src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRF_VdwCSTab_GeomW3W3_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_avx_128_fma_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_VF_avx_128_fma_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_VF_avx_128_fma_double
  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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     int              vdwjidx1A,vdwjidx1B;
  75     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  76     int              vdwjidx2A,vdwjidx2B;
  77     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  78     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  79     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  80     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  81     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  83     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  84     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  85     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  86     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  94     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  95     __m128i          vfitab;
  96     __m128i          ifour       = _mm_set1_epi32(4);
  97     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
  98     real             *vftab;
  99     __m128d          dummy_mask,cutoff_mask;
 100     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 101     __m128d          one     = _mm_set1_pd(1.0);
 102     __m128d          two     = _mm_set1_pd(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm_set1_pd(fr->epsfac);
 115     charge           = mdatoms->chargeA;
 116     krf              = _mm_set1_pd(fr->ic->k_rf);
 117     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 118     crf              = _mm_set1_pd(fr->ic->c_rf);
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     vftab            = kernel_data->table_vdw->data;
 124     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 129     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 130     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 131     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 132
 133     jq0              = _mm_set1_pd(charge[inr+0]);
 134     jq1              = _mm_set1_pd(charge[inr+1]);
 135     jq2              = _mm_set1_pd(charge[inr+2]);
 136     vdwjidx0A        = 2*vdwtype[inr+0];
 137     qq00             = _mm_mul_pd(iq0,jq0);
 138     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 139     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 140     qq01             = _mm_mul_pd(iq0,jq1);
 141     qq02             = _mm_mul_pd(iq0,jq2);
 142     qq10             = _mm_mul_pd(iq1,jq0);
 143     qq11             = _mm_mul_pd(iq1,jq1);
 144     qq12             = _mm_mul_pd(iq1,jq2);
 145     qq20             = _mm_mul_pd(iq2,jq0);
 146     qq21             = _mm_mul_pd(iq2,jq1);
 147     qq22             = _mm_mul_pd(iq2,jq2);
 148
 149     /* Avoid stupid compiler warnings */
 150     jnrA = jnrB = 0;
 151     j_coord_offsetA = 0;
 152     j_coord_offsetB = 0;
 153
 154     outeriter        = 0;
 155     inneriter        = 0;
 156
 157     /* Start outer loop over neighborlists */
 158     for(iidx=0; iidx<nri; iidx++)
 159     {
 160         /* Load shift vector for this list */
 161         i_shift_offset   = DIM*shiftidx[iidx];
 162
 163         /* Load limits for loop over neighbors */
 164         j_index_start    = jindex[iidx];
 165         j_index_end      = jindex[iidx+1];
 166
 167         /* Get outer coordinate index */
 168         inr              = iinr[iidx];
 169         i_coord_offset   = DIM*inr;
 170
 171         /* Load i particle coords and add shift vector */
 172         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 173                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 174
 175         fix0             = _mm_setzero_pd();
 176         fiy0             = _mm_setzero_pd();
 177         fiz0             = _mm_setzero_pd();
 178         fix1             = _mm_setzero_pd();
 179         fiy1             = _mm_setzero_pd();
 180         fiz1             = _mm_setzero_pd();
 181         fix2             = _mm_setzero_pd();
 182         fiy2             = _mm_setzero_pd();
 183         fiz2             = _mm_setzero_pd();
 184
 185         /* Reset potential sums */
 186         velecsum         = _mm_setzero_pd();
 187         vvdwsum          = _mm_setzero_pd();
 188
 189         /* Start inner kernel loop */
 190         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 191         {
 192
 193             /* Get j neighbor index, and coordinate index */
 194             jnrA             = jjnr[jidx];
 195             jnrB             = jjnr[jidx+1];
 196             j_coord_offsetA  = DIM*jnrA;
 197             j_coord_offsetB  = DIM*jnrB;
 198
 199             /* load j atom coordinates */
 200             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 201                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 202
 203             /* Calculate displacement vector */
 204             dx00             = _mm_sub_pd(ix0,jx0);
 205             dy00             = _mm_sub_pd(iy0,jy0);
 206             dz00             = _mm_sub_pd(iz0,jz0);
 207             dx01             = _mm_sub_pd(ix0,jx1);
 208             dy01             = _mm_sub_pd(iy0,jy1);
 209             dz01             = _mm_sub_pd(iz0,jz1);
 210             dx02             = _mm_sub_pd(ix0,jx2);
 211             dy02             = _mm_sub_pd(iy0,jy2);
 212             dz02             = _mm_sub_pd(iz0,jz2);
 213             dx10             = _mm_sub_pd(ix1,jx0);
 214             dy10             = _mm_sub_pd(iy1,jy0);
 215             dz10             = _mm_sub_pd(iz1,jz0);
 216             dx11             = _mm_sub_pd(ix1,jx1);
 217             dy11             = _mm_sub_pd(iy1,jy1);
 218             dz11             = _mm_sub_pd(iz1,jz1);
 219             dx12             = _mm_sub_pd(ix1,jx2);
 220             dy12             = _mm_sub_pd(iy1,jy2);
 221             dz12             = _mm_sub_pd(iz1,jz2);
 222             dx20             = _mm_sub_pd(ix2,jx0);
 223             dy20             = _mm_sub_pd(iy2,jy0);
 224             dz20             = _mm_sub_pd(iz2,jz0);
 225             dx21             = _mm_sub_pd(ix2,jx1);
 226             dy21             = _mm_sub_pd(iy2,jy1);
 227             dz21             = _mm_sub_pd(iz2,jz1);
 228             dx22             = _mm_sub_pd(ix2,jx2);
 229             dy22             = _mm_sub_pd(iy2,jy2);
 230             dz22             = _mm_sub_pd(iz2,jz2);
 231
 232             /* Calculate squared distance and things based on it */
 233             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 234             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 235             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 236             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 237             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 238             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 239             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 240             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 241             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 242
 243             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 244             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 245             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 246             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 247             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 248             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 249             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 250             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 251             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 252
 253             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 254             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 255             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 256             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 257             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 258             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 259             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 260             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 261             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 262
 263             fjx0             = _mm_setzero_pd();
 264             fjy0             = _mm_setzero_pd();
 265             fjz0             = _mm_setzero_pd();
 266             fjx1             = _mm_setzero_pd();
 267             fjy1             = _mm_setzero_pd();
 268             fjz1             = _mm_setzero_pd();
 269             fjx2             = _mm_setzero_pd();
 270             fjy2             = _mm_setzero_pd();
 271             fjz2             = _mm_setzero_pd();
 272
 273             /**************************
 274              * CALCULATE INTERACTIONS *
 275              **************************/
 276
