src/gromacs/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              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     __m128           dummy_mask,cutoff_mask;
  90     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  91     __m128           one     = _mm_set1_ps(1.0);
  92     __m128           two     = _mm_set1_ps(2.0);
  93     x                = xx[0];
  94     f                = ff[0];
  95
  96     nri              = nlist->nri;
  97     iinr             = nlist->iinr;
  98     jindex           = nlist->jindex;
  99     jjnr             = nlist->jjnr;
 100     shiftidx         = nlist->shift;
 101     gid              = nlist->gid;
 102     shiftvec         = fr->shift_vec[0];
 103     fshift           = fr->fshift[0];
 104     facel            = _mm_set1_ps(fr->epsfac);
 105     charge           = mdatoms->chargeA;
 106     krf              = _mm_set1_ps(fr->ic->k_rf);
 107     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 108     crf              = _mm_set1_ps(fr->ic->c_rf);
 109
 110     /* Setup water-specific parameters */
 111     inr              = nlist->iinr[0];
 112     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 113     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 114     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 115
 116     jq0              = _mm_set1_ps(charge[inr+0]);
 117     jq1              = _mm_set1_ps(charge[inr+1]);
 118     jq2              = _mm_set1_ps(charge[inr+2]);
 119     qq00             = _mm_mul_ps(iq0,jq0);
 120     qq01             = _mm_mul_ps(iq0,jq1);
 121     qq02             = _mm_mul_ps(iq0,jq2);
 122     qq10             = _mm_mul_ps(iq1,jq0);
 123     qq11             = _mm_mul_ps(iq1,jq1);
 124     qq12             = _mm_mul_ps(iq1,jq2);
 125     qq20             = _mm_mul_ps(iq2,jq0);
 126     qq21             = _mm_mul_ps(iq2,jq1);
 127     qq22             = _mm_mul_ps(iq2,jq2);
 128
 129     /* Avoid stupid compiler warnings */
 130     jnrA = jnrB = jnrC = jnrD = 0;
 131     j_coord_offsetA = 0;
 132     j_coord_offsetB = 0;
 133     j_coord_offsetC = 0;
 134     j_coord_offsetD = 0;
 135
 136     outeriter        = 0;
 137     inneriter        = 0;
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144         shX              = shiftvec[i_shift_offset+XX];
 145         shY              = shiftvec[i_shift_offset+YY];
 146         shZ              = shiftvec[i_shift_offset+ZZ];
 147
 148         /* Load limits for loop over neighbors */
 149         j_index_start    = jindex[iidx];
 150         j_index_end      = jindex[iidx+1];
 151
 152         /* Get outer coordinate index */
 153         inr              = iinr[iidx];
 154         i_coord_offset   = DIM*inr;
 155
 156         /* Load i particle coords and add shift vector */
 157         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
 158         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
 159         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
 160         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
 161         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
 162         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
 163         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
 164         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
 165         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
 166
 167         fix0             = _mm_setzero_ps();
 168         fiy0             = _mm_setzero_ps();
 169         fiz0             = _mm_setzero_ps();
 170         fix1             = _mm_setzero_ps();
 171         fiy1             = _mm_setzero_ps();
 172         fiz1             = _mm_setzero_ps();
 173         fix2             = _mm_setzero_ps();
 174         fiy2             = _mm_setzero_ps();
 175         fiz2             = _mm_setzero_ps();
 176
 177         /* Reset potential sums */
 178         velecsum         = _mm_setzero_ps();
 179
 180         /* Start inner kernel loop */
 181         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 182         {
 183
 184             /* Get j neighbor index, and coordinate index */
 185             jnrA             = jjnr[jidx];
 186             jnrB             = jjnr[jidx+1];
 187             jnrC             = jjnr[jidx+2];
 188             jnrD             = jjnr[jidx+3];
 189
 190             j_coord_offsetA  = DIM*jnrA;
 191             j_coord_offsetB  = DIM*jnrB;
 192             j_coord_offsetC  = DIM*jnrC;
 193             j_coord_offsetD  = DIM*jnrD;
 194
 195             /* load j atom coordinates */
 196             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 197                                               x+j_coord_offsetC,x+j_coord_offsetD,
 198                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 199
 200             /* Calculate displacement vector */
 201             dx00             = _mm_sub_ps(ix0,jx0);
 202             dy00             = _mm_sub_ps(iy0,jy0);
 203             dz00             = _mm_sub_ps(iz0,jz0);
 204             dx01             = _mm_sub_ps(ix0,jx1);
 205             dy01             = _mm_sub_ps(iy0,jy1);
 206             dz01             = _mm_sub_ps(iz0,jz1);
 207             dx02             = _mm_sub_ps(ix0,jx2);
 208             dy02             = _mm_sub_ps(iy0,jy2);
 209             dz02             = _mm_sub_ps(iz0,jz2);
 210             dx10             = _mm_sub_ps(ix1,jx0);
 211             dy10             = _mm_sub_ps(iy1,jy0);
 212             dz10             = _mm_sub_ps(iz1,jz0);
 213             dx11             = _mm_sub_ps(ix1,jx1);
 214             dy11             = _mm_sub_ps(iy1,jy1);
 215             dz11             = _mm_sub_ps(iz1,jz1);
 216             dx12             = _mm_sub_ps(ix1,jx2);
 217             dy12             = _mm_sub_ps(iy1,jy2);
 218             dz12             = _mm_sub_ps(iz1,jz2);
 219             dx20             = _mm_sub_ps(ix2,jx0);
 220             dy20             = _mm_sub_ps(iy2,jy0);
 221             dz20             = _mm_sub_ps(iz2,jz0);
 222             dx21             = _mm_sub_ps(ix2,jx1);
 223             dy21             = _mm_sub_ps(iy2,jy1);
 224             dz21             = _mm_sub_ps(iz2,jz1);
 225             dx22             = _mm_sub_ps(ix2,jx2);
 226             dy22             = _mm_sub_ps(iy2,jy2);
 227             dz22             = _mm_sub_ps(iz2,jz2);
 228
 229             /* Calculate squared distance and things based on it */
 230             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 231             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 232             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 233             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 234             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 235             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 236             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 237             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 238             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 239
