| File: | gromacs/gmxlib/nonbonded/nb_kernel_c/nb_kernel_ElecEw_VdwLJ_GeomW4W4_c.c |
| Location: | line 125, column 5 |
| Description: | Value stored to 'sh_ewald' is never read |
| 1 | /* |
| 2 | * This file is part of the GROMACS molecular simulation package. |
| 3 | * |
| 4 | * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by |
| 5 | * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl, |
| 6 | * and including many others, as listed in the AUTHORS file in the |
| 7 | * top-level source directory and at http://www.gromacs.org. |
| 8 | * |
| 9 | * GROMACS is free software; you can redistribute it and/or |
| 10 | * modify it under the terms of the GNU Lesser General Public License |
| 11 | * as published by the Free Software Foundation; either version 2.1 |
| 12 | * of the License, or (at your option) any later version. |
| 13 | * |
| 14 | * GROMACS is distributed in the hope that it will be useful, |
| 15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 17 | * Lesser General Public License for more details. |
| 18 | * |
| 19 | * You should have received a copy of the GNU Lesser General Public |
| 20 | * License along with GROMACS; if not, see |
| 21 | * http://www.gnu.org/licenses, or write to the Free Software Foundation, |
| 22 | * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |
| 23 | * |
| 24 | * If you want to redistribute modifications to GROMACS, please |
| 25 | * consider that scientific software is very special. Version |
| 26 | * control is crucial - bugs must be traceable. We will be happy to |
| 27 | * consider code for inclusion in the official distribution, but |
| 28 | * derived work must not be called official GROMACS. Details are found |
| 29 | * in the README & COPYING files - if they are missing, get the |
| 30 | * official version at http://www.gromacs.org. |
| 31 | * |
| 32 | * To help us fund GROMACS development, we humbly ask that you cite |
| 33 | * the research papers on the package. Check out http://www.gromacs.org. |
| 34 | */ |
| 35 | /* |
| 36 | * Note: this file was generated by the GROMACS c kernel generator. |
| 37 | */ |
| 38 | #ifdef HAVE_CONFIG_H1 |
| 39 | #include <config.h> |
| 40 | #endif |
| 41 | |
| 42 | #include <math.h> |
| 43 | |
| 44 | #include "../nb_kernel.h" |
| 45 | #include "types/simple.h" |
| 46 | #include "gromacs/math/vec.h" |
| 47 | #include "nrnb.h" |
| 48 | |
| 49 | /* |
| 50 | * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwLJ_GeomW4W4_VF_c |
| 51 | * Electrostatics interaction: Ewald |
| 52 | * VdW interaction: LennardJones |
| 53 | * Geometry: Water4-Water4 |
| 54 | * Calculate force/pot: PotentialAndForce |
| 55 | */ |
| 56 | void |
| 57 | nb_kernel_ElecEw_VdwLJ_GeomW4W4_VF_c |
| 58 | (t_nblist * gmx_restrict__restrict nlist, |
| 59 | rvec * gmx_restrict__restrict xx, |
| 60 | rvec * gmx_restrict__restrict ff, |
| 61 | t_forcerec * gmx_restrict__restrict fr, |
| 62 | t_mdatoms * gmx_restrict__restrict mdatoms, |
| 63 | nb_kernel_data_t gmx_unused__attribute__ ((unused)) * gmx_restrict__restrict kernel_data, |
| 64 | t_nrnb * gmx_restrict__restrict nrnb) |
| 65 | { |
| 66 | int i_shift_offset,i_coord_offset,j_coord_offset; |
| 67 | int j_index_start,j_index_end; |
| 68 | int nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter; |
| 69 | real shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2; |
| 70 | int *iinr,*jindex,*jjnr,*shiftidx,*gid; |
| 71 | real *shiftvec,*fshift,*x,*f; |
| 72 | int vdwioffset0; |
| 73 | real ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0; |
| 74 | int vdwioffset1; |
| 75 | real ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1; |
| 76 | int vdwioffset2; |
| 77 | real ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2; |
| 78 | int vdwioffset3; |
| 79 | real ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3; |
| 80 | int vdwjidx0; |
| 81 | real jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0; |
| 82 | int vdwjidx1; |
| 83 | real jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1; |
| 84 | int vdwjidx2; |
| 85 | real jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2; |
| 86 | int vdwjidx3; |
| 87 | real jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3; |
| 88 | real dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00; |
| 89 | real dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11; |
| 90 | real dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12; |
| 91 | real dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13; |
| 92 | real dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21; |
| 93 | real dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22; |
| 94 | real dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23; |
| 95 | real dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31; |
| 96 | real dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32; |
| 97 | real dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33; |
| 98 | real velec,felec,velecsum,facel,crf,krf,krf2; |
| 99 | real *charge; |
| 100 | int nvdwtype; |
| 101 | real rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6; |
| 102 | int *vdwtype; |
| 103 | real *vdwparam; |
| 104 | int ewitab; |
| 105 | real ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace; |
| 106 | real *ewtab; |
| 107 | |
| 108 | x = xx[0]; |
| 109 | f = ff[0]; |
| 110 | |
| 111 | nri = nlist->nri; |
| 112 | iinr = nlist->iinr; |
| 113 | jindex = nlist->jindex; |
| 114 | jjnr = nlist->jjnr; |
| 115 | shiftidx = nlist->shift; |
| 116 | gid = nlist->gid; |
| 117 | shiftvec = fr->shift_vec[0]; |
| 118 | fshift = fr->fshift[0]; |
| 119 | facel = fr->epsfac; |
