src/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecRF_VdwNone_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_ElecRF_VdwNone_GeomW3W3_VF_sse2_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            None
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwNone_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              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwioffset1;
  72     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  73     int              vdwioffset2;
  74     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  75     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  76     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  77     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  78     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  79     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  80     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  81     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  82     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  83     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  84     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  85     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  86     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  87     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  88     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  89     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m128           dummy_mask,cutoff_mask;
  93     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  94     __m128           one     = _mm_set1_ps(1.0);
  95     __m128           two     = _mm_set1_ps(2.0);
  96     x                = xx[0];
  97     f                = ff[0];
  98
  99     nri              = nlist->nri;
 100     iinr             = nlist->iinr;
 101     jindex           = nlist->jindex;
 102     jjnr             = nlist->jjnr;
 103     shiftidx         = nlist->shift;
 104     gid              = nlist->gid;
 105     shiftvec         = fr->shift_vec[0];
 106     fshift           = fr->fshift[0];
 107     facel            = _mm_set1_ps(fr->epsfac);
 108     charge           = mdatoms->chargeA;
 109     krf              = _mm_set1_ps(fr->ic->k_rf);
 110     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 111     crf              = _mm_set1_ps(fr->ic->c_rf);
 112
 113     /* Setup water-specific parameters */
 114     inr              = nlist->iinr[0];
 115     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 116     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 117     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 118
 119     jq0              = _mm_set1_ps(charge[inr+0]);
 120     jq1              = _mm_set1_ps(charge[inr+1]);
 121     jq2              = _mm_set1_ps(charge[inr+2]);
 122     qq00             = _mm_mul_ps(iq0,jq0);
 123     qq01             = _mm_mul_ps(iq0,jq1);
 124     qq02             = _mm_mul_ps(iq0,jq2);
 125     qq10             = _mm_mul_ps(iq1,jq0);
 126     qq11             = _mm_mul_ps(iq1,jq1);
 127     qq12             = _mm_mul_ps(iq1,jq2);
 128     qq20             = _mm_mul_ps(iq2,jq0);
 129     qq21             = _mm_mul_ps(iq2,jq1);
 130     qq22             = _mm_mul_ps(iq2,jq2);
 131
 132     /* Avoid stupid compiler warnings */
 133     jnrA = jnrB = jnrC = jnrD = 0;
 134     j_coord_offsetA = 0;
 135     j_coord_offsetB = 0;
 136     j_coord_offsetC = 0;
 137     j_coord_offsetD = 0;
 138
 139     outeriter        = 0;
 140     inneriter        = 0;
 141
 142     for(iidx=0;iidx<4*DIM;iidx++)
 143     {
 144         scratch[iidx] = 0.0;
 145     }
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 163                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 164
 165         fix0             = _mm_setzero_ps();
 166         fiy0             = _mm_setzero_ps();
 167         fiz0             = _mm_setzero_ps();
 168         fix1             = _mm_setzero_ps();
 169         fiy1             = _mm_setzero_ps();
 170         fiz1             = _mm_setzero_ps();
 171         fix2             = _mm_setzero_ps();
 172         fiy2             = _mm_setzero_ps();
 173         fiz2             = _mm_setzero_ps();
 174
 175         /* Reset potential sums */
 176         velecsum         = _mm_setzero_ps();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             jnrC             = jjnr[jidx+2];
 186             jnrD             = jjnr[jidx+3];
 187             j_coord_offsetA  = DIM*jnrA;
 188             j_coord_offsetB  = DIM*jnrB;
 189             j_coord_offsetC  = DIM*jnrC;
 190             j_coord_offsetD  = DIM*jnrD;
 191
 192             /* load j atom coordinates */
 193             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 194                                               x+j_coord_offsetC,x+j_coord_offsetD,
 195                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 196
 197             /* Calculate displacement vector */
 198             dx00             = _mm_sub_ps(ix0,jx0);
 199             dy00             = _mm_sub_ps(iy0,jy0);
 200             dz00             = _mm_sub_ps(iz0,jz0);
 201             dx01             = _mm_sub_ps(ix0,jx1);
 202             dy01             = _mm_sub_ps(iy0,jy1);
 203             dz01             = _mm_sub_ps(iz0,jz1);
 204             dx02             = _mm_sub_ps(ix0,jx2);
 205             dy02             = _mm_sub_ps(iy0,jy2);
 206             dz02             = _mm_sub_ps(iz0,jz2);
 207             dx10             = _mm_sub_ps(ix1,jx0);
 208             dy10             = _mm_sub_ps(iy1,jy0);
 209             dz10             = _mm_sub_ps(iz1,jz0);
 210             dx11             = _mm_sub_ps(ix1,jx1);
 211             dy11             = _mm_sub_ps(iy1,jy1);
 212             dz11             = _mm_sub_ps(iz1,jz1);
 213             dx12             = _mm_sub_ps(ix1,jx2);
 214             dy12             = _mm_sub_ps(iy1,jy2);
 215             dz12             = _mm_sub_ps(iz1,jz2);
 216             dx20             = _mm_sub_ps(ix2,jx0);
 217             dy20             = _mm_sub_ps(iy2,jy0);
 218             dz20             = _mm_sub_ps(iz2,jz0);
 219             dx21             = _mm_sub_ps(ix2,jx1);
 220             dy21             = _mm_sub_ps(iy2,jy1);
 221             dz21             = _mm_sub_ps(iz2,jz1);
 222             dx22             = _mm_sub_ps(ix2,jx2);
 223             dy22             = _mm_sub_ps(iy2,jy2);
 224             dz22             = _mm_sub_ps(iz2,jz2);
 225
 226             /* Calculate squared distance and things based on it */
 227             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 228             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 229             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 230             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 231             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 232             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 233             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 234             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 235             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 236
 237             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 238             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 239             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 240             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 241             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 242             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 243             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 244             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 245             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 246
 247             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 248             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 249             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 250             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 251             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 252             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 253             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 254             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 255             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 256
 257             fjx0             = _mm_setzero_ps();
 258             fjy0             = _mm_setzero_ps();
 259             fjz0             = _mm_setzero_ps();
 260             fjx1             = _mm_setzero_ps();
 261             fjy1             = _mm_setzero_ps();
 262             fjz1             = _mm_setzero_ps();
 263             fjx2             = _mm_setzero_ps();
 264             fjy2             = _mm_setzero_ps();
 265             fjz2             = _mm_setzero_ps();
 266
 267             /**************************
 268              * CALCULATE INTERACTIONS *
 269              **************************/
 270
 271             /* REACTION-FIELD ELECTROSTATICS */
 272             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 273             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 274
 275             /* Update potential sum for this i atom from the interaction with this j atom. */
 276             velecsum         = _mm_add_ps(velecsum,velec);
 277
 278             fscal            = felec;
 279
 280             /* Calculate temporary vectorial force */
 281             tx               = _mm_mul_ps(fscal,dx00);
 282             ty               = _mm_mul_ps(fscal,dy00);
 283             tz               = _mm_mul_ps(fscal,dz00);
 284
 285             /* Update vectorial force */
 286             fix0             = _mm_add_ps(fix0,tx);
 287             fiy0             = _mm_add_ps(fiy0,ty);
 288             fiz0             = _mm_add_ps(fiz0,tz);
 289
 290             fjx0             = _mm_add_ps(fjx0,tx);
 291             fjy0             = _mm_add_ps(fjy0,ty);
 292             fjz0             = _mm_add_ps(fjz0,tz);
 293
 294             /**************************
