2 * This file is part of the GROMACS molecular simulation package.
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.
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.
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.
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.
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.
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.
36 * Note: this file was generated by the GROMACS avx_256_double kernel generator.
44 #include "../nb_kernel.h"
45 #include "gromacs/legacyheaders/types/simple.h"
46 #include "gromacs/math/vec.h"
47 #include "gromacs/legacyheaders/nrnb.h"
49 #include "gromacs/simd/math_x86_avx_256_double.h"
50 #include "kernelutil_x86_avx_256_double.h"
53 * Gromacs nonbonded kernel: nb_kernel_ElecEwSh_VdwLJSh_GeomW3P1_VF_avx_256_double
54 * Electrostatics interaction: Ewald
55 * VdW interaction: LennardJones
56 * Geometry: Water3-Particle
57 * Calculate force/pot: PotentialAndForce
60 nb_kernel_ElecEwSh_VdwLJSh_GeomW3P1_VF_avx_256_double
61 (t_nblist * gmx_restrict nlist,
62 rvec * gmx_restrict xx,
63 rvec * gmx_restrict ff,
64 t_forcerec * gmx_restrict fr,
65 t_mdatoms * gmx_restrict mdatoms,
66 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
67 t_nrnb * gmx_restrict nrnb)
69 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
70 * just 0 for non-waters.
71 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
72 * jnr indices corresponding to data put in the four positions in the SIMD register.
74 int i_shift_offset,i_coord_offset,outeriter,inneriter;
75 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
76 int jnrA,jnrB,jnrC,jnrD;
77 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
78 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
79 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
80 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
82 real *shiftvec,*fshift,*x,*f;
83 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
85 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
86 real * vdwioffsetptr0;
87 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
88 real * vdwioffsetptr1;
89 __m256d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
90 real * vdwioffsetptr2;
91 __m256d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
92 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
93 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
94 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
95 __m256d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
96 __m256d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
97 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
100 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
103 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
104 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
106 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
107 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
109 __m256d dummy_mask,cutoff_mask;
110 __m128 tmpmask0,tmpmask1;
111 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
112 __m256d one = _mm256_set1_pd(1.0);
113 __m256d two = _mm256_set1_pd(2.0);
119 jindex = nlist->jindex;
121 shiftidx = nlist->shift;
123 shiftvec = fr->shift_vec[0];
124 fshift = fr->fshift[0];
125 facel = _mm256_set1_pd(fr->epsfac);
126 charge = mdatoms->chargeA;
127 nvdwtype = fr->ntype;
129 vdwtype = mdatoms->typeA;
131 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
132 beta = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
133 beta2 = _mm256_mul_pd(beta,beta);
134 beta3 = _mm256_mul_pd(beta,beta2);
136 ewtab = fr->ic->tabq_coul_FDV0;
137 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
138 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
140 /* Setup water-specific parameters */
141 inr = nlist->iinr[0];
142 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
143 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
144 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
145 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
147 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
148 rcutoff_scalar = fr->rcoulomb;
149 rcutoff = _mm256_set1_pd(rcutoff_scalar);
150 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
152 sh_vdw_invrcut6 = _mm256_set1_pd(fr->ic->sh_invrc6);
153 rvdw = _mm256_set1_pd(fr->rvdw);
155 /* Avoid stupid compiler warnings */
156 jnrA = jnrB = jnrC = jnrD = 0;
165 for(iidx=0;iidx<4*DIM;iidx++)
170 /* Start outer loop over neighborlists */
171 for(iidx=0; iidx<nri; iidx++)
173 /* Load shift vector for this list */
