src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_sse2_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse2_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_sse2_single.h"
  34 #include "kernelutil_x86_sse2_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_VF_sse2_single
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_VF_sse2_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 SSE, 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              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             shX,shY,shZ,rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  73     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  75     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  76     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  77     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  78     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  79     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  80     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  81     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  83     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  84     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  85     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  86     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  87     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  94     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  95     __m128           dummy_mask,cutoff_mask;
  96     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  97     __m128           one     = _mm_set1_ps(1.0);
  98     __m128           two     = _mm_set1_ps(2.0);
  99     x                = xx[0];
 100     f                = ff[0];
 101
 102     nri              = nlist->nri;
 103     iinr             = nlist->iinr;
 104     jindex           = nlist->jindex;
 105     jjnr             = nlist->jjnr;
 106     shiftidx         = nlist->shift;
 107     gid              = nlist->gid;
 108     shiftvec         = fr->shift_vec[0];
 109     fshift           = fr->fshift[0];
 110     facel            = _mm_set1_ps(fr->epsfac);
 111     charge           = mdatoms->chargeA;
 112     nvdwtype         = fr->ntype;
 113     vdwparam         = fr->nbfp;
 114     vdwtype          = mdatoms->typeA;
 115
 116     /* Setup water-specific parameters */
 117     inr              = nlist->iinr[0];
 118     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 119     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 120     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 121     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 122
 123     jq0              = _mm_set1_ps(charge[inr+0]);
 124     jq1              = _mm_set1_ps(charge[inr+1]);
 125     jq2              = _mm_set1_ps(charge[inr+2]);
 126     vdwjidx0A        = 2*vdwtype[inr+0];
 127     qq00             = _mm_mul_ps(iq0,jq0);
 128     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 129     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 130     qq01             = _mm_mul_ps(iq0,jq1);
 131     qq02             = _mm_mul_ps(iq0,jq2);
 132     qq10             = _mm_mul_ps(iq1,jq0);
 133     qq11             = _mm_mul_ps(iq1,jq1);
 134     qq12             = _mm_mul_ps(iq1,jq2);
 135     qq20             = _mm_mul_ps(iq2,jq0);
 136     qq21             = _mm_mul_ps(iq2,jq1);
 137     qq22             = _mm_mul_ps(iq2,jq2);
 138
 139     /* Avoid stupid compiler warnings */
 140     jnrA = jnrB = jnrC = jnrD = 0;
 141     j_coord_offsetA = 0;
 142     j_coord_offsetB = 0;
 143     j_coord_offsetC = 0;
 144     j_coord_offsetD = 0;
 145
 146     outeriter        = 0;
 147     inneriter        = 0;
 148
 149     /* Start outer loop over neighborlists */
 150     for(iidx=0; iidx<nri; iidx++)
 151     {
 152         /* Load shift vector for this list */
 153         i_shift_offset   = DIM*shiftidx[iidx];
 154         shX              = shiftvec[i_shift_offset+XX];
 155         shY              = shiftvec[i_shift_offset+YY];
 156         shZ              = shiftvec[i_shift_offset+ZZ];
 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         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 168         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 169         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 170         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 171         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 172         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 173         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 174         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 175         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 176
 177         fix0             = _mm_setzero_ps();
 178         fiy0             = _mm_setzero_ps();
 179         fiz0             = _mm_setzero_ps();
 180         fix1             = _mm_setzero_ps();
 181         fiy1             = _mm_setzero_ps();
 182         fiz1             = _mm_setzero_ps();
 183         fix2             = _mm_setzero_ps();
 184         fiy2             = _mm_setzero_ps();
 185         fiz2             = _mm_setzero_ps();
 186
 187         /* Reset potential sums */
 188         velecsum         = _mm_setzero_ps();
 189         vvdwsum          = _mm_setzero_ps();
 190
 191         /* Start inner kernel loop */
 192         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 193         {
 194
 195             /* Get j neighbor index, and coordinate index */
 196             jnrA             = jjnr[jidx];
 197             jnrB             = jjnr[jidx+1];
 198             jnrC             = jjnr[jidx+2];
 199             jnrD             = jjnr[jidx+3];
 200
 201             j_coord_offsetA  = DIM*jnrA;
 202             j_coord_offsetB  = DIM*jnrB;
 203             j_coord_offsetC  = DIM*jnrC;
 204             j_coord_offsetD  = DIM*jnrD;
 205
 206             /* load j atom coordinates */
 207             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 208                                               x+j_coord_offsetC,x+j_coord_offsetD,
 209                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 210
 211             /* Calculate displacement vector */
 212             dx00             = _mm_sub_ps(ix0,jx0);
 213             dy00             = _mm_sub_ps(iy0,jy0);
 214             dz00             = _mm_sub_ps(iz0,jz0);
 215             dx01             = _mm_sub_ps(ix0,jx1);
 216             dy01             = _mm_sub_ps(iy0,jy1);
 217             dz01             = _mm_sub_ps(iz0,jz1);
 218             dx02             = _mm_sub_ps(ix0,jx2);
 219             dy02             = _mm_sub_ps(iy0,jy2);
 220             dz02             = _mm_sub_ps(iz0,jz2);
 221             dx10             = _mm_sub_ps(ix1,jx0);
 222             dy10             = _mm_sub_ps(iy1,jy0);
 223             dz10             = _mm_sub_ps(iz1,jz0);
 224             dx11             = _mm_sub_ps(ix1,jx1);
 225             dy11             = _mm_sub_ps(iy1,jy1);
 226             dz11             = _mm_sub_ps(iz1,jz1);
 227             dx12             = _mm_sub_ps(ix1,jx2);
 228             dy12             = _mm_sub_ps(iy1,jy2);
 229             dz12             = _mm_sub_ps(iz1,jz2);
 230             dx20             = _mm_sub_ps(ix2,jx0);
 231             dy20             = _mm_sub_ps(iy2,jy0);
 232             dz20             = _mm_sub_ps(iz2,jz0);
 233             dx21             = _mm_sub_ps(ix2,jx1);
 234             dy21             = _mm_sub_ps(iy2,jy1);
 235             dz21             = _mm_sub_ps(iz2,jz1);
 236             dx22             = _mm_sub_ps(ix2,jx2);
 237             dy22             = _mm_sub_ps(iy2,jy2);
 238             dz22             = _mm_sub_ps(iz2,jz2);
 239
 240             /* Calculate squared distance and things based on it */
 241             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 242             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 243             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 244             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 245             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 246             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 247             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 248             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 249             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 250
 251             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 252             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 253             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 254             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 255             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 256             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 257             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 258             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 259             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 260
 261             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 262             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 263             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 264             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 265             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 266             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 267             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 268             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 269             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 270
 271             fjx0             = _mm_setzero_ps();
 272             fjy0             = _mm_setzero_ps();
 273             fjz0             = _mm_setzero_ps();
 274             fjx1             = _mm_setzero_ps();
 275             fjy1             = _mm_setzero_ps();
 276             fjz1             = _mm_setzero_ps();
 277             fjx2             = _mm_setzero_ps();
 278             fjy2             = _mm_setzero_ps();
 279             fjz2             = _mm_setzero_ps();
 280
 281             /**************************
 282              * CALCULATE INTERACTIONS *
 283              **************************/
 284
 285             /* COULOMB ELECTROSTATICS */
 286             velec            = _mm_mul_ps(qq00,rinv00);
 287             felec            = _mm_mul_ps(velec,rinvsq00);
 288
 289             /* LENNARD-JONES DISPERSION/REPULSION */
 290
 291             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 292             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 293             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 294             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 295             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 296