 295              * CALCULATE INTERACTIONS *
 296              **************************/
 297
 298             /* REACTION-FIELD ELECTROSTATICS */
 299             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 300             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 301
 302             /* Update potential sum for this i atom from the interaction with this j atom. */
 303             velecsum         = _mm_add_ps(velecsum,velec);
 304
 305             fscal            = felec;
 306
 307             /* Calculate temporary vectorial force */
 308             tx               = _mm_mul_ps(fscal,dx01);
 309             ty               = _mm_mul_ps(fscal,dy01);
 310             tz               = _mm_mul_ps(fscal,dz01);
 311
 312             /* Update vectorial force */
 313             fix0             = _mm_add_ps(fix0,tx);
 314             fiy0             = _mm_add_ps(fiy0,ty);
 315             fiz0             = _mm_add_ps(fiz0,tz);
 316
 317             fjx1             = _mm_add_ps(fjx1,tx);
 318             fjy1             = _mm_add_ps(fjy1,ty);
 319             fjz1             = _mm_add_ps(fjz1,tz);
 320
 321             /**************************
 322              * CALCULATE INTERACTIONS *
 323              **************************/
 324
 325             /* REACTION-FIELD ELECTROSTATICS */
 326             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 327             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 328
 329             /* Update potential sum for this i atom from the interaction with this j atom. */
 330             velecsum         = _mm_add_ps(velecsum,velec);
 331
 332             fscal            = felec;
 333
 334             /* Calculate temporary vectorial force */
 335             tx               = _mm_mul_ps(fscal,dx02);
 336             ty               = _mm_mul_ps(fscal,dy02);
 337             tz               = _mm_mul_ps(fscal,dz02);
 338
 339             /* Update vectorial force */
 340             fix0             = _mm_add_ps(fix0,tx);
 341             fiy0             = _mm_add_ps(fiy0,ty);
 342             fiz0             = _mm_add_ps(fiz0,tz);
 343
 344             fjx2             = _mm_add_ps(fjx2,tx);
 345             fjy2             = _mm_add_ps(fjy2,ty);
 346             fjz2             = _mm_add_ps(fjz2,tz);
 347
 348             /**************************
 349              * CALCULATE INTERACTIONS *
 350              **************************/
 351
 352             /* REACTION-FIELD ELECTROSTATICS */
 353             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 354             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 355
 356             /* Update potential sum for this i atom from the interaction with this j atom. */
 357             velecsum         = _mm_add_ps(velecsum,velec);
 358
 359             fscal            = felec;
 360
 361             /* Calculate temporary vectorial force */
 362             tx               = _mm_mul_ps(fscal,dx10);
 363             ty               = _mm_mul_ps(fscal,dy10);
 364             tz               = _mm_mul_ps(fscal,dz10);
 365
 366             /* Update vectorial force */
 367             fix1             = _mm_add_ps(fix1,tx);
 368             fiy1             = _mm_add_ps(fiy1,ty);
 369             fiz1             = _mm_add_ps(fiz1,tz);
 370
 371             fjx0             = _mm_add_ps(fjx0,tx);
 372             fjy0             = _mm_add_ps(fjy0,ty);
 373             fjz0             = _mm_add_ps(fjz0,tz);
 374
 375             /**************************
 376              * CALCULATE INTERACTIONS *
 377              **************************/
 378
 379             /* REACTION-FIELD ELECTROSTATICS */
 380             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 381             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velecsum         = _mm_add_ps(velecsum,velec);
 385
 386             fscal            = felec;
 387
 388             /* Calculate temporary vectorial force */
 389             tx               = _mm_mul_ps(fscal,dx11);
 390             ty               = _mm_mul_ps(fscal,dy11);
 391             tz               = _mm_mul_ps(fscal,dz11);
 392
 393             /* Update vectorial force */
 394             fix1             = _mm_add_ps(fix1,tx);
 395             fiy1             = _mm_add_ps(fiy1,ty);
 396             fiz1             = _mm_add_ps(fiz1,tz);
 397
 398             fjx1             = _mm_add_ps(fjx1,tx);
 399             fjy1             = _mm_add_ps(fjy1,ty);
 400             fjz1             = _mm_add_ps(fjz1,tz);
 401
 402             /**************************
 403              * CALCULATE INTERACTIONS *
 404              **************************/
 405
 406             /* REACTION-FIELD ELECTROSTATICS */
 407             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 408             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 409
 410             /* Update potential sum for this i atom from the interaction with this j atom. */
 411             velecsum         = _mm_add_ps(velecsum,velec);
 412
 413             fscal            = felec;
 414
 415             /* Calculate temporary vectorial force */
 416             tx               = _mm_mul_ps(fscal,dx12);
 417             ty               = _mm_mul_ps(fscal,dy12);
 418             tz               = _mm_mul_ps(fscal,dz12);
 419
 420             /* Update vectorial force */
 421             fix1             = _mm_add_ps(fix1,tx);
 422             fiy1             = _mm_add_ps(fiy1,ty);
 423             fiz1             = _mm_add_ps(fiz1,tz);
 424
 425             fjx2             = _mm_add_ps(fjx2,tx);
 426             fjy2             = _mm_add_ps(fjy2,ty);
 427             fjz2             = _mm_add_ps(fjz2,tz);
 428
 429             /**************************
 430              * CALCULATE INTERACTIONS *
 431              **************************/
 432
 433             /* REACTION-FIELD ELECTROSTATICS */
 434             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 435             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 436
 437             /* Update potential sum for this i atom from the interaction with this j atom. */
 438             velecsum         = _mm_add_ps(velecsum,velec);
 439
 440             fscal            = felec;
 441
 442             /* Calculate temporary vectorial force */
 443             tx               = _mm_mul_ps(fscal,dx20);
 444             ty               = _mm_mul_ps(fscal,dy20);
 445             tz               = _mm_mul_ps(fscal,dz20);
 446
 447             /* Update vectorial force */
 448             fix2             = _mm_add_ps(fix2,tx);
 449             fiy2             = _mm_add_ps(fiy2,ty);
 450             fiz2             = _mm_add_ps(fiz2,tz);
 451
 452             fjx0             = _mm_add_ps(fjx0,tx);
 453             fjy0             = _mm_add_ps(fjy0,ty);
 454             fjz0             = _mm_add_ps(fjz0,tz);
 455
 456             /**************************
 457              * CALCULATE INTERACTIONS *
 458              **************************/
 459
 460             /* REACTION-FIELD ELECTROSTATICS */
 461             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 462             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 463
 464             /* Update potential sum for this i atom from the interaction with this j atom. */
 465             velecsum         = _mm_add_ps(velecsum,velec);
 466
 467             fscal            = felec;
 468
 469             /* Calculate temporary vectorial force */
 470             tx               = _mm_mul_ps(fscal,dx21);
 471             ty               = _mm_mul_ps(fscal,dy21);
 472             tz               = _mm_mul_ps(fscal,dz21);
 473
 474             /* Update vectorial force */
 475             fix2             = _mm_add_ps(fix2,tx);
 476             fiy2             = _mm_add_ps(fiy2,ty);
 477             fiz2             = _mm_add_ps(fiz2,tz);
 478
 479             fjx1             = _mm_add_ps(fjx1,tx);
 480             fjy1             = _mm_add_ps(fjy1,ty);
 481             fjz1             = _mm_add_ps(fjz1,tz);
 482
 483             /**************************
 484              * CALCULATE INTERACTIONS *
 485              **************************/
 486
 487             /* REACTION-FIELD ELECTROSTATICS */
 488             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 489             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 490
 491             /* Update potential sum for this i atom from the interaction with this j atom. */
 492             velecsum         = _mm_add_ps(velecsum,velec);
 493
 494             fscal            = felec;
 495
 496             /* Calculate temporary vectorial force */
 497             tx               = _mm_mul_ps(fscal,dx22);
 498             ty               = _mm_mul_ps(fscal,dy22);
 499             tz               = _mm_mul_ps(fscal,dz22);
 500
 501             /* Update vectorial force */
 502             fix2             = _mm_add_ps(fix2,tx);
 503             fiy2             = _mm_add_ps(fiy2,ty);
 504             fiz2             = _mm_add_ps(fiz2,tz);
 505
 506             fjx2             = _mm_add_ps(fjx2,tx);
 507             fjy2             = _mm_add_ps(fjy2,ty);
 508             fjz2             = _mm_add_ps(fjz2,tz);
 509
 510             fjptrA             = f+j_coord_offsetA;
 511             fjptrB             = f+j_coord_offsetB;
 512             fjptrC             = f+j_coord_offsetC;
 513             fjptrD             = f+j_coord_offsetD;
 514
 515             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 516                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 517
 518             /* Inner loop uses 288 flops */
 519         }
 520
 521         if(jidx<j_index_end)
 522         {
 523
 524             /* Get j neighbor index, and coordinate index */
 525             jnrlistA         = jjnr[jidx];
 526             jnrlistB         = jjnr[jidx+1];
 527             jnrlistC         = jjnr[jidx+2];
 528             jnrlistD         = jjnr[jidx+3];
 529             /* Sign of each element will be negative for non-real atoms.