| 120 | charge = mdatoms->chargeA; |
| 121 | nvdwtype = fr->ntype; |
| 122 | vdwparam = fr->nbfp; |
| 123 | vdwtype = mdatoms->typeA; |
| 124 | |
| 125 | sh_ewald = fr->ic->sh_ewald; |
Value stored to 'sh_ewald' is never read | |
| 126 | ewtab = fr->ic->tabq_coul_FDV0; |
| 127 | ewtabscale = fr->ic->tabq_scale; |
| 128 | ewtabhalfspace = 0.5/ewtabscale; |
| 129 | |
| 130 | /* Setup water-specific parameters */ |
| 131 | inr = nlist->iinr[0]; |
| 132 | iq1 = facel*charge[inr+1]; |
| 133 | iq2 = facel*charge[inr+2]; |
| 134 | iq3 = facel*charge[inr+3]; |
| 135 | vdwioffset0 = 2*nvdwtype*vdwtype[inr+0]; |
| 136 | |
| 137 | jq1 = charge[inr+1]; |
| 138 | jq2 = charge[inr+2]; |
| 139 | jq3 = charge[inr+3]; |
| 140 | vdwjidx0 = 2*vdwtype[inr+0]; |
| 141 | c6_00 = vdwparam[vdwioffset0+vdwjidx0]; |
| 142 | c12_00 = vdwparam[vdwioffset0+vdwjidx0+1]; |
| 143 | qq11 = iq1*jq1; |
| 144 | qq12 = iq1*jq2; |
| 145 | qq13 = iq1*jq3; |
| 146 | qq21 = iq2*jq1; |
| 147 | qq22 = iq2*jq2; |
| 148 | qq23 = iq2*jq3; |
| 149 | qq31 = iq3*jq1; |
| 150 | qq32 = iq3*jq2; |
| 151 | qq33 = iq3*jq3; |
| 152 | |
| 153 | outeriter = 0; |
| 154 | inneriter = 0; |
| 155 | |
| 156 | /* Start outer loop over neighborlists */ |
| 157 | for(iidx=0; iidx<nri; iidx++) |
| 158 | { |
| 159 | /* Load shift vector for this list */ |
| 160 | i_shift_offset = DIM3*shiftidx[iidx]; |
| 161 | shX = shiftvec[i_shift_offset+XX0]; |
| 162 | shY = shiftvec[i_shift_offset+YY1]; |
| 163 | shZ = shiftvec[i_shift_offset+ZZ2]; |
| 164 | |
| 165 | /* Load limits for loop over neighbors */ |
| 166 | j_index_start = jindex[iidx]; |
| 167 | j_index_end = jindex[iidx+1]; |
| 168 | |
| 169 | /* Get outer coordinate index */ |
| 170 | inr = iinr[iidx]; |
| 171 | i_coord_offset = DIM3*inr; |
| 172 | |
| 173 | /* Load i particle coords and add shift vector */ |
| 174 | ix0 = shX + x[i_coord_offset+DIM3*0+XX0]; |
| 175 | iy0 = shY + x[i_coord_offset+DIM3*0+YY1]; |
| 176 | iz0 = shZ + x[i_coord_offset+DIM3*0+ZZ2]; |
| 177 | ix1 = shX + x[i_coord_offset+DIM3*1+XX0]; |
| 178 | iy1 = shY + x[i_coord_offset+DIM3*1+YY1]; |
| 179 | iz1 = shZ + x[i_coord_offset+DIM3*1+ZZ2]; |
| 180 | ix2 = shX + x[i_coord_offset+DIM3*2+XX0]; |
| 181 | iy2 = shY + x[i_coord_offset+DIM3*2+YY1]; |
| 182 | iz2 = shZ + x[i_coord_offset+DIM3*2+ZZ2]; |
| 183 | ix3 = shX + x[i_coord_offset+DIM3*3+XX0]; |
| 184 | iy3 = shY + x[i_coord_offset+DIM3*3+YY1]; |
| 185 | iz3 = shZ + x[i_coord_offset+DIM3*3+ZZ2]; |
| 186 | |
| 187 | fix0 = 0.0; |
| 188 | fiy0 = 0.0; |
| 189 | fiz0 = 0.0; |
| 190 | fix1 = 0.0; |
| 191 | fiy1 = 0.0; |
| 192 | fiz1 = 0.0; |
| 193 | fix2 = 0.0; |
| 194 | fiy2 = 0.0; |
| 195 | fiz2 = 0.0; |
| 196 | fix3 = 0.0; |
| 197 | fiy3 = 0.0; |
| 198 | fiz3 = 0.0; |
| 199 | |
| 200 | /* Reset potential sums */ |
| 201 | velecsum = 0.0; |
| 202 | vvdwsum = 0.0; |
| 203 | |
| 204 | /* Start inner kernel loop */ |
| 205 | for(jidx=j_index_start; jidx<j_index_end; jidx++) |
| 206 | { |
| 207 | /* Get j neighbor index, and coordinate index */ |
| 208 | jnr = jjnr[jidx]; |
| 209 | j_coord_offset = DIM3*jnr; |
| 210 | |
| 211 | /* load j atom coordinates */ |
| 212 | jx0 = x[j_coord_offset+DIM3*0+XX0]; |
| 213 | jy0 = x[j_coord_offset+DIM3*0+YY1]; |
| 214 | jz0 = x[j_coord_offset+DIM3*0+ZZ2]; |
| 215 | jx1 = x[j_coord_offset+DIM3*1+XX0]; |
| 216 | jy1 = x[j_coord_offset+DIM3*1+YY1]; |
| 217 | jz1 = x[j_coord_offset+DIM3*1+ZZ2]; |
| 218 | jx2 = x[j_coord_offset+DIM3*2+XX0]; |
| 219 | jy2 = x[j_coord_offset+DIM3*2+YY1]; |
| 220 | jz2 = x[j_coord_offset+DIM3*2+ZZ2]; |
| 221 | jx3 = x[j_coord_offset+DIM3*3+XX0]; |
| 222 | jy3 = x[j_coord_offset+DIM3*3+YY1]; |
| 223 | jz3 = x[j_coord_offset+DIM3*3+ZZ2]; |
| 224 | |
| 225 | /* Calculate displacement vector */ |
| 226 | dx00 = ix0 - jx0; |
| 227 | dy00 = iy0 - jy0; |
| 228 | dz00 = iz0 - jz0; |
| 229 | dx11 = ix1 - jx1; |
| 230 | dy11 = iy1 - jy1; |
| 231 | dz11 = iz1 - jz1; |
| 232 | dx12 = ix1 - jx2; |
| 233 | dy12 = iy1 - jy2; |
| 234 | dz12 = iz1 - jz2; |
| 235 | dx13 = ix1 - jx3; |
| 236 | dy13 = iy1 - jy3; |
| 237 | dz13 = iz1 - jz3; |
| 238 | dx21 = ix2 - jx1; |
| 239 | dy21 = iy2 - jy1; |
| 240 | dz21 = iz2 - jz1; |
| 241 | dx22 = ix2 - jx2; |
| 242 | dy22 = iy2 - jy2; |
| 243 | dz22 = iz2 - jz2; |
| 244 | dx23 = ix2 - jx3; |
| 245 | dy23 = iy2 - jy3; |
| 246 | dz23 = iz2 - jz3; |
| 247 | dx31 = ix3 - jx1; |
| 248 | dy31 = iy3 - jy1; |
| 249 | dz31 = iz3 - jz1; |
| 250 | dx32 = ix3 - jx2; |
| 251 | dy32 = iy3 - jy2; |
| 252 | dz32 = iz3 - jz2; |
| 253 | dx33 = ix3 - jx3; |
| 254 | dy33 = iy3 - jy3; |
| 255 | dz33 = iz3 - jz3; |
| 256 | |
| 257 | /* Calculate squared distance and things based on it */ |
| 258 | rsq00 = dx00*dx00+dy00*dy00+dz00*dz00; |
| 259 | rsq11 = dx11*dx11+dy11*dy11+dz11*dz11; |
| 260 | rsq12 = dx12*dx12+dy12*dy12+dz12*dz12; |
| 261 | rsq13 = dx13*dx13+dy13*dy13+dz13*dz13; |
| 262 | rsq21 = dx21*dx21+dy21*dy21+dz21*dz21; |
| 263 | rsq22 = dx22*dx22+dy22*dy22+dz22*dz22; |
| 264 | rsq23 = dx23*dx23+dy23*dy23+dz23*dz23; |
| 265 | rsq31 = dx31*dx31+dy31*dy31+dz31*dz31; |
| 266 | rsq32 = dx32*dx32+dy32*dy32+dz32*dz32; |
| 267 | rsq33 = dx33*dx33+dy33*dy33+dz33*dz33; |
| 268 | |
| 269 | rinv11 = gmx_invsqrt(rsq11)gmx_software_invsqrt(rsq11); |
| 270 | rinv12 = gmx_invsqrt(rsq12)gmx_software_invsqrt(rsq12); |
| 271 | rinv13 = gmx_invsqrt(rsq13)gmx_software_invsqrt(rsq13); |
| 272 | rinv21 = gmx_invsqrt(rsq21)gmx_software_invsqrt(rsq21); |
| 273 | rinv22 = gmx_invsqrt(rsq22)gmx_software_invsqrt(rsq22); |
| 274 | rinv23 = gmx_invsqrt(rsq23)gmx_software_invsqrt(rsq23); |
| 275 | rinv31 = gmx_invsqrt(rsq31)gmx_software_invsqrt(rsq31); |
