src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecEwSw_VdwNone_GeomP1P1_sse2_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse2_single kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse2_single.h"
  50 #include "kernelutil_x86_sse2_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_sse2_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            None
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_sse2_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128i          ewitab;
  93     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  94     real             *ewtab;
  95     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  96     real             rswitch_scalar,d_scalar;
  97     __m128           dummy_mask,cutoff_mask;
  98     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  99     __m128           one     = _mm_set1_ps(1.0);
 100     __m128           two     = _mm_set1_ps(2.0);
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     facel            = _mm_set1_ps(fr->epsfac);
 113     charge           = mdatoms->chargeA;
 114
 115     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 116     ewtab            = fr->ic->tabq_coul_FDV0;
 117     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 118     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 119
 120     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 121     rcutoff_scalar   = fr->rcoulomb;
 122     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 123     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 124
 125     rswitch_scalar   = fr->rcoulomb_switch;
 126     rswitch          = _mm_set1_ps(rswitch_scalar);
 127     /* Setup switch parameters */
 128     d_scalar         = rcutoff_scalar-rswitch_scalar;
 129     d                = _mm_set1_ps(d_scalar);
 130     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 131     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 132     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 133     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 134     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 135     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 136
 137     /* Avoid stupid compiler warnings */
 138     jnrA = jnrB = jnrC = jnrD = 0;
 139     j_coord_offsetA = 0;
 140     j_coord_offsetB = 0;
 141     j_coord_offsetC = 0;
 142     j_coord_offsetD = 0;
 143
 144     outeriter        = 0;
 145     inneriter        = 0;
 146
 147     for(iidx=0;iidx<4*DIM;iidx++)
 148     {
 149         scratch[iidx] = 0.0;
 150     }
 151
 152     /* Start outer loop over neighborlists */
 153     for(iidx=0; iidx<nri; iidx++)
 154     {
 155         /* Load shift vector for this list */
 156         i_shift_offset   = DIM*shiftidx[iidx];
 157
 158         /* Load limits for loop over neighbors */
 159         j_index_start    = jindex[iidx];
 160         j_index_end      = jindex[iidx+1];
 161
 162         /* Get outer coordinate index */
 163         inr              = iinr[iidx];
 164         i_coord_offset   = DIM*inr;
 165
 166         /* Load i particle coords and add shift vector */
 167         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 168
 169         fix0             = _mm_setzero_ps();
 170         fiy0             = _mm_setzero_ps();
 171         fiz0             = _mm_setzero_ps();
 172
 173         /* Load parameters for i particles */
 174         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 175
 176         /* Reset potential sums */
 177         velecsum         = _mm_setzero_ps();
 178
 179         /* Start inner kernel loop */
 180         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 181         {
 182
 183             /* Get j neighbor index, and coordinate index */
 184             jnrA             = jjnr[jidx];
 185             jnrB             = jjnr[jidx+1];
 186             jnrC             = jjnr[jidx+2];
 187             jnrD             = jjnr[jidx+3];
 188             j_coord_offsetA  = DIM*jnrA;
 189             j_coord_offsetB  = DIM*jnrB;
 190             j_coord_offsetC  = DIM*jnrC;
 191             j_coord_offsetD  = DIM*jnrD;
 192
 193             /* load j atom coordinates */
 194             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 195                                               x+j_coord_offsetC,x+j_coord_offsetD,
 196                                               &jx0,&jy0,&jz0);
 197
 198             /* Calculate displacement vector */
 199             dx00             = _mm_sub_ps(ix0,jx0);
 200             dy00             = _mm_sub_ps(iy0,jy0);
 201             dz00             = _mm_sub_ps(iz0,jz0);
 202
 203             /* Calculate squared distance and things based on it */
 204             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 205
 206             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 207
 208             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 209
 210             /* Load parameters for j particles */
 211             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 212                                                               charge+jnrC+0,charge+jnrD+0);
 213
 214             /**************************
 215              * CALCULATE INTERACTIONS *
 216              **************************/
 217
 218             if (gmx_mm_any_lt(rsq00,rcutoff2))
 219             {
 220
 221             r00              = _mm_mul_ps(rsq00,rinv00);
 222
 223             /* Compute parameters for interactions between i and j atoms */
 224             qq00             = _mm_mul_ps(iq0,jq0);
 225
 226             /* EWALD ELECTROSTATICS */
 227
 228             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 229             ewrt             = _mm_mul_ps(r00,ewtabscale);
 230             ewitab           = _mm_cvttps_epi32(ewrt);
 231             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 232             ewitab           = _mm_slli_epi32(ewitab,2);
 233             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 234             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 235             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 236             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 237             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 238             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 239             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 240             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 241             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 242
