src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwNone_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_VdwNone_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: GeneralizedBorn
  39  * VdW interaction:            None
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecGB_VdwNone_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     __m128i          vfitab;
  81     __m128i          ifour       = _mm_set1_epi32(4);
  82     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  83     real             *vftab;
  84     __m128           dummy_mask,cutoff_mask;
  85     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  86     __m128           one     = _mm_set1_ps(1.0);
  87     __m128           two     = _mm_set1_ps(2.0);
  88     x                = xx[0];
  89     f                = ff[0];
  90
  91     nri              = nlist->nri;
  92     iinr             = nlist->iinr;
  93     jindex           = nlist->jindex;
  94     jjnr             = nlist->jjnr;
  95     shiftidx         = nlist->shift;
  96     gid              = nlist->gid;
  97     shiftvec         = fr->shift_vec[0];
  98     fshift           = fr->fshift[0];
  99     facel            = _mm_set1_ps(fr->epsfac);
 100     charge           = mdatoms->chargeA;
 101
 102     invsqrta         = fr->invsqrta;
 103     dvda             = fr->dvda;
 104     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 105     gbtab            = fr->gbtab.data;
 106     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 107
 108     /* Avoid stupid compiler warnings */
 109     jnrA = jnrB = jnrC = jnrD = 0;
 110     j_coord_offsetA = 0;
 111     j_coord_offsetB = 0;
 112     j_coord_offsetC = 0;
 113     j_coord_offsetD = 0;
 114
 115     outeriter        = 0;
 116     inneriter        = 0;
 117
 118     for(iidx=0;iidx<4*DIM;iidx++)
 119     {
 120         scratch[iidx] = 0.0;
 121     }
 122
 123     /* Start outer loop over neighborlists */
 124     for(iidx=0; iidx<nri; iidx++)
 125     {
 126         /* Load shift vector for this list */
 127         i_shift_offset   = DIM*shiftidx[iidx];
 128
 129         /* Load limits for loop over neighbors */
 130         j_index_start    = jindex[iidx];
 131         j_index_end      = jindex[iidx+1];
 132
 133         /* Get outer coordinate index */
 134         inr              = iinr[iidx];
 135         i_coord_offset   = DIM*inr;
 136
 137         /* Load i particle coords and add shift vector */
 138         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 139
 140         fix0             = _mm_setzero_ps();
 141         fiy0             = _mm_setzero_ps();
 142         fiz0             = _mm_setzero_ps();
 143
 144         /* Load parameters for i particles */
 145         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 146         isai0            = _mm_load1_ps(invsqrta+inr+0);
 147
 148         /* Reset potential sums */
 149         velecsum         = _mm_setzero_ps();
 150         vgbsum           = _mm_setzero_ps();
 151         dvdasum          = _mm_setzero_ps();
 152
 153         /* Start inner kernel loop */
 154         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 155         {
 156
 157             /* Get j neighbor index, and coordinate index */
 158             jnrA             = jjnr[jidx];
 159             jnrB             = jjnr[jidx+1];
 160             jnrC             = jjnr[jidx+2];
 161             jnrD             = jjnr[jidx+3];
 162             j_coord_offsetA  = DIM*jnrA;
 163             j_coord_offsetB  = DIM*jnrB;
 164             j_coord_offsetC  = DIM*jnrC;
 165             j_coord_offsetD  = DIM*jnrD;
 166
 167             /* load j atom coordinates */
 168             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 169                                               x+j_coord_offsetC,x+j_coord_offsetD,
 170                                               &jx0,&jy0,&jz0);
 171
 172             /* Calculate displacement vector */
 173             dx00             = _mm_sub_ps(ix0,jx0);
 174             dy00             = _mm_sub_ps(iy0,jy0);
 175             dz00             = _mm_sub_ps(iz0,jz0);
 176
 177             /* Calculate squared distance and things based on it */
 178             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 179
 180             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 181
 182             /* Load parameters for j particles */
 183             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 184                                                               charge+jnrC+0,charge+jnrD+0);
 185             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 186                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 187
 188             /**************************
 189              * CALCULATE INTERACTIONS *
 190              **************************/
 191
 192             r00              = _mm_mul_ps(rsq00,rinv00);
 193
 194             /* Compute parameters for interactions between i and j atoms */
 195             qq00             = _mm_mul_ps(iq0,jq0);
 196
 197             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 198             isaprod          = _mm_mul_ps(isai0,isaj0);
 199             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 200             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 201
 202             /* Calculate generalized born table index - this is a separate table from the normal one,
 203              * but we use the same procedure by multiplying r with scale and truncating to integer.