174 i_shift_offset = DIM*shiftidx[iidx];
176 /* Load limits for loop over neighbors */
177 j_index_start = jindex[iidx];
178 j_index_end = jindex[iidx+1];
180 /* Get outer coordinate index */
182 i_coord_offset = DIM*inr;
184 /* Load i particle coords and add shift vector */
185 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
186 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
188 fix0 = _mm256_setzero_pd();
189 fiy0 = _mm256_setzero_pd();
190 fiz0 = _mm256_setzero_pd();
191 fix1 = _mm256_setzero_pd();
192 fiy1 = _mm256_setzero_pd();
193 fiz1 = _mm256_setzero_pd();
194 fix2 = _mm256_setzero_pd();
195 fiy2 = _mm256_setzero_pd();
196 fiz2 = _mm256_setzero_pd();
198 /* Reset potential sums */
199 velecsum = _mm256_setzero_pd();
200 vvdwsum = _mm256_setzero_pd();
202 /* Start inner kernel loop */
203 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
206 /* Get j neighbor index, and coordinate index */
211 j_coord_offsetA = DIM*jnrA;
212 j_coord_offsetB = DIM*jnrB;
213 j_coord_offsetC = DIM*jnrC;
214 j_coord_offsetD = DIM*jnrD;
216 /* load j atom coordinates */
217 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
218 x+j_coord_offsetC,x+j_coord_offsetD,
221 /* Calculate displacement vector */
222 dx00 = _mm256_sub_pd(ix0,jx0);
223 dy00 = _mm256_sub_pd(iy0,jy0);
224 dz00 = _mm256_sub_pd(iz0,jz0);
225 dx10 = _mm256_sub_pd(ix1,jx0);
226 dy10 = _mm256_sub_pd(iy1,jy0);
227 dz10 = _mm256_sub_pd(iz1,jz0);
228 dx20 = _mm256_sub_pd(ix2,jx0);
229 dy20 = _mm256_sub_pd(iy2,jy0);
230 dz20 = _mm256_sub_pd(iz2,jz0);
232 /* Calculate squared distance and things based on it */
233 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
234 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
235 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
237 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
238 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
239 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
241 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
242 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
243 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
245 /* Load parameters for j particles */
246 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
247 charge+jnrC+0,charge+jnrD+0);
248 vdwjidx0A = 2*vdwtype[jnrA+0];
249 vdwjidx0B = 2*vdwtype[jnrB+0];
250 vdwjidx0C = 2*vdwtype[jnrC+0];
251 vdwjidx0D = 2*vdwtype[jnrD+0];
253 fjx0 = _mm256_setzero_pd();
254 fjy0 = _mm256_setzero_pd();
255 fjz0 = _mm256_setzero_pd();
257 /**************************
258 * CALCULATE INTERACTIONS *
259 **************************/
261 if (gmx_mm256_any_lt(rsq00,rcutoff2))
264 r00 = _mm256_mul_pd(rsq00,rinv00);
266 /* Compute parameters for interactions between i and j atoms */
267 qq00 = _mm256_mul_pd(iq0,jq0);
268 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
269 vdwioffsetptr0+vdwjidx0B,
270 vdwioffsetptr0+vdwjidx0C,
271 vdwioffsetptr0+vdwjidx0D,
274 /* EWALD ELECTROSTATICS */
276 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
277 ewrt = _mm256_mul_pd(r00,ewtabscale);
278 ewitab = _mm256_cvttpd_epi32(ewrt);
279 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
280 ewitab = _mm_slli_epi32(ewitab,2);
281 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
282 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
283 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
284 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
285 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
286 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
287 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
288 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
289 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
291 /* LENNARD-JONES DISPERSION/REPULSION */
293 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
294 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
295 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
296 vvdw = _mm256_sub_pd(_mm256_mul_pd( _mm256_sub_pd(vvdw12 , _mm256_mul_pd(c12_00,_mm256_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
297 _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
298 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
300 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
302 /* Update potential sum for this i atom from the interaction with this j atom. */
303 velec = _mm256_and_pd(velec,cutoff_mask);
304 velecsum = _mm256_add_pd(velecsum,velec);
305 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