 243             d                = _mm_sub_ps(r00,rswitch);
 244             d                = _mm_max_ps(d,_mm_setzero_ps());
 245             d2               = _mm_mul_ps(d,d);
 246             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 247
 248             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 249
 250             /* Evaluate switch function */
 251             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 252             felec            = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 253             velec            = _mm_mul_ps(velec,sw);
 254             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 255
 256             /* Update potential sum for this i atom from the interaction with this j atom. */
 257             velec            = _mm_and_ps(velec,cutoff_mask);
 258             velecsum         = _mm_add_ps(velecsum,velec);
 259
 260             fscal            = felec;
 261
 262             fscal            = _mm_and_ps(fscal,cutoff_mask);
 263
 264             /* Calculate temporary vectorial force */
 265             tx               = _mm_mul_ps(fscal,dx00);
 266             ty               = _mm_mul_ps(fscal,dy00);
 267             tz               = _mm_mul_ps(fscal,dz00);
 268
 269             /* Update vectorial force */
 270             fix0             = _mm_add_ps(fix0,tx);
 271             fiy0             = _mm_add_ps(fiy0,ty);
 272             fiz0             = _mm_add_ps(fiz0,tz);
 273
 274             fjptrA             = f+j_coord_offsetA;
 275             fjptrB             = f+j_coord_offsetB;
 276             fjptrC             = f+j_coord_offsetC;
 277             fjptrD             = f+j_coord_offsetD;
 278             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 279
 280             }
 281
 282             /* Inner loop uses 65 flops */
 283         }
 284
 285         if(jidx<j_index_end)
 286         {
 287
 288             /* Get j neighbor index, and coordinate index */
 289             jnrlistA         = jjnr[jidx];
 290             jnrlistB         = jjnr[jidx+1];
 291             jnrlistC         = jjnr[jidx+2];
 292             jnrlistD         = jjnr[jidx+3];
 293             /* Sign of each element will be negative for non-real atoms.
 294              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 295              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 296              */
 297             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 298             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 299             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 300             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 301             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 302             j_coord_offsetA  = DIM*jnrA;
 303             j_coord_offsetB  = DIM*jnrB;
 304             j_coord_offsetC  = DIM*jnrC;
 305             j_coord_offsetD  = DIM*jnrD;
 306
 307             /* load j atom coordinates */
 308             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 309                                               x+j_coord_offsetC,x+j_coord_offsetD,
 310                                               &jx0,&jy0,&jz0);
 311
 312             /* Calculate displacement vector */
 313             dx00             = _mm_sub_ps(ix0,jx0);
 314             dy00             = _mm_sub_ps(iy0,jy0);
 315             dz00             = _mm_sub_ps(iz0,jz0);
 316
 317             /* Calculate squared distance and things based on it */
 318             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 319
 320             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 321
 322             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 323
 324             /* Load parameters for j particles */
 325             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 326                                                               charge+jnrC+0,charge+jnrD+0);
 327
 328             /**************************
 329              * CALCULATE INTERACTIONS *
 330              **************************/
 331
 332             if (gmx_mm_any_lt(rsq00,rcutoff2))
 333             {
 334
 335             r00              = _mm_mul_ps(rsq00,rinv00);
 336             r00              = _mm_andnot_ps(dummy_mask,r00);
 337
 338             /* Compute parameters for interactions between i and j atoms */
 339             qq00             = _mm_mul_ps(iq0,jq0);
 340
 341             /* EWALD ELECTROSTATICS */
 342
 343             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 344             ewrt             = _mm_mul_ps(r00,ewtabscale);
 345             ewitab           = _mm_cvttps_epi32(ewrt);
 346             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 347             ewitab           = _mm_slli_epi32(ewitab,2);
 348             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 349             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 350             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 351             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 352             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 353             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 354             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 355             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 356             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 357
 358             d                = _mm_sub_ps(r00,rswitch);
 359             d                = _mm_max_ps(d,_mm_setzero_ps());
 360             d2               = _mm_mul_ps(d,d);
 361             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 362
 363             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 364
 365             /* Evaluate switch function */
 366             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 367             felec            = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 368             velec            = _mm_mul_ps(velec,sw);