| 276 | rinv32 = gmx_invsqrt(rsq32)gmx_software_invsqrt(rsq32); |
| 277 | rinv33 = gmx_invsqrt(rsq33)gmx_software_invsqrt(rsq33); |
| 278 | |
| 279 | rinvsq00 = 1.0/rsq00; |
| 280 | rinvsq11 = rinv11*rinv11; |
| 281 | rinvsq12 = rinv12*rinv12; |
| 282 | rinvsq13 = rinv13*rinv13; |
| 283 | rinvsq21 = rinv21*rinv21; |
| 284 | rinvsq22 = rinv22*rinv22; |
| 285 | rinvsq23 = rinv23*rinv23; |
| 286 | rinvsq31 = rinv31*rinv31; |
| 287 | rinvsq32 = rinv32*rinv32; |
| 288 | rinvsq33 = rinv33*rinv33; |
| 289 | |
| 290 | /************************** |
| 291 | * CALCULATE INTERACTIONS * |
| 292 | **************************/ |
| 293 | |
| 294 | /* LENNARD-JONES DISPERSION/REPULSION */ |
| 295 | |
| 296 | rinvsix = rinvsq00*rinvsq00*rinvsq00; |
| 297 | vvdw6 = c6_00*rinvsix; |
| 298 | vvdw12 = c12_00*rinvsix*rinvsix; |
| 299 | vvdw = vvdw12*(1.0/12.0) - vvdw6*(1.0/6.0); |
| 300 | fvdw = (vvdw12-vvdw6)*rinvsq00; |
| 301 | |
| 302 | /* Update potential sums from outer loop */ |
| 303 | vvdwsum += vvdw; |
| 304 | |
| 305 | fscal = fvdw; |
| 306 | |
| 307 | /* Calculate temporary vectorial force */ |
| 308 | tx = fscal*dx00; |
| 309 | ty = fscal*dy00; |
| 310 | tz = fscal*dz00; |
| 311 | |
| 312 | /* Update vectorial force */ |
| 313 | fix0 += tx; |
| 314 | fiy0 += ty; |
| 315 | fiz0 += tz; |
| 316 | f[j_coord_offset+DIM3*0+XX0] -= tx; |
| 317 | f[j_coord_offset+DIM3*0+YY1] -= ty; |
| 318 | f[j_coord_offset+DIM3*0+ZZ2] -= tz; |
| 319 | |
| 320 | /************************** |
| 321 | * CALCULATE INTERACTIONS * |
| 322 | **************************/ |
| 323 | |
| 324 | r11 = rsq11*rinv11; |
| 325 | |
| 326 | /* EWALD ELECTROSTATICS */ |
| 327 | |
| 328 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 329 | ewrt = r11*ewtabscale; |
| 330 | ewitab = ewrt; |
| 331 | eweps = ewrt-ewitab; |
| 332 | ewitab = 4*ewitab; |
| 333 | felec = ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 334 | velec = qq11*(rinv11-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec))); |
| 335 | felec = qq11*rinv11*(rinvsq11-felec); |
| 336 | |
| 337 | /* Update potential sums from outer loop */ |
| 338 | velecsum += velec; |
| 339 | |
| 340 | fscal = felec; |
| 341 | |
| 342 | /* Calculate temporary vectorial force */ |
| 343 | tx = fscal*dx11; |
| 344 | ty = fscal*dy11; |
| 345 | tz = fscal*dz11; |
| 346 | |
| 347 | /* Update vectorial force */ |
| 348 | fix1 += tx; |
| 349 | fiy1 += ty; |
| 350 | fiz1 += tz; |
| 351 | f[j_coord_offset+DIM3*1+XX0] -= tx; |
| 352 | f[j_coord_offset+DIM3*1+YY1] -= ty; |
| 353 | f[j_coord_offset+DIM3*1+ZZ2] -= tz; |
| 354 | |
| 355 | /************************** |
| 356 | * CALCULATE INTERACTIONS * |
| 357 | **************************/ |
| 358 | |
| 359 | r12 = rsq12*rinv12; |
| 360 | |
| 361 | /* EWALD ELECTROSTATICS */ |
| 362 | |
| 363 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 364 | ewrt = r12*ewtabscale; |
| 365 | ewitab = ewrt; |
| 366 | eweps = ewrt-ewitab; |
| 367 | ewitab = 4*ewitab; |
| 368 | felec = ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 369 | velec = qq12*(rinv12-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec))); |
| 370 | felec = qq12*rinv12*(rinvsq12-felec); |
| 371 | |
| 372 | /* Update potential sums from outer loop */ |
| 373 | velecsum += velec; |
| 374 | |
| 375 | fscal = felec; |
| 376 | |
| 377 | /* Calculate temporary vectorial force */ |
| 378 | tx = fscal*dx12; |
| 379 | ty = fscal*dy12; |
| 380 | tz = fscal*dz12; |
| 381 | |
| 382 | /* Update vectorial force */ |
| 383 | fix1 += tx; |
| 384 | fiy1 += ty; |
| 385 | fiz1 += tz; |
| 386 | f[j_coord_offset+DIM3*2+XX0] -= tx; |
| 387 | f[j_coord_offset+DIM3*2+YY1] -= ty; |
| 388 | f[j_coord_offset+DIM3*2+ZZ2] -= tz; |
| 389 | |
| 390 | /************************** |
| 391 | * CALCULATE INTERACTIONS * |
| 392 | **************************/ |
| 393 | |
| 394 | r13 = rsq13*rinv13; |
| 395 | |
| 396 | /* EWALD ELECTROSTATICS */ |
| 397 | |
| 398 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 399 | ewrt = r13*ewtabscale; |
| 400 | ewitab = ewrt; |
| 401 | eweps = ewrt-ewitab; |
| 402 | ewitab = 4*ewitab; |
| 403 | felec = ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 404 | velec = qq13*(rinv13-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec))); |
| 405 | felec = qq13*rinv13*(rinvsq13-felec); |
| 406 | |
| 407 | /* Update potential sums from outer loop */ |
| 408 | velecsum += velec; |
| 409 | |
| 410 | fscal = felec; |
| 411 | |
| 412 | /* Calculate temporary vectorial force */ |
| 413 | tx = fscal*dx13; |
| 414 | ty = fscal*dy13; |
| 415 | tz = fscal*dz13; |
| 416 | |
| 417 | /* Update vectorial force */ |
| 418 | fix1 += tx; |
| 419 | fiy1 += ty; |
| 420 | fiz1 += tz; |
| 421 | f[j_coord_offset+DIM3*3+XX0] -= tx; |
| 422 | f[j_coord_offset+DIM3*3+YY1] -= ty; |
| 423 | f[j_coord_offset+DIM3*3+ZZ2] -= tz; |
| 424 | |
| 425 | /************************** |
| 426 | * CALCULATE INTERACTIONS * |
| 427 | **************************/ |
| 428 | |
| 429 | r21 = rsq21*rinv21; |
| 430 | |
| 431 | /* EWALD ELECTROSTATICS */ |
| 432 | |
| 433 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 434 | ewrt = r21*ewtabscale; |
| 435 | ewitab = ewrt; |
| 436 | eweps = ewrt-ewitab; |
| 437 | ewitab = 4*ewitab; |
| 438 | felec = ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 439 | velec = qq21*(rinv21-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec))); |
| 440 | felec = qq21*rinv21*(rinvsq21-felec); |
| 441 | |
| 442 | /* Update potential sums from outer loop */ |
| 443 | velecsum += velec; |
| 444 | |
| 445 | fscal = felec; |
| 446 | |
| 447 | /* Calculate temporary vectorial force */ |