 277             r00              = _mm_mul_pd(rsq00,rinv00);
 278
 279             /* Calculate table index by multiplying r with table scale and truncate to integer */
 280             rt               = _mm_mul_pd(r00,vftabscale);
 281             vfitab           = _mm_cvttpd_epi32(rt);
 282 #ifdef __XOP__
 283             vfeps            = _mm_frcz_pd(rt);
 284 #else
 285             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 286 #endif
 287             twovfeps         = _mm_add_pd(vfeps,vfeps);
 288             vfitab           = _mm_slli_epi32(vfitab,3);
 289
 290             /* REACTION-FIELD ELECTROSTATICS */
 291             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 292             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 293
 294             /* CUBIC SPLINE TABLE DISPERSION */
 295             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 296             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 297             GMX_MM_TRANSPOSE2_PD(Y,F);
 298             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 299             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 300             GMX_MM_TRANSPOSE2_PD(G,H);
 301             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 302             VV               = _mm_macc_pd(vfeps,Fp,Y);
 303             vvdw6            = _mm_mul_pd(c6_00,VV);
 304             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 305             fvdw6            = _mm_mul_pd(c6_00,FF);
 306
 307             /* CUBIC SPLINE TABLE REPULSION */
 308             vfitab           = _mm_add_epi32(vfitab,ifour);
 309             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 310             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 311             GMX_MM_TRANSPOSE2_PD(Y,F);
 312             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 313             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 314             GMX_MM_TRANSPOSE2_PD(G,H);
 315             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 316             VV               = _mm_macc_pd(vfeps,Fp,Y);
 317             vvdw12           = _mm_mul_pd(c12_00,VV);
 318             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 319             fvdw12           = _mm_mul_pd(c12_00,FF);
 320             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 321             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 322
 323             /* Update potential sum for this i atom from the interaction with this j atom. */
 324             velecsum         = _mm_add_pd(velecsum,velec);
 325             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 326
 327             fscal            = _mm_add_pd(felec,fvdw);
 328
 329             /* Update vectorial force */
 330             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 331             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 332             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 333
 334             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 335             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 336             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             /* REACTION-FIELD ELECTROSTATICS */
 343             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 344             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             velecsum         = _mm_add_pd(velecsum,velec);
 348
 349             fscal            = felec;
 350
 351             /* Update vectorial force */
 352             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 353             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 354             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 355
 356             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 357             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 358             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 359
 360             /**************************
 361              * CALCULATE INTERACTIONS *
 362              **************************/
 363
 364             /* REACTION-FIELD ELECTROSTATICS */
 365             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 366             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 367
 368             /* Update potential sum for this i atom from the interaction with this j atom. */
 369             velecsum         = _mm_add_pd(velecsum,velec);
 370
 371             fscal            = felec;
 372
 373             /* Update vectorial force */
 374             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 375             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 376             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 377
 378             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 379             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 380             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 381
 382             /**************************
 383              * CALCULATE INTERACTIONS *
 384              **************************/
 385
 386             /* REACTION-FIELD ELECTROSTATICS */
 387             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 388             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 389
 390             /* Update potential sum for this i atom from the interaction with this j atom. */
 391             velecsum         = _mm_add_pd(velecsum,velec);
 392
 393             fscal            = felec;
 394
 395             /* Update vectorial force */
 396             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 397             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 398             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 399
 400             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 401             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 402             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 403
 404             /**************************
 405              * CALCULATE INTERACTIONS *
 406              **************************/
 407
 408             /* REACTION-FIELD ELECTROSTATICS */
 409             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 410             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 411
 412             /* Update potential sum for this i atom from the interaction with this j atom. */
 413             velecsum         = _mm_add_pd(velecsum,velec);
 414
 415             fscal            = felec;
 416
 417             /* Update vectorial force */
 418             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 419             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 420             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 421
 422             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 423             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 424             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 425
 426             /**************************
 427              * CALCULATE INTERACTIONS *
 428              **************************/
 429
 430             /* REACTION-FIELD ELECTROSTATICS */
 431             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 432             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 433
 434             /* Update potential sum for this i atom from the interaction with this j atom. */
 435             velecsum         = _mm_add_pd(velecsum,velec);
 436
 437             fscal            = felec;
 438
 439             /* Update vectorial force */
 440             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 441             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 442             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 443
 444             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 445             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 446             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 447
 448             /**************************
 449              * CALCULATE INTERACTIONS *
 450              **************************/
 451
 452             /* REACTION-FIELD ELECTROSTATICS */
 453             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 454             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 455
 456             /* Update potential sum for this i atom from the interaction with this j atom. */
 457             velecsum         = _mm_add_pd(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             /* Update vectorial force */
 462             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 463             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 464             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 465
 466             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 467             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 468             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 469
 470             /**************************
 471              * CALCULATE INTERACTIONS *
 472              **************************/
 473
 474             /* REACTION-FIELD ELECTROSTATICS */