306 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
308 fscal = _mm256_add_pd(felec,fvdw);
310 fscal = _mm256_and_pd(fscal,cutoff_mask);
312 /* Calculate temporary vectorial force */
313 tx = _mm256_mul_pd(fscal,dx00);
314 ty = _mm256_mul_pd(fscal,dy00);
315 tz = _mm256_mul_pd(fscal,dz00);
317 /* Update vectorial force */
318 fix0 = _mm256_add_pd(fix0,tx);
319 fiy0 = _mm256_add_pd(fiy0,ty);
320 fiz0 = _mm256_add_pd(fiz0,tz);
322 fjx0 = _mm256_add_pd(fjx0,tx);
323 fjy0 = _mm256_add_pd(fjy0,ty);
324 fjz0 = _mm256_add_pd(fjz0,tz);
328 /**************************
329 * CALCULATE INTERACTIONS *
330 **************************/
332 if (gmx_mm256_any_lt(rsq10,rcutoff2))
335 r10 = _mm256_mul_pd(rsq10,rinv10);
337 /* Compute parameters for interactions between i and j atoms */
338 qq10 = _mm256_mul_pd(iq1,jq0);
340 /* EWALD ELECTROSTATICS */
342 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
343 ewrt = _mm256_mul_pd(r10,ewtabscale);
344 ewitab = _mm256_cvttpd_epi32(ewrt);
345 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
346 ewitab = _mm_slli_epi32(ewitab,2);
347 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
348 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
349 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
350 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
351 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
352 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
353 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
354 velec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_sub_pd(rinv10,sh_ewald),velec));
355 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
357 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
359 /* Update potential sum for this i atom from the interaction with this j atom. */
360 velec = _mm256_and_pd(velec,cutoff_mask);
361 velecsum = _mm256_add_pd(velecsum,velec);
365 fscal = _mm256_and_pd(fscal,cutoff_mask);
367 /* Calculate temporary vectorial force */
368 tx = _mm256_mul_pd(fscal,dx10);
369 ty = _mm256_mul_pd(fscal,dy10);
370 tz = _mm256_mul_pd(fscal,dz10);
372 /* Update vectorial force */
373 fix1 = _mm256_add_pd(fix1,tx);
374 fiy1 = _mm256_add_pd(fiy1,ty);
375 fiz1 = _mm256_add_pd(fiz1,tz);
377 fjx0 = _mm256_add_pd(fjx0,tx);
378 fjy0 = _mm256_add_pd(fjy0,ty);
379 fjz0 = _mm256_add_pd(fjz0,tz);
383 /**************************
384 * CALCULATE INTERACTIONS *
385 **************************/
387 if (gmx_mm256_any_lt(rsq20,rcutoff2))
390 r20 = _mm256_mul_pd(rsq20,rinv20);
392 /* Compute parameters for interactions between i and j atoms */
393 qq20 = _mm256_mul_pd(iq2,jq0);
395 /* EWALD ELECTROSTATICS */
397 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
398 ewrt = _mm256_mul_pd(r20,ewtabscale);
399 ewitab = _mm256_cvttpd_epi32(ewrt);
400 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
401 ewitab = _mm_slli_epi32(ewitab,2);
402 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
403 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
404 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
405 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
406 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
407 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
408 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
409 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_sub_pd(rinv20,sh_ewald),velec));
410 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
412 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
414 /* Update potential sum for this i atom from the interaction with this j atom. */
415 velec = _mm256_and_pd(velec,cutoff_mask);
416 velecsum = _mm256_add_pd(velecsum,velec);
420 fscal = _mm256_and_pd(fscal,cutoff_mask);
422 /* Calculate temporary vectorial force */
423 tx = _mm256_mul_pd(fscal,dx20);
424 ty = _mm256_mul_pd(fscal,dy20);
425 tz = _mm256_mul_pd(fscal,dz20);
427 /* Update vectorial force */
428 fix2 = _mm256_add_pd(fix2,tx);
429 fiy2 = _mm256_add_pd(fiy2,ty);
430 fiz2 = _mm256_add_pd(fiz2,tz);
432 fjx0 = _mm256_add_pd(fjx0,tx);
433 fjy0 = _mm256_add_pd(fjy0,ty);
434 fjz0 = _mm256_add_pd(fjz0,tz);
438 fjptrA = f+j_coord_offsetA;
439 fjptrB = f+j_coord_offsetB;
440 fjptrC = f+j_coord_offsetC;
441 fjptrD = f+j_coord_offsetD;
443 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
445 /* Inner loop uses 159 flops */
451 /* Get j neighbor index, and coordinate index */
452 jnrlistA = jjnr[jidx];
453 jnrlistB = jjnr[jidx+1];
454 jnrlistC = jjnr[jidx+2];
455 jnrlistD = jjnr[jidx+3];
456 /* Sign of each element will be negative for non-real atoms.