| 448 | tx = fscal*dx21; |
| 449 | ty = fscal*dy21; |
| 450 | tz = fscal*dz21; |
| 451 | |
| 452 | /* Update vectorial force */ |
| 453 | fix2 += tx; |
| 454 | fiy2 += ty; |
| 455 | fiz2 += tz; |
| 456 | f[j_coord_offset+DIM3*1+XX0] -= tx; |
| 457 | f[j_coord_offset+DIM3*1+YY1] -= ty; |
| 458 | f[j_coord_offset+DIM3*1+ZZ2] -= tz; |
| 459 | |
| 460 | /************************** |
| 461 | * CALCULATE INTERACTIONS * |
| 462 | **************************/ |
| 463 | |
| 464 | r22 = rsq22*rinv22; |
| 465 | |
| 466 | /* EWALD ELECTROSTATICS */ |
| 467 | |
| 468 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 469 | ewrt = r22*ewtabscale; |
| 470 | ewitab = ewrt; |
| 471 | eweps = ewrt-ewitab; |
| 472 | ewitab = 4*ewitab; |
| 473 | felec = ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 474 | velec = qq22*(rinv22-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec))); |
| 475 | felec = qq22*rinv22*(rinvsq22-felec); |
| 476 | |
| 477 | /* Update potential sums from outer loop */ |
| 478 | velecsum += velec; |
| 479 | |
| 480 | fscal = felec; |
| 481 | |
| 482 | /* Calculate temporary vectorial force */ |
| 483 | tx = fscal*dx22; |
| 484 | ty = fscal*dy22; |
| 485 | tz = fscal*dz22; |
| 486 | |
| 487 | /* Update vectorial force */ |
| 488 | fix2 += tx; |
| 489 | fiy2 += ty; |
| 490 | fiz2 += tz; |
| 491 | f[j_coord_offset+DIM3*2+XX0] -= tx; |
| 492 | f[j_coord_offset+DIM3*2+YY1] -= ty; |
| 493 | f[j_coord_offset+DIM3*2+ZZ2] -= tz; |
| 494 | |
| 495 | /************************** |
| 496 | * CALCULATE INTERACTIONS * |
| 497 | **************************/ |
| 498 | |
| 499 | r23 = rsq23*rinv23; |
| 500 | |
| 501 | /* EWALD ELECTROSTATICS */ |
| 502 | |
| 503 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 504 | ewrt = r23*ewtabscale; |
| 505 | ewitab = ewrt; |
| 506 | eweps = ewrt-ewitab; |
| 507 | ewitab = 4*ewitab; |
| 508 | felec = ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 509 | velec = qq23*(rinv23-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec))); |
| 510 | felec = qq23*rinv23*(rinvsq23-felec); |
| 511 | |
| 512 | /* Update potential sums from outer loop */ |
| 513 | velecsum += velec; |
| 514 | |
| 515 | fscal = felec; |
| 516 | |
| 517 | /* Calculate temporary vectorial force */ |
| 518 | tx = fscal*dx23; |
| 519 | ty = fscal*dy23; |
| 520 | tz = fscal*dz23; |
| 521 | |
| 522 | /* Update vectorial force */ |
| 523 | fix2 += tx; |
| 524 | fiy2 += ty; |
| 525 | fiz2 += tz; |
| 526 | f[j_coord_offset+DIM3*3+XX0] -= tx; |
| 527 | f[j_coord_offset+DIM3*3+YY1] -= ty; |
| 528 | f[j_coord_offset+DIM3*3+ZZ2] -= tz; |
| 529 | |
| 530 | /************************** |
| 531 | * CALCULATE INTERACTIONS * |
| 532 | **************************/ |
| 533 | |
| 534 | r31 = rsq31*rinv31; |
| 535 | |
| 536 | /* EWALD ELECTROSTATICS */ |
| 537 | |
| 538 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 539 | ewrt = r31*ewtabscale; |
| 540 | ewitab = ewrt; |
| 541 | eweps = ewrt-ewitab; |
| 542 | ewitab = 4*ewitab; |
| 543 | felec = ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 544 | velec = qq31*(rinv31-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec))); |
| 545 | felec = qq31*rinv31*(rinvsq31-felec); |
| 546 | |
| 547 | /* Update potential sums from outer loop */ |
| 548 | velecsum += velec; |
| 549 | |
| 550 | fscal = felec; |
| 551 | |
| 552 | /* Calculate temporary vectorial force */ |
| 553 | tx = fscal*dx31; |
| 554 | ty = fscal*dy31; |
| 555 | tz = fscal*dz31; |
| 556 | |
| 557 | /* Update vectorial force */ |
| 558 | fix3 += tx; |
| 559 | fiy3 += ty; |
| 560 | fiz3 += tz; |
| 561 | f[j_coord_offset+DIM3*1+XX0] -= tx; |
| 562 | f[j_coord_offset+DIM3*1+YY1] -= ty; |
| 563 | f[j_coord_offset+DIM3*1+ZZ2] -= tz; |
| 564 | |
| 565 | /************************** |
| 566 | * CALCULATE INTERACTIONS * |
| 567 | **************************/ |
| 568 | |
| 569 | r32 = rsq32*rinv32; |
| 570 | |
| 571 | /* EWALD ELECTROSTATICS */ |
| 572 | |
| 573 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 574 | ewrt = r32*ewtabscale; |
| 575 | ewitab = ewrt; |
| 576 | eweps = ewrt-ewitab; |
| 577 | ewitab = 4*ewitab; |
| 578 | felec = ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 579 | velec = qq32*(rinv32-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec))); |
| 580 | felec = qq32*rinv32*(rinvsq32-felec); |
| 581 | |
| 582 | /* Update potential sums from outer loop */ |
| 583 | velecsum += velec; |
| 584 | |
| 585 | fscal = felec; |
| 586 | |
| 587 | /* Calculate temporary vectorial force */ |
| 588 | tx = fscal*dx32; |
| 589 | ty = fscal*dy32; |
| 590 | tz = fscal*dz32; |
| 591 | |
| 592 | /* Update vectorial force */ |
| 593 | fix3 += tx; |
| 594 | fiy3 += ty; |
| 595 | fiz3 += tz; |
| 596 | f[j_coord_offset+DIM3*2+XX0] -= tx; |
| 597 | f[j_coord_offset+DIM3*2+YY1] -= ty; |
| 598 | f[j_coord_offset+DIM3*2+ZZ2] -= tz; |
| 599 | |
| 600 | /************************** |
| 601 | * CALCULATE INTERACTIONS * |
| 602 | **************************/ |
| 603 | |
| 604 | r33 = rsq33*rinv33; |
| 605 | |
| 606 | /* EWALD ELECTROSTATICS */ |
| 607 | |
| 608 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 609 | ewrt = r33*ewtabscale; |
| 610 | ewitab = ewrt; |
| 611 | eweps = ewrt-ewitab; |
| 612 | ewitab = 4*ewitab; |
| 613 | felec = ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 614 | velec = qq33*(rinv33-(ewtab[ewitab+2]-ewtabhalfspace*eweps*(ewtab[ewitab]+felec))); |
| 615 | felec = qq33*rinv33*(rinvsq33-felec); |
| 616 | |
| 617 | /* Update potential sums from outer loop */ |
| 618 | velecsum += velec; |
| 619 | |
| 620 | fscal = felec; |
| 621 | |
| 622 | /* Calculate temporary vectorial force */ |