 369             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velec            = _mm_and_ps(velec,cutoff_mask);
 373             velec            = _mm_andnot_ps(dummy_mask,velec);
 374             velecsum         = _mm_add_ps(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378             fscal            = _mm_and_ps(fscal,cutoff_mask);
 379
 380             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 381
 382             /* Calculate temporary vectorial force */
 383             tx               = _mm_mul_ps(fscal,dx00);
 384             ty               = _mm_mul_ps(fscal,dy00);
 385             tz               = _mm_mul_ps(fscal,dz00);
 386
 387             /* Update vectorial force */
 388             fix0             = _mm_add_ps(fix0,tx);
 389             fiy0             = _mm_add_ps(fiy0,ty);
 390             fiz0             = _mm_add_ps(fiz0,tz);
 391
 392             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 393             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 394             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 395             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 396             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 397
 398             }
 399
 400             /* Inner loop uses 66 flops */
 401         }
 402
 403         /* End of innermost loop */
 404
 405         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 406                                               f+i_coord_offset,fshift+i_shift_offset);
 407
 408         ggid                        = gid[iidx];
 409         /* Update potential energies */
 410         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 411
 412         /* Increment number of inner iterations */
 413         inneriter                  += j_index_end - j_index_start;
 414
 415         /* Outer loop uses 8 flops */
 416     }
 417
 418     /* Increment number of outer iterations */
 419     outeriter        += nri;
 420
 421     /* Update outer/inner flops */
 422
 423     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*66);
 424 }
 425 /*
 426  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_sse2_single
 427  * Electrostatics interaction: Ewald
 428  * VdW interaction:            None
 429  * Geometry:                   Particle-Particle
 430  * Calculate force/pot:        Force
 431  */
 432 void
 433 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_sse2_single
 434                     (t_nblist                    * gmx_restrict       nlist,
 435                      rvec                        * gmx_restrict          xx,
 436                      rvec                        * gmx_restrict          ff,
 437                      t_forcerec                  * gmx_restrict          fr,
 438                      t_mdatoms                   * gmx_restrict     mdatoms,
 439                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 440                      t_nrnb                      * gmx_restrict        nrnb)
 441 {
 442     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 443      * just 0 for non-waters.
 444      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 445      * jnr indices corresponding to data put in the four positions in the SIMD register.
 446      */
 447     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 448     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 449     int              jnrA,jnrB,jnrC,jnrD;
 450     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 451     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 452     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 453     real             rcutoff_scalar;
 454     real             *shiftvec,*fshift,*x,*f;
 455     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 456     real             scratch[4*DIM];
 457     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 458     int              vdwioffset0;
 459     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 460     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 461     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 462     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 463     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 464     real             *charge;
 465     __m128i          ewitab;
 466     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 467     real             *ewtab;
 468     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 469     real             rswitch_scalar,d_scalar;
 470     __m128           dummy_mask,cutoff_mask;
 471     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 472     __m128           one     = _mm_set1_ps(1.0);
 473     __m128           two     = _mm_set1_ps(2.0);
 474     x                = xx[0];
 475     f                = ff[0];
 476
 477     nri              = nlist->nri;
 478     iinr             = nlist->iinr;
 479     jindex           = nlist->jindex;
 480     jjnr             = nlist->jjnr;
 481     shiftidx         = nlist->shift;
 482     gid              = nlist->gid;
 483     shiftvec         = fr->shift_vec[0];
 484     fshift           = fr->fshift[0];
 485     facel            = _mm_set1_ps(fr->epsfac);
 486     charge           = mdatoms->chargeA;
 487
 488     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 489     ewtab            = fr->ic->tabq_coul_FDV0;
 490     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 491     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 492
 493     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 494     rcutoff_scalar   = fr->rcoulomb;
 495     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 496     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 497
 498     rswitch_scalar   = fr->rcoulomb_switch;
 499     rswitch          = _mm_set1_ps(rswitch_scalar);
 500     /* Setup switch parameters */
 501     d_scalar         = rcutoff_scalar-rswitch_scalar;
 502     d                = _mm_set1_ps(d_scalar);
 503     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 504     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 505     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 506     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 507     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 508     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 509