 240             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 241             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 242             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 243             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 244             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 245             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 246             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 247             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 248             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 249
 250             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 251             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 252             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 253             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 254             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 255             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 256             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 257             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 258             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 259
 260             fjx0             = _mm_setzero_ps();
 261             fjy0             = _mm_setzero_ps();
 262             fjz0             = _mm_setzero_ps();
 263             fjx1             = _mm_setzero_ps();
 264             fjy1             = _mm_setzero_ps();
 265             fjz1             = _mm_setzero_ps();
 266             fjx2             = _mm_setzero_ps();
 267             fjy2             = _mm_setzero_ps();
 268             fjz2             = _mm_setzero_ps();
 269
 270             /**************************
 271              * CALCULATE INTERACTIONS *
 272              **************************/
 273
 274             /* REACTION-FIELD ELECTROSTATICS */
 275             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 276             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 277
 278             /* Update potential sum for this i atom from the interaction with this j atom. */
 279             velecsum         = _mm_add_ps(velecsum,velec);
 280
 281             fscal            = felec;
 282
 283             /* Calculate temporary vectorial force */
 284             tx               = _mm_mul_ps(fscal,dx00);
 285             ty               = _mm_mul_ps(fscal,dy00);
 286             tz               = _mm_mul_ps(fscal,dz00);
 287
 288             /* Update vectorial force */
 289             fix0             = _mm_add_ps(fix0,tx);
 290             fiy0             = _mm_add_ps(fiy0,ty);
 291             fiz0             = _mm_add_ps(fiz0,tz);
 292
 293             fjx0             = _mm_add_ps(fjx0,tx);
 294             fjy0             = _mm_add_ps(fjy0,ty);
 295             fjz0             = _mm_add_ps(fjz0,tz);
 296
 297             /**************************
 298              * CALCULATE INTERACTIONS *
 299              **************************/
 300
 301             /* REACTION-FIELD ELECTROSTATICS */
 302             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 303             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 304
 305             /* Update potential sum for this i atom from the interaction with this j atom. */
 306             velecsum         = _mm_add_ps(velecsum,velec);
 307
 308             fscal            = felec;
 309
 310             /* Calculate temporary vectorial force */
 311             tx               = _mm_mul_ps(fscal,dx01);
 312             ty               = _mm_mul_ps(fscal,dy01);
 313             tz               = _mm_mul_ps(fscal,dz01);
 314
 315             /* Update vectorial force */
 316             fix0             = _mm_add_ps(fix0,tx);
 317             fiy0             = _mm_add_ps(fiy0,ty);
 318             fiz0             = _mm_add_ps(fiz0,tz);
 319
 320             fjx1             = _mm_add_ps(fjx1,tx);
 321             fjy1             = _mm_add_ps(fjy1,ty);
 322             fjz1             = _mm_add_ps(fjz1,tz);
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             /* REACTION-FIELD ELECTROSTATICS */
 329             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 330             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 331
 332             /* Update potential sum for this i atom from the interaction with this j atom. */
 333             velecsum         = _mm_add_ps(velecsum,velec);
 334
 335             fscal            = felec;
 336
 337             /* Calculate temporary vectorial force */
 338             tx               = _mm_mul_ps(fscal,dx02);
 339             ty               = _mm_mul_ps(fscal,dy02);
 340             tz               = _mm_mul_ps(fscal,dz02);
 341
 342             /* Update vectorial force */
 343             fix0             = _mm_add_ps(fix0,tx);
 344             fiy0             = _mm_add_ps(fiy0,ty);
 345             fiz0             = _mm_add_ps(fiz0,tz);
 346
 347             fjx2             = _mm_add_ps(fjx2,tx);
 348             fjy2             = _mm_add_ps(fjy2,ty);
 349             fjz2             = _mm_add_ps(fjz2,tz);
 350
 351             /**************************
 352              * CALCULATE INTERACTIONS *
 353              **************************/
 354
 355             /* REACTION-FIELD ELECTROSTATICS */
 356             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 357             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             velecsum         = _mm_add_ps(velecsum,velec);
 361
 362             fscal            = felec;
 363
 364             /* Calculate temporary vectorial force */
 365             tx               = _mm_mul_ps(fscal,dx10);
 366             ty               = _mm_mul_ps(fscal,dy10);
 367             tz               = _mm_mul_ps(fscal,dz10);
 368
 369             /* Update vectorial force */
 370             fix1             = _mm_add_ps(fix1,tx);
 371             fiy1             = _mm_add_ps(fiy1,ty);
 372             fiz1             = _mm_add_ps(fiz1,tz);
 373
 374             fjx0             = _mm_add_ps(fjx0,tx);
 375             fjy0             = _mm_add_ps(fjy0,ty);
 376             fjz0             = _mm_add_ps(fjz0,tz);
 377
 378             /**************************
 379              * CALCULATE INTERACTIONS *
 380              **************************/
 381
 382             /* REACTION-FIELD ELECTROSTATICS */
 383             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 384             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 385
 386             /* Update potential sum for this i atom from the interaction with this j atom. */
 387             velecsum         = _mm_add_ps(velecsum,velec);