| 623 | tx = fscal*dx33; |
| 624 | ty = fscal*dy33; |
| 625 | tz = fscal*dz33; |
| 626 | |
| 627 | /* Update vectorial force */ |
| 628 | fix3 += tx; |
| 629 | fiy3 += ty; |
| 630 | fiz3 += tz; |
| 631 | f[j_coord_offset+DIM3*3+XX0] -= tx; |
| 632 | f[j_coord_offset+DIM3*3+YY1] -= ty; |
| 633 | f[j_coord_offset+DIM3*3+ZZ2] -= tz; |
| 634 | |
| 635 | /* Inner loop uses 392 flops */ |
| 636 | } |
| 637 | /* End of innermost loop */ |
| 638 | |
| 639 | tx = ty = tz = 0; |
| 640 | f[i_coord_offset+DIM3*0+XX0] += fix0; |
| 641 | f[i_coord_offset+DIM3*0+YY1] += fiy0; |
| 642 | f[i_coord_offset+DIM3*0+ZZ2] += fiz0; |
| 643 | tx += fix0; |
| 644 | ty += fiy0; |
| 645 | tz += fiz0; |
| 646 | f[i_coord_offset+DIM3*1+XX0] += fix1; |
| 647 | f[i_coord_offset+DIM3*1+YY1] += fiy1; |
| 648 | f[i_coord_offset+DIM3*1+ZZ2] += fiz1; |
| 649 | tx += fix1; |
| 650 | ty += fiy1; |
| 651 | tz += fiz1; |
| 652 | f[i_coord_offset+DIM3*2+XX0] += fix2; |
| 653 | f[i_coord_offset+DIM3*2+YY1] += fiy2; |
| 654 | f[i_coord_offset+DIM3*2+ZZ2] += fiz2; |
| 655 | tx += fix2; |
| 656 | ty += fiy2; |
| 657 | tz += fiz2; |
| 658 | f[i_coord_offset+DIM3*3+XX0] += fix3; |
| 659 | f[i_coord_offset+DIM3*3+YY1] += fiy3; |
| 660 | f[i_coord_offset+DIM3*3+ZZ2] += fiz3; |
| 661 | tx += fix3; |
| 662 | ty += fiy3; |
| 663 | tz += fiz3; |
| 664 | fshift[i_shift_offset+XX0] += tx; |
| 665 | fshift[i_shift_offset+YY1] += ty; |
| 666 | fshift[i_shift_offset+ZZ2] += tz; |
| 667 | |
| 668 | ggid = gid[iidx]; |
| 669 | /* Update potential energies */ |
| 670 | kernel_data->energygrp_elec[ggid] += velecsum; |
| 671 | kernel_data->energygrp_vdw[ggid] += vvdwsum; |
| 672 | |
| 673 | /* Increment number of inner iterations */ |
| 674 | inneriter += j_index_end - j_index_start; |
| 675 | |
| 676 | /* Outer loop uses 41 flops */ |
| 677 | } |
| 678 | |
| 679 | /* Increment number of outer iterations */ |
| 680 | outeriter += nri; |
| 681 | |
| 682 | /* Update outer/inner flops */ |
| 683 | |
| 684 | inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*41 + inneriter*392)(nrnb)->n[eNR_NBKERNEL_ELEC_VDW_W4W4_VF] += outeriter*41 + inneriter*392; |
| 685 | } |
| 686 | /* |
| 687 | * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwLJ_GeomW4W4_F_c |
| 688 | * Electrostatics interaction: Ewald |
| 689 | * VdW interaction: LennardJones |
| 690 | * Geometry: Water4-Water4 |
| 691 | * Calculate force/pot: Force |
| 692 | */ |
| 693 | void |
| 694 | nb_kernel_ElecEw_VdwLJ_GeomW4W4_F_c |
| 695 | (t_nblist * gmx_restrict__restrict nlist, |
| 696 | rvec * gmx_restrict__restrict xx, |
| 697 | rvec * gmx_restrict__restrict ff, |
| 698 | t_forcerec * gmx_restrict__restrict fr, |
| 699 | t_mdatoms * gmx_restrict__restrict mdatoms, |
| 700 | nb_kernel_data_t gmx_unused__attribute__ ((unused)) * gmx_restrict__restrict kernel_data, |
| 701 | t_nrnb * gmx_restrict__restrict nrnb) |
| 702 | { |
| 703 | int i_shift_offset,i_coord_offset,j_coord_offset; |
| 704 | int j_index_start,j_index_end; |
| 705 | int nri,inr,ggid,iidx,jidx,jnr,outeriter,inneriter; |
| 706 | real shX,shY,shZ,tx,ty,tz,fscal,rcutoff,rcutoff2; |
| 707 | int *iinr,*jindex,*jjnr,*shiftidx,*gid; |
| 708 | real *shiftvec,*fshift,*x,*f; |
| 709 | int vdwioffset0; |
| 710 | real ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0; |
| 711 | int vdwioffset1; |
| 712 | real ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1; |
| 713 | int vdwioffset2; |
| 714 | real ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2; |
| 715 | int vdwioffset3; |
| 716 | real ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3; |
| 717 | int vdwjidx0; |
| 718 | real jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0; |
| 719 | int vdwjidx1; |
| 720 | real jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1; |
| 721 | int vdwjidx2; |
| 722 | real jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2; |
| 723 | int vdwjidx3; |
| 724 | real jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3; |
| 725 | real dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00,cexp1_00,cexp2_00; |
| 726 | real dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11,cexp1_11,cexp2_11; |
| 727 | real dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12,cexp1_12,cexp2_12; |
| 728 | real dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13,cexp1_13,cexp2_13; |
| 729 | real dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21,cexp1_21,cexp2_21; |
| 730 | real dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22,cexp1_22,cexp2_22; |
| 731 | real dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23,cexp1_23,cexp2_23; |
| 732 | real dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31,cexp1_31,cexp2_31; |
| 733 | real dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32,cexp1_32,cexp2_32; |
| 734 | real dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33,cexp1_33,cexp2_33; |
| 735 | real velec,felec,velecsum,facel,crf,krf,krf2; |
| 736 | real *charge; |
| 737 | int nvdwtype; |
| 738 | real rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,br,vvdwexp,sh_vdw_invrcut6; |
| 739 | int *vdwtype; |
| 740 | real *vdwparam; |
| 741 | int ewitab; |
| 742 | real ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace; |
| 743 | real *ewtab; |
| 744 | |
| 745 | x = xx[0]; |
| 746 | f = ff[0]; |
| 747 | |
| 748 | nri = nlist->nri; |
| 749 | iinr = nlist->iinr; |
| 750 | jindex = nlist->jindex; |
| 751 | jjnr = nlist->jjnr; |
| 752 | shiftidx = nlist->shift; |
| 753 | gid = nlist->gid; |
| 754 | shiftvec = fr->shift_vec[0]; |
| 755 | fshift = fr->fshift[0]; |
| 756 | facel = fr->epsfac; |
