2*vdwtype[jnrD+0];
 204
 205             /**************************
 206              * CALCULATE INTERACTIONS *
 207              **************************/
 208
 209             r00              = _mm_mul_ps(rsq00,rinv00);
 210
 211             /* Compute parameters for interactions between i and j atoms */
 212             qq00             = _mm_mul_ps(iq0,jq0);
 213             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 214                                          vdwparam+vdwioffset0+vdwjidx0B,
 215                                          vdwparam+vdwioffset0+vdwjidx0C,
 216                                          vdwparam+vdwioffset0+vdwjidx0D,
 217                                          &c6_00,&c12_00);
 218
 219             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 220             isaprod          = _mm_mul_ps(isai0,isaj0);
 221             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 222             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 223
 224             /* Calculate generalized born table index - this is a separate table from the normal one,
 225              * but we use the same procedure by multiplying r with scale and truncating to integer.
 226              */
 227             rt               = _mm_mul_ps(r00,gbscale);
 228             gbitab           = _mm_cvttps_epi32(rt);
 229 #ifdef __XOP__
 230             gbeps            = _mm_frcz_ps(rt);
 231 #else
 232             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 233 #endif
 234             gbitab           = _mm_slli_epi32(gbitab,2);
 235
 236             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 237             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 238             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 239             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 240             _MM_TRANSPOSE4_PS(Y,F,G,H);
 241             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 242             VV               = _mm_macc_ps(gbeps,Fp,Y);
 243             vgb              = _mm_mul_ps(gbqqfactor,VV);
 244
 245             twogbeps         = _mm_add_ps(gbeps,gbeps);
 246             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 247             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 248             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 249             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 250             fjptrA           = dvda+jnrA;
 251             fjptrB           = dvda+jnrB;
 252             fjptrC           = dvda+jnrC;
 253             fjptrD           = dvda+jnrD;
 254             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 255             velec            = _mm_mul_ps(qq00,rinv00);
 256             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 257
 258             /* LENNARD-JONES DISPERSION/REPULSION */
 259
 260             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 261             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 262             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 263             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 264             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 265
 266             /* Update potential sum for this i atom from the interaction with this j atom. */
 267             velecsum         = _mm_add_ps(velecsum,velec);
 268             vgbsum           = _mm_add_ps(vgbsum,vgb);
 269             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 270
 271             fscal            = _mm_add_ps(felec,fvdw);
 272
 273              /* Update vectorial force */
 274             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 275             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 276             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 277
 278             fjptrA             = f+j_coord_offsetA;
 279             fjptrB             = f+j_coord_offsetB;
 280             fjptrC             = f+j_coord_offsetC;
 281             fjptrD             = f+j_coord_offsetD;
 282             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 283                                                    _mm_mul_ps(dx00,fscal),
 284                                                    _mm_mul_ps(dy00,fscal),
 285                                                    _mm_mul_ps(dz00,fscal));
 286
 287             /* Inner loop uses 74 flops */
 288         }
 289
 290         if(jidx<j_index_end)
 291         {
 292
 293             /* Get j neighbor index, and coordinate index */
 294             jnrlistA         = jjnr[jidx];
 295             jnrlistB         = jjnr[jidx+1];
 296             jnrlistC         = jjnr[jidx+2];
 297             jnrlistD         = jjnr[jidx+3];
 298             /* Sign of each element will be negative for non-real atoms.