 388
 389             fscal            = felec;
 390
 391             /* Calculate temporary vectorial force */
 392             tx               = _mm_mul_ps(fscal,dx11);
 393             ty               = _mm_mul_ps(fscal,dy11);
 394             tz               = _mm_mul_ps(fscal,dz11);
 395
 396             /* Update vectorial force */
 397             fix1             = _mm_add_ps(fix1,tx);
 398             fiy1             = _mm_add_ps(fiy1,ty);
 399             fiz1             = _mm_add_ps(fiz1,tz);
 400
 401             fjx1             = _mm_add_ps(fjx1,tx);
 402             fjy1             = _mm_add_ps(fjy1,ty);
 403             fjz1             = _mm_add_ps(fjz1,tz);
 404
 405             /**************************
 406              * CALCULATE INTERACTIONS *
 407              **************************/
 408
 409             /* REACTION-FIELD ELECTROSTATICS */
 410             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 411             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 412
 413             /* Update potential sum for this i atom from the interaction with this j atom. */
 414             velecsum         = _mm_add_ps(velecsum,velec);
 415
 416             fscal            = felec;
 417
 418             /* Calculate temporary vectorial force */
 419             tx               = _mm_mul_ps(fscal,dx12);
 420             ty               = _mm_mul_ps(fscal,dy12);
 421             tz               = _mm_mul_ps(fscal,dz12);
 422
 423             /* Update vectorial force */
 424             fix1             = _mm_add_ps(fix1,tx);
 425             fiy1             = _mm_add_ps(fiy1,ty);
 426             fiz1             = _mm_add_ps(fiz1,tz);
 427
 428             fjx2             = _mm_add_ps(fjx2,tx);
 429             fjy2             = _mm_add_ps(fjy2,ty);
 430             fjz2             = _mm_add_ps(fjz2,tz);
 431
 432             /**************************
 433              * CALCULATE INTERACTIONS *
 434              **************************/
 435
 436             /* REACTION-FIELD ELECTROSTATICS */
 437             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 438             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 439
 440             /* Update potential sum for this i atom from the interaction with this j atom. */
 441             velecsum         = _mm_add_ps(velecsum,velec);
 442
 443             fscal            = felec;
 444
 445             /* Calculate temporary vectorial force */
 446             tx               = _mm_mul_ps(fscal,dx20);
 447             ty               = _mm_mul_ps(fscal,dy20);
 448             tz               = _mm_mul_ps(fscal,dz20);
 449
 450             /* Update vectorial force */
 451             fix2             = _mm_add_ps(fix2,tx);
 452             fiy2             = _mm_add_ps(fiy2,ty);
 453             fiz2             = _mm_add_ps(fiz2,tz);
 454
 455             fjx0             = _mm_add_ps(fjx0,tx);
 456             fjy0             = _mm_add_ps(fjy0,ty);
 457             fjz0             = _mm_add_ps(fjz0,tz);
 458
 459             /**************************
 460              * CALCULATE INTERACTIONS *
 461              **************************/
 462
 463             /* REACTION-FIELD ELECTROSTATICS */
 464             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 465             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 466
 467             /* Update potential sum for this i atom from the interaction with this j atom. */
 468             velecsum         = _mm_add_ps(velecsum,velec);
 469
 470             fscal            = felec;
 471
 472             /* Calculate temporary vectorial force */
 473             tx               = _mm_mul_ps(fscal,dx21);
 474             ty               = _mm_mul_ps(fscal,dy21);
 475             tz               = _mm_mul_ps(fscal,dz21);
 476
 477             /* Update vectorial force */
 478             fix2             = _mm_add_ps(fix2,tx);
 479             fiy2             = _mm_add_ps(fiy2,ty);
 480             fiz2             = _mm_add_ps(fiz2,tz);
 481
 482             fjx1             = _mm_add_ps(fjx1,tx);
 483             fjy1             = _mm_add_ps(fjy1,ty);
 484             fjz1             = _mm_add_ps(fjz1,tz);
 485
 486             /**************************
 487              * CALCULATE INTERACTIONS *
 488              **************************/
 489
 490             /* REACTION-FIELD ELECTROSTATICS */
 491             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 492             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 493
 494             /* Update potential sum for this i atom from the interaction with this j atom. */
 495             velecsum         = _mm_add_ps(velecsum,velec);
 496
 497             fscal            = felec;
 498
 499             /* Calculate temporary vectorial force */
 500             tx               = _mm_mul_ps(fscal,dx22);
 501             ty               = _mm_mul_ps(fscal,dy22);
 502             tz               = _mm_mul_ps(fscal,dz22);
 503
 504             /* Update vectorial force */
 505             fix2             = _mm_add_ps(fix2,tx);
 506             fiy2             = _mm_add_ps(fiy2,ty);
 507             fiz2             = _mm_add_ps(fiz2,tz);
 508
 509             fjx2             = _mm_add_ps(fjx2,tx);
 510             fjy2             = _mm_add_ps(fjy2,ty);
 511             fjz2             = _mm_add_ps(fjz2,tz);
 512
 513             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 514                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 515                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 516
 517             /* Inner loop uses 288 flops */
 518         }
 519
 520         if(jidx<j_index_end)
 521         {
 522
 523             /* Get j neighbor index, and coordinate index */
 524             jnrA             = jjnr[jidx];
 525             jnrB             = jjnr[jidx+1];
 526             jnrC             = jjnr[jidx+2];
 527             jnrD             = jjnr[jidx+3];
 528
 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       = (jnrA>=0) ? jnrA : 0;
 535             jnrB       = (jnrB>=0) ? jnrB : 0;
 536             jnrC       = (jnrC>=0) ? jnrC : 0;
 537             jnrD       = (jnrD>=0) ? jnrD : 0;
 538
 539             j_coord_offsetA  = DIM*jnrA;
 540             j_coord_offsetB  = DIM*jnrB;
 541             j_coord_offsetC  = DIM*jnrC;
 542             j_coord_offsetD  = DIM*jnrD;
 543
 544             /* load j atom coordinates */
 545             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 546                                               x+j_coord_offsetC,x+j_coord_offsetD,
 547                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 548
 549             /* Calculate displacement vector */
 550             dx00             = _mm_sub_ps(ix0,jx0);
 551             dy00             = _mm_sub_ps(iy0,jy0);