 475             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 476             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 477
 478             /* Update potential sum for this i atom from the interaction with this j atom. */
 479             velecsum         = _mm_add_pd(velecsum,velec);
 480
 481             fscal            = felec;
 482
 483             /* Update vectorial force */
 484             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 485             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 486             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 487
 488             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 489             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 490             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 491
 492             /**************************
 493              * CALCULATE INTERACTIONS *
 494              **************************/
 495
 496             /* REACTION-FIELD ELECTROSTATICS */
 497             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 498             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 499
 500             /* Update potential sum for this i atom from the interaction with this j atom. */
 501             velecsum         = _mm_add_pd(velecsum,velec);
 502
 503             fscal            = felec;
 504
 505             /* Update vectorial force */
 506             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 507             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 508             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 509
 510             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 511             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 512             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 513
 514             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 515
 516             /* Inner loop uses 350 flops */
 517         }
 518
 519         if(jidx<j_index_end)
 520         {
 521
 522             jnrA             = jjnr[jidx];
 523             j_coord_offsetA  = DIM*jnrA;
 524
 525             /* load j atom coordinates */
 526             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 527                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 528
 529             /* Calculate displacement vector */
 530             dx00             = _mm_sub_pd(ix0,jx0);
 531             dy00             = _mm_sub_pd(iy0,jy0);
 532             dz00             = _mm_sub_pd(iz0,jz0);
 533             dx01             = _mm_sub_pd(ix0,jx1);
 534             dy01             = _mm_sub_pd(iy0,jy1);
 535             dz01             = _mm_sub_pd(iz0,jz1);
 536             dx02             = _mm_sub_pd(ix0,jx2);
 537             dy02             = _mm_sub_pd(iy0,jy2);
 538             dz02             = _mm_sub_pd(iz0,jz2);
 539             dx10             = _mm_sub_pd(ix1,jx0);
 540             dy10             = _mm_sub_pd(iy1,jy0);
 541             dz10             = _mm_sub_pd(iz1,jz0);
 542             dx11             = _mm_sub_pd(ix1,jx1);
 543             dy11             = _mm_sub_pd(iy1,jy1);
 544             dz11             = _mm_sub_pd(iz1,jz1);
 545             dx12             = _mm_sub_pd(ix1,jx2);
 546             dy12             = _mm_sub_pd(iy1,jy2);
 547             dz12             = _mm_sub_pd(iz1,jz2);
 548             dx20             = _mm_sub_pd(ix2,jx0);
 549             dy20             = _mm_sub_pd(iy2,jy0);
 550             dz20             = _mm_sub_pd(iz2,jz0);
 551             dx21             = _mm_sub_pd(ix2,jx1);
 552             dy21             = _mm_sub_pd(iy2,jy1);
 553             dz21             = _mm_sub_pd(iz2,jz1);
 554             dx22             = _mm_sub_pd(ix2,jx2);
 555             dy22             = _mm_sub_pd(iy2,jy2);
 556             dz22             = _mm_sub_pd(iz2,jz2);
 557
 558             /* Calculate squared distance and things based on it */
 559             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 560             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 561             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 562             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 563             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 564             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 565             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 566             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 567             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 568
 569             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 570             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 571             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 572             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 573             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 574             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 575             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 576             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 577             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 578
 579             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 580             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 581             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 582             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 583             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 584             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 585             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 586             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 587             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 588
 589             fjx0             = _mm_setzero_pd();
 590             fjy0             = _mm_setzero_pd();
 591             fjz0             = _mm_setzero_pd();
 592             fjx1             = _mm_setzero_pd();
 593             fjy1             = _mm_setzero_pd();
 594             fjz1             = _mm_setzero_pd();
 595             fjx2             = _mm_setzero_pd();
 596             fjy2             = _mm_setzero_pd();
 597             fjz2             = _mm_setzero_pd();
 598
 599             /**************************
 600              * CALCULATE INTERACTIONS *
 601              **************************/
 602
 603             r00              = _mm_mul_pd(rsq00,rinv00);
 604
 605             /* Calculate table index by multiplying r with table scale and truncate to integer */
 606             rt               = _mm_mul_pd(r00,vftabscale);
 607             vfitab           = _mm_cvttpd_epi32(rt);
 608 #ifdef __XOP__
 609             vfeps            = _mm_frcz_pd(rt);
 610 #else
 611             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 612 #endif
 613             twovfeps         = _mm_add_pd(vfeps,vfeps);
 614             vfitab           = _mm_slli_epi32(vfitab,3);
 615
 616             /* REACTION-FIELD ELECTROSTATICS */
 617             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 618             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 619
 620             /* CUBIC SPLINE TABLE DISPERSION */
 621             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 622             F                = _mm_setzero_pd();
 623             GMX_MM_TRANSPOSE2_PD(Y,F);
 624             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 625             H                = _mm_setzero_pd();
 626             GMX_MM_TRANSPOSE2_PD(G,H);
 627             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 628             VV               = _mm_macc_pd(vfeps,Fp,Y);
 629             vvdw6            = _mm_mul_pd(c6_00,VV);
 630             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 631             fvdw6            = _mm_mul_pd(c6_00,FF);
 632
 633             /* CUBIC SPLINE TABLE REPULSION */
 634             vfitab           = _mm_add_epi32(vfitab,ifour);
 635             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 636             F                = _mm_setzero_pd();
 637             GMX_MM_TRANSPOSE2_PD(Y,F);
 638             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 639             H                = _mm_setzero_pd();
 640             GMX_MM_TRANSPOSE2_PD(G,H);
 641             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 642             VV               = _mm_macc_pd(vfeps,Fp,Y);
 643             vvdw12           = _mm_mul_pd(c12_00,VV);
 644             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 645             fvdw12           = _mm_mul_pd(c12_00,FF);
 646             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 647             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 648
 649             /* Update potential sum for this i atom from the interaction with this j atom. */
 650             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 651             velecsum         = _mm_add_pd(velecsum,velec);
 652             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 653             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 654
