src/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             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 385             /* 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. */
 386             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 387             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 388             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 389             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 390             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 391             velec            = _mm_mul_ps(qq00,rinv00);
 392             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 393
 394             /* LENNARD-JONES DISPERSION/REPULSION */
 395
 396             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 397             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 398             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 399             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 400             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 401
 402             /* Update potential sum for this i atom from the interaction with this j atom. */
 403             velec            = _mm_andnot_ps(dummy_mask,velec);
 404             velecsum         = _mm_add_ps(velecsum,velec);
 405             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 406             vgbsum           = _mm_add_ps(vgbsum,vgb);
 407             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 408             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 409
 410             fscal            = _mm_add_ps(felec,fvdw);
 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             /* Inner loop uses 75 flops */
 429         }
 430
 431         /* End of innermost loop */
 432
 433         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 434                                               f+i_coord_offset,fshift+i_shift_offset);
 435
 436         ggid                        = gid[iidx];
 437         /* Update potential energies */
 438         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 439         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 440         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 441         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 442         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 443
 444         /* Increment number of inner iterations */
 445         inneriter                  += j_index_end - j_index_start;
 446
 447         /* Outer loop uses 10 flops */
 448     }
 449
 450     /* Increment number of outer iterations */
 451     outeriter        += nri;
 452
 453     /* Update outer/inner flops */
 454
 455     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*75);
 456 }
 457 /*
 458  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_single
 459  * Electrostatics interaction: GeneralizedBorn
 460  * VdW interaction:            LennardJones
 461  * Geometry:                   Particle-Particle
 462  * Calculate force/pot:        Force
 463  */
 464 void
 465 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_128_fma_single
 466                     (t_nblist * gmx_restrict                nlist,
 467                      rvec * gmx_restrict                    xx,
 468                      rvec * gmx_restrict                    ff,
 469                      t_forcerec * gmx_restrict              fr,
 470                      t_mdatoms * gmx_restrict               mdatoms,
 471                      nb_kernel_data_t * gmx_restrict        kernel_data,
 472                      t_nrnb * gmx_restrict                  nrnb)
 473 {
 474     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 475      * just 0 for non-waters.
 476      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 477      * jnr indices corresponding to data put in the four positions in the SIMD register.
 478      */
 479     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 480     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 481     int              jnrA,jnrB,jnrC,jnrD;
 482     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 483     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 484     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 485     real             rcutoff_scalar;
 486     real             *shiftvec,*fshift,*x,*f;
 487     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 488     real             scratch[4*DIM];
 489     __m128           fscal,rcutoff,rcutoff2,jidxall;
 490     int              vdwioffset0;
 491     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 492     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 493     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 494     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 495     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 496     real             *charge;
 497     __m128i          gbitab;
 498     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 499     __m128           minushalf = _mm_set1_ps(-0.5);
 500     real             *invsqrta,*dvda,*gbtab;
 501     int              nvdwtype;
 502     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 503     int              *vdwtype;
 504     real             *vdwparam;
 505     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 506     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 507     __m128i          vfitab;
 508     __m128i          ifour       = _mm_set1_epi32(4);
 509     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 510     real             *vftab;
 511     __m128           dummy_mask,cutoff_mask;
 512     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 513     __m128           one     = _mm_set1_ps(1.0);
 514     __m128           two     = _mm_set1_ps(2.0);
 515     x                = xx[0];
 516     f                = ff[0];
 517
 518     nri              = nlist->nri;
 519     iinr             = nlist->iinr;
 520     jindex           = nlist->jindex;
 521     jjnr             = nlist->jjnr;
 522     shiftidx         = nlist->shift;
 523     gid              = nlist->gid;
 524     shiftvec         = fr->shift_vec[0];
 525     fshift           = fr->fshift[0];
 526     facel            = _mm_set1_ps(fr->epsfac);
 527     charge           = mdatoms->chargeA;
 528     nvdwtype         = fr->ntype;
 529     vdwparam         = fr->nbfp;
 530     vdwtype          = mdatoms->typeA;
 531
 532     invsqrta         = fr->invsqrta;
 533     dvda             = fr->dvda;
 534     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 535     gbtab            = fr->gbtab.data;