 552             dz00             = _mm_sub_ps(iz0,jz0);
 553             dx01             = _mm_sub_ps(ix0,jx1);
 554             dy01             = _mm_sub_ps(iy0,jy1);
 555             dz01             = _mm_sub_ps(iz0,jz1);
 556             dx02             = _mm_sub_ps(ix0,jx2);
 557             dy02             = _mm_sub_ps(iy0,jy2);
 558             dz02             = _mm_sub_ps(iz0,jz2);
 559             dx10             = _mm_sub_ps(ix1,jx0);
 560             dy10             = _mm_sub_ps(iy1,jy0);
 561             dz10             = _mm_sub_ps(iz1,jz0);
 562             dx11             = _mm_sub_ps(ix1,jx1);
 563             dy11             = _mm_sub_ps(iy1,jy1);
 564             dz11             = _mm_sub_ps(iz1,jz1);
 565             dx12             = _mm_sub_ps(ix1,jx2);
 566             dy12             = _mm_sub_ps(iy1,jy2);
 567             dz12             = _mm_sub_ps(iz1,jz2);
 568             dx20             = _mm_sub_ps(ix2,jx0);
 569             dy20             = _mm_sub_ps(iy2,jy0);
 570             dz20             = _mm_sub_ps(iz2,jz0);
 571             dx21             = _mm_sub_ps(ix2,jx1);
 572             dy21             = _mm_sub_ps(iy2,jy1);
 573             dz21             = _mm_sub_ps(iz2,jz1);
 574             dx22             = _mm_sub_ps(ix2,jx2);
 575             dy22             = _mm_sub_ps(iy2,jy2);
 576             dz22             = _mm_sub_ps(iz2,jz2);
 577
 578             /* Calculate squared distance and things based on it */
 579             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 580             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 581             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 582             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 583             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 584             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 585             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 586             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 587             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 588
 589             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 590             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 591             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 592             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 593             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 594             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 595             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 596             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 597             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 598
 599             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 600             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 601             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 602             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 603             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 604             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 605             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 606             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 607             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 608
 609             fjx0             = _mm_setzero_ps();
 610             fjy0             = _mm_setzero_ps();
 611             fjz0             = _mm_setzero_ps();
 612             fjx1             = _mm_setzero_ps();
 613             fjy1             = _mm_setzero_ps();
 614             fjz1             = _mm_setzero_ps();
 615             fjx2             = _mm_setzero_ps();
 616             fjy2             = _mm_setzero_ps();
 617             fjz2             = _mm_setzero_ps();
 618
 619             /**************************
 620              * CALCULATE INTERACTIONS *
 621              **************************/
 622
 623             /* REACTION-FIELD ELECTROSTATICS */
 624             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 625             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 626
 627             /* Update potential sum for this i atom from the interaction with this j atom. */
 628             velec            = _mm_andnot_ps(dummy_mask,velec);
 629             velecsum         = _mm_add_ps(velecsum,velec);
 630
 631             fscal            = felec;
 632
 633             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 634
 635             /* Calculate temporary vectorial force */
 636             tx               = _mm_mul_ps(fscal,dx00);
 637             ty               = _mm_mul_ps(fscal,dy00);
 638             tz               = _mm_mul_ps(fscal,dz00);
 639
 640             /* Update vectorial force */
 641             fix0             = _mm_add_ps(fix0,tx);
 642             fiy0             = _mm_add_ps(fiy0,ty);
 643             fiz0             = _mm_add_ps(fiz0,tz);
 644
 645             fjx0             = _mm_add_ps(fjx0,tx);
 646             fjy0             = _mm_add_ps(fjy0,ty);
 647             fjz0             = _mm_add_ps(fjz0,tz);
 648
 649             /**************************
 650              * CALCULATE INTERACTIONS *
 651              **************************/
 652
 653             /* REACTION-FIELD ELECTROSTATICS */
 654             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 655             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 656
 657             /* Update potential sum for this i atom from the interaction with this j atom. */
 658             velec            = _mm_andnot_ps(dummy_mask,velec);
 659             velecsum         = _mm_add_ps(velecsum,velec);
 660
 661             fscal            = felec;
 662
 663             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 664
 665             /* Calculate temporary vectorial force */
 666             tx               = _mm_mul_ps(fscal,dx01);
 667             ty               = _mm_mul_ps(fscal,dy01);
 668             tz               = _mm_mul_ps(fscal,dz01);
 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             fjx1             = _mm_add_ps(fjx1,tx);
 676             fjy1             = _mm_add_ps(fjy1,ty);
 677             fjz1             = _mm_add_ps(fjz1,tz);
 678
 679             /**************************
 680              * CALCULATE INTERACTIONS *
 681              **************************/
 682
 683             /* REACTION-FIELD ELECTROSTATICS */
 684             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 685             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 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,dx02);
 697             ty               = _mm_mul_ps(fscal,dy02);
 698             tz               = _mm_mul_ps(fscal,dz02);
 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             fjx2             = _mm_add_ps(fjx2,tx);