 297             /* Update potential sum for this i atom from the interaction with this j atom. */
 298             velecsum         = _mm_add_ps(velecsum,velec);
 299             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 300
 301             fscal            = _mm_add_ps(felec,fvdw);
 302
 303             /* Calculate temporary vectorial force */
 304             tx               = _mm_mul_ps(fscal,dx00);
 305             ty               = _mm_mul_ps(fscal,dy00);
 306             tz               = _mm_mul_ps(fscal,dz00);
 307
 308             /* Update vectorial force */
 309             fix0             = _mm_add_ps(fix0,tx);
 310             fiy0             = _mm_add_ps(fiy0,ty);
 311             fiz0             = _mm_add_ps(fiz0,tz);
 312
 313             fjx0             = _mm_add_ps(fjx0,tx);
 314             fjy0             = _mm_add_ps(fjy0,ty);
 315             fjz0             = _mm_add_ps(fjz0,tz);
 316
 317             /**************************
 318              * CALCULATE INTERACTIONS *
 319              **************************/
 320
 321             /* COULOMB ELECTROSTATICS */
 322             velec            = _mm_mul_ps(qq01,rinv01);
 323             felec            = _mm_mul_ps(velec,rinvsq01);
 324
 325             /* Update potential sum for this i atom from the interaction with this j atom. */
 326             velecsum         = _mm_add_ps(velecsum,velec);
 327
 328             fscal            = felec;
 329
 330             /* Calculate temporary vectorial force */
 331             tx               = _mm_mul_ps(fscal,dx01);
 332             ty               = _mm_mul_ps(fscal,dy01);
 333             tz               = _mm_mul_ps(fscal,dz01);
 334
 335             /* Update vectorial force */
 336             fix0             = _mm_add_ps(fix0,tx);
 337             fiy0             = _mm_add_ps(fiy0,ty);
 338             fiz0             = _mm_add_ps(fiz0,tz);
 339
 340             fjx1             = _mm_add_ps(fjx1,tx);
 341             fjy1             = _mm_add_ps(fjy1,ty);
 342             fjz1             = _mm_add_ps(fjz1,tz);
 343
 344             /**************************
 345              * CALCULATE INTERACTIONS *
 346              **************************/
 347
 348             /* COULOMB ELECTROSTATICS */
 349             velec            = _mm_mul_ps(qq02,rinv02);
 350             felec            = _mm_mul_ps(velec,rinvsq02);
 351
 352             /* Update potential sum for this i atom from the interaction with this j atom. */
 353             velecsum         = _mm_add_ps(velecsum,velec);
 354
 355             fscal            = felec;
 356
 357             /* Calculate temporary vectorial force */
 358             tx               = _mm_mul_ps(fscal,dx02);
 359             ty               = _mm_mul_ps(fscal,dy02);
 360             tz               = _mm_mul_ps(fscal,dz02);
 361
 362             /* Update vectorial force */
 363             fix0             = _mm_add_ps(fix0,tx);
 364             fiy0             = _mm_add_ps(fiy0,ty);
 365             fiz0             = _mm_add_ps(fiz0,tz);
 366
 367             fjx2             = _mm_add_ps(fjx2,tx);
 368             fjy2             = _mm_add_ps(fjy2,ty);
 369             fjz2             = _mm_add_ps(fjz2,tz);
 370
 371             /**************************
 372              * CALCULATE INTERACTIONS *
 373              **************************/
 374
 375             /* COULOMB ELECTROSTATICS */
 376             velec            = _mm_mul_ps(qq10,rinv10);
 377             felec            = _mm_mul_ps(velec,rinvsq10);
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velecsum         = _mm_add_ps(velecsum,velec);
 381
 382             fscal            = felec;
 383
 384             /* Calculate temporary vectorial force */
 385             tx               = _mm_mul_ps(fscal,dx10);
 386             ty               = _mm_mul_ps(fscal,dy10);
 387             tz               = _mm_mul_ps(fscal,dz10);
 388
 389             /* Update vectorial force */
 390             fix1             = _mm_add_ps(fix1,tx);
 391             fiy1             = _mm_add_ps(fiy1,ty);
 392             fiz1             = _mm_add_ps(fiz1,tz);
 393
 394             fjx0             = _mm_add_ps(fjx0,tx);
 395             fjy0             = _mm_add_ps(fjy0,ty);
 396             fjz0             = _mm_add_ps(fjz0,tz);
 397
 398             /**************************
 399              * CALCULATE INTERACTIONS *
 400              **************************/
 401
 402             /* COULOMB ELECTROSTATICS */
 403             velec            = _mm_mul_ps(qq11,rinv11);
 404             felec            = _mm_mul_ps(velec,rinvsq11);
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velecsum         = _mm_add_ps(velecsum,velec);
 408
 409             fscal            = felec;
 410
 411             /* Calculate temporary vectorial force */
 412             tx               = _mm_mul_ps(fscal,dx11);
 413             ty               = _mm_mul_ps(fscal,dy11);
 414             tz               = _mm_mul_ps(fscal,dz11);
 415
 416             /* Update vectorial force */
 417             fix1             = _mm_add_ps(fix1,tx);
 418             fiy1             = _mm_add_ps(fiy1,ty);
 419             fiz1             = _mm_add_ps(fiz1,tz);
 420
 421             fjx1             = _mm_add_ps(fjx1,tx);
 422             fjy1             = _mm_add_ps(fjy1,ty);
 423             fjz1             = _mm_add_ps(fjz1,tz);
 424
 425             /**************************
 426              * CALCULATE INTERACTIONS *
 427              **************************/
 428
 429             /* COULOMB ELECTROSTATICS */
 430             velec            = _mm_mul_ps(qq12,rinv12);
 431             felec            = _mm_mul_ps(velec,rinvsq12);
 432
 433             /* Update potential sum for this i atom from the interaction with this j atom. */
 434             velecsum         = _mm_add_ps(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             /* Calculate temporary vectorial force */
 439             tx               = _mm_mul_ps(fscal,dx12);
 440             ty               = _mm_mul_ps(fscal,dy12);
 441             tz               = _mm_mul_ps(fscal,dz12);
 442
 443             /* Update vectorial force */
 444             fix1             = _mm_add_ps(fix1,tx);
 445             fiy1             = _mm_add_ps(fiy1,ty);
 446             fiz1             = _mm_add_ps(fiz1,tz);
 447
 448             fjx2             = _mm_add_ps(fjx2,tx);
 449             fjy2             = _mm_add_ps(fjy2,ty);
 450             fjz2             = _mm_add_ps(fjz2,tz);
 451
 452             /**************************
 453              * CALCULATE INTERACTIONS *
 454              **************************/
 455
 456             /* COULOMB ELECTROSTATICS */
 457             velec            = _mm_mul_ps(qq20,rinv20);
 458             felec            = _mm_mul_ps(velec,rinvsq20);
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velecsum         = _mm_add_ps(velecsum,velec);
 462
 463             fscal            = felec;
 464
 465             /* Calculate temporary vectorial force */
 466             tx               = _mm_mul_ps(fscal,dx20);
 467             ty               = _mm_mul_ps(fscal,dy20);
 468             tz               = _mm_mul_ps(fscal,dz20);
 469
 470             /* Update vectorial force */
 471             fix2             = _mm_add_ps(fix2,tx);
 472             fiy2             = _mm_add_ps(fiy2,ty);
 473             fiz2             = _mm_add_ps(fiz2,tz);
 474
 475             fjx0             = _mm_add_ps(fjx0,tx);
 476             fjy0             = _mm_add_ps(fjy0,ty);
 477             fjz0             = _mm_add_ps(fjz0,tz);
 478
 479             /**************************
 480              * CALCULATE INTERACTIONS *
 481              **************************/
 482
 483             /* COULOMB ELECTROSTATICS */
 484             velec            = _mm_mul_ps(qq21,rinv21);
 485             felec            = _mm_mul_ps(velec,rinvsq21);
 486
 487             /* Update potential sum for this i atom from the interaction with this j atom. */
 488             velecsum         = _mm_add_ps(velecsum,velec);
 489
 490             fscal            = felec;
 491
 492             /* Calculate temporary vectorial force */
 493             tx               = _mm_mul_ps(fscal,dx21);
 494             ty               = _mm_mul_ps(fscal,dy21);
 495             tz               = _mm_mul_ps(fscal,dz21);
 496
 497             /* Update vectorial force */
 498             fix2             = _mm_add_ps(fix2,tx);
 499             fiy2             = _mm_add_ps(fiy2,ty);
 500             fiz2             = _mm_add_ps(fiz2,tz);
 501
 502             fjx1             = _mm_add_ps(fjx1,tx);
 503             fjy1             = _mm_add_ps(fjy1,ty);
 504             fjz1             = _mm_add_ps(fjz1,tz);
 505
 506             /**************************
 507              * CALCULATE INTERACTIONS *
 508              **************************/
 509
 510             /* COULOMB ELECTROSTATICS */
 511             velec            = _mm_mul_ps(qq22,rinv22);
 512             felec            = _mm_mul_ps(velec,rinvsq22);
 513
 514             /* Update potential sum for this i atom from the interaction with this j atom. */
 515             velecsum         = _mm_add_ps(velecsum,velec);
 516
 517             fscal            = felec;
 518
 519             /* Calculate temporary vectorial force */
 520             tx               = _mm_mul_ps(fscal,dx22);
 521             ty               = _mm_mul_ps(fscal,dy22);
 522             tz               = _mm_mul_ps(fscal,dz22);
 523
 524             /* Update vectorial force */
 525             fix2             = _mm_add_ps(fix2,tx);
 526             fiy2             = _mm_add_ps(fiy2,ty);
 527             fiz2             = _mm_add_ps(fiz2,tz);
 528
 529             fjx2             = _mm_add_ps(fjx2,tx);
 530             fjy2             = _mm_add_ps(fjy2,ty);
 531             fjz2             = _mm_add_ps(fjz2,tz);
 532
 533             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 534                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 535                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 536
 537             /* Inner loop uses 264 flops */
 538         }
 539
 540         if(jidx<j_index_end)
 541         {
 542
 543             /* Get j neighbor index, and coordinate index */
 544             jnrA             = jjnr[jidx];
 545             jnrB             = jjnr[jidx+1];
 546             jnrC             = jjnr[jidx+2];
 547             jnrD             = jjnr[jidx+3];
 548
 549             /* Sign of each element will be negative for non-real atoms.