 655             fscal            = _mm_add_pd(felec,fvdw);
 656
 657             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 658
 659             /* Update vectorial force */
 660             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 661             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 662             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 663
 664             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 665             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 666             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 667
 668             /**************************
 669              * CALCULATE INTERACTIONS *
 670              **************************/
 671
 672             /* REACTION-FIELD ELECTROSTATICS */
 673             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 674             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 675
 676             /* Update potential sum for this i atom from the interaction with this j atom. */
 677             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 678             velecsum         = _mm_add_pd(velecsum,velec);
 679
 680             fscal            = felec;
 681
 682             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 683
 684             /* Update vectorial force */
 685             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 686             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 687             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 688
 689             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 690             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 691             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 692
 693             /**************************
 694              * CALCULATE INTERACTIONS *
 695              **************************/
 696
 697             /* REACTION-FIELD ELECTROSTATICS */
 698             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 699             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 700
 701             /* Update potential sum for this i atom from the interaction with this j atom. */
 702             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 703             velecsum         = _mm_add_pd(velecsum,velec);
 704
 705             fscal            = felec;
 706
 707             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 708
 709             /* Update vectorial force */
 710             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 711             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 712             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 713
 714             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 715             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 716             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 717
 718             /**************************
 719              * CALCULATE INTERACTIONS *
 720              **************************/
 721
 722             /* REACTION-FIELD ELECTROSTATICS */
 723             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 724             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 725
 726             /* Update potential sum for this i atom from the interaction with this j atom. */
 727             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 728             velecsum         = _mm_add_pd(velecsum,velec);
 729
 730             fscal            = felec;
 731
 732             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 733
 734             /* Update vectorial force */
 735             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 736             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 737             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 738
 739             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 740             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 741             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 742
 743             /**************************
 744              * CALCULATE INTERACTIONS *
 745              **************************/
 746
 747             /* REACTION-FIELD ELECTROSTATICS */
 748             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 749             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 750
 751             /* Update potential sum for this i atom from the interaction with this j atom. */
 752             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 753             velecsum         = _mm_add_pd(velecsum,velec);
 754
 755             fscal            = felec;
 756
 757             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 758
 759             /* Update vectorial force */
 760             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 761             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 762             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 763
 764             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 765             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 766             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 767
 768             /**************************
 769              * CALCULATE INTERACTIONS *
 770              **************************/
 771
 772             /* REACTION-FIELD ELECTROSTATICS */
 773             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 774             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 775
 776             /* Update potential sum for this i atom from the interaction with this j atom. */
 777             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 778             velecsum         = _mm_add_pd(velecsum,velec);
 779
 780             fscal            = felec;
 781
 782             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 783
 784             /* Update vectorial force */
 785             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 786             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 787             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 788
 789             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 790             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 791             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 792
 793             /**************************
 794              * CALCULATE INTERACTIONS *
 795              **************************/
 796
 797             /* REACTION-FIELD ELECTROSTATICS */
 798             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 799             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 800
 801             /* Update potential sum for this i atom from the interaction with this j atom. */
 802             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 803             velecsum         = _mm_add_pd(velecsum,velec);
 804
 805             fscal            = felec;
 806
 807             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 808
 809             /* Update vectorial force */
 810             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 811             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 812             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 813
 814             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 815             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 816             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 817
 818             /**************************
 819              * CALCULATE INTERACTIONS *
 820              **************************/
 821
 822             /* REACTION-FIELD ELECTROSTATICS */
 823             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 824             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 825
 826             /* Update potential sum for this i atom from the interaction with this j atom. */
 827             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 828             velecsum         = _mm_add_pd(velecsum,velec);
 829
 830             fscal            = felec;
 831
 832             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 833
 834             /* Update vectorial force */
 835             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 836             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 837             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 838
 839             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 840             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 841             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 842
 843             /**************************
 844              * CALCULATE INTERACTIONS *
 845              **************************/
 846
 847             /* REACTION-FIELD ELECTROSTATICS */
 848             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 849             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 850
 851             /* Update potential sum for this i atom from the interaction with this j atom. */