 530              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 531              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 532              */
 533             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 534             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 535             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 536             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 537             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 538             j_coord_offsetA  = DIM*jnrA;
 539             j_coord_offsetB  = DIM*jnrB;
 540             j_coord_offsetC  = DIM*jnrC;
 541             j_coord_offsetD  = DIM*jnrD;
 542
 543             /* load j atom coordinates */
 544             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 545                                               x+j_coord_offsetC,x+j_coord_offsetD,
 546                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 547
 548             /* Calculate displacement vector */
 549             dx00             = _mm_sub_ps(ix0,jx0);
 550             dy00             = _mm_sub_ps(iy0,jy0);
 551             dz00             = _mm_sub_ps(iz0,jz0);
 552             dx01             = _mm_sub_ps(ix0,jx1);
 553             dy01             = _mm_sub_ps(iy0,jy1);
 554             dz01             = _mm_sub_ps(iz0,jz1);
 555             dx02             = _mm_sub_ps(ix0,jx2);
 556             dy02             = _mm_sub_ps(iy0,jy2);
 557             dz02             = _mm_sub_ps(iz0,jz2);
 558             dx10             = _mm_sub_ps(ix1,jx0);
 559             dy10             = _mm_sub_ps(iy1,jy0);
 560             dz10             = _mm_sub_ps(iz1,jz0);
 561             dx11             = _mm_sub_ps(ix1,jx1);
 562             dy11             = _mm_sub_ps(iy1,jy1);
 563             dz11             = _mm_sub_ps(iz1,jz1);
 564             dx12             = _mm_sub_ps(ix1,jx2);
 565             dy12             = _mm_sub_ps(iy1,jy2);
 566             dz12             = _mm_sub_ps(iz1,jz2);
 567             dx20             = _mm_sub_ps(ix2,jx0);
 568             dy20             = _mm_sub_ps(iy2,jy0);
 569             dz20             = _mm_sub_ps(iz2,jz0);
 570             dx21             = _mm_sub_ps(ix2,jx1);
 571             dy21             = _mm_sub_ps(iy2,jy1);
 572             dz21             = _mm_sub_ps(iz2,jz1);
 573             dx22             = _mm_sub_ps(ix2,jx2);
 574             dy22             = _mm_sub_ps(iy2,jy2);
 575             dz22             = _mm_sub_ps(iz2,jz2);
 576
 577             /* Calculate squared distance and things based on it */
 578             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 579             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 580             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 581             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 582             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 583             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 584             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 585             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 586             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 587
 588             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 589             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 590             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 591             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 592             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 593             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 594             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 595             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 596             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 597
 598             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 599             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 600             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 601             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 602             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 603             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 604             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 605             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 606             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 607
 608             fjx0             = _mm_setzero_ps();
 609             fjy0             = _mm_setzero_ps();
 610             fjz0             = _mm_setzero_ps();
 611             fjx1             = _mm_setzero_ps();
 612             fjy1             = _mm_setzero_ps();
 613             fjz1             = _mm_setzero_ps();
 614             fjx2             = _mm_setzero_ps();
 615             fjy2             = _mm_setzero_ps();
 616             fjz2             = _mm_setzero_ps();
 617
 618             /**************************
 619              * CALCULATE INTERACTIONS *
 620              **************************/
 621
 622             /* REACTION-FIELD ELECTROSTATICS */
 623             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 624             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 625
 626             /* Update potential sum for this i atom from the interaction with this j atom. */
 627             velec            = _mm_andnot_ps(dummy_mask,velec);
 628             velecsum         = _mm_add_ps(velecsum,velec);
 629
 630             fscal            = felec;
 631
 632             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 633
 634             /* Calculate temporary vectorial force */
 635             tx               = _mm_mul_ps(fscal,dx00);
 636             ty               = _mm_mul_ps(fscal,dy00);
 637             tz               = _mm_mul_ps(fscal,dz00);
 638
 639             /* Update vectorial force */
 640             fix0             = _mm_add_ps(fix0,tx);
 641             fiy0             = _mm_add_ps(fiy0,ty);
 642             fiz0             = _mm_add_ps(fiz0,tz);
 643
 644             fjx0             = _mm_add_ps(fjx0,tx);
 645             fjy0             = _mm_add_ps(fjy0,ty);
 646             fjz0             = _mm_add_ps(fjz0,tz);
 647
 648             /**************************
 649              * CALCULATE INTERACTIONS *
 650              **************************/
 651
 652             /* REACTION-FIELD ELECTROSTATICS */
 653             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 654             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 655
 656             /* Update potential sum for this i atom from the interaction with this j atom. */
 657             velec            = _mm_andnot_ps(dummy_mask,velec);
 658             velecsum         = _mm_add_ps(velecsum,velec);
 659
 660             fscal            = felec;
 661
 662             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 663
 664             /* Calculate temporary vectorial force */
 665             tx               = _mm_mul_ps(fscal,dx01);
 666             ty               = _mm_mul_ps(fscal,dy01);
 667             tz               = _mm_mul_ps(fscal,dz01);
 668
 669             /* Update vectorial force */
 670             fix0             = _mm_add_ps(fix0,tx);
 671             fiy0             = _mm_add_ps(fiy0,ty);
 672             fiz0             = _mm_add_ps(fiz0,tz);
 673
 674             fjx1             = _mm_add_ps(fjx1,tx);
 675             fjy1             = _mm_add_ps(fjy1,ty);
 676             fjz1             = _mm_add_ps(fjz1,tz);
 677
 678             /**************************
 679              * CALCULATE INTERACTIONS *
 680              **************************/
 681
 682             /* REACTION-FIELD ELECTROSTATICS */
 683             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 684             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 685
 686             /* Update potential sum for this i atom from the interaction with this j atom. */
 687             velec            = _mm_andnot_ps(dummy_mask,velec);
 688             velecsum         = _mm_add_ps(velecsum,velec);
 689
 690             fscal            = felec;
 691
 692             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 693
 694             /* Calculate temporary vectorial force */
 695             tx               = _mm_mul_ps(fscal,dx02);
 696             ty               = _mm_mul_ps(fscal,dy02);
 697             tz               = _mm_mul_ps(fscal,dz02);
 698
 699             /* Update vectorial force */
 700             fix0             = _mm_add_ps(fix0,tx);
 701             fiy0             = _mm_add_ps(fiy0,ty);
 702             fiz0             = _mm_add_ps(fiz0,tz);
 703
 704             fjx2             = _mm_add_ps(fjx2,tx);
 705             fjy2             = _mm_add_ps(fjy2,ty);
 706             fjz2             = _mm_add_ps(fjz2,tz);
 707
 708             /**************************
 709              * CALCULATE INTERACTIONS *
 710              **************************/
 711
 712             /* REACTION-FIELD ELECTROSTATICS */
 713             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 714             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 715
 716             /* Update potential sum for this i atom from the interaction with this j atom. */
 717             velec            = _mm_andnot_ps(dummy_mask,velec);
 718             velecsum         = _mm_add_ps(velecsum,velec);
 719
 720             fscal            = felec;