 550              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 551              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 552              */
 553             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 554             jnrA       = (jnrA>=0) ? jnrA : 0;
 555             jnrB       = (jnrB>=0) ? jnrB : 0;
 556             jnrC       = (jnrC>=0) ? jnrC : 0;
 557             jnrD       = (jnrD>=0) ? jnrD : 0;
 558
 559             j_coord_offsetA  = DIM*jnrA;
 560             j_coord_offsetB  = DIM*jnrB;
 561             j_coord_offsetC  = DIM*jnrC;
 562             j_coord_offsetD  = DIM*jnrD;
 563
 564             /* load j atom coordinates */
 565             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 566                                               x+j_coord_offsetC,x+j_coord_offsetD,
 567                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 568
 569             /* Calculate displacement vector */
 570             dx00             = _mm_sub_ps(ix0,jx0);
 571             dy00             = _mm_sub_ps(iy0,jy0);
 572             dz00             = _mm_sub_ps(iz0,jz0);
 573             dx01             = _mm_sub_ps(ix0,jx1);
 574             dy01             = _mm_sub_ps(iy0,jy1);
 575             dz01             = _mm_sub_ps(iz0,jz1);
 576             dx02             = _mm_sub_ps(ix0,jx2);
 577             dy02             = _mm_sub_ps(iy0,jy2);
 578             dz02             = _mm_sub_ps(iz0,jz2);
 579             dx10             = _mm_sub_ps(ix1,jx0);
 580             dy10             = _mm_sub_ps(iy1,jy0);
 581             dz10             = _mm_sub_ps(iz1,jz0);
 582             dx11             = _mm_sub_ps(ix1,jx1);
 583             dy11             = _mm_sub_ps(iy1,jy1);
 584             dz11             = _mm_sub_ps(iz1,jz1);
 585             dx12             = _mm_sub_ps(ix1,jx2);
 586             dy12             = _mm_sub_ps(iy1,jy2);
 587             dz12             = _mm_sub_ps(iz1,jz2);
 588             dx20             = _mm_sub_ps(ix2,jx0);
 589             dy20             = _mm_sub_ps(iy2,jy0);
 590             dz20             = _mm_sub_ps(iz2,jz0);
 591             dx21             = _mm_sub_ps(ix2,jx1);
 592             dy21             = _mm_sub_ps(iy2,jy1);
 593             dz21             = _mm_sub_ps(iz2,jz1);
 594             dx22             = _mm_sub_ps(ix2,jx2);
 595             dy22             = _mm_sub_ps(iy2,jy2);
 596             dz22             = _mm_sub_ps(iz2,jz2);
 597
 598             /* Calculate squared distance and things based on it */
 599             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 600             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 601             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 602             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 603             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 604             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 605             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 606             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 607             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 608
 609             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 610             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 611             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 612             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 613             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 614             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 615             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 616             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 617             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 618
 619             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 620             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 621             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 622             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 623             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 624             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 625             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 626             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 627             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 628
 629             fjx0             = _mm_setzero_ps();
 630             fjy0             = _mm_setzero_ps();
 631             fjz0             = _mm_setzero_ps();
 632             fjx1             = _mm_setzero_ps();
 633             fjy1             = _mm_setzero_ps();
 634             fjz1             = _mm_setzero_ps();
 635             fjx2             = _mm_setzero_ps();
 636             fjy2             = _mm_setzero_ps();
 637             fjz2             = _mm_setzero_ps();
 638
 639             /**************************
 640              * CALCULATE INTERACTIONS *
 641              **************************/
 642
 643             /* COULOMB ELECTROSTATICS */
 644             velec            = _mm_mul_ps(qq00,rinv00);
 645             felec            = _mm_mul_ps(velec,rinvsq00);
 646
 647             /* LENNARD-JONES DISPERSION/REPULSION */
 648
 649             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 650             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 651             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 652             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 653             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 654
 655             /* Update potential sum for this i atom from the interaction with this j atom. */
 656             velec            = _mm_andnot_ps(dummy_mask,velec);
 657             velecsum         = _mm_add_ps(velecsum,velec);
 658             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 659             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 660
 661             fscal            = _mm_add_ps(felec,fvdw);
 662
 663             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 664
 665             /* Calculate temporary vectorial force */
 666             tx               = _mm_mul_ps(fscal,dx00);
 667             ty               = _mm_mul_ps(fscal,dy00);
 668             tz               = _mm_mul_ps(fscal,dz00);
 669
 670             /* Update vectorial force */
 671             fix0             = _mm_add_ps(fix0,tx);
 672             fiy0             = _mm_add_ps(fiy0,ty);
 673             fiz0             = _mm_add_ps(fiz0,tz);
 674
 675             fjx0             = _mm_add_ps(fjx0,tx);
 676             fjy0             = _mm_add_ps(fjy0,ty);
 677             fjz0             = _mm_add_ps(fjz0,tz);
 678
 679             /**************************
 680              * CALCULATE INTERACTIONS *
 681              **************************/
 682
 683             /* COULOMB ELECTROSTATICS */
 684             velec            = _mm_mul_ps(qq01,rinv01);
 685             felec            = _mm_mul_ps(velec,rinvsq01);
 686
 687             /* Update potential sum for this i atom from the interaction with this j atom. */
 688             velec            = _mm_andnot_ps(dummy_mask,velec);
 689             velecsum         = _mm_add_ps(velecsum,velec);
 690
 691             fscal            = felec;
 692
 693             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 694
 695             /* Calculate temporary vectorial force */
 696             tx               = _mm_mul_ps(fscal,dx01);
 697             ty               = _mm_mul_ps(fscal,dy01);
 698             tz               = _mm_mul_ps(fscal,dz01);
 699
 700             /* Update vectorial force */
 701             fix0             = _mm_add_ps(fix0,tx);
 702             fiy0             = _mm_add_ps(fiy0,ty);
 703             fiz0             = _mm_add_ps(fiz0,tz);
 704
 705             fjx1             = _mm_add_ps(fjx1,tx);
 706             fjy1             = _mm_add_ps(fjy1,ty);
 707             fjz1             = _mm_add_ps(fjz1,tz);
 708
 709             /**************************
 710              * CALCULATE INTERACTIONS *
 711              **************************/
 712
 713             /* COULOMB ELECTROSTATICS */
 714             velec            = _mm_mul_ps(qq02,rinv02);
 715             felec            = _mm_mul_ps(velec,rinvsq02);
 716
 717             /* Update potential sum for this i atom from the interaction with this j atom. */
 718             velec            = _mm_andnot_ps(dummy_mask,velec);
 719             velecsum         = _mm_add_ps(velecsum,velec);
 720
 721             fscal            = felec;
 722
 723             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 724
 725             /* Calculate temporary vectorial force */
 726             tx               = _mm_mul_ps(fscal,dx02);
 727             ty               = _mm_mul_ps(fscal,dy02);
 728             tz               = _mm_mul_ps(fscal,dz02);
 729
 730             /* Update vectorial force */
 731             fix0             = _mm_add_ps(fix0,tx);
 732             fiy0             = _mm_add_ps(fiy0,ty);
 733             fiz0             = _mm_add_ps(fiz0,tz);
 734
 735             fjx2             = _mm_add_ps(fjx2,tx);
 736             fjy2             = _mm_add_ps(fjy2,ty);
 737             fjz2             = _mm_add_ps(fjz2,tz);
 738
 739             /**************************
 740              * CALCULATE INTERACTIONS *
 741              **************************/