 299              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 300              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 301              */
 302             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 303             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 304             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 305             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 306             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 307             j_coord_offsetA  = DIM*jnrA;
 308             j_coord_offsetB  = DIM*jnrB;
 309             j_coord_offsetC  = DIM*jnrC;
 310             j_coord_offsetD  = DIM*jnrD;
 311
 312             /* load j atom coordinates */
 313             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 314                                               x+j_coord_offsetC,x+j_coord_offsetD,
 315                                               &jx0,&jy0,&jz0);
 316
 317             /* Calculate displacement vector */
 318             dx00             = _mm_sub_ps(ix0,jx0);
 319             dy00             = _mm_sub_ps(iy0,jy0);
 320             dz00             = _mm_sub_ps(iz0,jz0);
 321
 322             /* Calculate squared distance and things based on it */
 323             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 324
 325             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 326
 327             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 328
 329             /* Load parameters for j particles */
 330             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 331                                                               charge+jnrC+0,charge+jnrD+0);
 332             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 333                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 334             vdwjidx0A        = 2*vdwtype[jnrA+0];
 335             vdwjidx0B        = 2*vdwtype[jnrB+0];
 336             vdwjidx0C        = 2*vdwtype[jnrC+0];
 337             vdwjidx0D        = 2*vdwtype[jnrD+0];
 338
 339             /**************************
 340              * CALCULATE INTERACTIONS *
 341              **************************/
 342
 343             r00              = _mm_mul_ps(rsq00,rinv00);
 344             r00              = _mm_andnot_ps(dummy_mask,r00);
 345
 346             /* Compute parameters for interactions between i and j atoms */
 347             qq00             = _mm_mul_ps(iq0,jq0);
 348             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 349                                          vdwparam+vdwioffset0+vdwjidx0B,
 350                                          vdwparam+vdwioffset0+vdwjidx0C,
 351                                          vdwparam+vdwioffset0+vdwjidx0D,
 352                                          &c6_00,&c12_00);
 353
 354             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 355             isaprod          = _mm_mul_ps(isai0,isaj0);
 356             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 357             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 358
 359             /* Calculate generalized born table index - this is a separate table from the normal one,
 360              * but we use the same procedure by multiplying r with scale and truncating to integer.
 361              */
 362             rt               = _mm_mul_ps(r00,gbscale);
 363             gbitab           = _mm_cvttps_epi32(rt);
 364 #ifdef __XOP__
 365             gbeps            = _mm_frcz_ps(rt);
 366 #else
 367             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 368 #endif
 369             gbitab           = _mm_slli_epi32(gbitab,2);
 370
 371             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 372             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 373             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 374             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 375             _MM_TRANSPOSE4_PS(Y,F,G,H);
 376             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 377             VV               = _mm_macc_ps(gbeps,Fp,Y);
 378             vgb              = _mm_mul_ps(gbqqfactor,VV);
 379
 380             twogbeps         = _mm_add_ps(gbeps,gbeps);
 381             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 382             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 383             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 384             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 385             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 386             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 387             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 388             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 389             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 390             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 391             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 392             velec            = _mm_mul_ps(qq00,rinv00);
 393             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 394
 395             /* LENNARD-JONES DISPERSION/REPULSION */
 396
 397             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 398             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 399             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 400             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 401             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 402
 403             /* Update potential sum for this i atom from the interaction with this j atom. */
 404             velec            = _mm_andnot_ps(dummy_mask,velec);
 405             velecsum         = _mm_add_ps(velecsum,velec);
 406             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 407             vgbsum           = _mm_add_ps(vgbsum,vgb);
 408             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 409             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 410
 411             fscal            = _mm_add_ps(felec,fvdw);
 412
 413             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 414
 415              /* Update vectorial force */
 416             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 417             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 418             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 419
 420             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 421             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 422             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 423             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 424             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 425                                                    _mm_mul_ps(dx00,fscal),
 426                                                    _mm_mul_ps(dy00,fscal),
 427                                                    _mm_mul_ps(dz00,fscal));
 428
 429             /* Inner loop uses 75 flops */
 430         }
 431
 432         /* End of innermost loop */
 433
 434         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 435                                               f+i_coord_offset,fshift+i_shift_offset);
 436
 437         ggid                        = gid[iidx];
 438         /* Update potential energies */
 439         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 440         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 441         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 442         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 443         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 444
 445         /* Increment number of inner iterations */
 446         inneriter                  += j_index_end - j_index_start;
 447
 448         /* Outer loop uses 10 flops */
 449     }
 450
 451     /* Increment number of outer iterations */
 452     outeriter        += nri;
 453
 454     /* Update outer/inner flops */
 455
 456     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*75);
 457 }
 458 /*
 459  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_single
 460  * Electrostatics interaction: GeneralizedBorn
 461  * VdW interaction:            LennardJones
 462  * Geometry:                   Particle-Particle
 463  * Calculate force/pot:        Force
 464  */
 465 void
 466 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_single
 467                     (t_nblist * gmx_restrict                nlist,
 468                      rvec * gmx_restrict                    xx,
 469                      rvec * gmx_restrict                    ff,
 470                      t_forcerec * gmx_restrict              fr,
 471                      t_mdatoms * gmx_restrict               mdatoms,
 472                      nb_kernel_data_t * gmx_restrict        kernel_data,
 473                      t_nrnb * gmx_restrict                  nrnb)
 474 {
 475     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 476      * just 0 for non-waters.