 852             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 853             velecsum         = _mm_add_pd(velecsum,velec);
 854
 855             fscal            = felec;
 856
 857             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 858
 859             /* Update vectorial force */
 860             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 861             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 862             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 863
 864             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 865             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 866             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 867
 868             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 869
 870             /* Inner loop uses 350 flops */
 871         }
 872
 873         /* End of innermost loop */
 874
 875         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 876                                               f+i_coord_offset,fshift+i_shift_offset);
 877
 878         ggid                        = gid[iidx];
 879         /* Update potential energies */
 880         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 881         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 882
 883         /* Increment number of inner iterations */
 884         inneriter                  += j_index_end - j_index_start;
 885
 886         /* Outer loop uses 20 flops */
 887     }
 888
 889     /* Increment number of outer iterations */
 890     outeriter        += nri;
 891
 892     /* Update outer/inner flops */
 893
 894     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*350);
 895 }
 896 /*
 897  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_avx_128_fma_double
 898  * Electrostatics interaction: ReactionField
 899  * VdW interaction:            CubicSplineTable
 900  * Geometry:                   Water3-Water3
 901  * Calculate force/pot:        Force
 902  */
 903 void
 904 nb_kernel_ElecRF_VdwCSTab_GeomW3W3_F_avx_128_fma_double
 905                     (t_nblist * gmx_restrict                nlist,
 906                      rvec * gmx_restrict                    xx,
 907                      rvec * gmx_restrict                    ff,
 908                      t_forcerec * gmx_restrict              fr,
 909                      t_mdatoms * gmx_restrict               mdatoms,
 910                      nb_kernel_data_t * gmx_restrict        kernel_data,
 911                      t_nrnb * gmx_restrict                  nrnb)
 912 {
 913     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 914      * just 0 for non-waters.
 915      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 916      * jnr indices corresponding to data put in the four positions in the SIMD register.
 917      */
 918     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 919     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 920     int              jnrA,jnrB;
 921     int              j_coord_offsetA,j_coord_offsetB;
 922     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 923     real             rcutoff_scalar;
 924     real             *shiftvec,*fshift,*x,*f;
 925     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 926     int              vdwioffset0;
 927     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 928     int              vdwioffset1;
 929     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 930     int              vdwioffset2;
 931     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 932     int              vdwjidx0A,vdwjidx0B;
 933     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 934     int              vdwjidx1A,vdwjidx1B;
 935     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 936     int              vdwjidx2A,vdwjidx2B;
 937     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 938     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 939     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 940     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 941     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 942     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 943     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 944     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 945     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 946     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 947     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 948     real             *charge;
 949     int              nvdwtype;
 950     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 951     int              *vdwtype;
 952     real             *vdwparam;
 953     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 954     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 955     __m128i          vfitab;
 956     __m128i          ifour       = _mm_set1_epi32(4);
 957     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 958     real             *vftab;
 959     __m128d          dummy_mask,cutoff_mask;
 960     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 961     __m128d          one     = _mm_set1_pd(1.0);
 962     __m128d          two     = _mm_set1_pd(2.0);
 963     x                = xx[0];
 964     f                = ff[0];
 965
 966     nri              = nlist->nri;
 967     iinr             = nlist->iinr;
 968     jindex           = nlist->jindex;
 969     jjnr             = nlist->jjnr;
 970     shiftidx         = nlist->shift;
 971     gid              = nlist->gid;
 972     shiftvec         = fr->shift_vec[0];
 973     fshift           = fr->fshift[0];
 974     facel            = _mm_set1_pd(fr->epsfac);
 975     charge           = mdatoms->chargeA;
 976     krf              = _mm_set1_pd(fr->ic->k_rf);
 977     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 978     crf              = _mm_set1_pd(fr->ic->c_rf);
 979     nvdwtype         = fr->ntype;
 980     vdwparam         = fr->nbfp;
 981     vdwtype          = mdatoms->typeA;
 982
 983     vftab            = kernel_data->table_vdw->data;
 984     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 985
 986     /* Setup water-specific parameters */
 987     inr              = nlist->iinr[0];
 988     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 989     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 990     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 991     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 992
 993     jq0              = _mm_set1_pd(charge[inr+0]);
 994     jq1              = _mm_set1_pd(charge[inr+1]);
 995     jq2              = _mm_set1_pd(charge[inr+2]);
 996     vdwjidx0A        = 2*vdwtype[inr+0];
 997     qq00             = _mm_mul_pd(iq0,jq0);
 998     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 999     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1000     qq01             = _mm_mul_pd(iq0,jq1);
1001     qq02             = _mm_mul_pd(iq0,jq2);
1002     qq10             = _mm_mul_pd(iq1,jq0);
1003     qq11             = _mm_mul_pd(iq1,jq1);
1004     qq12             = _mm_mul_pd(iq1,jq2);
1005     qq20             = _mm_mul_pd(iq2,jq0);
1006     qq21             = _mm_mul_pd(iq2,jq1);
1007     qq22             = _mm_mul_pd(iq2,jq2);
1008
1009     /* Avoid stupid compiler warnings */
1010     jnrA = jnrB = 0;
1011     j_coord_offsetA = 0;
1012     j_coord_offsetB = 0;
1013
1014     outeriter        = 0;
1015     inneriter        = 0;
1016
1017     /* Start outer loop over neighborlists */
1018     for(iidx=0; iidx<nri; iidx++)
1019     {
1020         /* Load shift vector for this list */
1021         i_shift_offset   = DIM*shiftidx[iidx];
1022
1023         /* Load limits for loop over neighbors */
1024         j_index_start    = jindex[iidx];
1025         j_index_end      = jindex[iidx+1];
1026
1027         /* Get outer coordinate index */
1028         inr              = iinr[iidx];
1029         i_coord_offset   = DIM*inr;
1030
1031         /* Load i particle coords and add shift vector */
1032         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1033                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1034
1035         fix0             = _mm_setzero_pd();
1036         fiy0             = _mm_setzero_pd();
1037         fiz0             = _mm_setzero_pd();
1038         fix1             = _mm_setzero_pd();
1039         fiy1             = _mm_setzero_pd();
1040         fiz1             = _mm_setzero_pd();
1041         fix2             = _mm_setzero_pd();
1042         fiy2             = _mm_setzero_pd();
1043         fiz2             = _mm_setzero_pd();
1044
1045         /* Start inner kernel loop */
1046         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1047         {
1048
1049             /* Get j neighbor index, and coordinate index */
1050             jnrA             = jjnr[jidx];
1051             jnrB             = jjnr[jidx+1];
1052             j_coord_offsetA  = DIM*jnrA;
1053             j_coord_offsetB  = DIM*jnrB;
1054
1055             /* load j atom coordinates */
1056             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1057                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1058
1059             /* Calculate displacement vector */