 721
 722             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 723
 724             /* Calculate temporary vectorial force */
 725             tx               = _mm_mul_ps(fscal,dx10);
 726             ty               = _mm_mul_ps(fscal,dy10);
 727             tz               = _mm_mul_ps(fscal,dz10);
 728
 729             /* Update vectorial force */
 730             fix1             = _mm_add_ps(fix1,tx);
 731             fiy1             = _mm_add_ps(fiy1,ty);
 732             fiz1             = _mm_add_ps(fiz1,tz);
 733
 734             fjx0             = _mm_add_ps(fjx0,tx);
 735             fjy0             = _mm_add_ps(fjy0,ty);
 736             fjz0             = _mm_add_ps(fjz0,tz);
 737
 738             /**************************
 739              * CALCULATE INTERACTIONS *
 740              **************************/
 741
 742             /* REACTION-FIELD ELECTROSTATICS */
 743             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 744             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 745
 746             /* Update potential sum for this i atom from the interaction with this j atom. */
 747             velec            = _mm_andnot_ps(dummy_mask,velec);
 748             velecsum         = _mm_add_ps(velecsum,velec);
 749
 750             fscal            = felec;
 751
 752             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 753
 754             /* Calculate temporary vectorial force */
 755             tx               = _mm_mul_ps(fscal,dx11);
 756             ty               = _mm_mul_ps(fscal,dy11);
 757             tz               = _mm_mul_ps(fscal,dz11);
 758
 759             /* Update vectorial force */
 760             fix1             = _mm_add_ps(fix1,tx);
 761             fiy1             = _mm_add_ps(fiy1,ty);
 762             fiz1             = _mm_add_ps(fiz1,tz);
 763
 764             fjx1             = _mm_add_ps(fjx1,tx);
 765             fjy1             = _mm_add_ps(fjy1,ty);
 766             fjz1             = _mm_add_ps(fjz1,tz);
 767
 768             /**************************
 769              * CALCULATE INTERACTIONS *
 770              **************************/
 771
 772             /* REACTION-FIELD ELECTROSTATICS */
 773             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 774             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 775
 776             /* Update potential sum for this i atom from the interaction with this j atom. */
 777             velec            = _mm_andnot_ps(dummy_mask,velec);
 778             velecsum         = _mm_add_ps(velecsum,velec);
 779
 780             fscal            = felec;
 781
 782             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 783
 784             /* Calculate temporary vectorial force */
 785             tx               = _mm_mul_ps(fscal,dx12);
 786             ty               = _mm_mul_ps(fscal,dy12);
 787             tz               = _mm_mul_ps(fscal,dz12);
 788
 789             /* Update vectorial force */
 790             fix1             = _mm_add_ps(fix1,tx);
 791             fiy1             = _mm_add_ps(fiy1,ty);
 792             fiz1             = _mm_add_ps(fiz1,tz);
 793
 794             fjx2             = _mm_add_ps(fjx2,tx);
 795             fjy2             = _mm_add_ps(fjy2,ty);
 796             fjz2             = _mm_add_ps(fjz2,tz);
 797
 798             /**************************
 799              * CALCULATE INTERACTIONS *
 800              **************************/
 801
 802             /* REACTION-FIELD ELECTROSTATICS */
 803             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 804             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 805
 806             /* Update potential sum for this i atom from the interaction with this j atom. */
 807             velec            = _mm_andnot_ps(dummy_mask,velec);
 808             velecsum         = _mm_add_ps(velecsum,velec);
 809
 810             fscal            = felec;
 811
 812             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 813
 814             /* Calculate temporary vectorial force */
 815             tx               = _mm_mul_ps(fscal,dx20);
 816             ty               = _mm_mul_ps(fscal,dy20);
 817             tz               = _mm_mul_ps(fscal,dz20);
 818
 819             /* Update vectorial force */
 820             fix2             = _mm_add_ps(fix2,tx);
 821             fiy2             = _mm_add_ps(fiy2,ty);
 822             fiz2             = _mm_add_ps(fiz2,tz);
 823
 824             fjx0             = _mm_add_ps(fjx0,tx);
 825             fjy0             = _mm_add_ps(fjy0,ty);
 826             fjz0             = _mm_add_ps(fjz0,tz);
 827
 828             /**************************
 829              * CALCULATE INTERACTIONS *
 830              **************************/
 831
 832             /* REACTION-FIELD ELECTROSTATICS */
 833             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 834             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 835
 836             /* Update potential sum for this i atom from the interaction with this j atom. */
 837             velec            = _mm_andnot_ps(dummy_mask,velec);
 838             velecsum         = _mm_add_ps(velecsum,velec);
 839
 840             fscal            = felec;
 841
 842             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 843
 844             /* Calculate temporary vectorial force */
 845             tx               = _mm_mul_ps(fscal,dx21);
 846             ty               = _mm_mul_ps(fscal,dy21);
 847             tz               = _mm_mul_ps(fscal,dz21);
 848
 849             /* Update vectorial force */
 850             fix2             = _mm_add_ps(fix2,tx);
 851             fiy2             = _mm_add_ps(fiy2,ty);
 852             fiz2             = _mm_add_ps(fiz2,tz);
 853
 854             fjx1             = _mm_add_ps(fjx1,tx);
 855             fjy1             = _mm_add_ps(fjy1,ty);
 856             fjz1             = _mm_add_ps(fjz1,tz);
 857
 858             /**************************
 859              * CALCULATE INTERACTIONS *
 860              **************************/
 861
 862             /* REACTION-FIELD ELECTROSTATICS */
 863             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 864             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 865
 866             /* Update potential sum for this i atom from the interaction with this j atom. */
 867             velec            = _mm_andnot_ps(dummy_mask,velec);
 868             velecsum         = _mm_add_ps(velecsum,velec);
 869
 870             fscal            = felec;
 871
 872             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 873
 874             /* Calculate temporary vectorial force */
 875             tx               = _mm_mul_ps(fscal,dx22);
 876             ty               = _mm_mul_ps(fscal,dy22);
 877             tz               = _mm_mul_ps(fscal,dz22);
 878
 879             /* Update vectorial force */
 880             fix2             = _mm_add_ps(fix2,tx);
 881             fiy2             = _mm_add_ps(fiy2,ty);
 882             fiz2             = _mm_add_ps(fiz2,tz);
 883
 884             fjx2             = _mm_add_ps(fjx2,tx);
 885             fjy2             = _mm_add_ps(fjy2,ty);
 886             fjz2             = _mm_add_ps(fjz2,tz);
 887
 888             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 889             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 890             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 891             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 892
 893             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 894                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 895
 896             /* Inner loop uses 288 flops */
 897         }
 898
 899         /* End of innermost loop */
 900
 901         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 902                                               f+i_coord_offset,fshift+i_shift_offset);
 903
 904         ggid                        = gid[iidx];
 905         /* Update potential energies */
 906         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 907
 908         /* Increment number of inner iterations */
 909         inneriter                  += j_index_end - j_index_start;
 910
 911         /* Outer loop uses 19 flops */
 912     }
 913
 914     /* Increment number of outer iterations */
 915     outeriter        += nri;
 916
 917     /* Update outer/inner flops */
 918
 919     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*288);
 920 }
 921 /*
 922  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse2_single
 923  * Electrostatics interaction: ReactionField
 924  * VdW interaction:            None
 925  * Geometry:                   Water3-Water3
 926  * Calculate force/pot:        Force
 927  */
 928 void
 929 nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse2_single
 930                     (t_nblist * gmx_restrict                nlist,
 931                      rvec * gmx_restrict                    xx,
 932                      rvec * gmx_restrict                    ff,
 933                      t_forcerec * gmx_restrict              fr,
 934                      t_mdatoms * gmx_restrict               mdatoms,
 935                      nb_kernel_data_t * gmx_restrict        kernel_data,
 936                      t_nrnb * gmx_restrict                  nrnb)
 937 {
 938     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 939      * just 0 for non-waters.