| 757 | charge = mdatoms->chargeA; |
| 758 | nvdwtype = fr->ntype; |
| 759 | vdwparam = fr->nbfp; |
| 760 | vdwtype = mdatoms->typeA; |
| 761 | |
| 762 | sh_ewald = fr->ic->sh_ewald; |
| 763 | ewtab = fr->ic->tabq_coul_F; |
| 764 | ewtabscale = fr->ic->tabq_scale; |
| 765 | ewtabhalfspace = 0.5/ewtabscale; |
| 766 | |
| 767 | /* Setup water-specific parameters */ |
| 768 | inr = nlist->iinr[0]; |
| 769 | iq1 = facel*charge[inr+1]; |
| 770 | iq2 = facel*charge[inr+2]; |
| 771 | iq3 = facel*charge[inr+3]; |
| 772 | vdwioffset0 = 2*nvdwtype*vdwtype[inr+0]; |
| 773 | |
| 774 | jq1 = charge[inr+1]; |
| 775 | jq2 = charge[inr+2]; |
| 776 | jq3 = charge[inr+3]; |
| 777 | vdwjidx0 = 2*vdwtype[inr+0]; |
| 778 | c6_00 = vdwparam[vdwioffset0+vdwjidx0]; |
| 779 | c12_00 = vdwparam[vdwioffset0+vdwjidx0+1]; |
| 780 | qq11 = iq1*jq1; |
| 781 | qq12 = iq1*jq2; |
| 782 | qq13 = iq1*jq3; |
| 783 | qq21 = iq2*jq1; |
| 784 | qq22 = iq2*jq2; |
| 785 | qq23 = iq2*jq3; |
| 786 | qq31 = iq3*jq1; |
| 787 | qq32 = iq3*jq2; |
| 788 | qq33 = iq3*jq3; |
| 789 | |
| 790 | outeriter = 0; |
| 791 | inneriter = 0; |
| 792 | |
| 793 | /* Start outer loop over neighborlists */ |
| 794 | for(iidx=0; iidx<nri; iidx++) |
| 795 | { |
| 796 | /* Load shift vector for this list */ |
| 797 | i_shift_offset = DIM3*shiftidx[iidx]; |
| 798 | shX = shiftvec[i_shift_offset+XX0]; |
| 799 | shY = shiftvec[i_shift_offset+YY1]; |
| 800 | shZ = shiftvec[i_shift_offset+ZZ2]; |
| 801 | |
| 802 | /* Load limits for loop over neighbors */ |
| 803 | j_index_start = jindex[iidx]; |
| 804 | j_index_end = jindex[iidx+1]; |
| 805 | |
| 806 | /* Get outer coordinate index */ |
| 807 | inr = iinr[iidx]; |
| 808 | i_coord_offset = DIM3*inr; |
| 809 | |
| 810 | /* Load i particle coords and add shift vector */ |
| 811 | ix0 = shX + x[i_coord_offset+DIM3*0+XX0]; |
| 812 | iy0 = shY + x[i_coord_offset+DIM3*0+YY1]; |
| 813 | iz0 = shZ + x[i_coord_offset+DIM3*0+ZZ2]; |
| 814 | ix1 = shX + x[i_coord_offset+DIM3*1+XX0]; |
| 815 | iy1 = shY + x[i_coord_offset+DIM3*1+YY1]; |
| 816 | iz1 = shZ + x[i_coord_offset+DIM3*1+ZZ2]; |
| 817 | ix2 = shX + x[i_coord_offset+DIM3*2+XX0]; |
| 818 | iy2 = shY + x[i_coord_offset+DIM3*2+YY1]; |
| 819 | iz2 = shZ + x[i_coord_offset+DIM3*2+ZZ2]; |
| 820 | ix3 = shX + x[i_coord_offset+DIM3*3+XX0]; |
| 821 | iy3 = shY + x[i_coord_offset+DIM3*3+YY1]; |
| 822 | iz3 = shZ + x[i_coord_offset+DIM3*3+ZZ2]; |
| 823 | |
| 824 | fix0 = 0.0; |
| 825 | fiy0 = 0.0; |
| 826 | fiz0 = 0.0; |
| 827 | fix1 = 0.0; |
| 828 | fiy1 = 0.0; |
| 829 | fiz1 = 0.0; |
| 830 | fix2 = 0.0; |
| 831 | fiy2 = 0.0; |
| 832 | fiz2 = 0.0; |
| 833 | fix3 = 0.0; |
| 834 | fiy3 = 0.0; |
| 835 | fiz3 = 0.0; |
| 836 | |
| 837 | /* Start inner kernel loop */ |
| 838 | for(jidx=j_index_start; jidx<j_index_end; jidx++) |
| 839 | { |
| 840 | /* Get j neighbor index, and coordinate index */ |
| 841 | jnr = jjnr[jidx]; |
| 842 | j_coord_offset = DIM3*jnr; |
| 843 | |
| 844 | /* load j atom coordinates */ |
| 845 | jx0 = x[j_coord_offset+DIM3*0+XX0]; |
| 846 | jy0 = x[j_coord_offset+DIM3*0+YY1]; |
| 847 | jz0 = x[j_coord_offset+DIM3*0+ZZ2]; |
| 848 | jx1 = x[j_coord_offset+DIM3*1+XX0]; |
| 849 | jy1 = x[j_coord_offset+DIM3*1+YY1]; |
| 850 | jz1 = x[j_coord_offset+DIM3*1+ZZ2]; |
| 851 | jx2 = x[j_coord_offset+DIM3*2+XX0]; |
| 852 | jy2 = x[j_coord_offset+DIM3*2+YY1]; |
| 853 | jz2 = x[j_coord_offset+DIM3*2+ZZ2]; |
| 854 | jx3 = x[j_coord_offset+DIM3*3+XX0]; |
| 855 | jy3 = x[j_coord_offset+DIM3*3+YY1]; |
| 856 | jz3 = x[j_coord_offset+DIM3*3+ZZ2]; |
| 857 | |
| 858 | /* Calculate displacement vector */ |
| 859 | dx00 = ix0 - jx0; |
| 860 | dy00 = iy0 - jy0; |
| 861 | dz00 = iz0 - jz0; |
| 862 | dx11 = ix1 - jx1; |
| 863 | dy11 = iy1 - jy1; |
| 864 | dz11 = iz1 - jz1; |
| 865 | dx12 = ix1 - jx2; |
| 866 | dy12 = iy1 - jy2; |
| 867 | dz12 = iz1 - jz2; |
| 868 | dx13 = ix1 - jx3; |
| 869 | dy13 = iy1 - jy3; |
| 870 | dz13 = iz1 - jz3; |
| 871 | dx21 = ix2 - jx1; |
| 872 | dy21 = iy2 - jy1; |
| 873 | dz21 = iz2 - jz1; |
| 874 | dx22 = ix2 - jx2; |
| 875 | dy22 = iy2 - jy2; |
| 876 | dz22 = iz2 - jz2; |
| 877 | dx23 = ix2 - jx3; |
| 878 | dy23 = iy2 - jy3; |
| 879 | dz23 = iz2 - jz3; |
| 880 | dx31 = ix3 - jx1; |
| 881 | dy31 = iy3 - jy1; |
| 882 | dz31 = iz3 - jz1; |
| 883 | dx32 = ix3 - jx2; |
| 884 | dy32 = iy3 - jy2; |
| 885 | dz32 = iz3 - jz2; |
| 886 | dx33 = ix3 - jx3; |
| 887 | dy33 = iy3 - jy3; |
| 888 | dz33 = iz3 - jz3; |
| 889 | |
| 890 | /* Calculate squared distance and things based on it */ |
| 891 | rsq00 = dx00*dx00+dy00*dy00+dz00*dz00; |
| 892 | rsq11 = dx11*dx11+dy11*dy11+dz11*dz11; |
| 893 | rsq12 = dx12*dx12+dy12*dy12+dz12*dz12; |
| 894 | rsq13 = dx13*dx13+dy13*dy13+dz13*dz13; |
| 895 | rsq21 = dx21*dx21+dy21*dy21+dz21*dz21; |
| 896 | rsq22 = dx22*dx22+dy22*dy22+dz22*dz22; |
| 897 | rsq23 = dx23*dx23+dy23*dy23+dz23*dz23; |
| 898 | rsq31 = dx31*dx31+dy31*dy31+dz31*dz31; |
| 899 | rsq32 = dx32*dx32+dy32*dy32+dz32*dz32; |
| 900 | rsq33 = dx33*dx33+dy33*dy33+dz33*dz33; |
| 901 | |
| 902 | rinv11 = gmx_invsqrt(rsq11)gmx_software_invsqrt(rsq11); |
| 903 | rinv12 = gmx_invsqrt(rsq12)gmx_software_invsqrt(rsq12); |
| 904 | rinv13 = gmx_invsqrt(rsq13)gmx_software_invsqrt(rsq13); |
| 905 | rinv21 = gmx_invsqrt(rsq21)gmx_software_invsqrt(rsq21); |
| 906 | rinv22 = gmx_invsqrt(rsq22)gmx_software_invsqrt(rsq22); |
| 907 | rinv23 = gmx_invsqrt(rsq23)gmx_software_invsqrt(rsq23); |
| 908 | rinv31 = gmx_invsqrt(rsq31)gmx_software_invsqrt(rsq31); |
| 909 | rinv32 = gmx_invsqrt(rsq32)gmx_software_invsqrt(rsq32); |
| 910 | rinv33 = gmx_invsqrt(rsq33)gmx_software_invsqrt(rsq33); |
| 911 | |