 204              */
 205             rt               = _mm_mul_ps(r00,gbscale);
 206             gbitab           = _mm_cvttps_epi32(rt);
 207 #ifdef __XOP__
 208             gbeps            = _mm_frcz_ps(rt);
 209 #else
 210             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 211 #endif
 212             gbitab           = _mm_slli_epi32(gbitab,2);
 213
 214             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 215             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 216             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 217             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 218             _MM_TRANSPOSE4_PS(Y,F,G,H);
 219             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 220             VV               = _mm_macc_ps(gbeps,Fp,Y);
 221             vgb              = _mm_mul_ps(gbqqfactor,VV);
 222
 223             twogbeps         = _mm_add_ps(gbeps,gbeps);
 224             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 225             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 226             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 227             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 228             fjptrA           = dvda+jnrA;
 229             fjptrB           = dvda+jnrB;
 230             fjptrC           = dvda+jnrC;
 231             fjptrD           = dvda+jnrD;
 232             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 233             velec            = _mm_mul_ps(qq00,rinv00);
 234             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 235
 236             /* Update potential sum for this i atom from the interaction with this j atom. */
 237             velecsum         = _mm_add_ps(velecsum,velec);
 238             vgbsum           = _mm_add_ps(vgbsum,vgb);
 239
 240             fscal            = felec;
 241
 242              /* Update vectorial force */
 243             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 244             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 245             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 246
 247             fjptrA             = f+j_coord_offsetA;
 248             fjptrB             = f+j_coord_offsetB;
 249             fjptrC             = f+j_coord_offsetC;
 250             fjptrD             = f+j_coord_offsetD;
 251             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 252                                                    _mm_mul_ps(dx00,fscal),
 253                                                    _mm_mul_ps(dy00,fscal),
 254                                                    _mm_mul_ps(dz00,fscal));
 255
 256             /* Inner loop uses 61 flops */
 257         }
 258
 259         if(jidx<j_index_end)
 260         {
 261
 262             /* Get j neighbor index, and coordinate index */
 263             jnrlistA         = jjnr[jidx];
 264             jnrlistB         = jjnr[jidx+1];
 265             jnrlistC         = jjnr[jidx+2];
 266             jnrlistD         = jjnr[jidx+3];
 267             /* Sign of each element will be negative for non-real atoms.
 268              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 269              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 270              */
 271             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 272             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 273             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 274             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 275             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 276             j_coord_offsetA  = DIM*jnrA;
 277             j_coord_offsetB  = DIM*jnrB;
 278             j_coord_offsetC  = DIM*jnrC;
 279             j_coord_offsetD  = DIM*jnrD;
 280
 281             /* load j atom coordinates */
 282             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 283                                               x+j_coord_offsetC,x+j_coord_offsetD,
 284                                               &jx0,&jy0,&jz0);
 285
 286             /* Calculate displacement vector */
 287             dx00             = _mm_sub_ps(ix0,jx0);
 288             dy00             = _mm_sub_ps(iy0,jy0);
 289             dz00             = _mm_sub_ps(iz0,jz0);
 290
 291             /* Calculate squared distance and things based on it */
 292             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 293
 294             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 295
 296             /* Load parameters for j particles */
 297             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 298                                                               charge+jnrC+0,charge+jnrD+0);
 299             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 300                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 301
 302             /**************************
 303              * CALCULATE INTERACTIONS *
 304              **************************/
 305
 306             r00              = _mm_mul_ps(rsq00,rinv00);
 307             r00              = _mm_andnot_ps(dummy_mask,r00);
 308
 309             /* Compute parameters for interactions between i and j atoms */
 310             qq00             = _mm_mul_ps(iq0,jq0);
 311
 312             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 313             isaprod          = _mm_mul_ps(isai0,isaj0);
 314             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 315             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 316
 317             /* Calculate generalized born table index - this is a separate table from the normal one,
 318              * but we use the same procedure by multiplying r with scale and truncating to integer.