 477      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 478      * jnr indices corresponding to data put in the four positions in the SIMD register.
 479      */
 480     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 481     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 482     int              jnrA,jnrB,jnrC,jnrD;
 483     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 484     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 485     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 486     real             rcutoff_scalar;
 487     real             *shiftvec,*fshift,*x,*f;
 488     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 489     real             scratch[4*DIM];
 490     __m128           fscal,rcutoff,rcutoff2,jidxall;
 491     int              vdwioffset0;
 492     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 493     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 494     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 495     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 496     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 497     real             *charge;
 498     __m128i          gbitab;
 499     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 500     __m128           minushalf = _mm_set1_ps(-0.5);
 501     real             *invsqrta,*dvda,*gbtab;
 502     int              nvdwtype;
 503     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 504     int              *vdwtype;
 505     real             *vdwparam;
 506     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 507     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 508     __m128i          vfitab;
 509     __m128i          ifour       = _mm_set1_epi32(4);
 510     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 511     real             *vftab;
 512     __m128           dummy_mask,cutoff_mask;
 513     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 514     __m128           one     = _mm_set1_ps(1.0);
 515     __m128           two     = _mm_set1_ps(2.0);
 516     x                = xx[0];
 517     f                = ff[0];
 518
 519     nri              = nlist->nri;
 520     iinr             = nlist->iinr;
 521     jindex           = nlist->jindex;
 522     jjnr             = nlist->jjnr;
 523     shiftidx         = nlist->shift;
 524     gid              = nlist->gid;
 525     shiftvec         = fr->shift_vec[0];
 526     fshift           = fr->fshift[0];
 527     facel            = _mm_set1_ps(fr->epsfac);
 528     charge           = mdatoms->chargeA;
 529     nvdwtype         = fr->ntype;
 530     vdwparam         = fr->nbfp;
 531     vdwtype          = mdatoms->typeA;
 532
 533     invsqrta         = fr->invsqrta;
 534     dvda             = fr->dvda;
 535     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 536     gbtab            = fr->gbtab.data;
 537     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 538
 539     /* Avoid stupid compiler warnings */
 540     jnrA = jnrB = jnrC = jnrD = 0;
 541     j_coord_offsetA = 0;
 542     j_coord_offsetB = 0;
 543     j_coord_offsetC = 0;
 544     j_coord_offsetD = 0;
 545
 546     outeriter        = 0;
 547     inneriter        = 0;
 548
 549     for(iidx=0;iidx<4*DIM;iidx++)
 550     {
 551         scratch[iidx] = 0.0;
 552     }
 553
 554     /* Start outer loop over neighborlists */
 555     for(iidx=0; iidx<nri; iidx++)
 556     {
 557         /* Load shift vector for this list */
 558         i_shift_offset   = DIM*shiftidx[iidx];
 559
 560         /* Load limits for loop over neighbors */
 561         j_index_start    = jindex[iidx];
 562         j_index_end      = jindex[iidx+1];
 563
 564         /* Get outer coordinate index */
 565         inr              = iinr[iidx];
 566         i_coord_offset   = DIM*inr;
 567
 568         /* Load i particle coords and add shift vector */
 569         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 570
 571         fix0             = _mm_setzero_ps();
 572         fiy0             = _mm_setzero_ps();
 573         fiz0             = _mm_setzero_ps();
 574
 575         /* Load parameters for i particles */
 576         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 577         isai0            = _mm_load1_ps(invsqrta+inr+0);
 578         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 579
 580         dvdasum          = _mm_setzero_ps();
 581
 582         /* Start inner kernel loop */
 583         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 584         {
 585
 586             /* Get j neighbor index, and coordinate index */
 587             jnrA             = jjnr[jidx];
 588             jnrB             = jjnr[jidx+1];
 589             jnrC             = jjnr[jidx+2];
 590             jnrD             = jjnr[jidx+3];
 591             j_coord_offsetA  = DIM*jnrA;
 592             j_coord_offsetB  = DIM*jnrB;
 593             j_coord_offsetC  = DIM*jnrC;
 594             j_coord_offsetD  = DIM*jnrD;
 595
 596             /* load j atom coordinates */
 597             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 598                                               x+j_coord_offsetC,x+j_coord_offsetD,
 599                                               &jx0,&jy0,&jz0);
 600
 601             /* Calculate displacement vector */
 602             dx00             = _mm_sub_ps(ix0,jx0);