 706             fjy2             = _mm_add_ps(fjy2,ty);
 707             fjz2             = _mm_add_ps(fjz2,tz);
 708
 709             /**************************
 710              * CALCULATE INTERACTIONS *
 711              **************************/
 712
 713             /* REACTION-FIELD ELECTROSTATICS */
 714             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 715             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 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,dx10);
 727             ty               = _mm_mul_ps(fscal,dy10);
 728             tz               = _mm_mul_ps(fscal,dz10);
 729
 730             /* Update vectorial force */
 731             fix1             = _mm_add_ps(fix1,tx);
 732             fiy1             = _mm_add_ps(fiy1,ty);
 733             fiz1             = _mm_add_ps(fiz1,tz);
 734
 735             fjx0             = _mm_add_ps(fjx0,tx);
 736             fjy0             = _mm_add_ps(fjy0,ty);
 737             fjz0             = _mm_add_ps(fjz0,tz);
 738
 739             /**************************
 740              * CALCULATE INTERACTIONS *
 741              **************************/
 742
 743             /* REACTION-FIELD ELECTROSTATICS */
 744             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 745             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 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,dx11);
 757             ty               = _mm_mul_ps(fscal,dy11);
 758             tz               = _mm_mul_ps(fscal,dz11);
 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             fjx1             = _mm_add_ps(fjx1,tx);
 766             fjy1             = _mm_add_ps(fjy1,ty);
 767             fjz1             = _mm_add_ps(fjz1,tz);
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             /* REACTION-FIELD ELECTROSTATICS */
 774             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 775             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 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,dx12);
 787             ty               = _mm_mul_ps(fscal,dy12);
 788             tz               = _mm_mul_ps(fscal,dz12);
 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             fjx2             = _mm_add_ps(fjx2,tx);
 796             fjy2             = _mm_add_ps(fjy2,ty);
 797             fjz2             = _mm_add_ps(fjz2,tz);
 798
 799             /**************************
 800              * CALCULATE INTERACTIONS *
 801              **************************/
 802
 803             /* REACTION-FIELD ELECTROSTATICS */
 804             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 805             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 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,dx20);
 817             ty               = _mm_mul_ps(fscal,dy20);
 818             tz               = _mm_mul_ps(fscal,dz20);
 819
 820             /* Update vectorial force */
 821             fix2             = _mm_add_ps(fix2,tx);
 822             fiy2             = _mm_add_ps(fiy2,ty);
 823             fiz2             = _mm_add_ps(fiz2,tz);
 824
 825             fjx0             = _mm_add_ps(fjx0,tx);
 826             fjy0             = _mm_add_ps(fjy0,ty);
 827             fjz0             = _mm_add_ps(fjz0,tz);
 828
 829             /**************************
 830              * CALCULATE INTERACTIONS *
 831              **************************/
 832
 833             /* REACTION-FIELD ELECTROSTATICS */
 834             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 835             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 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,dx21);
 847             ty               = _mm_mul_ps(fscal,dy21);
 848             tz               = _mm_mul_ps(fscal,dz21);
 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             fjx1             = _mm_add_ps(fjx1,tx);
 856             fjy1             = _mm_add_ps(fjy1,ty);
 857             fjz1             = _mm_add_ps(fjz1,tz);
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             /* REACTION-FIELD ELECTROSTATICS */
 864             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 865             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 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,dx22);
 877             ty               = _mm_mul_ps(fscal,dy22);
 878             tz               = _mm_mul_ps(fscal,dz22);
 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             fjx2             = _mm_add_ps(fjx2,tx);
 886             fjy2             = _mm_add_ps(fjy2,ty);
 887             fjz2             = _mm_add_ps(fjz2,tz);
 888
 889             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
 890                                                    f+j_coord_offsetC,f+j_coord_offsetD,
 891                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 892
 893             /* Inner loop uses 288 flops */
 894         }
 895
 896         /* End of innermost loop */
 897
 898         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 899                                               f+i_coord_offset,fshift+i_shift_offset);
 900
 901         ggid                        = gid[iidx];
 902         /* Update potential energies */
 903         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 904
 905         /* Increment number of inner iterations */
 906         inneriter                  += j_index_end - j_index_start;
 907
 908         /* Outer loop uses 28 flops */
 909     }
 910
 911     /* Increment number of outer iterations */
 912     outeriter        += nri;
 913
 914     /* Update outer/inner flops */
 915
 916     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*28 + inneriter*288);
 917 }
 918 /*
 919  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse2_single
 920  * Electrostatics interaction: ReactionField
 921  * VdW interaction:            None
 922  * Geometry:                   Water3-Water3
 923  * Calculate force/pot:        Force
 924  */
 925 void
 926 nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse2_single
 927                     (t_nblist * gmx_restrict                nlist,
 928                      rvec * gmx_restrict                    xx,
 929                      rvec * gmx_restrict                    ff,
 930                      t_forcerec * gmx_restrict              fr,
 931                      t_mdatoms * gmx_restrict               mdatoms,
 932                      nb_kernel_data_t * gmx_restrict        kernel_data,
 933                      t_nrnb * gmx_restrict                  nrnb)
 934 {
 935     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 936      * just 0 for non-waters.
 937      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 938      * jnr indices corresponding to data put in the four positions in the SIMD register.