 742
 743             /* COULOMB ELECTROSTATICS */
 744             velec            = _mm_mul_ps(qq10,rinv10);
 745             felec            = _mm_mul_ps(velec,rinvsq10);
 746
 747             /* Update potential sum for this i atom from the interaction with this j atom. */
 748             velec            = _mm_andnot_ps(dummy_mask,velec);
 749             velecsum         = _mm_add_ps(velecsum,velec);
 750
 751             fscal            = felec;
 752
 753             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 754
 755             /* Calculate temporary vectorial force */
 756             tx               = _mm_mul_ps(fscal,dx10);
 757             ty               = _mm_mul_ps(fscal,dy10);
 758             tz               = _mm_mul_ps(fscal,dz10);
 759
 760             /* Update vectorial force */
 761             fix1             = _mm_add_ps(fix1,tx);
 762             fiy1             = _mm_add_ps(fiy1,ty);
 763             fiz1             = _mm_add_ps(fiz1,tz);
 764
 765             fjx0             = _mm_add_ps(fjx0,tx);
 766             fjy0             = _mm_add_ps(fjy0,ty);
 767             fjz0             = _mm_add_ps(fjz0,tz);
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             /* COULOMB ELECTROSTATICS */
 774             velec            = _mm_mul_ps(qq11,rinv11);
 775             felec            = _mm_mul_ps(velec,rinvsq11);
 776
 777             /* Update potential sum for this i atom from the interaction with this j atom. */
 778             velec            = _mm_andnot_ps(dummy_mask,velec);
 779             velecsum         = _mm_add_ps(velecsum,velec);
 780
 781             fscal            = felec;
 782
 783             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 784
 785             /* Calculate temporary vectorial force */
 786             tx               = _mm_mul_ps(fscal,dx11);
 787             ty               = _mm_mul_ps(fscal,dy11);
 788             tz               = _mm_mul_ps(fscal,dz11);
 789
 790             /* Update vectorial force */
 791             fix1             = _mm_add_ps(fix1,tx);
 792             fiy1             = _mm_add_ps(fiy1,ty);
 793             fiz1             = _mm_add_ps(fiz1,tz);
 794
 795             fjx1             = _mm_add_ps(fjx1,tx);
 796             fjy1             = _mm_add_ps(fjy1,ty);
 797             fjz1             = _mm_add_ps(fjz1,tz);
 798
 799             /**************************
 800              * CALCULATE INTERACTIONS *
 801              **************************/
 802
 803             /* COULOMB ELECTROSTATICS */
 804             velec            = _mm_mul_ps(qq12,rinv12);
 805             felec            = _mm_mul_ps(velec,rinvsq12);
 806
 807             /* Update potential sum for this i atom from the interaction with this j atom. */
 808             velec            = _mm_andnot_ps(dummy_mask,velec);
 809             velecsum         = _mm_add_ps(velecsum,velec);
 810
 811             fscal            = felec;
 812
 813             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 814
 815             /* Calculate temporary vectorial force */
 816             tx               = _mm_mul_ps(fscal,dx12);
 817             ty               = _mm_mul_ps(fscal,dy12);
 818             tz               = _mm_mul_ps(fscal,dz12);
 819
 820             /* Update vectorial force */
 821             fix1             = _mm_add_ps(fix1,tx);
 822             fiy1             = _mm_add_ps(fiy1,ty);
 823             fiz1             = _mm_add_ps(fiz1,tz);
 824
 825             fjx2             = _mm_add_ps(fjx2,tx);
 826             fjy2             = _mm_add_ps(fjy2,ty);
 827             fjz2             = _mm_add_ps(fjz2,tz);
 828
 829             /**************************
 830              * CALCULATE INTERACTIONS *
 831              **************************/
 832
 833             /* COULOMB ELECTROSTATICS */
 834             velec            = _mm_mul_ps(qq20,rinv20);
 835             felec            = _mm_mul_ps(velec,rinvsq20);
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _mm_andnot_ps(dummy_mask,velec);
 839             velecsum         = _mm_add_ps(velecsum,velec);
 840
 841             fscal            = felec;
 842
 843             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 844
 845             /* Calculate temporary vectorial force */
 846             tx               = _mm_mul_ps(fscal,dx20);
 847             ty               = _mm_mul_ps(fscal,dy20);
 848             tz               = _mm_mul_ps(fscal,dz20);
 849
 850             /* Update vectorial force */
 851             fix2             = _mm_add_ps(fix2,tx);
 852             fiy2             = _mm_add_ps(fiy2,ty);
 853             fiz2             = _mm_add_ps(fiz2,tz);
 854
 855             fjx0             = _mm_add_ps(fjx0,tx);
 856             fjy0             = _mm_add_ps(fjy0,ty);
 857             fjz0             = _mm_add_ps(fjz0,tz);
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             /* COULOMB ELECTROSTATICS */
 864             velec            = _mm_mul_ps(qq21,rinv21);
 865             felec            = _mm_mul_ps(velec,rinvsq21);
 866
 867             /* Update potential sum for this i atom from the interaction with this j atom. */
 868             velec            = _mm_andnot_ps(dummy_mask,velec);
 869             velecsum         = _mm_add_ps(velecsum,velec);
 870
 871             fscal            = felec;
 872
 873             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 874
 875             /* Calculate temporary vectorial force */
 876             tx               = _mm_mul_ps(fscal,dx21);
 877             ty               = _mm_mul_ps(fscal,dy21);
 878             tz               = _mm_mul_ps(fscal,dz21);
 879
 880             /* Update vectorial force */
 881             fix2             = _mm_add_ps(fix2,tx);
 882             fiy2             = _mm_add_ps(fiy2,ty);
 883             fiz2             = _mm_add_ps(fiz2,tz);
 884
 885             fjx1             = _mm_add_ps(fjx1,tx);
 886             fjy1             = _mm_add_ps(fjy1,ty);
 887             fjz1             = _mm_add_ps(fjz1,tz);
 888
 889             /**************************
 890              * CALCULATE INTERACTIONS *
 891              **************************/
 892
 893             /* COULOMB ELECTROSTATICS */
 894             velec            = _mm_mul_ps(qq22,rinv22);
 895             felec            = _mm_mul_ps(velec,rinvsq22);
 896
 897             /* Update potential sum for this i atom from the interaction with this j atom. */
 898             velec            = _mm_andnot_ps(dummy_mask,velec);
 899             velecsum         = _mm_add_ps(velecsum,velec);
 900
 901             fscal            = felec;
 902
 903             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 904
 905             /* Calculate temporary vectorial force */
 906             tx               = _mm_mul_ps(fscal,dx22);
 907             ty               = _mm_mul_ps(fscal,dy22);
 908             tz               = _mm_mul_ps(fscal,dz22);
 909
 910             /* Update vectorial force */
 911             fix2             = _mm_add_ps(fix2,tx);
 912             fiy2             = _mm_add_ps(fiy2,ty);
 913             fiz2             = _mm_add_ps(fiz2,tz);
 914
 915             fjx2             = _mm_add_ps(fjx2,tx);
 916             fjy2             = _mm_add_ps(fjy2,ty);
 917             fjz2             = _mm_add_ps(fjz2,tz);
 918
 919             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 920                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 921                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 922
 923             /* Inner loop uses 264 flops */
 924         }
 925
 926         /* End of innermost loop */
 927
 928         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 929                                               f+i_coord_offset,fshift+i_shift_offset);
 930
 931         ggid                        = gid[iidx];
 932         /* Update potential energies */
 933         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 934         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 935
 936         /* Increment number of inner iterations */
 937         inneriter                  += j_index_end - j_index_start;
 938
 939         /* Outer loop uses 29 flops */
 940     }
 941
 942     /* Increment number of outer iterations */
 943     outeriter        += nri;
 944
 945     /* Update outer/inner flops */
 946
 947     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*29 + inneriter*264);
 948 }
 949 /*
 950  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse2_single
 951  * Electrostatics interaction: Coulomb
 952  * VdW interaction:            LennardJones
 953  * Geometry:                   Water3-Water3
 954  * Calculate force/pot:        Force
 955  */
 956 void
 957 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse2_single
 958                     (t_nblist * gmx_restrict                nlist,
 959                      rvec * gmx_restrict                    xx,
 960                      rvec * gmx_restrict                    ff,
 961                      t_forcerec * gmx_restrict              fr,
 962                      t_mdatoms * gmx_restrict               mdatoms,
 963                      nb_kernel_data_t * gmx_restrict        kernel_data,
 964                      t_nrnb * gmx_restrict                  nrnb)
 965 {
 966     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 967      * just 0 for non-waters.