1060             dx00             = _mm_sub_pd(ix0,jx0);
1061             dy00             = _mm_sub_pd(iy0,jy0);
1062             dz00             = _mm_sub_pd(iz0,jz0);
1063             dx01             = _mm_sub_pd(ix0,jx1);
1064             dy01             = _mm_sub_pd(iy0,jy1);
1065             dz01             = _mm_sub_pd(iz0,jz1);
1066             dx02             = _mm_sub_pd(ix0,jx2);
1067             dy02             = _mm_sub_pd(iy0,jy2);
1068             dz02             = _mm_sub_pd(iz0,jz2);
1069             dx10             = _mm_sub_pd(ix1,jx0);
1070             dy10             = _mm_sub_pd(iy1,jy0);
1071             dz10             = _mm_sub_pd(iz1,jz0);
1072             dx11             = _mm_sub_pd(ix1,jx1);
1073             dy11             = _mm_sub_pd(iy1,jy1);
1074             dz11             = _mm_sub_pd(iz1,jz1);
1075             dx12             = _mm_sub_pd(ix1,jx2);
1076             dy12             = _mm_sub_pd(iy1,jy2);
1077             dz12             = _mm_sub_pd(iz1,jz2);
1078             dx20             = _mm_sub_pd(ix2,jx0);
1079             dy20             = _mm_sub_pd(iy2,jy0);
1080             dz20             = _mm_sub_pd(iz2,jz0);
1081             dx21             = _mm_sub_pd(ix2,jx1);
1082             dy21             = _mm_sub_pd(iy2,jy1);
1083             dz21             = _mm_sub_pd(iz2,jz1);
1084             dx22             = _mm_sub_pd(ix2,jx2);
1085             dy22             = _mm_sub_pd(iy2,jy2);
1086             dz22             = _mm_sub_pd(iz2,jz2);
1087
1088             /* Calculate squared distance and things based on it */
1089             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1090             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1091             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1092             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1093             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1094             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1095             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1096             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1097             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1098
1099             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1100             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1101             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1102             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1103             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1104             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1105             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1106             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1107             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1108
1109             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1110             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1111             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1112             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1113             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1114             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1115             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1116             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1117             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1118
1119             fjx0             = _mm_setzero_pd();
1120             fjy0             = _mm_setzero_pd();
1121             fjz0             = _mm_setzero_pd();
1122             fjx1             = _mm_setzero_pd();
1123             fjy1             = _mm_setzero_pd();
1124             fjz1             = _mm_setzero_pd();
1125             fjx2             = _mm_setzero_pd();
1126             fjy2             = _mm_setzero_pd();
1127             fjz2             = _mm_setzero_pd();
1128
1129             /**************************
1130              * CALCULATE INTERACTIONS *
1131              **************************/
1132
1133             r00              = _mm_mul_pd(rsq00,rinv00);
1134
1135             /* Calculate table index by multiplying r with table scale and truncate to integer */
1136             rt               = _mm_mul_pd(r00,vftabscale);
1137             vfitab           = _mm_cvttpd_epi32(rt);
1138 #ifdef __XOP__
1139             vfeps            = _mm_frcz_pd(rt);
1140 #else
1141             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1142 #endif
1143             twovfeps         = _mm_add_pd(vfeps,vfeps);
1144             vfitab           = _mm_slli_epi32(vfitab,3);
1145
1146             /* REACTION-FIELD ELECTROSTATICS */
1147             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1148
1149             /* CUBIC SPLINE TABLE DISPERSION */
1150             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1151             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1152             GMX_MM_TRANSPOSE2_PD(Y,F);
1153             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1154             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1155             GMX_MM_TRANSPOSE2_PD(G,H);
1156             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1157             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1158             fvdw6            = _mm_mul_pd(c6_00,FF);
1159
1160             /* CUBIC SPLINE TABLE REPULSION */
1161             vfitab           = _mm_add_epi32(vfitab,ifour);
1162             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1163             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1164             GMX_MM_TRANSPOSE2_PD(Y,F);
1165             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1166             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1167             GMX_MM_TRANSPOSE2_PD(G,H);
1168             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1169             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1170             fvdw12           = _mm_mul_pd(c12_00,FF);
1171             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1172
1173             fscal            = _mm_add_pd(felec,fvdw);
1174
1175             /* Update vectorial force */
1176             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1177             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1178             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1179
1180             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1181             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1182             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1183
1184             /**************************
1185              * CALCULATE INTERACTIONS *
1186              **************************/
1187
1188             /* REACTION-FIELD ELECTROSTATICS */
1189             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1190
1191             fscal            = felec;
1192
1193             /* Update vectorial force */
1194             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1195             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1196             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1197
1198             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1199             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1200             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1201
1202             /**************************
1203              * CALCULATE INTERACTIONS *
1204              **************************/
1205
1206             /* REACTION-FIELD ELECTROSTATICS */
1207             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1208
1209             fscal            = felec;
1210
1211             /* Update vectorial force */
1212             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1213             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1214             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1215
1216             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1217             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1218             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1219
1220             /**************************
1221              * CALCULATE INTERACTIONS *
1222              **************************/
1223
1224             /* REACTION-FIELD ELECTROSTATICS */
1225             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1226
1227             fscal            = felec;
1228
1229             /* Update vectorial force */
1230             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1231             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1232             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1233
1234             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1235             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1236             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1237
1238             /**************************
1239              * CALCULATE INTERACTIONS *
1240              **************************/
1241
1242             /* REACTION-FIELD ELECTROSTATICS */