 939      */
 940     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 941     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 942     int              jnrA,jnrB,jnrC,jnrD;
 943     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 944     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 945     real             shX,shY,shZ,rcutoff_scalar;
 946     real             *shiftvec,*fshift,*x,*f;
 947     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 948     int              vdwioffset0;
 949     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 950     int              vdwioffset1;
 951     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 952     int              vdwioffset2;
 953     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 954     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 955     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 956     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 957     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 958     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 959     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 960     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 961     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 962     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 963     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 964     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 965     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 966     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 967     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 968     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 969     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 970     real             *charge;
 971     __m128           dummy_mask,cutoff_mask;
 972     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 973     __m128           one     = _mm_set1_ps(1.0);
 974     __m128           two     = _mm_set1_ps(2.0);
 975     x                = xx[0];
 976     f                = ff[0];
 977
 978     nri              = nlist->nri;
 979     iinr             = nlist->iinr;
 980     jindex           = nlist->jindex;
 981     jjnr             = nlist->jjnr;
 982     shiftidx         = nlist->shift;
 983     gid              = nlist->gid;
 984     shiftvec         = fr->shift_vec[0];
 985     fshift           = fr->fshift[0];
 986     facel            = _mm_set1_ps(fr->epsfac);
 987     charge           = mdatoms->chargeA;
 988     krf              = _mm_set1_ps(fr->ic->k_rf);
 989     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 990     crf              = _mm_set1_ps(fr->ic->c_rf);
 991
 992     /* Setup water-specific parameters */
 993     inr              = nlist->iinr[0];
 994     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 995     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 996     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 997
 998     jq0              = _mm_set1_ps(charge[inr+0]);
 999     jq1              = _mm_set1_ps(charge[inr+1]);
1000     jq2              = _mm_set1_ps(charge[inr+2]);
1001     qq00             = _mm_mul_ps(iq0,jq0);
1002     qq01             = _mm_mul_ps(iq0,jq1);
1003     qq02             = _mm_mul_ps(iq0,jq2);
1004     qq10             = _mm_mul_ps(iq1,jq0);
1005     qq11             = _mm_mul_ps(iq1,jq1);
1006     qq12             = _mm_mul_ps(iq1,jq2);
1007     qq20             = _mm_mul_ps(iq2,jq0);
1008     qq21             = _mm_mul_ps(iq2,jq1);
1009     qq22             = _mm_mul_ps(iq2,jq2);
1010
1011     /* Avoid stupid compiler warnings */
1012     jnrA = jnrB = jnrC = jnrD = 0;
1013     j_coord_offsetA = 0;
1014     j_coord_offsetB = 0;
1015     j_coord_offsetC = 0;
1016     j_coord_offsetD = 0;
1017
1018     outeriter        = 0;
1019     inneriter        = 0;
1020
1021     /* Start outer loop over neighborlists */
1022     for(iidx=0; iidx<nri; iidx++)
1023     {
1024         /* Load shift vector for this list */
1025         i_shift_offset   = DIM*shiftidx[iidx];
1026         shX              = shiftvec[i_shift_offset+XX];
1027         shY              = shiftvec[i_shift_offset+YY];
1028         shZ              = shiftvec[i_shift_offset+ZZ];
1029
1030         /* Load limits for loop over neighbors */
1031         j_index_start    = jindex[iidx];
1032         j_index_end      = jindex[iidx+1];
1033
1034         /* Get outer coordinate index */
1035         inr              = iinr[iidx];
1036         i_coord_offset   = DIM*inr;
1037
1038         /* Load i particle coords and add shift vector */
1039         ix0              = _mm_set1_ps(shX + x[i_coord_offset+DIM*0+XX]);
1040         iy0              = _mm_set1_ps(shY + x[i_coord_offset+DIM*0+YY]);
1041         iz0              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*0+ZZ]);
1042         ix1              = _mm_set1_ps(shX + x[i_coord_offset+DIM*1+XX]);
1043         iy1              = _mm_set1_ps(shY + x[i_coord_offset+DIM*1+YY]);
1044         iz1              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*1+ZZ]);
1045         ix2              = _mm_set1_ps(shX + x[i_coord_offset+DIM*2+XX]);
1046         iy2              = _mm_set1_ps(shY + x[i_coord_offset+DIM*2+YY]);
1047         iz2              = _mm_set1_ps(shZ + x[i_coord_offset+DIM*2+ZZ]);
1048
1049         fix0             = _mm_setzero_ps();
1050         fiy0             = _mm_setzero_ps();
1051         fiz0             = _mm_setzero_ps();
1052         fix1             = _mm_setzero_ps();
1053         fiy1             = _mm_setzero_ps();
1054         fiz1             = _mm_setzero_ps();
1055         fix2             = _mm_setzero_ps();
1056         fiy2             = _mm_setzero_ps();
1057         fiz2             = _mm_setzero_ps();
1058
1059         /* Start inner kernel loop */
1060         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1061         {
1062
1063             /* Get j neighbor index, and coordinate index */
1064             jnrA             = jjnr[jidx];
1065             jnrB             = jjnr[jidx+1];
1066             jnrC             = jjnr[jidx+2];
1067             jnrD             = jjnr[jidx+3];
1068
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             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1357                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1358                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1359
1360             /* Inner loop uses 243 flops */
1361         }
1362
1363         if(jidx<j_index_end)
1364         {
1365
1366             /* Get j neighbor index, and coordinate index */
1367             jnrA             = jjnr[jidx];
1368             jnrB             = jjnr[jidx+1];
1369             jnrC             = jjnr[jidx+2];
1370             jnrD             = jjnr[jidx+3];
1371
1372             /* Sign of each element will be negative for non-real atoms.