 510     /* Avoid stupid compiler warnings */
 511     jnrA = jnrB = jnrC = jnrD = 0;
 512     j_coord_offsetA = 0;
 513     j_coord_offsetB = 0;
 514     j_coord_offsetC = 0;
 515     j_coord_offsetD = 0;
 516
 517     outeriter        = 0;
 518     inneriter        = 0;
 519
 520     for(iidx=0;iidx<4*DIM;iidx++)
 521     {
 522         scratch[iidx] = 0.0;
 523     }
 524
 525     /* Start outer loop over neighborlists */
 526     for(iidx=0; iidx<nri; iidx++)
 527     {
 528         /* Load shift vector for this list */
 529         i_shift_offset   = DIM*shiftidx[iidx];
 530
 531         /* Load limits for loop over neighbors */
 532         j_index_start    = jindex[iidx];
 533         j_index_end      = jindex[iidx+1];
 534
 535         /* Get outer coordinate index */
 536         inr              = iinr[iidx];
 537         i_coord_offset   = DIM*inr;
 538
 539         /* Load i particle coords and add shift vector */
 540         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 541
 542         fix0             = _mm_setzero_ps();
 543         fiy0             = _mm_setzero_ps();
 544         fiz0             = _mm_setzero_ps();
 545
 546         /* Load parameters for i particles */
 547         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 548
 549         /* Start inner kernel loop */
 550         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 551         {
 552
 553             /* Get j neighbor index, and coordinate index */
 554             jnrA             = jjnr[jidx];
 555             jnrB             = jjnr[jidx+1];
 556             jnrC             = jjnr[jidx+2];
 557             jnrD             = jjnr[jidx+3];
 558             j_coord_offsetA  = DIM*jnrA;
 559             j_coord_offsetB  = DIM*jnrB;
 560             j_coord_offsetC  = DIM*jnrC;
 561             j_coord_offsetD  = DIM*jnrD;
 562
 563             /* load j atom coordinates */
 564             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 565                                               x+j_coord_offsetC,x+j_coord_offsetD,
 566                                               &jx0,&jy0,&jz0);
 567
 568             /* Calculate displacement vector */
 569             dx00             = _mm_sub_ps(ix0,jx0);
 570             dy00             = _mm_sub_ps(iy0,jy0);
 571             dz00             = _mm_sub_ps(iz0,jz0);
 572
 573             /* Calculate squared distance and things based on it */
 574             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 575
 576             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 577
 578             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 579
 580             /* Load parameters for j particles */
 581             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 582                                                               charge+jnrC+0,charge+jnrD+0);
 583
 584             /**************************
 585              * CALCULATE INTERACTIONS *
 586              **************************/
 587
 588             if (gmx_mm_any_lt(rsq00,rcutoff2))
 589             {
 590
 591             r00              = _mm_mul_ps(rsq00,rinv00);
 592
 593             /* Compute parameters for interactions between i and j atoms */
 594             qq00             = _mm_mul_ps(iq0,jq0);
 595
 596             /* EWALD ELECTROSTATICS */
 597
 598             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 599             ewrt             = _mm_mul_ps(r00,ewtabscale);
 600             ewitab           = _mm_cvttps_epi32(ewrt);
 601             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 602             ewitab           = _mm_slli_epi32(ewitab,2);
 603             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 604             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 605             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 606             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 607             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 608             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 609             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 610             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 611             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 612
 613             d                = _mm_sub_ps(r00,rswitch);
 614             d                = _mm_max_ps(d,_mm_setzero_ps());
 615             d2               = _mm_mul_ps(d,d);
 616             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 617
 618             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 619
 620             /* Evaluate switch function */
 621             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 622             felec            = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 623             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 624
 625             fscal            = felec;
 626
 627             fscal            = _mm_and_ps(fscal,cutoff_mask);
 628
 629             /* Calculate temporary vectorial force */
 630             tx               = _mm_mul_ps(fscal,dx00);
 631             ty               = _mm_mul_ps(fscal,dy00);
 632             tz               = _mm_mul_ps(fscal,dz00);
 633
 634             /* Update vectorial force */
 635             fix0             = _mm_add_ps(fix0,tx);
 636             fiy0             = _mm_add_ps(fiy0,ty);
 637             fiz0             = _mm_add_ps(fiz0,tz);
 638
 639             fjptrA             = f+j_coord_offsetA;
 640             fjptrB             = f+j_coord_offsetB;
 641             fjptrC             = f+j_coord_offsetC;
 642             fjptrD             = f+j_coord_offsetD;
 643             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 644
 645             }
 646
 647             /* Inner loop uses 62 flops */
 648         }
 649
 650         if(jidx<j_index_end)
 651         {
 652
 653             /* Get j neighbor index, and coordinate index */
 654             jnrlistA         = jjnr[jidx];
 655             jnrlistB         = jjnr[jidx+1];
 656             jnrlistC         = jjnr[jidx+2];
 657             jnrlistD         = jjnr[jidx+3];
 658             /* Sign of each element will be negative for non-real atoms.