457 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
458 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
460 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
462 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
463 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
464 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
466 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
467 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
468 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
469 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
470 j_coord_offsetA = DIM*jnrA;
471 j_coord_offsetB = DIM*jnrB;
472 j_coord_offsetC = DIM*jnrC;
473 j_coord_offsetD = DIM*jnrD;
475 /* load j atom coordinates */
476 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
477 x+j_coord_offsetC,x+j_coord_offsetD,
480 /* Calculate displacement vector */
481 dx00 = _mm256_sub_pd(ix0,jx0);
482 dy00 = _mm256_sub_pd(iy0,jy0);
483 dz00 = _mm256_sub_pd(iz0,jz0);
484 dx10 = _mm256_sub_pd(ix1,jx0);
485 dy10 = _mm256_sub_pd(iy1,jy0);
486 dz10 = _mm256_sub_pd(iz1,jz0);
487 dx20 = _mm256_sub_pd(ix2,jx0);
488 dy20 = _mm256_sub_pd(iy2,jy0);
489 dz20 = _mm256_sub_pd(iz2,jz0);
491 /* Calculate squared distance and things based on it */
492 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
493 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
494 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
496 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
497 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
498 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
500 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
501 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
502 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
504 /* Load parameters for j particles */
505 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
506 charge+jnrC+0,charge+jnrD+0);
507 vdwjidx0A = 2*vdwtype[jnrA+0];
508 vdwjidx0B = 2*vdwtype[jnrB+0];
509 vdwjidx0C = 2*vdwtype[jnrC+0];
510 vdwjidx0D = 2*vdwtype[jnrD+0];
512 fjx0 = _mm256_setzero_pd();
513 fjy0 = _mm256_setzero_pd();
514 fjz0 = _mm256_setzero_pd();
516 /**************************
517 * CALCULATE INTERACTIONS *
518 **************************/
520 if (gmx_mm256_any_lt(rsq00,rcutoff2))
523 r00 = _mm256_mul_pd(rsq00,rinv00);
524 r00 = _mm256_andnot_pd(dummy_mask,r00);
526 /* Compute parameters for interactions between i and j atoms */
527 qq00 = _mm256_mul_pd(iq0,jq0);
528 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
529 vdwioffsetptr0+vdwjidx0B,
530 vdwioffsetptr0+vdwjidx0C,
531 vdwioffsetptr0+vdwjidx0D,
534 /* EWALD ELECTROSTATICS */
536 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
537 ewrt = _mm256_mul_pd(r00,ewtabscale);
538 ewitab = _mm256_cvttpd_epi32(ewrt);
539 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
540 ewitab = _mm_slli_epi32(ewitab,2);
541 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
542 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
543 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
544 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
545 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
546 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
547 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
548 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
549 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
551 /* LENNARD-JONES DISPERSION/REPULSION */
553 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
554 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
555 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
556 vvdw = _mm256_sub_pd(_mm256_mul_pd( _mm256_sub_pd(vvdw12 , _mm256_mul_pd(c12_00,_mm256_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
557 _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
558 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
560 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
562 /* Update potential sum for this i atom from the interaction with this j atom. */
563 velec = _mm256_and_pd(velec,cutoff_mask);
564 velec = _mm256_andnot_pd(dummy_mask,velec);
565 velecsum = _mm256_add_pd(velecsum,velec);
566 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
567 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
568 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