 968      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 969      * jnr indices corresponding to data put in the four positions in the SIMD register.
 970      */
 971     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 972     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 973     int              jnrA,jnrB,jnrC,jnrD;
 974     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 975     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 976     real             shX,shY,shZ,rcutoff_scalar;
 977     real             *shiftvec,*fshift,*x,*f;
 978     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 979     int              vdwioffset0;
 980     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 981     int              vdwioffset1;
 982     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 983     int              vdwioffset2;
 984     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 985     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 986     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 987     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 988     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 989     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 990     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 991     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 992     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 993     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 994     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 995     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 996     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 997     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 998     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 999     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1000     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1001     real             *charge;
1002     int              nvdwtype;
1003     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1004     int              *vdwtype;
1005     real             *vdwparam;
1006     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1007     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1008     __m128           dummy_mask,cutoff_mask;
1009     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1010     __m128           one     = _mm_set1_ps(1.0);
1011     __m128           two     = _mm_set1_ps(2.0);
1012     x                = xx[0];
1013     f                = ff[0];
1014
1015     nri              = nlist->nri;
1016     iinr             = nlist->iinr;
1017     jindex           = nlist->jindex;
1018     jjnr             = nlist->jjnr;
1019     shiftidx         = nlist->shift;
1020     gid              = nlist->gid;
1021     shiftvec         = fr->shift_vec[0];
1022     fshift           = fr->fshift[0];
1023     facel            = _mm_set1_ps(fr->epsfac);
1024     charge           = mdatoms->chargeA;
1025     nvdwtype         = fr->ntype;
1026     vdwparam         = fr->nbfp;
1027     vdwtype          = mdatoms->typeA;
1028
1029     /* Setup water-specific parameters */
1030     inr              = nlist->iinr[0];
1031     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1032     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1033     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1034     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1035
1036     jq0              = _mm_set1_ps(charge[inr+0]);
1037     jq1              = _mm_set1_ps(charge[inr+1]);
1038     jq2              = _mm_set1_ps(charge[inr+2]);
1039     vdwjidx0A        = 2*vdwtype[inr+0];
1040     qq00             = _mm_mul_ps(iq0,jq0);
1041     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1042     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1043     qq01             = _mm_mul_ps(iq0,jq1);
1044     qq02             = _mm_mul_ps(iq0,jq2);
1045     qq10             = _mm_mul_ps(iq1,jq0);
1046     qq11             = _mm_mul_ps(iq1,jq1);
1047     qq12             = _mm_mul_ps(iq1,jq2);
1048     qq20             = _mm_mul_ps(iq2,jq0);
1049     qq21             = _mm_mul_ps(iq2,jq1);
1050     qq22             = _mm_mul_ps(iq2,jq2);
1051
1052     /* Avoid stupid compiler warnings */
1053     jnrA = jnrB = jnrC = jnrD = 0;
1054     j_coord_offsetA = 0;
1055     j_coord_offsetB = 0;
1056     j_coord_offsetC = 0;
1057     j_coord_offsetD = 0;
1058
1059     outeriter        = 0;
1060     inneriter        = 0;
1061
1062     /* Start outer loop over neighborlists */
1063     for(iidx=0; iidx<nri; iidx++)
1064     {
1065         /* Load shift vector for this list */
1066         i_shift_offset   = DIM*shiftidx[iidx];
1067         shX              = shiftvec[i_shift_offset+XX];
1068         shY              = shiftvec[i_shift_offset+YY];
1069         shZ              = shiftvec[i_shift_offset+ZZ];
1070
1071         /* Load limits for loop over neighbors */
1072         j_index_start    = jindex[iidx];
1073         j_index_end      = jindex[iidx+1];
1074
1075         /* Get outer coordinate index */
1076         inr              = iinr[iidx];
1077         i_coord_offset   = DIM*inr;
1078
1079         /* Load i particle coords and add shift vector */
1080         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
1081         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
1082         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
1083         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1084         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1085         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1086         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1087         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1088         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1089
1090         fix0             = _mm_setzero_ps();
1091         fiy0             = _mm_setzero_ps();
1092         fiz0             = _mm_setzero_ps();
1093         fix1             = _mm_setzero_ps();
1094         fiy1             = _mm_setzero_ps();
1095         fiz1             = _mm_setzero_ps();
1096         fix2             = _mm_setzero_ps();
1097         fiy2             = _mm_setzero_ps();
1098         fiz2             = _mm_setzero_ps();
1099
1100         /* Start inner kernel loop */
1101         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1102         {
1103
1104             /* Get j neighbor index, and coordinate index */
1105             jnrA             = jjnr[jidx];
1106             jnrB             = jjnr[jidx+1];
1107             jnrC             = jjnr[jidx+2];
1108             jnrD             = jjnr[jidx+3];
1109
1110             j_coord_offsetA  = DIM*jnrA;
1111             j_coord_offsetB  = DIM*jnrB;
1112             j_coord_offsetC  = DIM*jnrC;
1113             j_coord_offsetD  = DIM*jnrD;
1114
1115             /* load j atom coordinates */
1116             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1117                                               x+j_coord_offsetC,x+j_coord_offsetD,
1118                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1119
1120             /* Calculate displacement vector */
1121             dx00             = _mm_sub_ps(ix0,jx0);
1122             dy00             = _mm_sub_ps(iy0,jy0);
1123             dz00             = _mm_sub_ps(iz0,jz0);
1124             dx01             = _mm_sub_ps(ix0,jx1);
1125             dy01             = _mm_sub_ps(iy0,jy1);
1126             dz01             = _mm_sub_ps(iz0,jz1);
1127             dx02             = _mm_sub_ps(ix0,jx2);
1128             dy02             = _mm_sub_ps(iy0,jy2);
1129             dz02             = _mm_sub_ps(iz0,jz2);
1130             dx10             = _mm_sub_ps(ix1,jx0);
1131             dy10             = _mm_sub_ps(iy1,jy0);
1132             dz10             = _mm_sub_ps(iz1,jz0);
1133             dx11             = _mm_sub_ps(ix1,jx1);
1134             dy11             = _mm_sub_ps(iy1,jy1);
1135             dz11             = _mm_sub_ps(iz1,jz1);
1136             dx12             = _mm_sub_ps(ix1,jx2);
1137             dy12             = _mm_sub_ps(iy1,jy2);
1138             dz12             = _mm_sub_ps(iz1,jz2);
1139             dx20             = _mm_sub_ps(ix2,jx0);
1140             dy20             = _mm_sub_ps(iy2,jy0);
1141             dz20             = _mm_sub_ps(iz2,jz0);
1142             dx21             = _mm_sub_ps(ix2,jx1);
1143             dy21             = _mm_sub_ps(iy2,jy1);
1144             dz21             = _mm_sub_ps(iz2,jz1);
1145             dx22             = _mm_sub_ps(ix2,jx2);
1146             dy22             = _mm_sub_ps(iy2,jy2);
1147             dz22             = _mm_sub_ps(iz2,jz2);
1148
1149             /* Calculate squared distance and things based on it */
1150             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1151             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1152             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1153             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1154             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1155             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1156             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1157             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1158             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1159