 319              */
 320             rt               = _mm_mul_ps(r00,gbscale);
 321             gbitab           = _mm_cvttps_epi32(rt);
 322 #ifdef __XOP__
 323             gbeps            = _mm_frcz_ps(rt);
 324 #else
 325             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 326 #endif
 327             gbitab           = _mm_slli_epi32(gbitab,2);
 328
 329             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 330             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 331             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 332             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 333             _MM_TRANSPOSE4_PS(Y,F,G,H);
 334             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 335             VV               = _mm_macc_ps(gbeps,Fp,Y);
 336             vgb              = _mm_mul_ps(gbqqfactor,VV);
 337
 338             twogbeps         = _mm_add_ps(gbeps,gbeps);
 339             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 340             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 341             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 342             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 343             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 344             /* 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. */
 345             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 346             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 347             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 348             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 349             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 350             velec            = _mm_mul_ps(qq00,rinv00);
 351             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 352
 353             /* Update potential sum for this i atom from the interaction with this j atom. */
 354             velec            = _mm_andnot_ps(dummy_mask,velec);
 355             velecsum         = _mm_add_ps(velecsum,velec);
 356             vgb              = _mm_andnot_ps(dummy_mask,vgb);
 357             vgbsum           = _mm_add_ps(vgbsum,vgb);
 358
 359             fscal            = felec;
 360
 361             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 362
 363              /* Update vectorial force */
 364             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 365             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 366             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 367
 368             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 369             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 370             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 371             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 372             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 373                                                    _mm_mul_ps(dx00,fscal),
 374                                                    _mm_mul_ps(dy00,fscal),
 375                                                    _mm_mul_ps(dz00,fscal));
 376
 377             /* Inner loop uses 62 flops */
 378         }
 379
 380         /* End of innermost loop */
 381
 382         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 383                                               f+i_coord_offset,fshift+i_shift_offset);
 384
 385         ggid                        = gid[iidx];
 386         /* Update potential energies */
 387         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 388         gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
 389         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 390         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 391
 392         /* Increment number of inner iterations */
 393         inneriter                  += j_index_end - j_index_start;
 394
 395         /* Outer loop uses 9 flops */
 396     }
 397
 398     /* Increment number of outer iterations */
 399     outeriter        += nri;
 400
 401     /* Update outer/inner flops */
 402
 403     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*62);
 404 }
 405 /*
 406  * Gromacs nonbonded kernel:   nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 407  * Electrostatics interaction: GeneralizedBorn
 408  * VdW interaction:            None
 409  * Geometry:                   Particle-Particle
 410  * Calculate force/pot:        Force
 411  */
 412 void
 413 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_128_fma_single
 414                     (t_nblist * gmx_restrict                nlist,
 415                      rvec * gmx_restrict                    xx,
 416                      rvec * gmx_restrict                    ff,
 417                      t_forcerec * gmx_restrict              fr,
 418                      t_mdatoms * gmx_restrict               mdatoms,
 419                      nb_kernel_data_t * gmx_restrict        kernel_data,
 420                      t_nrnb * gmx_restrict                  nrnb)
 421 {
 422     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 423      * just 0 for non-waters.