 603             dy00             = _mm_sub_ps(iy0,jy0);
 604             dz00             = _mm_sub_ps(iz0,jz0);
 605
 606             /* Calculate squared distance and things based on it */
 607             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 608
 609             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 610
 611             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 612
 613             /* Load parameters for j particles */
 614             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 615                                                               charge+jnrC+0,charge+jnrD+0);
 616             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 617                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 618             vdwjidx0A        = 2*vdwtype[jnrA+0];
 619             vdwjidx0B        = 2*vdwtype[jnrB+0];
 620             vdwjidx0C        = 2*vdwtype[jnrC+0];
 621             vdwjidx0D        = 2*vdwtype[jnrD+0];
 622
 623             /**************************
 624              * CALCULATE INTERACTIONS *
 625              **************************/
 626
 627             r00              = _mm_mul_ps(rsq00,rinv00);
 628
 629             /* Compute parameters for interactions between i and j atoms */
 630             qq00             = _mm_mul_ps(iq0,jq0);
 631             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 632                                          vdwparam+vdwioffset0+vdwjidx0B,
 633                                          vdwparam+vdwioffset0+vdwjidx0C,
 634                                          vdwparam+vdwioffset0+vdwjidx0D,
 635                                          &c6_00,&c12_00);
 636
 637             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 638             isaprod          = _mm_mul_ps(isai0,isaj0);
 639             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 640             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 641
 642             /* Calculate generalized born table index - this is a separate table from the normal one,
 643              * but we use the same procedure by multiplying r with scale and truncating to integer.
 644              */
 645             rt               = _mm_mul_ps(r00,gbscale);
 646             gbitab           = _mm_cvttps_epi32(rt);
 647 #ifdef __XOP__
 648             gbeps            = _mm_frcz_ps(rt);
 649 #else
 650             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 651 #endif
 652             gbitab           = _mm_slli_epi32(gbitab,2);
 653
 654             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 655             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 656             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 657             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 658             _MM_TRANSPOSE4_PS(Y,F,G,H);
 659             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 660             VV               = _mm_macc_ps(gbeps,Fp,Y);
 661             vgb              = _mm_mul_ps(gbqqfactor,VV);
 662
 663             twogbeps         = _mm_add_ps(gbeps,gbeps);
 664             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 665             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 666             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 667             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 668             fjptrA           = dvda+jnrA;
 669             fjptrB           = dvda+jnrB;
 670             fjptrC           = dvda+jnrC;
 671             fjptrD           = dvda+jnrD;
 672             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 673             velec            = _mm_mul_ps(qq00,rinv00);
 674             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 675
 676             /* LENNARD-JONES DISPERSION/REPULSION */
 677
 678             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 679             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 680
 681             fscal            = _mm_add_ps(felec,fvdw);
 682
 683              /* Update vectorial force */
 684             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 685             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 686             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 687
 688             fjptrA             = f+j_coord_offsetA;
 689             fjptrB             = f+j_coord_offsetB;
 690             fjptrC             = f+j_coord_offsetC;
 691             fjptrD             = f+j_coord_offsetD;
 692             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 693                                                    _mm_mul_ps(dx00,fscal),
 694                                                    _mm_mul_ps(dy00,fscal),
 695                                                    _mm_mul_ps(dz00,fscal));
 696
 697             /* Inner loop uses 67 flops */
 698         }
 699
 700         if(jidx<j_index_end)
 701         {
 702
 703             /* Get j neighbor index, and coordinate index */
 704             jnrlistA         = jjnr[jidx];
 705             jnrlistB         = jjnr[jidx+1];
 706             jnrlistC         = jjnr[jidx+2];
 707             jnrlistD         = jjnr[jidx+3];
 708             /* Sign of each element will be negative for non-real atoms.