1243             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1244
1245             fscal            = felec;
1246
1247             /* Update vectorial force */
1248             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1249             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1250             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1251
1252             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1253             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1254             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1255
1256             /**************************
1257              * CALCULATE INTERACTIONS *
1258              **************************/
1259
1260             /* REACTION-FIELD ELECTROSTATICS */
1261             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1262
1263             fscal            = felec;
1264
1265             /* Update vectorial force */
1266             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1267             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1268             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1269
1270             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1271             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1272             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1273
1274             /**************************
1275              * CALCULATE INTERACTIONS *
1276              **************************/
1277
1278             /* REACTION-FIELD ELECTROSTATICS */
1279             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1280
1281             fscal            = felec;
1282
1283             /* Update vectorial force */
1284             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1285             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1286             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1287
1288             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1289             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1290             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             /* REACTION-FIELD ELECTROSTATICS */
1297             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1298
1299             fscal            = felec;
1300
1301             /* Update vectorial force */
1302             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1303             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1304             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1305
1306             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1307             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1308             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1309
1310             /**************************
1311              * CALCULATE INTERACTIONS *
1312              **************************/
1313
1314             /* REACTION-FIELD ELECTROSTATICS */
1315             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1316
1317             fscal            = felec;
1318
1319             /* Update vectorial force */
1320             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1321             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1322             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1323
1324             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1325             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1326             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1327
1328             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1329
1330             /* Inner loop uses 297 flops */
1331         }
1332
1333         if(jidx<j_index_end)
1334         {
1335
1336             jnrA             = jjnr[jidx];
1337             j_coord_offsetA  = DIM*jnrA;
1338
1339             /* load j atom coordinates */
1340             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1341                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1342
1343             /* Calculate displacement vector */
1344             dx00             = _mm_sub_pd(ix0,jx0);
1345             dy00             = _mm_sub_pd(iy0,jy0);
1346             dz00             = _mm_sub_pd(iz0,jz0);
1347             dx01             = _mm_sub_pd(ix0,jx1);
1348             dy01             = _mm_sub_pd(iy0,jy1);
1349             dz01             = _mm_sub_pd(iz0,jz1);
1350             dx02             = _mm_sub_pd(ix0,jx2);
1351             dy02             = _mm_sub_pd(iy0,jy2);
1352             dz02             = _mm_sub_pd(iz0,jz2);
1353             dx10             = _mm_sub_pd(ix1,jx0);
1354             dy10             = _mm_sub_pd(iy1,jy0);
1355             dz10             = _mm_sub_pd(iz1,jz0);
1356             dx11             = _mm_sub_pd(ix1,jx1);
1357             dy11             = _mm_sub_pd(iy1,jy1);
1358             dz11             = _mm_sub_pd(iz1,jz1);
1359             dx12             = _mm_sub_pd(ix1,jx2);
1360             dy12             = _mm_sub_pd(iy1,jy2);
1361             dz12             = _mm_sub_pd(iz1,jz2);
1362             dx20             = _mm_sub_pd(ix2,jx0);
1363             dy20             = _mm_sub_pd(iy2,jy0);
1364             dz20             = _mm_sub_pd(iz2,jz0);
1365             dx21             = _mm_sub_pd(ix2,jx1);
1366             dy21             = _mm_sub_pd(iy2,jy1);
1367             dz21             = _mm_sub_pd(iz2,jz1);
1368             dx22             = _mm_sub_pd(ix2,jx2);
1369             dy22             = _mm_sub_pd(iy2,jy2);
1370             dz22             = _mm_sub_pd(iz2,jz2);
1371
1372             /* Calculate squared distance and things based on it */
1373             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1374             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1375             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1376             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1377             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1378             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1379             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1380             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1381             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1382
1383             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1384             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1385             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1386             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1387             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1388             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1389             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1390             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1391             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1392
1393             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1394             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1395             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1396             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1397             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1398             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1399             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1400             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1401             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1402
1403             fjx0             = _mm_setzero_pd();
1404             fjy0             = _mm_setzero_pd();
1405             fjz0             = _mm_setzero_pd();
1406             fjx1             = _mm_setzero_pd();
1407             fjy1             = _mm_setzero_pd();
1408             fjz1             = _mm_setzero_pd();
1409             fjx2             = _mm_setzero_pd();
1410             fjy2             = _mm_setzero_pd();
1411             fjz2             = _mm_setzero_pd();
1412
1413             /**************************
1414              * CALCULATE INTERACTIONS *
1415              **************************/
1416
1417             r00              = _mm_mul_pd(rsq00,rinv00);
1418
1419             /* Calculate table index by multiplying r with table scale and truncate to integer */
1420             rt               = _mm_mul_pd(r00,vftabscale);
1421             vfitab           = _mm_cvttpd_epi32(rt);
1422 #ifdef __XOP__
1423             vfeps            = _mm_frcz_pd(rt);
1424 #else
1425             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1426 #endif
1427             twovfeps         = _mm_add_pd(vfeps,vfeps);
1428             vfitab           = _mm_slli_epi32(vfitab,3);
1429
1430             /* REACTION-FIELD ELECTROSTATICS */
1431             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1432
1433             /* CUBIC SPLINE TABLE DISPERSION */
1434             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1435             F                = _mm_setzero_pd();
1436             GMX_MM_TRANSPOSE2_PD(Y,F);
1437             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1438             H                = _mm_setzero_pd();