 940      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 941      * jnr indices corresponding to data put in the four positions in the SIMD register.
 942      */
 943     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 944     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 945     int              jnrA,jnrB,jnrC,jnrD;
 946     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 947     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 948     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 949     real             rcutoff_scalar;
 950     real             *shiftvec,*fshift,*x,*f;
 951     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 952     real             scratch[4*DIM];
 953     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 954     int              vdwioffset0;
 955     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 956     int              vdwioffset1;
 957     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 958     int              vdwioffset2;
 959     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 960     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 961     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 962     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 963     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 964     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 965     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 966     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 967     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 968     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 969     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 970     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 971     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 972     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 973     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 974     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 975     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 976     real             *charge;
 977     __m128           dummy_mask,cutoff_mask;
 978     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 979     __m128           one     = _mm_set1_ps(1.0);
 980     __m128           two     = _mm_set1_ps(2.0);
 981     x                = xx[0];
 982     f                = ff[0];
 983
 984     nri              = nlist->nri;
 985     iinr             = nlist->iinr;
 986     jindex           = nlist->jindex;
 987     jjnr             = nlist->jjnr;
 988     shiftidx         = nlist->shift;
 989     gid              = nlist->gid;
 990     shiftvec         = fr->shift_vec[0];
 991     fshift           = fr->fshift[0];
 992     facel            = _mm_set1_ps(fr->epsfac);
 993     charge           = mdatoms->chargeA;
 994     krf              = _mm_set1_ps(fr->ic->k_rf);
 995     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 996     crf              = _mm_set1_ps(fr->ic->c_rf);
 997
 998     /* Setup water-specific parameters */
 999     inr              = nlist->iinr[0];
1000     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1001     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1002     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1003
1004     jq0              = _mm_set1_ps(charge[inr+0]);
1005     jq1              = _mm_set1_ps(charge[inr+1]);
1006     jq2              = _mm_set1_ps(charge[inr+2]);
1007     qq00             = _mm_mul_ps(iq0,jq0);
1008     qq01             = _mm_mul_ps(iq0,jq1);
1009     qq02             = _mm_mul_ps(iq0,jq2);
1010     qq10             = _mm_mul_ps(iq1,jq0);
1011     qq11             = _mm_mul_ps(iq1,jq1);
1012     qq12             = _mm_mul_ps(iq1,jq2);
1013     qq20             = _mm_mul_ps(iq2,jq0);
1014     qq21             = _mm_mul_ps(iq2,jq1);
1015     qq22             = _mm_mul_ps(iq2,jq2);
1016
1017     /* Avoid stupid compiler warnings */
1018     jnrA = jnrB = jnrC = jnrD = 0;
1019     j_coord_offsetA = 0;
1020     j_coord_offsetB = 0;
1021     j_coord_offsetC = 0;
1022     j_coord_offsetD = 0;
1023
1024     outeriter        = 0;
1025     inneriter        = 0;
1026
1027     for(iidx=0;iidx<4*DIM;iidx++)
1028     {
1029         scratch[iidx] = 0.0;
1030     }
1031
1032     /* Start outer loop over neighborlists */
1033     for(iidx=0; iidx<nri; iidx++)
1034     {
1035         /* Load shift vector for this list */
1036         i_shift_offset   = DIM*shiftidx[iidx];
1037
1038         /* Load limits for loop over neighbors */
1039         j_index_start    = jindex[iidx];
1040         j_index_end      = jindex[iidx+1];
1041
1042         /* Get outer coordinate index */
1043         inr              = iinr[iidx];
1044         i_coord_offset   = DIM*inr;
1045
1046         /* Load i particle coords and add shift vector */
1047         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1048                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1049
1050         fix0             = _mm_setzero_ps();
1051         fiy0             = _mm_setzero_ps();
1052         fiz0             = _mm_setzero_ps();
1053         fix1             = _mm_setzero_ps();
1054         fiy1             = _mm_setzero_ps();
1055         fiz1             = _mm_setzero_ps();
1056         fix2             = _mm_setzero_ps();
1057         fiy2             = _mm_setzero_ps();
1058         fiz2             = _mm_setzero_ps();
1059
1060         /* Start inner kernel loop */
1061         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1062         {
1063
1064             /* Get j neighbor index, and coordinate index */
1065             jnrA             = jjnr[jidx];
1066             jnrB             = jjnr[jidx+1];
1067             jnrC             = jjnr[jidx+2];
1068             jnrD             = jjnr[jidx+3];
1069             j_coord_offsetA  = DIM*jnrA;
1070             j_coord_offsetB  = DIM*jnrB;
1071             j_coord_offsetC  = DIM*jnrC;
1072             j_coord_offsetD  = DIM*jnrD;
1073
1074             /* load j atom coordinates */
1075             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1076                                               x+j_coord_offsetC,x+j_coord_offsetD,
1077                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1078
1079             /* Calculate displacement vector */
1080             dx00             = _mm_sub_ps(ix0,jx0);
1081             dy00             = _mm_sub_ps(iy0,jy0);
1082             dz00             = _mm_sub_ps(iz0,jz0);
1083             dx01             = _mm_sub_ps(ix0,jx1);
1084             dy01             = _mm_sub_ps(iy0,jy1);
1085             dz01             = _mm_sub_ps(iz0,jz1);
1086             dx02             = _mm_sub_ps(ix0,jx2);
1087             dy02             = _mm_sub_ps(iy0,jy2);
1088             dz02             = _mm_sub_ps(iz0,jz2);
1089             dx10             = _mm_sub_ps(ix1,jx0);
1090             dy10             = _mm_sub_ps(iy1,jy0);
1091             dz10             = _mm_sub_ps(iz1,jz0);
1092             dx11             = _mm_sub_ps(ix1,jx1);
1093             dy11             = _mm_sub_ps(iy1,jy1);
1094             dz11             = _mm_sub_ps(iz1,jz1);
1095             dx12             = _mm_sub_ps(ix1,jx2);
1096             dy12             = _mm_sub_ps(iy1,jy2);
1097             dz12             = _mm_sub_ps(iz1,jz2);
1098             dx20             = _mm_sub_ps(ix2,jx0);
1099             dy20             = _mm_sub_ps(iy2,jy0);
1100             dz20             = _mm_sub_ps(iz2,jz0);
1101             dx21             = _mm_sub_ps(ix2,jx1);
1102             dy21             = _mm_sub_ps(iy2,jy1);
1103             dz21             = _mm_sub_ps(iz2,jz1);
1104             dx22             = _mm_sub_ps(ix2,jx2);
1105             dy22             = _mm_sub_ps(iy2,jy2);
1106             dz22             = _mm_sub_ps(iz2,jz2);
1107
1108             /* Calculate squared distance and things based on it */
1109             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1110             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1111             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1112             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1113             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1114             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1115             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1116             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1117             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1118