570 fscal = _mm256_add_pd(felec,fvdw);
572 fscal = _mm256_and_pd(fscal,cutoff_mask);
574 fscal = _mm256_andnot_pd(dummy_mask,fscal);
576 /* Calculate temporary vectorial force */
577 tx = _mm256_mul_pd(fscal,dx00);
578 ty = _mm256_mul_pd(fscal,dy00);
579 tz = _mm256_mul_pd(fscal,dz00);
581 /* Update vectorial force */
582 fix0 = _mm256_add_pd(fix0,tx);
583 fiy0 = _mm256_add_pd(fiy0,ty);
584 fiz0 = _mm256_add_pd(fiz0,tz);
586 fjx0 = _mm256_add_pd(fjx0,tx);
587 fjy0 = _mm256_add_pd(fjy0,ty);
588 fjz0 = _mm256_add_pd(fjz0,tz);
592 /**************************
593 * CALCULATE INTERACTIONS *
594 **************************/
596 if (gmx_mm256_any_lt(rsq10,rcutoff2))
599 r10 = _mm256_mul_pd(rsq10,rinv10);
600 r10 = _mm256_andnot_pd(dummy_mask,r10);
602 /* Compute parameters for interactions between i and j atoms */
603 qq10 = _mm256_mul_pd(iq1,jq0);
605 /* EWALD ELECTROSTATICS */
607 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
608 ewrt = _mm256_mul_pd(r10,ewtabscale);
609 ewitab = _mm256_cvttpd_epi32(ewrt);
610 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
611 ewitab = _mm_slli_epi32(ewitab,2);
612 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
613 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
614 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
615 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
616 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
617 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
618 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
619 velec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_sub_pd(rinv10,sh_ewald),velec));
620 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
622 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
624 /* Update potential sum for this i atom from the interaction with this j atom. */
625 velec = _mm256_and_pd(velec,cutoff_mask);
626 velec = _mm256_andnot_pd(dummy_mask,velec);
627 velecsum = _mm256_add_pd(velecsum,velec);
631 fscal = _mm256_and_pd(fscal,cutoff_mask);
633 fscal = _mm256_andnot_pd(dummy_mask,fscal);
635 /* Calculate temporary vectorial force */
636 tx = _mm256_mul_pd(fscal,dx10);
637 ty = _mm256_mul_pd(fscal,dy10);
638 tz = _mm256_mul_pd(fscal,dz10);
640 /* Update vectorial force */
641 fix1 = _mm256_add_pd(fix1,tx);
642 fiy1 = _mm256_add_pd(fiy1,ty);
643 fiz1 = _mm256_add_pd(fiz1,tz);
645 fjx0 = _mm256_add_pd(fjx0,tx);
646 fjy0 = _mm256_add_pd(fjy0,ty);
647 fjz0 = _mm256_add_pd(fjz0,tz);
651 /**************************
652 * CALCULATE INTERACTIONS *
653 **************************/
655 if (gmx_mm256_any_lt(rsq20,rcutoff2))
658 r20 = _mm256_mul_pd(rsq20,rinv20);
659 r20 = _mm256_andnot_pd(dummy_mask,r20);
661 /* Compute parameters for interactions between i and j atoms */
662 qq20 = _mm256_mul_pd(iq2,jq0);
664 /* EWALD ELECTROSTATICS */
666 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
667 ewrt = _mm256_mul_pd(r20,ewtabscale);
668 ewitab = _mm256_cvttpd_epi32(ewrt);
669 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
670 ewitab = _mm_slli_epi32(ewitab,2);
671 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
672 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
673 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
674 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
675 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
676 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
677 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
678 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_sub_pd(rinv20,sh_ewald),velec));
679 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
681 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
683 /* Update potential sum for this i atom from the interaction with this j atom. */
684 velec = _mm256_and_pd(velec,cutoff_mask);
685 velec = _mm256_andnot_pd(dummy_mask,velec);
686 velecsum = _mm256_add_pd(velecsum,velec);
690 fscal = _mm256_and_pd(fscal,cutoff_mask);
692 fscal = _mm256_andnot_pd(dummy_mask,fscal);
694 /* Calculate temporary vectorial force */
695 tx = _mm256_mul_pd(fscal,dx20);
696 ty = _mm256_mul_pd(fscal,dy20);
697 tz = _mm256_mul_pd(fscal,dz20);
699 /* Update vectorial force */
700 fix2 = _mm256_add_pd(fix2,tx);
701 fiy2 = _mm256_add_pd(fiy2,ty);
702 fiz2 = _mm256_add_pd(fiz2,tz);
704 fjx0 = _mm256_add_pd(fjx0,tx);
705 fjy0 = _mm256_add_pd(fjy0,ty);