 424      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 425      * jnr indices corresponding to data put in the four positions in the SIMD register.
 426      */
 427     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 428     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 429     int              jnrA,jnrB,jnrC,jnrD;
 430     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 431     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 432     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 433     real             rcutoff_scalar;
 434     real             *shiftvec,*fshift,*x,*f;
 435     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 436     real             scratch[4*DIM];
 437     __m128           fscal,rcutoff,rcutoff2,jidxall;
 438     int              vdwioffset0;
 439     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 440     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 441     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 442     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 443     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 444     real             *charge;
 445     __m128i          gbitab;
 446     __m128           vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,twogbeps,dvdatmp;
 447     __m128           minushalf = _mm_set1_ps(-0.5);
 448     real             *invsqrta,*dvda,*gbtab;
 449     __m128i          vfitab;
 450     __m128i          ifour       = _mm_set1_epi32(4);
 451     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 452     real             *vftab;
 453     __m128           dummy_mask,cutoff_mask;
 454     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 455     __m128           one     = _mm_set1_ps(1.0);
 456     __m128           two     = _mm_set1_ps(2.0);
 457     x                = xx[0];
 458     f                = ff[0];
 459
 460     nri              = nlist->nri;
 461     iinr             = nlist->iinr;
 462     jindex           = nlist->jindex;
 463     jjnr             = nlist->jjnr;
 464     shiftidx         = nlist->shift;
 465     gid              = nlist->gid;
 466     shiftvec         = fr->shift_vec[0];
 467     fshift           = fr->fshift[0];
 468     facel            = _mm_set1_ps(fr->epsfac);
 469     charge           = mdatoms->chargeA;
 470
 471     invsqrta         = fr->invsqrta;
 472     dvda             = fr->dvda;
 473     gbtabscale       = _mm_set1_ps(fr->gbtab.scale);
 474     gbtab            = fr->gbtab.data;
 475     gbinvepsdiff     = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
 476
 477     /* Avoid stupid compiler warnings */
 478     jnrA = jnrB = jnrC = jnrD = 0;
 479     j_coord_offsetA = 0;
 480     j_coord_offsetB = 0;
 481     j_coord_offsetC = 0;
 482     j_coord_offsetD = 0;
 483
 484     outeriter        = 0;
 485     inneriter        = 0;
 486
 487     for(iidx=0;iidx<4*DIM;iidx++)
 488     {
 489         scratch[iidx] = 0.0;
 490     }
 491
 492     /* Start outer loop over neighborlists */
 493     for(iidx=0; iidx<nri; iidx++)
 494     {
 495         /* Load shift vector for this list */
 496         i_shift_offset   = DIM*shiftidx[iidx];
 497
 498         /* Load limits for loop over neighbors */
 499         j_index_start    = jindex[iidx];
 500         j_index_end      = jindex[iidx+1];
 501
 502         /* Get outer coordinate index */
 503         inr              = iinr[iidx];
 504         i_coord_offset   = DIM*inr;
 505
 506         /* Load i particle coords and add shift vector */
 507         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 508
 509         fix0             = _mm_setzero_ps();
 510         fiy0             = _mm_setzero_ps();
 511         fiz0             = _mm_setzero_ps();
 512
 513         /* Load parameters for i particles */
 514         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 515         isai0            = _mm_load1_ps(invsqrta+inr+0);
 516
 517         dvdasum          = _mm_setzero_ps();
 518
 519         /* Start inner kernel loop */
 520         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 521         {
 522
 523             /* Get j neighbor index, and coordinate index */
 524             jnrA             = jjnr[jidx];
 525             jnrB             = jjnr[jidx+1];
 526             jnrC             = jjnr[jidx+2];
 527             jnrD             = jjnr[jidx+3];
 528             j_coord_offsetA  = DIM*jnrA;
 529             j_coord_offsetB  = DIM*jnrB;
 530             j_coord_offsetC  = DIM*jnrC;
 531             j_coord_offsetD  = DIM*jnrD;
 532
 533             /* load j atom coordinates */
 534             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 535                                               x+j_coord_offsetC,x+j_coord_offsetD,