1373              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1374              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1375              */
1376             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1377             jnrA       = (jnrA>=0) ? jnrA : 0;
1378             jnrB       = (jnrB>=0) ? jnrB : 0;
1379             jnrC       = (jnrC>=0) ? jnrC : 0;
1380             jnrD       = (jnrD>=0) ? jnrD : 0;
1381
1382             j_coord_offsetA  = DIM*jnrA;
1383             j_coord_offsetB  = DIM*jnrB;
1384             j_coord_offsetC  = DIM*jnrC;
1385             j_coord_offsetD  = DIM*jnrD;
1386
1387             /* load j atom coordinates */
1388             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1389                                               x+j_coord_offsetC,x+j_coord_offsetD,
1390                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1391
1392             /* Calculate displacement vector */
1393             dx00             = _mm_sub_ps(ix0,jx0);
1394             dy00             = _mm_sub_ps(iy0,jy0);
1395             dz00             = _mm_sub_ps(iz0,jz0);
1396             dx01             = _mm_sub_ps(ix0,jx1);
1397             dy01             = _mm_sub_ps(iy0,jy1);
1398             dz01             = _mm_sub_ps(iz0,jz1);
1399             dx02             = _mm_sub_ps(ix0,jx2);
1400             dy02             = _mm_sub_ps(iy0,jy2);
1401             dz02             = _mm_sub_ps(iz0,jz2);
1402             dx10             = _mm_sub_ps(ix1,jx0);
1403             dy10             = _mm_sub_ps(iy1,jy0);
1404             dz10             = _mm_sub_ps(iz1,jz0);
1405             dx11             = _mm_sub_ps(ix1,jx1);
1406             dy11             = _mm_sub_ps(iy1,jy1);
1407             dz11             = _mm_sub_ps(iz1,jz1);
1408             dx12             = _mm_sub_ps(ix1,jx2);
1409             dy12             = _mm_sub_ps(iy1,jy2);
1410             dz12             = _mm_sub_ps(iz1,jz2);
1411             dx20             = _mm_sub_ps(ix2,jx0);
1412             dy20             = _mm_sub_ps(iy2,jy0);
1413             dz20             = _mm_sub_ps(iz2,jz0);
1414             dx21             = _mm_sub_ps(ix2,jx1);
1415             dy21             = _mm_sub_ps(iy2,jy1);
1416             dz21             = _mm_sub_ps(iz2,jz1);
1417             dx22             = _mm_sub_ps(ix2,jx2);
1418             dy22             = _mm_sub_ps(iy2,jy2);
1419             dz22             = _mm_sub_ps(iz2,jz2);
1420
1421             /* Calculate squared distance and things based on it */
1422             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1423             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1424             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1425             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1426             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1427             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1428             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1429             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1430             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1431
1432             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1433             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1434             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1435             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1436             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1437             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1438             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1439             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1440             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1441
1442             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1443             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1444             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1445             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1446             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1447             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1448             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1449             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1450             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1451
1452             fjx0             = _mm_setzero_ps();
1453             fjy0             = _mm_setzero_ps();
1454             fjz0             = _mm_setzero_ps();
1455             fjx1             = _mm_setzero_ps();
1456             fjy1             = _mm_setzero_ps();
1457             fjz1             = _mm_setzero_ps();
1458             fjx2             = _mm_setzero_ps();
1459             fjy2             = _mm_setzero_ps();
1460             fjz2             = _mm_setzero_ps();
1461
1462             /**************************
1463              * CALCULATE INTERACTIONS *
1464              **************************/
1465
1466             /* REACTION-FIELD ELECTROSTATICS */
1467             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1468
1469             fscal            = felec;
1470
1471             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1472
1473             /* Calculate temporary vectorial force */
1474             tx               = _mm_mul_ps(fscal,dx00);
1475             ty               = _mm_mul_ps(fscal,dy00);
1476             tz               = _mm_mul_ps(fscal,dz00);
1477
1478             /* Update vectorial force */
1479             fix0             = _mm_add_ps(fix0,tx);
1480             fiy0             = _mm_add_ps(fiy0,ty);
1481             fiz0             = _mm_add_ps(fiz0,tz);
1482
1483             fjx0             = _mm_add_ps(fjx0,tx);
1484             fjy0             = _mm_add_ps(fjy0,ty);
1485             fjz0             = _mm_add_ps(fjz0,tz);
1486
1487             /**************************
1488              * CALCULATE INTERACTIONS *
1489              **************************/
1490
1491             /* REACTION-FIELD ELECTROSTATICS */
1492             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1493
1494             fscal            = felec;
1495
1496             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1497
1498             /* Calculate temporary vectorial force */
1499             tx               = _mm_mul_ps(fscal,dx01);
1500             ty               = _mm_mul_ps(fscal,dy01);
1501             tz               = _mm_mul_ps(fscal,dz01);
1502
1503             /* Update vectorial force */
1504             fix0             = _mm_add_ps(fix0,tx);
1505             fiy0             = _mm_add_ps(fiy0,ty);
1506             fiz0             = _mm_add_ps(fiz0,tz);
1507
1508             fjx1             = _mm_add_ps(fjx1,tx);
1509             fjy1             = _mm_add_ps(fjy1,ty);
1510             fjz1             = _mm_add_ps(fjz1,tz);
1511
1512             /**************************
1513              * CALCULATE INTERACTIONS *
1514              **************************/
1515
1516             /* REACTION-FIELD ELECTROSTATICS */
1517             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1518
1519             fscal            = felec;
1520
1521             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1522
1523             /* Calculate temporary vectorial force */
1524             tx               = _mm_mul_ps(fscal,dx02);
1525             ty               = _mm_mul_ps(fscal,dy02);