 659              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 660              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 661              */
 662             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 663             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 664             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 665             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 666             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 667             j_coord_offsetA  = DIM*jnrA;
 668             j_coord_offsetB  = DIM*jnrB;
 669             j_coord_offsetC  = DIM*jnrC;
 670             j_coord_offsetD  = DIM*jnrD;
 671
 672             /* load j atom coordinates */
 673             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 674                                               x+j_coord_offsetC,x+j_coord_offsetD,
 675                                               &jx0,&jy0,&jz0);
 676
 677             /* Calculate displacement vector */
 678             dx00             = _mm_sub_ps(ix0,jx0);
 679             dy00             = _mm_sub_ps(iy0,jy0);
 680             dz00             = _mm_sub_ps(iz0,jz0);
 681
 682             /* Calculate squared distance and things based on it */
 683             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 684
 685             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 686
 687             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 688
 689             /* Load parameters for j particles */
 690             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 691                                                               charge+jnrC+0,charge+jnrD+0);
 692
 693             /**************************
 694              * CALCULATE INTERACTIONS *
 695              **************************/
 696
 697             if (gmx_mm_any_lt(rsq00,rcutoff2))
 698             {
 699
 700             r00              = _mm_mul_ps(rsq00,rinv00);
 701             r00              = _mm_andnot_ps(dummy_mask,r00);
 702
 703             /* Compute parameters for interactions between i and j atoms */
 704             qq00             = _mm_mul_ps(iq0,jq0);
 705
 706             /* EWALD ELECTROSTATICS */
 707
 708             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 709             ewrt             = _mm_mul_ps(r00,ewtabscale);
 710             ewitab           = _mm_cvttps_epi32(ewrt);
 711             eweps            = _mm_sub_ps(ewrt,_mm_cvtepi32_ps(ewitab));
 712             ewitab           = _mm_slli_epi32(ewitab,2);
 713             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 714             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 715             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 716             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 717             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 718             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 719             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 720             velec            = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
 721             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 722
 723             d                = _mm_sub_ps(r00,rswitch);
 724             d                = _mm_max_ps(d,_mm_setzero_ps());
 725             d2               = _mm_mul_ps(d,d);
 726             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_add_ps(swV3,_mm_mul_ps(d,_mm_add_ps(swV4,_mm_mul_ps(d,swV5)))))));
 727
 728             dsw              = _mm_mul_ps(d2,_mm_add_ps(swF2,_mm_mul_ps(d,_mm_add_ps(swF3,_mm_mul_ps(d,swF4)))));
 729
 730             /* Evaluate switch function */
 731             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 732             felec            = _mm_sub_ps( _mm_mul_ps(felec,sw) , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 733             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 734
 735             fscal            = felec;
 736
 737             fscal            = _mm_and_ps(fscal,cutoff_mask);
 738
 739             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 740
 741             /* Calculate temporary vectorial force */
 742             tx               = _mm_mul_ps(fscal,dx00);
 743             ty               = _mm_mul_ps(fscal,dy00);
 744             tz               = _mm_mul_ps(fscal,dz00);
 745
 746             /* Update vectorial force */
 747             fix0             = _mm_add_ps(fix0,tx);
 748             fiy0             = _mm_add_ps(fiy0,ty);
 749             fiz0             = _mm_add_ps(fiz0,tz);
 750
 751             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 752             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 753             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 754             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 755             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 756
 757             }
 758
 759             /* Inner loop uses 63 flops */
 760         }
 761
 762         /* End of innermost loop */
 763
 764         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 765                                               f+i_coord_offset,fshift+i_shift_offset);
 766
 767         /* Increment number of inner iterations */
 768         inneriter                  += j_index_end - j_index_start;
 769
 770         /* Outer loop uses 7 flops */
 771     }
 772
 773     /* Increment number of outer iterations */
 774     outeriter        += nri;
 775
 776     /* Update outer/inner flops */
 777
 778     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*63);
 779 }