1160             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1161             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1162             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1163             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1164             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1165             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1166             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1167             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1168             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1169
1170             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1171             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1172             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1173             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1174             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1175             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1176             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1177             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1178             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1179
1180             fjx0             = _mm_setzero_ps();
1181             fjy0             = _mm_setzero_ps();
1182             fjz0             = _mm_setzero_ps();
1183             fjx1             = _mm_setzero_ps();
1184             fjy1             = _mm_setzero_ps();
1185             fjz1             = _mm_setzero_ps();
1186             fjx2             = _mm_setzero_ps();
1187             fjy2             = _mm_setzero_ps();
1188             fjz2             = _mm_setzero_ps();
1189
1190             /**************************
1191              * CALCULATE INTERACTIONS *
1192              **************************/
1193
1194             /* COULOMB ELECTROSTATICS */
1195             velec            = _mm_mul_ps(qq00,rinv00);
1196             felec            = _mm_mul_ps(velec,rinvsq00);
1197
1198             /* LENNARD-JONES DISPERSION/REPULSION */
1199
1200             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1201             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1202
1203             fscal            = _mm_add_ps(felec,fvdw);
1204
1205             /* Calculate temporary vectorial force */
1206             tx               = _mm_mul_ps(fscal,dx00);
1207             ty               = _mm_mul_ps(fscal,dy00);
1208             tz               = _mm_mul_ps(fscal,dz00);
1209
1210             /* Update vectorial force */
1211             fix0             = _mm_add_ps(fix0,tx);
1212             fiy0             = _mm_add_ps(fiy0,ty);
1213             fiz0             = _mm_add_ps(fiz0,tz);
1214
1215             fjx0             = _mm_add_ps(fjx0,tx);
1216             fjy0             = _mm_add_ps(fjy0,ty);
1217             fjz0             = _mm_add_ps(fjz0,tz);
1218
1219             /**************************
1220              * CALCULATE INTERACTIONS *
1221              **************************/
1222
1223             /* COULOMB ELECTROSTATICS */
1224             velec            = _mm_mul_ps(qq01,rinv01);
1225             felec            = _mm_mul_ps(velec,rinvsq01);
1226
1227             fscal            = felec;
1228
1229             /* Calculate temporary vectorial force */
1230             tx               = _mm_mul_ps(fscal,dx01);
1231             ty               = _mm_mul_ps(fscal,dy01);
1232             tz               = _mm_mul_ps(fscal,dz01);
1233
1234             /* Update vectorial force */
1235             fix0             = _mm_add_ps(fix0,tx);
1236             fiy0             = _mm_add_ps(fiy0,ty);
1237             fiz0             = _mm_add_ps(fiz0,tz);
1238
1239             fjx1             = _mm_add_ps(fjx1,tx);
1240             fjy1             = _mm_add_ps(fjy1,ty);
1241             fjz1             = _mm_add_ps(fjz1,tz);
1242
1243             /**************************
1244              * CALCULATE INTERACTIONS *
1245              **************************/
1246
1247             /* COULOMB ELECTROSTATICS */
1248             velec            = _mm_mul_ps(qq02,rinv02);
1249             felec            = _mm_mul_ps(velec,rinvsq02);
1250
1251             fscal            = felec;
1252
1253             /* Calculate temporary vectorial force */
1254             tx               = _mm_mul_ps(fscal,dx02);
1255             ty               = _mm_mul_ps(fscal,dy02);
1256             tz               = _mm_mul_ps(fscal,dz02);
1257
1258             /* Update vectorial force */
1259             fix0             = _mm_add_ps(fix0,tx);
1260             fiy0             = _mm_add_ps(fiy0,ty);
1261             fiz0             = _mm_add_ps(fiz0,tz);
1262
1263             fjx2             = _mm_add_ps(fjx2,tx);
1264             fjy2             = _mm_add_ps(fjy2,ty);
1265             fjz2             = _mm_add_ps(fjz2,tz);
1266
1267             /**************************
1268              * CALCULATE INTERACTIONS *
1269              **************************/
1270
1271             /* COULOMB ELECTROSTATICS */
1272             velec            = _mm_mul_ps(qq10,rinv10);
1273             felec            = _mm_mul_ps(velec,rinvsq10);
1274
1275             fscal            = felec;
1276
1277             /* Calculate temporary vectorial force */
1278             tx               = _mm_mul_ps(fscal,dx10);
1279             ty               = _mm_mul_ps(fscal,dy10);
1280             tz               = _mm_mul_ps(fscal,dz10);
1281
1282             /* Update vectorial force */
1283             fix1             = _mm_add_ps(fix1,tx);
1284             fiy1             = _mm_add_ps(fiy1,ty);
1285             fiz1             = _mm_add_ps(fiz1,tz);
1286
1287             fjx0             = _mm_add_ps(fjx0,tx);
1288             fjy0             = _mm_add_ps(fjy0,ty);
1289             fjz0             = _mm_add_ps(fjz0,tz);
1290
1291             /**************************
1292              * CALCULATE INTERACTIONS *
1293              **************************/
1294
1295             /* COULOMB ELECTROSTATICS */
1296             velec            = _mm_mul_ps(qq11,rinv11);
1297             felec            = _mm_mul_ps(velec,rinvsq11);
1298
1299             fscal            = felec;
1300
1301             /* Calculate temporary vectorial force */
1302             tx               = _mm_mul_ps(fscal,dx11);
1303             ty               = _mm_mul_ps(fscal,dy11);
1304             tz               = _mm_mul_ps(fscal,dz11);
1305
1306             /* Update vectorial force */
1307             fix1             = _mm_add_ps(fix1,tx);
1308             fiy1             = _mm_add_ps(fiy1,ty);
1309             fiz1             = _mm_add_ps(fiz1,tz);
1310
1311             fjx1             = _mm_add_ps(fjx1,tx);
1312             fjy1             = _mm_add_ps(fjy1,ty);
1313             fjz1             = _mm_add_ps(fjz1,tz);
1314
1315             /**************************
1316              * CALCULATE INTERACTIONS *
1317              **************************/
1318
1319             /* COULOMB ELECTROSTATICS */
1320             velec            = _mm_mul_ps(qq12,rinv12);
1321             felec            = _mm_mul_ps(velec,rinvsq12);
1322
1323             fscal            = felec;
1324
1325             /* Calculate temporary vectorial force */
1326             tx               = _mm_mul_ps(fscal,dx12);
1327             ty               = _mm_mul_ps(fscal,dy12);
1328             tz               = _mm_mul_ps(fscal,dz12);
1329
1330             /* Update vectorial force */
1331             fix1             = _mm_add_ps(fix1,tx);
1332             fiy1             = _mm_add_ps(fiy1,ty);
1333             fiz1             = _mm_add_ps(fiz1,tz);
1334
1335             fjx2             = _mm_add_ps(fjx2,tx);
1336             fjy2             = _mm_add_ps(fjy2,ty);
1337             fjz2             = _mm_add_ps(fjz2,tz);
1338
1339             /**************************
1340              * CALCULATE INTERACTIONS *
1341              **************************/
1342
1343             /* COULOMB ELECTROSTATICS */
1344             velec            = _mm_mul_ps(qq20,rinv20);
1345             felec            = _mm_mul_ps(velec,rinvsq20);
1346
1347             fscal            = felec;
1348
1349             /* Calculate temporary vectorial force */
1350             tx               = _mm_mul_ps(fscal,dx20);
1351             ty               = _mm_mul_ps(fscal,dy20);
1352             tz               = _mm_mul_ps(fscal,dz20);
1353
1354             /* Update vectorial force */
1355             fix2             = _mm_add_ps(fix2,tx);
1356             fiy2             = _mm_add_ps(fiy2,ty);
1357             fiz2             = _mm_add_ps(fiz2,tz);
1358
1359             fjx0             = _mm_add_ps(fjx0,tx);
1360             fjy0             = _mm_add_ps(fjy0,ty);
1361             fjz0             = _mm_add_ps(fjz0,tz);
1362