 536     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 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         isai0            = _mm_load1_ps(invsqrta+inr+0);
 577         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 578
 579         dvdasum          = _mm_setzero_ps();
 580
 581         /* Start inner kernel loop */
 582         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 583         {
 584
 585             /* Get j neighbor index, and coordinate index */
 586             jnrA             = jjnr[jidx];
 587             jnrB             = jjnr[jidx+1];
 588             jnrC             = jjnr[jidx+2];
 589             jnrD             = jjnr[jidx+3];
 590             j_coord_offsetA  = DIM*jnrA;
 591             j_coord_offsetB  = DIM*jnrB;
 592             j_coord_offsetC  = DIM*jnrC;
 593             j_coord_offsetD  = DIM*jnrD;
 594
 595             /* load j atom coordinates */
 596             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 597                                               x+j_coord_offsetC,x+j_coord_offsetD,
 598                                               &jx0,&jy0,&jz0);
 599
 600             /* Calculate displacement vector */
 601             dx00             = _mm_sub_ps(ix0,jx0);
 602             dy00             = _mm_sub_ps(iy0,jy0);
 603             dz00             = _mm_sub_ps(iz0,jz0);
 604
 605             /* Calculate squared distance and things based on it */
 606             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 607
 608             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 609
 610             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 611
 612             /* Load parameters for j particles */
 613             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 614                                                               charge+jnrC+0,charge+jnrD+0);
 615             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 616                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 617             vdwjidx0A        = 2*vdwtype[jnrA+0];
 618             vdwjidx0B        = 2*vdwtype[jnrB+0];
 619             vdwjidx0C        = 2*vdwtype[jnrC+0];
 620             vdwjidx0D        = 2*vdwtype[jnrD+0];
 621
 622             /**************************
 623              * CALCULATE INTERACTIONS *
 624              **************************/
 625
 626             r00              = _mm_mul_ps(rsq00,rinv00);
 627
 628             /* Compute parameters for interactions between i and j atoms */
 629             qq00             = _mm_mul_ps(iq0,jq0);
 630             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 631                                          vdwparam+vdwioffset0+vdwjidx0B,
 632                                          vdwparam+vdwioffset0+vdwjidx0C,
 633                                          vdwparam+vdwioffset0+vdwjidx0D,
 634                                          &c6_00,&c12_00);
 635
 636             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 637             isaprod          = _mm_mul_ps(isai0,isaj0);
 638             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 639             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 640
 641             /* Calculate generalized born table index - this is a separate table from the normal one,
 642              * but we use the same procedure by multiplying r with scale and truncating to integer.
 643              */
 644             rt               = _mm_mul_ps(r00,gbscale);
 645             gbitab           = _mm_cvttps_epi32(rt);
 646 #ifdef __XOP__
 647             gbeps            = _mm_frcz_ps(rt);
 648 #else
 649             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 650 #endif
 651             gbitab           = _mm_slli_epi32(gbitab,2);
 652
 653             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 654             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 655             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 656             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 657             _MM_TRANSPOSE4_PS(Y,F,G,H);
 658             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 659             VV               = _mm_macc_ps(gbeps,Fp,Y);
 660             vgb              = _mm_mul_ps(gbqqfactor,VV);
 661
 662             twogbeps         = _mm_add_ps(gbeps,gbeps);
 663             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 664             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 665             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 666             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 667             fjptrA           = dvda+jnrA;
 668             fjptrB           = dvda+jnrB;
 669             fjptrC           = dvda+jnrC;
 670             fjptrD           = dvda+jnrD;
 671             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 672             velec            = _mm_mul_ps(qq00,rinv00);
 673             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 674
 675             /* LENNARD-JONES DISPERSION/REPULSION */
 676
 677             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 678             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 679
 680             fscal            = _mm_add_ps(felec,fvdw);
 681
 682              /* Update vectorial force */
 683             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 684             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 685             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 686
 687             fjptrA             = f+j_coord_offsetA;
 688             fjptrB             = f+j_coord_offsetB;
 689             fjptrC             = f+j_coord_offsetC;
 690             fjptrD             = f+j_coord_offsetD;
 691             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 692                                                    _mm_mul_ps(dx00,fscal),
 693                                                    _mm_mul_ps(dy00,fscal),
 694                                                    _mm_mul_ps(dz00,fscal));
 695
 696             /* Inner loop uses 67 flops */
 697         }
 698
 699         if(jidx<j_index_end)
 700         {
 701
 702             /* Get j neighbor index, and coordinate index */
 703             jnrlistA         = jjnr[jidx];
 704             jnrlistB         = jjnr[jidx+1];
 705             jnrlistC         = jjnr[jidx+2];
 706             jnrlistD         = jjnr[jidx+3];
 707             /* Sign of each element will be negative for non-real atoms.