 709              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 710              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 711              */
 712             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 713             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 714             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 715             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 716             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 717             j_coord_offsetA  = DIM*jnrA;
 718             j_coord_offsetB  = DIM*jnrB;
 719             j_coord_offsetC  = DIM*jnrC;
 720             j_coord_offsetD  = DIM*jnrD;
 721
 722             /* load j atom coordinates */
 723             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 724                                               x+j_coord_offsetC,x+j_coord_offsetD,
 725                                               &jx0,&jy0,&jz0);
 726
 727             /* Calculate displacement vector */
 728             dx00             = _mm_sub_ps(ix0,jx0);
 729             dy00             = _mm_sub_ps(iy0,jy0);
 730             dz00             = _mm_sub_ps(iz0,jz0);
 731
 732             /* Calculate squared distance and things based on it */
 733             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 734
 735             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 736
 737             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 738
 739             /* Load parameters for j particles */
 740             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 741                                                               charge+jnrC+0,charge+jnrD+0);
 742             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 743                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 744             vdwjidx0A        = 2*vdwtype[jnrA+0];
 745             vdwjidx0B        = 2*vdwtype[jnrB+0];
 746             vdwjidx0C        = 2*vdwtype[jnrC+0];
 747             vdwjidx0D        = 2*vdwtype[jnrD+0];
 748
 749             /**************************
 750              * CALCULATE INTERACTIONS *
 751              **************************/
 752
 753             r00              = _mm_mul_ps(rsq00,rinv00);
 754             r00              = _mm_andnot_ps(dummy_mask,r00);
 755
 756             /* Compute parameters for interactions between i and j atoms */
 757             qq00             = _mm_mul_ps(iq0,jq0);
 758             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 759                                          vdwparam+vdwioffset0+vdwjidx0B,
 760                                          vdwparam+vdwioffset0+vdwjidx0C,
 761                                          vdwparam+vdwioffset0+vdwjidx0D,
 762                                          &c6_00,&c12_00);
 763
 764             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 765             isaprod          = _mm_mul_ps(isai0,isaj0);
 766             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 767             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 768
 769             /* Calculate generalized born table index - this is a separate table from the normal one,
 770              * but we use the same procedure by multiplying r with scale and truncating to integer.
 771              */
 772             rt               = _mm_mul_ps(r00,gbscale);
 773             gbitab           = _mm_cvttps_epi32(rt);
 774 #ifdef __XOP__
 775             gbeps            = _mm_frcz_ps(rt);
 776 #else
 777             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 778 #endif
 779             gbitab           = _mm_slli_epi32(gbitab,2);
 780
 781             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 782             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 783             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 784             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 785             _MM_TRANSPOSE4_PS(Y,F,G,H);
 786             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 787             VV               = _mm_macc_ps(gbeps,Fp,Y);
 788             vgb              = _mm_mul_ps(gbqqfactor,VV);
 789
 790             twogbeps         = _mm_add_ps(gbeps,gbeps);
 791             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 792             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 793             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 794             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 795             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 796             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
 797             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 798             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 799             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 800             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 801             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 802             velec            = _mm_mul_ps(qq00,rinv00);
 803             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 804
 805             /* LENNARD-JONES DISPERSION/REPULSION */
 806
 807             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 808             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 809
 810             fscal            = _mm_add_ps(felec,fvdw);
 811
 812             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 813
 814              /* Update vectorial force */
 815             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 816             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 817             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 818
 819             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 820             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 821             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 822             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 823             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 824                                                    _mm_mul_ps(dx00,fscal),
 825                                                    _mm_mul_ps(dy00,fscal),
 826                                                    _mm_mul_ps(dz00,fscal));
 827
 828             /* Inner loop uses 68 flops */
 829         }
 830
 831         /* End of innermost loop */
 832
 833         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 834                                               f+i_coord_offset,fshift+i_shift_offset);
 835
 836         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 837         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 838
 839         /* Increment number of inner iterations */
 840         inneriter                  += j_index_end - j_index_start;
 841
 842         /* Outer loop uses 7 flops */
 843     }
 844
 845     /* Increment number of outer iterations */
 846     outeriter        += nri;
 847
 848     /* Update outer/inner flops */
 849
 850     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*68);
 851 }