 536                                               &jx0,&jy0,&jz0);
 537
 538             /* Calculate displacement vector */
 539             dx00             = _mm_sub_ps(ix0,jx0);
 540             dy00             = _mm_sub_ps(iy0,jy0);
 541             dz00             = _mm_sub_ps(iz0,jz0);
 542
 543             /* Calculate squared distance and things based on it */
 544             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 545
 546             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 547
 548             /* Load parameters for j particles */
 549             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 550                                                               charge+jnrC+0,charge+jnrD+0);
 551             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 552                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 553
 554             /**************************
 555              * CALCULATE INTERACTIONS *
 556              **************************/
 557
 558             r00              = _mm_mul_ps(rsq00,rinv00);
 559
 560             /* Compute parameters for interactions between i and j atoms */
 561             qq00             = _mm_mul_ps(iq0,jq0);
 562
 563             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 564             isaprod          = _mm_mul_ps(isai0,isaj0);
 565             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 566             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 567
 568             /* Calculate generalized born table index - this is a separate table from the normal one,
 569              * but we use the same procedure by multiplying r with scale and truncating to integer.
 570              */
 571             rt               = _mm_mul_ps(r00,gbscale);
 572             gbitab           = _mm_cvttps_epi32(rt);
 573 #ifdef __XOP__
 574             gbeps            = _mm_frcz_ps(rt);
 575 #else
 576             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 577 #endif
 578             gbitab           = _mm_slli_epi32(gbitab,2);
 579
 580             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 581             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 582             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 583             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 584             _MM_TRANSPOSE4_PS(Y,F,G,H);
 585             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 586             VV               = _mm_macc_ps(gbeps,Fp,Y);
 587             vgb              = _mm_mul_ps(gbqqfactor,VV);
 588
 589             twogbeps         = _mm_add_ps(gbeps,gbeps);
 590             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 591             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 592             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 593             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 594             fjptrA           = dvda+jnrA;
 595             fjptrB           = dvda+jnrB;
 596             fjptrC           = dvda+jnrC;
 597             fjptrD           = dvda+jnrD;
 598             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 599             velec            = _mm_mul_ps(qq00,rinv00);
 600             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 601
 602             fscal            = felec;
 603
 604              /* Update vectorial force */
 605             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 606             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 607             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 608
 609             fjptrA             = f+j_coord_offsetA;
 610             fjptrB             = f+j_coord_offsetB;
 611             fjptrC             = f+j_coord_offsetC;
 612             fjptrD             = f+j_coord_offsetD;
 613             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 614                                                    _mm_mul_ps(dx00,fscal),
 615                                                    _mm_mul_ps(dy00,fscal),
 616                                                    _mm_mul_ps(dz00,fscal));
 617
 618             /* Inner loop uses 59 flops */
 619         }
 620
 621         if(jidx<j_index_end)
 622         {
 623
 624             /* Get j neighbor index, and coordinate index */
 625             jnrlistA         = jjnr[jidx];
 626             jnrlistB         = jjnr[jidx+1];
 627             jnrlistC         = jjnr[jidx+2];
 628             jnrlistD         = jjnr[jidx+3];
 629             /* Sign of each element will be negative for non-real atoms.