| 912 | rinvsq00 = 1.0/rsq00; |
| 913 | rinvsq11 = rinv11*rinv11; |
| 914 | rinvsq12 = rinv12*rinv12; |
| 915 | rinvsq13 = rinv13*rinv13; |
| 916 | rinvsq21 = rinv21*rinv21; |
| 917 | rinvsq22 = rinv22*rinv22; |
| 918 | rinvsq23 = rinv23*rinv23; |
| 919 | rinvsq31 = rinv31*rinv31; |
| 920 | rinvsq32 = rinv32*rinv32; |
| 921 | rinvsq33 = rinv33*rinv33; |
| 922 | |
| 923 | /************************** |
| 924 | * CALCULATE INTERACTIONS * |
| 925 | **************************/ |
| 926 | |
| 927 | /* LENNARD-JONES DISPERSION/REPULSION */ |
| 928 | |
| 929 | rinvsix = rinvsq00*rinvsq00*rinvsq00; |
| 930 | fvdw = (c12_00*rinvsix-c6_00)*rinvsix*rinvsq00; |
| 931 | |
| 932 | fscal = fvdw; |
| 933 | |
| 934 | /* Calculate temporary vectorial force */ |
| 935 | tx = fscal*dx00; |
| 936 | ty = fscal*dy00; |
| 937 | tz = fscal*dz00; |
| 938 | |
| 939 | /* Update vectorial force */ |
| 940 | fix0 += tx; |
| 941 | fiy0 += ty; |
| 942 | fiz0 += tz; |
| 943 | f[j_coord_offset+DIM3*0+XX0] -= tx; |
| 944 | f[j_coord_offset+DIM3*0+YY1] -= ty; |
| 945 | f[j_coord_offset+DIM3*0+ZZ2] -= tz; |
| 946 | |
| 947 | /************************** |
| 948 | * CALCULATE INTERACTIONS * |
| 949 | **************************/ |
| 950 | |
| 951 | r11 = rsq11*rinv11; |
| 952 | |
| 953 | /* EWALD ELECTROSTATICS */ |
| 954 | |
| 955 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 956 | ewrt = r11*ewtabscale; |
| 957 | ewitab = ewrt; |
| 958 | eweps = ewrt-ewitab; |
| 959 | felec = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 960 | felec = qq11*rinv11*(rinvsq11-felec); |
| 961 | |
| 962 | fscal = felec; |
| 963 | |
| 964 | /* Calculate temporary vectorial force */ |
| 965 | tx = fscal*dx11; |
| 966 | ty = fscal*dy11; |
| 967 | tz = fscal*dz11; |
| 968 | |
| 969 | /* Update vectorial force */ |
| 970 | fix1 += tx; |
| 971 | fiy1 += ty; |
| 972 | fiz1 += tz; |
| 973 | f[j_coord_offset+DIM3*1+XX0] -= tx; |
| 974 | f[j_coord_offset+DIM3*1+YY1] -= ty; |
| 975 | f[j_coord_offset+DIM3*1+ZZ2] -= tz; |
| 976 | |
| 977 | /************************** |
| 978 | * CALCULATE INTERACTIONS * |
| 979 | **************************/ |
| 980 | |
| 981 | r12 = rsq12*rinv12; |
| 982 | |
| 983 | /* EWALD ELECTROSTATICS */ |
| 984 | |
| 985 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 986 | ewrt = r12*ewtabscale; |
| 987 | ewitab = ewrt; |
| 988 | eweps = ewrt-ewitab; |
| 989 | felec = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 990 | felec = qq12*rinv12*(rinvsq12-felec); |
| 991 | |
| 992 | fscal = felec; |
| 993 | |
| 994 | /* Calculate temporary vectorial force */ |
| 995 | tx = fscal*dx12; |
| 996 | ty = fscal*dy12; |
| 997 | tz = fscal*dz12; |
| 998 | |
| 999 | /* Update vectorial force */ |
| 1000 | fix1 += tx; |
| 1001 | fiy1 += ty; |
| 1002 | fiz1 += tz; |
| 1003 | f[j_coord_offset+DIM3*2+XX0] -= tx; |
| 1004 | f[j_coord_offset+DIM3*2+YY1] -= ty; |
| 1005 | f[j_coord_offset+DIM3*2+ZZ2] -= tz; |
| 1006 | |
| 1007 | /************************** |
| 1008 | * CALCULATE INTERACTIONS * |
| 1009 | **************************/ |
| 1010 | |
| 1011 | r13 = rsq13*rinv13; |
| 1012 | |
| 1013 | /* EWALD ELECTROSTATICS */ |
| 1014 | |
| 1015 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 1016 | ewrt = r13*ewtabscale; |
| 1017 | ewitab = ewrt; |
| 1018 | eweps = ewrt-ewitab; |
| 1019 | felec = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 1020 | felec = qq13*rinv13*(rinvsq13-felec); |
| 1021 | |
| 1022 | fscal = felec; |
| 1023 | |
| 1024 | /* Calculate temporary vectorial force */ |
| 1025 | tx = fscal*dx13; |
| 1026 | ty = fscal*dy13; |
| 1027 | tz = fscal*dz13; |
| 1028 | |
| 1029 | /* Update vectorial force */ |
| 1030 | fix1 += tx; |
| 1031 | fiy1 += ty; |
| 1032 | fiz1 += tz; |
| 1033 | f[j_coord_offset+DIM3*3+XX0] -= tx; |
| 1034 | f[j_coord_offset+DIM3*3+YY1] -= ty; |
| 1035 | f[j_coord_offset+DIM3*3+ZZ2] -= tz; |
| 1036 | |
| 1037 | /************************** |
| 1038 | * CALCULATE INTERACTIONS * |
| 1039 | **************************/ |
| 1040 | |
| 1041 | r21 = rsq21*rinv21; |
| 1042 | |
| 1043 | /* EWALD ELECTROSTATICS */ |
| 1044 | |
| 1045 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 1046 | ewrt = r21*ewtabscale; |
| 1047 | ewitab = ewrt; |
| 1048 | eweps = ewrt-ewitab; |
| 1049 | felec = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 1050 | felec = qq21*rinv21*(rinvsq21-felec); |
| 1051 | |
| 1052 | fscal = felec; |
| 1053 | |
| 1054 | /* Calculate temporary vectorial force */ |
| 1055 | tx = fscal*dx21; |
| 1056 | ty = fscal*dy21; |
| 1057 | tz = fscal*dz21; |
| 1058 | |
| 1059 | /* Update vectorial force */ |
| 1060 | fix2 += tx; |
| 1061 | fiy2 += ty; |
| 1062 | fiz2 += tz; |
| 1063 | f[j_coord_offset+DIM3*1+XX0] -= tx; |
| 1064 | f[j_coord_offset+DIM3*1+YY1] -= ty; |
| 1065 | f[j_coord_offset+DIM3*1+ZZ2] -= tz; |
| 1066 | |
| 1067 | /************************** |
| 1068 | * CALCULATE INTERACTIONS * |
| 1069 | **************************/ |
| 1070 | |
| 1071 | r22 = rsq22*rinv22; |
| 1072 | |
| 1073 | /* EWALD ELECTROSTATICS */ |
| 1074 | |
| 1075 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 1076 | ewrt = r22*ewtabscale; |
| 1077 | ewitab = ewrt; |
| 1078 | eweps = ewrt-ewitab; |
| 1079 | felec = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 1080 | felec = qq22*rinv22*(rinvsq22-felec); |
| 1081 | |
| 1082 | fscal = felec; |
| 1083 | |
| 1084 | /* Calculate temporary vectorial force */ |
| 1085 | tx = fscal*dx22; |
| 1086 | ty = fscal*dy22; |
| 1087 | tz = fscal*dz22; |
| 1088 | |
| 1089 | /* Update vectorial force */ |
| 1090 | fix2 += tx; |
| 1091 | fiy2 += ty; |
| 1092 | fiz2 += tz; |