1439             GMX_MM_TRANSPOSE2_PD(G,H);
1440             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1441             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1442             fvdw6            = _mm_mul_pd(c6_00,FF);
1443
1444             /* CUBIC SPLINE TABLE REPULSION */
1445             vfitab           = _mm_add_epi32(vfitab,ifour);
1446             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1447             F                = _mm_setzero_pd();
1448             GMX_MM_TRANSPOSE2_PD(Y,F);
1449             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1450             H                = _mm_setzero_pd();
1451             GMX_MM_TRANSPOSE2_PD(G,H);
1452             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1453             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1454             fvdw12           = _mm_mul_pd(c12_00,FF);
1455             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1456
1457             fscal            = _mm_add_pd(felec,fvdw);
1458
1459             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1460
1461             /* Update vectorial force */
1462             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1463             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1464             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1465
1466             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1467             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1468             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1469
1470             /**************************
1471              * CALCULATE INTERACTIONS *
1472              **************************/
1473
1474             /* REACTION-FIELD ELECTROSTATICS */
1475             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1476
1477             fscal            = felec;
1478
1479             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1480
1481             /* Update vectorial force */
1482             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1483             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1484             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1485
1486             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1487             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1488             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1489
1490             /**************************
1491              * CALCULATE INTERACTIONS *
1492              **************************/
1493
1494             /* REACTION-FIELD ELECTROSTATICS */
1495             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1496
1497             fscal            = felec;
1498
1499             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1500
1501             /* Update vectorial force */
1502             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1503             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1504             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1505
1506             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1507             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1508             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1509
1510             /**************************
1511              * CALCULATE INTERACTIONS *
1512              **************************/
1513
1514             /* REACTION-FIELD ELECTROSTATICS */
1515             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1516
1517             fscal            = felec;
1518
1519             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1520
1521             /* Update vectorial force */
1522             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1523             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1524             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1525
1526             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1527             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1528             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1529
1530             /**************************
1531              * CALCULATE INTERACTIONS *
1532              **************************/
1533
1534             /* REACTION-FIELD ELECTROSTATICS */
1535             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1536
1537             fscal            = felec;
1538
1539             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1540
1541             /* Update vectorial force */
1542             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1543             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1544             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1545
1546             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1547             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1548             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1549
1550             /**************************
1551              * CALCULATE INTERACTIONS *
1552              **************************/
1553
1554             /* REACTION-FIELD ELECTROSTATICS */
1555             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1556
1557             fscal            = felec;
1558
1559             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1560
1561             /* Update vectorial force */
1562             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1563             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1564             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1565
1566             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1567             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1568             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1569
1570             /**************************
1571              * CALCULATE INTERACTIONS *
1572              **************************/
1573
1574             /* REACTION-FIELD ELECTROSTATICS */
1575             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1576
1577             fscal            = felec;
1578
1579             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1580
1581             /* Update vectorial force */
1582             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1583             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1584             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1585
1586             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1587             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1588             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1589
1590             /**************************
1591              * CALCULATE INTERACTIONS *
1592              **************************/
1593
1594             /* REACTION-FIELD ELECTROSTATICS */
1595             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1596
1597             fscal            = felec;
1598
1599             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1600
1601             /* Update vectorial force */
1602             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1603             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1604             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1605
1606             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1607             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1608             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1609
1610             /**************************
1611              * CALCULATE INTERACTIONS *
1612              **************************/
1613
1614             /* REACTION-FIELD ELECTROSTATICS */
1615             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1616
1617             fscal            = felec;
1618
1619             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1620
1621             /* Update vectorial force */
1622             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1623             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1624             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1625
1626             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1627             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1628             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1629
1630             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1631
1632             /* Inner loop uses 297 flops */
1633         }
1634
1635         /* End of innermost loop */
1636
1637         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1638                                               f+i_coord_offset,fshift+i_shift_offset);
1639
1640         /* Increment number of inner iterations */
1641         inneriter                  += j_index_end - j_index_start;
1642
1643         /* Outer loop uses 18 flops */
1644     }
1645
1646     /* Increment number of outer iterations */
1647     outeriter        += nri;
1648
1649     /* Update outer/inner flops */
1650
1651     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*297);
1652 }