 630              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 631              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 632              */
 633             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 634             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 635             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 636             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 637             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 638             j_coord_offsetA  = DIM*jnrA;
 639             j_coord_offsetB  = DIM*jnrB;
 640             j_coord_offsetC  = DIM*jnrC;
 641             j_coord_offsetD  = DIM*jnrD;
 642
 643             /* load j atom coordinates */
 644             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 645                                               x+j_coord_offsetC,x+j_coord_offsetD,
 646                                               &jx0,&jy0,&jz0);
 647
 648             /* Calculate displacement vector */
 649             dx00             = _mm_sub_ps(ix0,jx0);
 650             dy00             = _mm_sub_ps(iy0,jy0);
 651             dz00             = _mm_sub_ps(iz0,jz0);
 652
 653             /* Calculate squared distance and things based on it */
 654             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 655
 656             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 657
 658             /* Load parameters for j particles */
 659             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 660                                                               charge+jnrC+0,charge+jnrD+0);
 661             isaj0            = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
 662                                                               invsqrta+jnrC+0,invsqrta+jnrD+0);
 663
 664             /**************************
 665              * CALCULATE INTERACTIONS *
 666              **************************/
 667
 668             r00              = _mm_mul_ps(rsq00,rinv00);
 669             r00              = _mm_andnot_ps(dummy_mask,r00);
 670
 671             /* Compute parameters for interactions between i and j atoms */
 672             qq00             = _mm_mul_ps(iq0,jq0);
 673
 674             /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
 675             isaprod          = _mm_mul_ps(isai0,isaj0);
 676             gbqqfactor       = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
 677             gbscale          = _mm_mul_ps(isaprod,gbtabscale);
 678
 679             /* Calculate generalized born table index - this is a separate table from the normal one,
 680              * but we use the same procedure by multiplying r with scale and truncating to integer.
 681              */
 682             rt               = _mm_mul_ps(r00,gbscale);
 683             gbitab           = _mm_cvttps_epi32(rt);
 684 #ifdef __XOP__
 685             gbeps            = _mm_frcz_ps(rt);
 686 #else
 687             gbeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 688 #endif
 689             gbitab           = _mm_slli_epi32(gbitab,2);
 690
 691             Y                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,0) );
 692             F                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,1) );
 693             G                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,2) );
 694             H                = _mm_load_ps( gbtab + _mm_extract_epi32(gbitab,3) );
 695             _MM_TRANSPOSE4_PS(Y,F,G,H);
 696             Fp               = _mm_macc_ps(gbeps,_mm_macc_ps(gbeps,H,G),F);
 697             VV               = _mm_macc_ps(gbeps,Fp,Y);
 698             vgb              = _mm_mul_ps(gbqqfactor,VV);
 699
 700             twogbeps         = _mm_add_ps(gbeps,gbeps);
 701             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
 702             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
 703             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
 704             dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
 705             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
 706             /* 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. */
 707             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
 708             fjptrB             = (jnrlistB>=0) ? dvda+jnrB : scratch;
 709             fjptrC             = (jnrlistC>=0) ? dvda+jnrC : scratch;
 710             fjptrD             = (jnrlistD>=0) ? dvda+jnrD : scratch;
 711             gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
 712             velec            = _mm_mul_ps(qq00,rinv00);
 713             felec            = _mm_mul_ps(_mm_msub_ps(velec,rinv00,fgb),rinv00);
 714
 715             fscal            = felec;
 716
 717             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 718
 719              /* Update vectorial force */
 720             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 721             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 722             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 723
 724             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 725             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 726             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 727             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 728             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 729                                                    _mm_mul_ps(dx00,fscal),
 730                                                    _mm_mul_ps(dy00,fscal),
 731                                                    _mm_mul_ps(dz00,fscal));
 732
 733             /* Inner loop uses 60 flops */
 734         }
 735
 736         /* End of innermost loop */
 737
 738         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 739                                               f+i_coord_offset,fshift+i_shift_offset);
 740
 741         dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
 742         gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
 743
 744         /* Increment number of inner iterations */
 745         inneriter                  += j_index_end - j_index_start;
 746
 747         /* Outer loop uses 7 flops */
 748     }
 749
 750     /* Increment number of outer iterations */
 751     outeriter        += nri;
 752
 753     /* Update outer/inner flops */
 754
 755     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*60);
 756 }