1363             /**************************
1364              * CALCULATE INTERACTIONS *
1365              **************************/
1366
1367             /* COULOMB ELECTROSTATICS */
1368             velec            = _mm_mul_ps(qq21,rinv21);
1369             felec            = _mm_mul_ps(velec,rinvsq21);
1370
1371             fscal            = felec;
1372
1373             /* Calculate temporary vectorial force */
1374             tx               = _mm_mul_ps(fscal,dx21);
1375             ty               = _mm_mul_ps(fscal,dy21);
1376             tz               = _mm_mul_ps(fscal,dz21);
1377
1378             /* Update vectorial force */
1379             fix2             = _mm_add_ps(fix2,tx);
1380             fiy2             = _mm_add_ps(fiy2,ty);
1381             fiz2             = _mm_add_ps(fiz2,tz);
1382
1383             fjx1             = _mm_add_ps(fjx1,tx);
1384             fjy1             = _mm_add_ps(fjy1,ty);
1385             fjz1             = _mm_add_ps(fjz1,tz);
1386
1387             /**************************
1388              * CALCULATE INTERACTIONS *
1389              **************************/
1390
1391             /* COULOMB ELECTROSTATICS */
1392             velec            = _mm_mul_ps(qq22,rinv22);
1393             felec            = _mm_mul_ps(velec,rinvsq22);
1394
1395             fscal            = felec;
1396
1397             /* Calculate temporary vectorial force */
1398             tx               = _mm_mul_ps(fscal,dx22);
1399             ty               = _mm_mul_ps(fscal,dy22);
1400             tz               = _mm_mul_ps(fscal,dz22);
1401
1402             /* Update vectorial force */
1403             fix2             = _mm_add_ps(fix2,tx);
1404             fiy2             = _mm_add_ps(fiy2,ty);
1405             fiz2             = _mm_add_ps(fiz2,tz);
1406
1407             fjx2             = _mm_add_ps(fjx2,tx);
1408             fjy2             = _mm_add_ps(fjy2,ty);
1409             fjz2             = _mm_add_ps(fjz2,tz);
1410
1411             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1412                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1413                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1414
1415             /* Inner loop uses 250 flops */
1416         }
1417
1418         if(jidx<j_index_end)
1419         {
1420
1421             /* Get j neighbor index, and coordinate index */
1422             jnrA             = jjnr[jidx];
1423             jnrB             = jjnr[jidx+1];
1424             jnrC             = jjnr[jidx+2];
1425             jnrD             = jjnr[jidx+3];
1426
1427             /* Sign of each element will be negative for non-real atoms.
1428              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1429              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1430              */
1431             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1432             jnrA       = (jnrA>=0) ? jnrA : 0;
1433             jnrB       = (jnrB>=0) ? jnrB : 0;
1434             jnrC       = (jnrC>=0) ? jnrC : 0;
1435             jnrD       = (jnrD>=0) ? jnrD : 0;
1436
1437             j_coord_offsetA  = DIM*jnrA;
1438             j_coord_offsetB  = DIM*jnrB;
1439             j_coord_offsetC  = DIM*jnrC;
1440             j_coord_offsetD  = DIM*jnrD;
1441
1442             /* load j atom coordinates */
1443             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1444                                               x+j_coord_offsetC,x+j_coord_offsetD,
1445                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1446
1447             /* Calculate displacement vector */
1448             dx00             = _mm_sub_ps(ix0,jx0);
1449             dy00             = _mm_sub_ps(iy0,jy0);
1450             dz00             = _mm_sub_ps(iz0,jz0);
1451             dx01             = _mm_sub_ps(ix0,jx1);
1452             dy01             = _mm_sub_ps(iy0,jy1);
1453             dz01             = _mm_sub_ps(iz0,jz1);
1454             dx02             = _mm_sub_ps(ix0,jx2);
1455             dy02             = _mm_sub_ps(iy0,jy2);
1456             dz02             = _mm_sub_ps(iz0,jz2);
1457             dx10             = _mm_sub_ps(ix1,jx0);
1458             dy10             = _mm_sub_ps(iy1,jy0);
1459             dz10             = _mm_sub_ps(iz1,jz0);
1460             dx11             = _mm_sub_ps(ix1,jx1);
1461             dy11             = _mm_sub_ps(iy1,jy1);
1462             dz11             = _mm_sub_ps(iz1,jz1);
1463             dx12             = _mm_sub_ps(ix1,jx2);
1464             dy12             = _mm_sub_ps(iy1,jy2);
1465             dz12             = _mm_sub_ps(iz1,jz2);
1466             dx20             = _mm_sub_ps(ix2,jx0);
1467             dy20             = _mm_sub_ps(iy2,jy0);
1468             dz20             = _mm_sub_ps(iz2,jz0);
1469             dx21             = _mm_sub_ps(ix2,jx1);
1470             dy21             = _mm_sub_ps(iy2,jy1);
1471             dz21             = _mm_sub_ps(iz2,jz1);
1472             dx22             = _mm_sub_ps(ix2,jx2);
1473             dy22             = _mm_sub_ps(iy2,jy2);
1474             dz22             = _mm_sub_ps(iz2,jz2);
1475
1476             /* Calculate squared distance and things based on it */
1477             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1478             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1479             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1480             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1481             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1482             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1483             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1484             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1485             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1486
1487             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1488             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1489             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1490             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1491             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1492             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1493             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1494             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1495             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1496
1497             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1498             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1499             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1500             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1501             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1502             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1503             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1504             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1505             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1506
1507             fjx0             = _mm_setzero_ps();
1508             fjy0             = _mm_setzero_ps();
1509             fjz0             = _mm_setzero_ps();
1510             fjx1             = _mm_setzero_ps();
1511             fjy1             = _mm_setzero_ps();
1512             fjz1             = _mm_setzero_ps();
1513             fjx2             = _mm_setzero_ps();
1514             fjy2             = _mm_setzero_ps();
1515             fjz2             = _mm_setzero_ps();
1516
1517             /**************************
1518              * CALCULATE INTERACTIONS *
1519              **************************/
1520
1521             /* COULOMB ELECTROSTATICS */
1522             velec            = _mm_mul_ps(qq00,rinv00);
1523             felec            = _mm_mul_ps(velec,rinvsq00);
1524
1525             /* LENNARD-JONES DISPERSION/REPULSION */
1526
1527             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1528             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1529
1530             fscal            = _mm_add_ps(felec,fvdw);
1531
1532             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1533
1534             /* Calculate temporary vectorial force */
1535             tx               = _mm_mul_ps(fscal,dx00);
1536             ty               = _mm_mul_ps(fscal,dy00);
1537             tz               = _mm_mul_ps(fscal,dz00);
1538
1539             /* Update vectorial force */
1540             fix0             = _mm_add_ps(fix0,tx);
1541             fiy0             = _mm_add_ps(fiy0,ty);
1542             fiz0             = _mm_add_ps(fiz0,tz);
1543
1544             fjx0             = _mm_add_ps(fjx0,tx);
1545             fjy0             = _mm_add_ps(fjy0,ty);
1546             fjz0             = _mm_add_ps(fjz0,tz);
1547
1548             /**************************
1549              * CALCULATE INTERACTIONS *
1550              **************************/
1551
1552             /* COULOMB ELECTROSTATICS */
1553             velec            = _mm_mul_ps(qq01,rinv01);
1554             felec            = _mm_mul_ps(velec,rinvsq01);