 708              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 709              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 710              */
 711             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 712             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 713             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 714             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 715             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 716             j_coord_offsetA  = DIM*jnrA;
 717             j_coord_offsetB  = DIM*jnrB;
 718             j_coord_offsetC  = DIM*jnrC;
 719             j_coord_offsetD  = DIM*jnrD;
 720
 721             /* load j atom coordinates */
 722             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 723                                               x+j_coord_offsetC,x+j_coord_offsetD,
 724                                               &jx0,&jy0,&jz0);
 725
 726             /* Calculate displacement vector */
 727             dx00             = _mm_sub_ps(ix0,jx0);
 728             dy00             = _mm_sub_ps(iy0,jy0);
 729             dz00             = _mm_sub_ps(iz0,jz0);
 730
 731             /* Calculate squared distance and things based on it */
 732             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 733
 734             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 735
 736             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 737
 738             /* Load parameters for j particles */
 739             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 740                                                               charge+jnrC+0,charge+jnrD+0);
 741             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 742                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 743             vdwjidx0A        = 2*vdwtype[jnrA+0];
 744             vdwjidx0B        = 2*vdwtype[jnrB+0];
 745             vdwjidx0C        = 2*vdwtype[jnrC+0];
 746             vdwjidx0D        = 2*vdwtype[jnrD+0];
 747
 748             /**************************
 749              * CALCULATE INTERACTIONS *
 750              **************************/
 751
 752             r00              = _mm_mul_ps(rsq00,rinv00);
 753             r00              = _mm_andnot_ps(dummy_mask,r00);
 754
 755             /* Compute parameters for interactions between i and j atoms */
 756             qq00             = _mm_mul_ps(iq0,jq0);
 757             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 758                                          vdwparam+vdwioffset0+vdwjidx0B,
 759                                          vdwparam+vdwioffset0+vdwjidx0C,
 760                                          vdwparam+vdwioffset0+vdwjidx0D,
 761                                          &c6_00,&c12_00);
 762
 763             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 764             isaprod          = _mm_mul_ps(isai0,isaj0);
 765             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 766             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 767
 768             /* Calculate generalized born table index - this is a separate table from the normal one,
 769              * but we use the same procedure by multiplying r with scale and truncating to integer.
 770              */
 771             rt               = _mm_mul_ps(r00,gbscale);
 772             gbitab           = _mm_cvttps_epi32(rt);
 773 #ifdef __XOP__
 774             gbeps            = _mm_frcz_ps(rt);
 775 #else
 776             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 777 #endif
 778             gbitab           = _mm_slli_epi32(gbitab,2);
 779
 780             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 781             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 782             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 783             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 784             _MM_TRANSPOSE4_PS(Y,F,G,H);
 785             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 786             VV               = _mm_macc_ps(gbeps,Fp,Y);
 787             vgb              = _mm_mul_ps(gbqqfactor,VV);
 788
 789             twogbeps         = _mm_add_ps(gbeps,gbeps);
 790             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 791             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 792             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 793             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 794             /* 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. */
 795             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 796             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 797             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 798             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 799             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 800             velec            = _mm_mul_ps(qq00,rinv00);
 801             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 802
 803             /* LENNARD-JONES DISPERSION/REPULSION */
 804
 805             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 806             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 807
 808             fscal            = _mm_add_ps(felec,fvdw);
 809
 810             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 811
 812              /* Update vectorial force */
 813             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 814             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 815             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 816
 817             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 818             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 819             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 820             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 821             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 822                                                    _mm_mul_ps(dx00,fscal),
 823                                                    _mm_mul_ps(dy00,fscal),
 824                                                    _mm_mul_ps(dz00,fscal));
 825
 826             /* Inner loop uses 68 flops */
 827         }
 828
 829         /* End of innermost loop */
 830
 831         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 832                                               f+i_coord_offset,fshift+i_shift_offset);
 833
 834         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 835         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 836
 837         /* Increment number of inner iterations */
 838         inneriter                  += j_index_end - j_index_start;
 839
 840         /* Outer loop uses 7 flops */
 841     }
 842
 843     /* Increment number of outer iterations */
 844     outeriter        += nri;
 845
 846     /* Update outer/inner flops */
 847
 848     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*68);
 849 }