1526             tz               = _mm_mul_ps(fscal,dz02);
1527
1528             /* Update vectorial force */
1529             fix0             = _mm_add_ps(fix0,tx);
1530             fiy0             = _mm_add_ps(fiy0,ty);
1531             fiz0             = _mm_add_ps(fiz0,tz);
1532
1533             fjx2             = _mm_add_ps(fjx2,tx);
1534             fjy2             = _mm_add_ps(fjy2,ty);
1535             fjz2             = _mm_add_ps(fjz2,tz);
1536
1537             /**************************
1538              * CALCULATE INTERACTIONS *
1539              **************************/
1540
1541             /* REACTION-FIELD ELECTROSTATICS */
1542             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1543
1544             fscal            = felec;
1545
1546             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1547
1548             /* Calculate temporary vectorial force */
1549             tx               = _mm_mul_ps(fscal,dx10);
1550             ty               = _mm_mul_ps(fscal,dy10);
1551             tz               = _mm_mul_ps(fscal,dz10);
1552
1553             /* Update vectorial force */
1554             fix1             = _mm_add_ps(fix1,tx);
1555             fiy1             = _mm_add_ps(fiy1,ty);
1556             fiz1             = _mm_add_ps(fiz1,tz);
1557
1558             fjx0             = _mm_add_ps(fjx0,tx);
1559             fjy0             = _mm_add_ps(fjy0,ty);
1560             fjz0             = _mm_add_ps(fjz0,tz);
1561
1562             /**************************
1563              * CALCULATE INTERACTIONS *
1564              **************************/
1565
1566             /* REACTION-FIELD ELECTROSTATICS */
1567             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1568
1569             fscal            = felec;
1570
1571             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1572
1573             /* Calculate temporary vectorial force */
1574             tx               = _mm_mul_ps(fscal,dx11);
1575             ty               = _mm_mul_ps(fscal,dy11);
1576             tz               = _mm_mul_ps(fscal,dz11);
1577
1578             /* Update vectorial force */
1579             fix1             = _mm_add_ps(fix1,tx);
1580             fiy1             = _mm_add_ps(fiy1,ty);
1581             fiz1             = _mm_add_ps(fiz1,tz);
1582
1583             fjx1             = _mm_add_ps(fjx1,tx);
1584             fjy1             = _mm_add_ps(fjy1,ty);
1585             fjz1             = _mm_add_ps(fjz1,tz);
1586
1587             /**************************
1588              * CALCULATE INTERACTIONS *
1589              **************************/
1590
1591             /* REACTION-FIELD ELECTROSTATICS */
1592             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1593
1594             fscal            = felec;
1595
1596             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1597
1598             /* Calculate temporary vectorial force */
1599             tx               = _mm_mul_ps(fscal,dx12);
1600             ty               = _mm_mul_ps(fscal,dy12);
1601             tz               = _mm_mul_ps(fscal,dz12);
1602
1603             /* Update vectorial force */
1604             fix1             = _mm_add_ps(fix1,tx);
1605             fiy1             = _mm_add_ps(fiy1,ty);
1606             fiz1             = _mm_add_ps(fiz1,tz);
1607
1608             fjx2             = _mm_add_ps(fjx2,tx);
1609             fjy2             = _mm_add_ps(fjy2,ty);
1610             fjz2             = _mm_add_ps(fjz2,tz);
1611
1612             /**************************
1613              * CALCULATE INTERACTIONS *
1614              **************************/
1615
1616             /* REACTION-FIELD ELECTROSTATICS */
1617             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1618
1619             fscal            = felec;
1620
1621             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1622
1623             /* Calculate temporary vectorial force */
1624             tx               = _mm_mul_ps(fscal,dx20);
1625             ty               = _mm_mul_ps(fscal,dy20);
1626             tz               = _mm_mul_ps(fscal,dz20);
1627
1628             /* Update vectorial force */
1629             fix2             = _mm_add_ps(fix2,tx);
1630             fiy2             = _mm_add_ps(fiy2,ty);
1631             fiz2             = _mm_add_ps(fiz2,tz);
1632
1633             fjx0             = _mm_add_ps(fjx0,tx);
1634             fjy0             = _mm_add_ps(fjy0,ty);
1635             fjz0             = _mm_add_ps(fjz0,tz);
1636
1637             /**************************
1638              * CALCULATE INTERACTIONS *
1639              **************************/
1640
1641             /* REACTION-FIELD ELECTROSTATICS */
1642             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1643
1644             fscal            = felec;
1645
1646             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1647
1648             /* Calculate temporary vectorial force */
1649             tx               = _mm_mul_ps(fscal,dx21);
1650             ty               = _mm_mul_ps(fscal,dy21);
1651             tz               = _mm_mul_ps(fscal,dz21);
1652
1653             /* Update vectorial force */
1654             fix2             = _mm_add_ps(fix2,tx);
1655             fiy2             = _mm_add_ps(fiy2,ty);
1656             fiz2             = _mm_add_ps(fiz2,tz);
1657
1658             fjx1             = _mm_add_ps(fjx1,tx);
1659             fjy1             = _mm_add_ps(fjy1,ty);
1660             fjz1             = _mm_add_ps(fjz1,tz);
1661
1662             /**************************
1663              * CALCULATE INTERACTIONS *
1664              **************************/
1665
1666             /* REACTION-FIELD ELECTROSTATICS */
1667             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1668
1669             fscal            = felec;
1670
1671             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1672
1673             /* Calculate temporary vectorial force */
1674             tx               = _mm_mul_ps(fscal,dx22);
1675             ty               = _mm_mul_ps(fscal,dy22);
1676             tz               = _mm_mul_ps(fscal,dz22);
1677
1678             /* Update vectorial force */
1679             fix2             = _mm_add_ps(fix2,tx);
1680             fiy2             = _mm_add_ps(fiy2,ty);
1681             fiz2             = _mm_add_ps(fiz2,tz);
1682
1683             fjx2             = _mm_add_ps(fjx2,tx);
1684             fjy2             = _mm_add_ps(fjy2,ty);
1685             fjz2             = _mm_add_ps(fjz2,tz);
1686
1687             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(f+j_coord_offsetA,f+j_coord_offsetB,
1688                                                    f+j_coord_offsetC,f+j_coord_offsetD,
1689                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1690
1691             /* Inner loop uses 243 flops */
1692         }
1693
1694         /* End of innermost loop */
1695
1696         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1697                                               f+i_coord_offset,fshift+i_shift_offset);
1698
1699         /* Increment number of inner iterations */
1700         inneriter                  += j_index_end - j_index_start;
1701
1702         /* Outer loop uses 27 flops */
1703     }
1704
1705     /* Increment number of outer iterations */
1706     outeriter        += nri;
1707
1708     /* Update outer/inner flops */
1709
1710     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*27 + inneriter*243);
1711 }