1555
1556             fscal            = felec;
1557
1558             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1559
1560             /* Calculate temporary vectorial force */
1561             tx               = _mm_mul_ps(fscal,dx01);
1562             ty               = _mm_mul_ps(fscal,dy01);
1563             tz               = _mm_mul_ps(fscal,dz01);
1564
1565             /* Update vectorial force */
1566             fix0             = _mm_add_ps(fix0,tx);
1567             fiy0             = _mm_add_ps(fiy0,ty);
1568             fiz0             = _mm_add_ps(fiz0,tz);
1569
1570             fjx1             = _mm_add_ps(fjx1,tx);
1571             fjy1             = _mm_add_ps(fjy1,ty);
1572             fjz1             = _mm_add_ps(fjz1,tz);
1573
1574             /**************************
1575              * CALCULATE INTERACTIONS *
1576              **************************/
1577
1578             /* COULOMB ELECTROSTATICS */
1579             velec            = _mm_mul_ps(qq02,rinv02);
1580             felec            = _mm_mul_ps(velec,rinvsq02);
1581
1582             fscal            = felec;
1583
1584             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1585
1586             /* Calculate temporary vectorial force */
1587             tx               = _mm_mul_ps(fscal,dx02);
1588             ty               = _mm_mul_ps(fscal,dy02);
1589             tz               = _mm_mul_ps(fscal,dz02);
1590
1591             /* Update vectorial force */
1592             fix0             = _mm_add_ps(fix0,tx);
1593             fiy0             = _mm_add_ps(fiy0,ty);
1594             fiz0             = _mm_add_ps(fiz0,tz);
1595
1596             fjx2             = _mm_add_ps(fjx2,tx);
1597             fjy2             = _mm_add_ps(fjy2,ty);
1598             fjz2             = _mm_add_ps(fjz2,tz);
1599
1600             /**************************
1601              * CALCULATE INTERACTIONS *
1602              **************************/
1603
1604             /* COULOMB ELECTROSTATICS */
1605             velec            = _mm_mul_ps(qq10,rinv10);
1606             felec            = _mm_mul_ps(velec,rinvsq10);
1607
1608             fscal            = felec;
1609
1610             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1611
1612             /* Calculate temporary vectorial force */
1613             tx               = _mm_mul_ps(fscal,dx10);
1614             ty               = _mm_mul_ps(fscal,dy10);
1615             tz               = _mm_mul_ps(fscal,dz10);
1616
1617             /* Update vectorial force */
1618             fix1             = _mm_add_ps(fix1,tx);
1619             fiy1             = _mm_add_ps(fiy1,ty);
1620             fiz1             = _mm_add_ps(fiz1,tz);
1621
1622             fjx0             = _mm_add_ps(fjx0,tx);
1623             fjy0             = _mm_add_ps(fjy0,ty);
1624             fjz0             = _mm_add_ps(fjz0,tz);
1625
1626             /**************************
1627              * CALCULATE INTERACTIONS *
1628              **************************/
1629
1630             /* COULOMB ELECTROSTATICS */
1631             velec            = _mm_mul_ps(qq11,rinv11);
1632             felec            = _mm_mul_ps(velec,rinvsq11);
1633
1634             fscal            = felec;
1635
1636             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1637
1638             /* Calculate temporary vectorial force */
1639             tx               = _mm_mul_ps(fscal,dx11);
1640             ty               = _mm_mul_ps(fscal,dy11);
1641             tz               = _mm_mul_ps(fscal,dz11);
1642
1643             /* Update vectorial force */
1644             fix1             = _mm_add_ps(fix1,tx);
1645             fiy1             = _mm_add_ps(fiy1,ty);
1646             fiz1             = _mm_add_ps(fiz1,tz);
1647
1648             fjx1             = _mm_add_ps(fjx1,tx);
1649             fjy1             = _mm_add_ps(fjy1,ty);
1650             fjz1             = _mm_add_ps(fjz1,tz);
1651
1652             /**************************
1653              * CALCULATE INTERACTIONS *
1654              **************************/
1655
1656             /* COULOMB ELECTROSTATICS */
1657             velec            = _mm_mul_ps(qq12,rinv12);
1658             felec            = _mm_mul_ps(velec,rinvsq12);
1659
1660             fscal            = felec;
1661
1662             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1663
1664             /* Calculate temporary vectorial force */
1665             tx               = _mm_mul_ps(fscal,dx12);
1666             ty               = _mm_mul_ps(fscal,dy12);
1667             tz               = _mm_mul_ps(fscal,dz12);
1668
1669             /* Update vectorial force */
1670             fix1             = _mm_add_ps(fix1,tx);
1671             fiy1             = _mm_add_ps(fiy1,ty);
1672             fiz1             = _mm_add_ps(fiz1,tz);
1673
1674             fjx2             = _mm_add_ps(fjx2,tx);
1675             fjy2             = _mm_add_ps(fjy2,ty);
1676             fjz2             = _mm_add_ps(fjz2,tz);
1677
1678             /**************************
1679              * CALCULATE INTERACTIONS *
1680              **************************/
1681
1682             /* COULOMB ELECTROSTATICS */
1683             velec            = _mm_mul_ps(qq20,rinv20);
1684             felec            = _mm_mul_ps(velec,rinvsq20);
1685
1686             fscal            = felec;
1687
1688             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1689
1690             /* Calculate temporary vectorial force */
1691             tx               = _mm_mul_ps(fscal,dx20);
1692             ty               = _mm_mul_ps(fscal,dy20);
1693             tz               = _mm_mul_ps(fscal,dz20);
1694
1695             /* Update vectorial force */
1696             fix2             = _mm_add_ps(fix2,tx);
1697             fiy2             = _mm_add_ps(fiy2,ty);
1698             fiz2             = _mm_add_ps(fiz2,tz);
1699
1700             fjx0             = _mm_add_ps(fjx0,tx);
1701             fjy0             = _mm_add_ps(fjy0,ty);
1702             fjz0             = _mm_add_ps(fjz0,tz);
1703
1704             /**************************
1705              * CALCULATE INTERACTIONS *
1706              **************************/
1707
1708             /* COULOMB ELECTROSTATICS */
1709             velec            = _mm_mul_ps(qq21,rinv21);
1710             felec            = _mm_mul_ps(velec,rinvsq21);
1711
1712             fscal            = felec;
1713
1714             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1715
1716             /* Calculate temporary vectorial force */
1717             tx               = _mm_mul_ps(fscal,dx21);
1718             ty               = _mm_mul_ps(fscal,dy21);
1719             tz               = _mm_mul_ps(fscal,dz21);
1720
1721             /* Update vectorial force */
1722             fix2             = _mm_add_ps(fix2,tx);
1723             fiy2             = _mm_add_ps(fiy2,ty);
1724             fiz2             = _mm_add_ps(fiz2,tz);
1725
1726             fjx1             = _mm_add_ps(fjx1,tx);
1727             fjy1             = _mm_add_ps(fjy1,ty);
1728             fjz1             = _mm_add_ps(fjz1,tz);
1729
1730             /**************************
1731              * CALCULATE INTERACTIONS *
1732              **************************/
1733
1734             /* COULOMB ELECTROSTATICS */
1735             velec            = _mm_mul_ps(qq22,rinv22);
1736             felec            = _mm_mul_ps(velec,rinvsq22);
1737
1738             fscal            = felec;
1739
1740             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1741
1742             /* Calculate temporary vectorial force */
1743             tx               = _mm_mul_ps(fscal,dx22);
1744             ty               = _mm_mul_ps(fscal,dy22);
1745             tz               = _mm_mul_ps(fscal,dz22);
1746
1747             /* Update vectorial force */
1748             fix2             = _mm_add_ps(fix2,tx);
1749             fiy2             = _mm_add_ps(fiy2,ty);
1750             fiz2             = _mm_add_ps(fiz2,tz);
1751
1752             fjx2             = _mm_add_ps(fjx2,tx);
1753             fjy2             = _mm_add_ps(fjy2,ty);
1754             fjz2             = _mm_add_ps(fjz2,tz);
1755
1756             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1757                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1758                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1759
1760             /* Inner loop uses 250 flops */
1761         }
1762
1763         /* End of innermost loop */
1764
1765         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1766                                               f+i_coord_offset,fshift+i_shift_offset);
1767
1768         /* Increment number of inner iterations */
1769         inneriter                  += j_index_end - j_index_start;
1770
1771         /* Outer loop uses 27 flops */
1772     }
1773
1774     /* Increment number of outer iterations */
1775     outeriter        += nri;
1776
1777     /* Update outer/inner flops */
1778
1779     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*27 + inneriter*250);
1780 }