1119             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1120             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1121             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1122             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1123             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1124             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1125             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1126             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1127             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1128
1129             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1130             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1131             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1132             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1133             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1134             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1135             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1136             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1137             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1138
1139             fjx0             = _mm_setzero_ps();
1140             fjy0             = _mm_setzero_ps();
1141             fjz0             = _mm_setzero_ps();
1142             fjx1             = _mm_setzero_ps();
1143             fjy1             = _mm_setzero_ps();
1144             fjz1             = _mm_setzero_ps();
1145             fjx2             = _mm_setzero_ps();
1146             fjy2             = _mm_setzero_ps();
1147             fjz2             = _mm_setzero_ps();
1148
1149             /**************************
1150              * CALCULATE INTERACTIONS *
1151              **************************/
1152
1153             /* REACTION-FIELD ELECTROSTATICS */
1154             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1155
1156             fscal            = felec;
1157
1158             /* Calculate temporary vectorial force */
1159             tx               = _mm_mul_ps(fscal,dx00);
1160             ty               = _mm_mul_ps(fscal,dy00);
1161             tz               = _mm_mul_ps(fscal,dz00);
1162
1163             /* Update vectorial force */
1164             fix0             = _mm_add_ps(fix0,tx);
1165             fiy0             = _mm_add_ps(fiy0,ty);
1166             fiz0             = _mm_add_ps(fiz0,tz);
1167
1168             fjx0             = _mm_add_ps(fjx0,tx);
1169             fjy0             = _mm_add_ps(fjy0,ty);
1170             fjz0             = _mm_add_ps(fjz0,tz);
1171
1172             /**************************
1173              * CALCULATE INTERACTIONS *
1174              **************************/
1175
1176             /* REACTION-FIELD ELECTROSTATICS */
1177             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1178
1179             fscal            = felec;
1180
1181             /* Calculate temporary vectorial force */
1182             tx               = _mm_mul_ps(fscal,dx01);
1183             ty               = _mm_mul_ps(fscal,dy01);
1184             tz               = _mm_mul_ps(fscal,dz01);
1185
1186             /* Update vectorial force */
1187             fix0             = _mm_add_ps(fix0,tx);
1188             fiy0             = _mm_add_ps(fiy0,ty);
1189             fiz0             = _mm_add_ps(fiz0,tz);
1190
1191             fjx1             = _mm_add_ps(fjx1,tx);
1192             fjy1             = _mm_add_ps(fjy1,ty);
1193             fjz1             = _mm_add_ps(fjz1,tz);
1194
1195             /**************************
1196              * CALCULATE INTERACTIONS *
1197              **************************/
1198
1199             /* REACTION-FIELD ELECTROSTATICS */
1200             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1201
1202             fscal            = felec;
1203
1204             /* Calculate temporary vectorial force */
1205             tx               = _mm_mul_ps(fscal,dx02);
1206             ty               = _mm_mul_ps(fscal,dy02);
1207             tz               = _mm_mul_ps(fscal,dz02);
1208
1209             /* Update vectorial force */
1210             fix0             = _mm_add_ps(fix0,tx);
1211             fiy0             = _mm_add_ps(fiy0,ty);
1212             fiz0             = _mm_add_ps(fiz0,tz);
1213
1214             fjx2             = _mm_add_ps(fjx2,tx);
1215             fjy2             = _mm_add_ps(fjy2,ty);
1216             fjz2             = _mm_add_ps(fjz2,tz);
1217
1218             /**************************
1219              * CALCULATE INTERACTIONS *
1220              **************************/
1221
1222             /* REACTION-FIELD ELECTROSTATICS */
1223             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1224
1225             fscal            = felec;
1226
1227             /* Calculate temporary vectorial force */
1228             tx               = _mm_mul_ps(fscal,dx10);
1229             ty               = _mm_mul_ps(fscal,dy10);
1230             tz               = _mm_mul_ps(fscal,dz10);
1231
1232             /* Update vectorial force */
1233             fix1             = _mm_add_ps(fix1,tx);
1234             fiy1             = _mm_add_ps(fiy1,ty);
1235             fiz1             = _mm_add_ps(fiz1,tz);
1236
1237             fjx0             = _mm_add_ps(fjx0,tx);
1238             fjy0             = _mm_add_ps(fjy0,ty);
1239             fjz0             = _mm_add_ps(fjz0,tz);
1240
1241             /**************************
1242              * CALCULATE INTERACTIONS *
1243              **************************/
1244
1245             /* REACTION-FIELD ELECTROSTATICS */
1246             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1247
1248             fscal            = felec;
1249
1250             /* Calculate temporary vectorial force */
1251             tx               = _mm_mul_ps(fscal,dx11);
1252             ty               = _mm_mul_ps(fscal,dy11);
1253             tz               = _mm_mul_ps(fscal,dz11);
1254
1255             /* Update vectorial force */
1256             fix1             = _mm_add_ps(fix1,tx);
1257             fiy1             = _mm_add_ps(fiy1,ty);
1258             fiz1             = _mm_add_ps(fiz1,tz);
1259
1260             fjx1             = _mm_add_ps(fjx1,tx);
1261             fjy1             = _mm_add_ps(fjy1,ty);
1262             fjz1             = _mm_add_ps(fjz1,tz);
1263
1264             /**************************
1265              * CALCULATE INTERACTIONS *
1266              **************************/
1267
1268             /* REACTION-FIELD ELECTROSTATICS */
1269             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1270
1271             fscal            = felec;
1272
1273             /* Calculate temporary vectorial force */
1274             tx               = _mm_mul_ps(fscal,dx12);
1275             ty               = _mm_mul_ps(fscal,dy12);
1276             tz               = _mm_mul_ps(fscal,dz12);
1277
1278             /* Update vectorial force */
1279             fix1             = _mm_add_ps(fix1,tx);
1280             fiy1             = _mm_add_ps(fiy1,ty);
1281             fiz1             = _mm_add_ps(fiz1,tz);
1282
1283             fjx2             = _mm_add_ps(fjx2,tx);
1284             fjy2             = _mm_add_ps(fjy2,ty);
1285             fjz2             = _mm_add_ps(fjz2,tz);
1286
1287             /**************************
1288              * CALCULATE INTERACTIONS *
1289              **************************/
1290
1291             /* REACTION-FIELD ELECTROSTATICS */
1292             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1293
1294             fscal            = felec;
1295
1296             /* Calculate temporary vectorial force */
1297             tx               = _mm_mul_ps(fscal,dx20);
1298             ty               = _mm_mul_ps(fscal,dy20);
1299             tz               = _mm_mul_ps(fscal,dz20);
1300
1301             /* Update vectorial force */
1302             fix2             = _mm_add_ps(fix2,tx);
1303             fiy2             = _mm_add_ps(fiy2,ty);
1304             fiz2             = _mm_add_ps(fiz2,tz);
1305
1306             fjx0             = _mm_add_ps(fjx0,tx);
1307             fjy0             = _mm_add_ps(fjy0,ty);
1308             fjz0             = _mm_add_ps(fjz0,tz);
1309
1310             /**************************
1311              * CALCULATE INTERACTIONS *
1312              **************************/
1313
1314             /* REACTION-FIELD ELECTROSTATICS */
1315             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1316
1317             fscal            = felec;
1318
1319             /* Calculate temporary vectorial force */
1320             tx               = _mm_mul_ps(fscal,dx21);
1321             ty               = _mm_mul_ps(fscal,dy21);