| 1093 | f[j_coord_offset+DIM3*2+XX0] -= tx; |
| 1094 | f[j_coord_offset+DIM3*2+YY1] -= ty; |
| 1095 | f[j_coord_offset+DIM3*2+ZZ2] -= tz; |
| 1096 | |
| 1097 | /************************** |
| 1098 | * CALCULATE INTERACTIONS * |
| 1099 | **************************/ |
| 1100 | |
| 1101 | r23 = rsq23*rinv23; |
| 1102 | |
| 1103 | /* EWALD ELECTROSTATICS */ |
| 1104 | |
| 1105 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 1106 | ewrt = r23*ewtabscale; |
| 1107 | ewitab = ewrt; |
| 1108 | eweps = ewrt-ewitab; |
| 1109 | felec = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 1110 | felec = qq23*rinv23*(rinvsq23-felec); |
| 1111 | |
| 1112 | fscal = felec; |
| 1113 | |
| 1114 | /* Calculate temporary vectorial force */ |
| 1115 | tx = fscal*dx23; |
| 1116 | ty = fscal*dy23; |
| 1117 | tz = fscal*dz23; |
| 1118 | |
| 1119 | /* Update vectorial force */ |
| 1120 | fix2 += tx; |
| 1121 | fiy2 += ty; |
| 1122 | fiz2 += tz; |
| 1123 | f[j_coord_offset+DIM3*3+XX0] -= tx; |
| 1124 | f[j_coord_offset+DIM3*3+YY1] -= ty; |
| 1125 | f[j_coord_offset+DIM3*3+ZZ2] -= tz; |
| 1126 | |
| 1127 | /************************** |
| 1128 | * CALCULATE INTERACTIONS * |
| 1129 | **************************/ |
| 1130 | |
| 1131 | r31 = rsq31*rinv31; |
| 1132 | |
| 1133 | /* EWALD ELECTROSTATICS */ |
| 1134 | |
| 1135 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 1136 | ewrt = r31*ewtabscale; |
| 1137 | ewitab = ewrt; |
| 1138 | eweps = ewrt-ewitab; |
| 1139 | felec = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 1140 | felec = qq31*rinv31*(rinvsq31-felec); |
| 1141 | |
| 1142 | fscal = felec; |
| 1143 | |
| 1144 | /* Calculate temporary vectorial force */ |
| 1145 | tx = fscal*dx31; |
| 1146 | ty = fscal*dy31; |
| 1147 | tz = fscal*dz31; |
| 1148 | |
| 1149 | /* Update vectorial force */ |
| 1150 | fix3 += tx; |
| 1151 | fiy3 += ty; |
| 1152 | fiz3 += tz; |
| 1153 | f[j_coord_offset+DIM3*1+XX0] -= tx; |
| 1154 | f[j_coord_offset+DIM3*1+YY1] -= ty; |
| 1155 | f[j_coord_offset+DIM3*1+ZZ2] -= tz; |
| 1156 | |
| 1157 | /************************** |
| 1158 | * CALCULATE INTERACTIONS * |
| 1159 | **************************/ |
| 1160 | |
| 1161 | r32 = rsq32*rinv32; |
| 1162 | |
| 1163 | /* EWALD ELECTROSTATICS */ |
| 1164 | |
| 1165 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 1166 | ewrt = r32*ewtabscale; |
| 1167 | ewitab = ewrt; |
| 1168 | eweps = ewrt-ewitab; |
| 1169 | felec = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 1170 | felec = qq32*rinv32*(rinvsq32-felec); |
| 1171 | |
| 1172 | fscal = felec; |
| 1173 | |
| 1174 | /* Calculate temporary vectorial force */ |
| 1175 | tx = fscal*dx32; |
| 1176 | ty = fscal*dy32; |
| 1177 | tz = fscal*dz32; |
| 1178 | |
| 1179 | /* Update vectorial force */ |
| 1180 | fix3 += tx; |
| 1181 | fiy3 += ty; |
| 1182 | fiz3 += tz; |
| 1183 | f[j_coord_offset+DIM3*2+XX0] -= tx; |
| 1184 | f[j_coord_offset+DIM3*2+YY1] -= ty; |
| 1185 | f[j_coord_offset+DIM3*2+ZZ2] -= tz; |
| 1186 | |
| 1187 | /************************** |
| 1188 | * CALCULATE INTERACTIONS * |
| 1189 | **************************/ |
| 1190 | |
| 1191 | r33 = rsq33*rinv33; |
| 1192 | |
| 1193 | /* EWALD ELECTROSTATICS */ |
| 1194 | |
| 1195 | /* Calculate Ewald table index by multiplying r with scale and truncate to integer */ |
| 1196 | ewrt = r33*ewtabscale; |
| 1197 | ewitab = ewrt; |
| 1198 | eweps = ewrt-ewitab; |
| 1199 | felec = (1.0-eweps)*ewtab[ewitab]+eweps*ewtab[ewitab+1]; |
| 1200 | felec = qq33*rinv33*(rinvsq33-felec); |
| 1201 | |
| 1202 | fscal = felec; |
| 1203 | |
| 1204 | /* Calculate temporary vectorial force */ |
| 1205 | tx = fscal*dx33; |
| 1206 | ty = fscal*dy33; |
| 1207 | tz = fscal*dz33; |
| 1208 | |
| 1209 | /* Update vectorial force */ |
| 1210 | fix3 += tx; |
| 1211 | fiy3 += ty; |
| 1212 | fiz3 += tz; |
| 1213 | f[j_coord_offset+DIM3*3+XX0] -= tx; |
| 1214 | f[j_coord_offset+DIM3*3+YY1] -= ty; |
| 1215 | f[j_coord_offset+DIM3*3+ZZ2] -= tz; |
| 1216 | |
| 1217 | /* Inner loop uses 324 flops */ |
| 1218 | } |
| 1219 | /* End of innermost loop */ |
| 1220 | |
| 1221 | tx = ty = tz = 0; |
| 1222 | f[i_coord_offset+DIM3*0+XX0] += fix0; |
| 1223 | f[i_coord_offset+DIM3*0+YY1] += fiy0; |
| 1224 | f[i_coord_offset+DIM3*0+ZZ2] += fiz0; |
| 1225 | tx += fix0; |
| 1226 | ty += fiy0; |
| 1227 | tz += fiz0; |
| 1228 | f[i_coord_offset+DIM3*1+XX0] += fix1; |
| 1229 | f[i_coord_offset+DIM3*1+YY1] += fiy1; |
| 1230 | f[i_coord_offset+DIM3*1+ZZ2] += fiz1; |
| 1231 | tx += fix1; |
| 1232 | ty += fiy1; |
| 1233 | tz += fiz1; |
| 1234 | f[i_coord_offset+DIM3*2+XX0] += fix2; |
| 1235 | f[i_coord_offset+DIM3*2+YY1] += fiy2; |
| 1236 | f[i_coord_offset+DIM3*2+ZZ2] += fiz2; |
| 1237 | tx += fix2; |
| 1238 | ty += fiy2; |
| 1239 | tz += fiz2; |
| 1240 | f[i_coord_offset+DIM3*3+XX0] += fix3; |
| 1241 | f[i_coord_offset+DIM3*3+YY1] += fiy3; |
| 1242 | f[i_coord_offset+DIM3*3+ZZ2] += fiz3; |
| 1243 | tx += fix3; |
| 1244 | ty += fiy3; |
| 1245 | tz += fiz3; |
| 1246 | fshift[i_shift_offset+XX0] += tx; |
| 1247 | fshift[i_shift_offset+YY1] += ty; |
| 1248 | fshift[i_shift_offset+ZZ2] += tz; |
| 1249 | |
| 1250 | /* Increment number of inner iterations */ |
| 1251 | inneriter += j_index_end - j_index_start; |
| 1252 | |
| 1253 | /* Outer loop uses 39 flops */ |
| 1254 | } |
| 1255 | |
| 1256 | /* Increment number of outer iterations */ |
| 1257 | outeriter += nri; |
| 1258 | |
| 1259 | /* Update outer/inner flops */ |
| 1260 | |
| 1261 | inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*39 + inneriter*324)(nrnb)->n[eNR_NBKERNEL_ELEC_VDW_W4W4_F] += outeriter*39 + inneriter *324; |
| 1262 | } |