1322             tz               = _mm_mul_ps(fscal,dz21);
1323
1324             /* Update vectorial force */
1325             fix2             = _mm_add_ps(fix2,tx);
1326             fiy2             = _mm_add_ps(fiy2,ty);
1327             fiz2             = _mm_add_ps(fiz2,tz);
1328
1329             fjx1             = _mm_add_ps(fjx1,tx);
1330             fjy1             = _mm_add_ps(fjy1,ty);
1331             fjz1             = _mm_add_ps(fjz1,tz);
1332
1333             /**************************
1334              * CALCULATE INTERACTIONS *
1335              **************************/
1336
1337             /* REACTION-FIELD ELECTROSTATICS */
1338             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1339
1340             fscal            = felec;
1341
1342             /* Calculate temporary vectorial force */
1343             tx               = _mm_mul_ps(fscal,dx22);
1344             ty               = _mm_mul_ps(fscal,dy22);
1345             tz               = _mm_mul_ps(fscal,dz22);
1346
1347             /* Update vectorial force */
1348             fix2             = _mm_add_ps(fix2,tx);
1349             fiy2             = _mm_add_ps(fiy2,ty);
1350             fiz2             = _mm_add_ps(fiz2,tz);
1351
1352             fjx2             = _mm_add_ps(fjx2,tx);
1353             fjy2             = _mm_add_ps(fjy2,ty);
1354             fjz2             = _mm_add_ps(fjz2,tz);
1355
1356             fjptrA             = f+j_coord_offsetA;
1357             fjptrB             = f+j_coord_offsetB;
1358             fjptrC             = f+j_coord_offsetC;
1359             fjptrD             = f+j_coord_offsetD;
1360
1361             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1362                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1363
1364             /* Inner loop uses 243 flops */
1365         }
1366
1367         if(jidx<j_index_end)
1368         {
1369
1370             /* Get j neighbor index, and coordinate index */
1371             jnrlistA         = jjnr[jidx];
1372             jnrlistB         = jjnr[jidx+1];
1373             jnrlistC         = jjnr[jidx+2];
1374             jnrlistD         = jjnr[jidx+3];
1375             /* Sign of each element will be negative for non-real atoms.
1376              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1377              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1378              */
1379             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1380             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1381             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1382             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1383             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1384             j_coord_offsetA  = DIM*jnrA;
1385             j_coord_offsetB  = DIM*jnrB;
1386             j_coord_offsetC  = DIM*jnrC;
1387             j_coord_offsetD  = DIM*jnrD;
1388
1389             /* load j atom coordinates */
1390             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1391                                               x+j_coord_offsetC,x+j_coord_offsetD,
1392                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1393
1394             /* Calculate displacement vector */
1395             dx00             = _mm_sub_ps(ix0,jx0);
1396             dy00             = _mm_sub_ps(iy0,jy0);
1397             dz00             = _mm_sub_ps(iz0,jz0);
1398             dx01             = _mm_sub_ps(ix0,jx1);
1399             dy01             = _mm_sub_ps(iy0,jy1);
1400             dz01             = _mm_sub_ps(iz0,jz1);
1401             dx02             = _mm_sub_ps(ix0,jx2);
1402             dy02             = _mm_sub_ps(iy0,jy2);
1403             dz02             = _mm_sub_ps(iz0,jz2);
1404             dx10             = _mm_sub_ps(ix1,jx0);
1405             dy10             = _mm_sub_ps(iy1,jy0);
1406             dz10             = _mm_sub_ps(iz1,jz0);
1407             dx11             = _mm_sub_ps(ix1,jx1);
1408             dy11             = _mm_sub_ps(iy1,jy1);
1409             dz11             = _mm_sub_ps(iz1,jz1);
1410             dx12             = _mm_sub_ps(ix1,jx2);
1411             dy12             = _mm_sub_ps(iy1,jy2);
1412             dz12             = _mm_sub_ps(iz1,jz2);
1413             dx20             = _mm_sub_ps(ix2,jx0);
1414             dy20             = _mm_sub_ps(iy2,jy0);
1415             dz20             = _mm_sub_ps(iz2,jz0);
1416             dx21             = _mm_sub_ps(ix2,jx1);
1417             dy21             = _mm_sub_ps(iy2,jy1);
1418             dz21             = _mm_sub_ps(iz2,jz1);
1419             dx22             = _mm_sub_ps(ix2,jx2);
1420             dy22             = _mm_sub_ps(iy2,jy2);
1421             dz22             = _mm_sub_ps(iz2,jz2);
1422
1423             /* Calculate squared distance and things based on it */
1424             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1425             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1426             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1427             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1428             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1429             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1430             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1431             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1432             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1433
1434             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1435             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1436             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1437             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1438             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1439             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1440             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1441             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1442             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1443
1444             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1445             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1446             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1447             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1448             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1449             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1450             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1451             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1452             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1453
1454             fjx0             = _mm_setzero_ps();
1455             fjy0             = _mm_setzero_ps();
1456             fjz0             = _mm_setzero_ps();
1457             fjx1             = _mm_setzero_ps();
1458             fjy1             = _mm_setzero_ps();
1459             fjz1             = _mm_setzero_ps();
1460             fjx2             = _mm_setzero_ps();
1461             fjy2             = _mm_setzero_ps();
1462             fjz2             = _mm_setzero_ps();
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             /* REACTION-FIELD ELECTROSTATICS */
1469             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1470
1471             fscal            = felec;
1472
1473             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1474
1475             /* Calculate temporary vectorial force */
1476             tx               = _mm_mul_ps(fscal,dx00);
1477             ty               = _mm_mul_ps(fscal,dy00);
1478             tz               = _mm_mul_ps(fscal,dz00);
1479
1480             /* Update vectorial force */
1481             fix0             = _mm_add_ps(fix0,tx);
1482             fiy0             = _mm_add_ps(fiy0,ty);
1483             fiz0             = _mm_add_ps(fiz0,tz);
1484
1485             fjx0             = _mm_add_ps(fjx0,tx);
1486             fjy0             = _mm_add_ps(fjy0,ty);
1487             fjz0             = _mm_add_ps(fjz0,tz);
1488
1489             /**************************
1490              * CALCULATE INTERACTIONS *
1491              **************************/
1492
1493             /* REACTION-FIELD ELECTROSTATICS */
1494             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1495
1496             fscal            = felec;
1497
1498             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1499
1500             /* Calculate temporary vectorial force */
1501             tx               = _mm_mul_ps(fscal,dx01);
1502             ty               = _mm_mul_ps(fscal,dy01);
1503             tz               = _mm_mul_ps(fscal,dz01);
1504
1505             /* Update vectorial force */
1506             fix0             = _mm_add_ps(fix0,tx);
1507             fiy0             = _mm_add_ps(fiy0,ty);
1508             fiz0             = _mm_add_ps(fiz0,tz);
1509
1510             fjx1             = _mm_add_ps(fjx1,tx);
1511             fjy1             = _mm_add_ps(fjy1,ty);
1512             fjz1             = _mm_add_ps(fjz1,tz);
1513
1514             /**************************
1515              * CALCULATE INTERACTIONS *
1516              **************************/
1517
1518             /* REACTION-FIELD ELECTROSTATICS */
1519             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1520
1521             fscal            = felec;
1522
1523             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1524
1525             /* Calculate temporary vectorial force */
1526             tx               = _mm_mul_ps(fscal,dx02);
1527             ty               = _mm_mul_ps(fscal,dy02);
1528             tz               = _mm_mul_ps(fscal,dz02);
1529
1530             /* Update vectorial force */
1531             fix0             = _mm_add_ps(fix0,tx);
1532             fiy0             = _mm_add_ps(fiy0,ty);
1533             fiz0             = _mm_add_ps(fiz0,tz);
1534
1535             fjx2             = _mm_add_ps(fjx2,tx);
1536             fjy2             = _mm_add_ps(fjy2,ty);
1537             fjz2             = _mm_add_ps(fjz2,tz);
1538
1539             /**************************
1540              * CALCULATE INTERACTIONS *
1541              **************************/
1542
1543             /* REACTION-FIELD ELECTROSTATICS */
1544             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1545
1546             fscal            = felec;
1547
1548             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1549
1550             /* Calculate temporary vectorial force */
1551             tx               = _mm_mul_ps(fscal,dx10);
1552             ty               = _mm_mul_ps(fscal,dy10);
1553             tz               = _mm_mul_ps(fscal,dz10);
1554
1555             /* Update vectorial force */
1556             fix1             = _mm_add_ps(fix1,tx);
1557             fiy1             = _mm_add_ps(fiy1,ty);
1558             fiz1             = _mm_add_ps(fiz1,tz);
1559
1560             fjx0             = _mm_add_ps(fjx0,tx);
1561             fjy0             = _mm_add_ps(fjy0,ty);
1562             fjz0             = _mm_add_ps(fjz0,tz);
1563
1564             /**************************
1565              * CALCULATE INTERACTIONS *
1566              **************************/
1567
1568             /* REACTION-FIELD ELECTROSTATICS */
1569             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1570
1571             fscal            = felec;
1572
1573             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1574
1575             /* Calculate temporary vectorial force */
1576             tx               = _mm_mul_ps(fscal,dx11);
1577             ty               = _mm_mul_ps(fscal,dy11);
1578             tz               = _mm_mul_ps(fscal,dz11);
1579
1580             /* Update vectorial force */
1581             fix1             = _mm_add_ps(fix1,tx);
1582             fiy1             = _mm_add_ps(fiy1,ty);
1583             fiz1             = _mm_add_ps(fiz1,tz);
1584
1585             fjx1             = _mm_add_ps(fjx1,tx);
1586             fjy1             = _mm_add_ps(fjy1,ty);
1587             fjz1             = _mm_add_ps(fjz1,tz);
1588
1589             /**************************
1590              * CALCULATE INTERACTIONS *
1591              **************************/
1592
1593             /* REACTION-FIELD ELECTROSTATICS */
1594             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1595
1596             fscal            = felec;
1597
1598             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1599
1600             /* Calculate temporary vectorial force */
1601             tx               = _mm_mul_ps(fscal,dx12);
1602             ty               = _mm_mul_ps(fscal,dy12);
1603             tz               = _mm_mul_ps(fscal,dz12);
1604
1605             /* Update vectorial force */
1606             fix1             = _mm_add_ps(fix1,tx);
1607             fiy1             = _mm_add_ps(fiy1,ty);
1608             fiz1             = _mm_add_ps(fiz1,tz);
1609
1610             fjx2             = _mm_add_ps(fjx2,tx);
1611             fjy2             = _mm_add_ps(fjy2,ty);
1612             fjz2             = _mm_add_ps(fjz2,tz);
1613
1614             /**************************
1615              * CALCULATE INTERACTIONS *
1616              **************************/
1617
1618             /* REACTION-FIELD ELECTROSTATICS */
1619             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1620
1621             fscal            = felec;
1622
1623             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1624
1625             /* Calculate temporary vectorial force */
1626             tx               = _mm_mul_ps(fscal,dx20);
1627             ty               = _mm_mul_ps(fscal,dy20);
1628             tz               = _mm_mul_ps(fscal,dz20);
1629
1630             /* Update vectorial force */
1631             fix2             = _mm_add_ps(fix2,tx);
1632             fiy2             = _mm_add_ps(fiy2,ty);
1633             fiz2             = _mm_add_ps(fiz2,tz);
1634
1635             fjx0             = _mm_add_ps(fjx0,tx);
1636             fjy0             = _mm_add_ps(fjy0,ty);
1637             fjz0             = _mm_add_ps(fjz0,tz);
1638
1639             /**************************
1640              * CALCULATE INTERACTIONS *
1641              **************************/
1642
1643             /* REACTION-FIELD ELECTROSTATICS */
1644             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1645
1646             fscal            = felec;
1647
1648             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1649
1650             /* Calculate temporary vectorial force */
1651             tx               = _mm_mul_ps(fscal,dx21);
1652             ty               = _mm_mul_ps(fscal,dy21);
1653             tz               = _mm_mul_ps(fscal,dz21);
1654
1655             /* Update vectorial force */
1656             fix2             = _mm_add_ps(fix2,tx);
1657             fiy2             = _mm_add_ps(fiy2,ty);
1658             fiz2             = _mm_add_ps(fiz2,tz);
1659
1660             fjx1             = _mm_add_ps(fjx1,tx);
1661             fjy1             = _mm_add_ps(fjy1,ty);
1662             fjz1             = _mm_add_ps(fjz1,tz);
1663
1664             /**************************
1665              * CALCULATE INTERACTIONS *
1666              **************************/
1667
1668             /* REACTION-FIELD ELECTROSTATICS */
1669             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1670
1671             fscal            = felec;
1672
1673             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1674
1675             /* Calculate temporary vectorial force */
1676             tx               = _mm_mul_ps(fscal,dx22);
1677             ty               = _mm_mul_ps(fscal,dy22);
1678             tz               = _mm_mul_ps(fscal,dz22);
1679
1680             /* Update vectorial force */
1681             fix2             = _mm_add_ps(fix2,tx);
1682             fiy2             = _mm_add_ps(fiy2,ty);
1683             fiz2             = _mm_add_ps(fiz2,tz);
1684
1685             fjx2             = _mm_add_ps(fjx2,tx);
1686             fjy2             = _mm_add_ps(fjy2,ty);
1687             fjz2             = _mm_add_ps(fjz2,tz);
1688
1689             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1690             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1691             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1692             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1693
1694             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1695                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1696
1697             /* Inner loop uses 243 flops */
1698         }
1699
1700         /* End of innermost loop */
1701
1702         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1703                                               f+i_coord_offset,fshift+i_shift_offset);
1704
1705         /* Increment number of inner iterations */
1706         inneriter                  += j_index_end - j_index_start;
1707
1708         /* Outer loop uses 18 flops */
1709     }
1710
1711     /* Increment number of outer iterations */
1712     outeriter        += nri;
1713
1714     /* Update outer/inner flops */
1715
1716     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*243);
1717 }