706 fjz0 = _mm256_add_pd(fjz0,tz);
710 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
711 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
712 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
713 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
715 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
717 /* Inner loop uses 162 flops */
720 /* End of innermost loop */
722 gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
723 f+i_coord_offset,fshift+i_shift_offset);
726 /* Update potential energies */
727 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
728 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
730 /* Increment number of inner iterations */
731 inneriter += j_index_end - j_index_start;
733 /* Outer loop uses 20 flops */
736 /* Increment number of outer iterations */
739 /* Update outer/inner flops */
741 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*162);
744 * Gromacs nonbonded kernel: nb_kernel_ElecEwSh_VdwLJSh_GeomW3P1_F_avx_256_double
745 * Electrostatics interaction: Ewald
746 * VdW interaction: LennardJones
747 * Geometry: Water3-Particle
748 * Calculate force/pot: Force
751 nb_kernel_ElecEwSh_VdwLJSh_GeomW3P1_F_avx_256_double
752 (t_nblist * gmx_restrict nlist,
753 rvec * gmx_restrict xx,
754 rvec * gmx_restrict ff,
755 t_forcerec * gmx_restrict fr,
756 t_mdatoms * gmx_restrict mdatoms,
757 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
758 t_nrnb * gmx_restrict nrnb)
760 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
761 * just 0 for non-waters.
762 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
763 * jnr indices corresponding to data put in the four positions in the SIMD register.
765 int i_shift_offset,i_coord_offset,outeriter,inneriter;
766 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
767 int jnrA,jnrB,jnrC,jnrD;
768 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
769 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
770 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
771 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
773 real *shiftvec,*fshift,*x,*f;
774 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
776 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
777 real * vdwioffsetptr0;
778 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
779 real * vdwioffsetptr1;
780 __m256d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
781 real * vdwioffsetptr2;
782 __m256d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
783 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
784 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
785 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
786 __m256d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
787 __m256d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
788 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
791 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
794 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
795 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
797 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
798 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
800 __m256d dummy_mask,cutoff_mask;
801 __m128 tmpmask0,tmpmask1;
802 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
803 __m256d one = _mm256_set1_pd(1.0);
804 __m256d two = _mm256_set1_pd(2.0);
810 jindex = nlist->jindex;
812 shiftidx = nlist->shift;
814 shiftvec = fr->shift_vec[0];
815 fshift = fr->fshift[0];
816 facel = _mm256_set1_pd(fr->epsfac);
817 charge = mdatoms->chargeA;
818 nvdwtype = fr->ntype;
820 vdwtype = mdatoms->typeA;
822 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
823 beta = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
824 beta2 = _mm256_mul_pd(beta,beta);
825 beta3 = _mm256_mul_pd(beta,beta2);
827 ewtab = fr->ic->tabq_coul_F;
828 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
829 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
831 /* Setup water-specific parameters */
832 inr = nlist->iinr[0];
833 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
834 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
835 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
836 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
838 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
839 rcutoff_scalar = fr->rcoulomb;
840 rcutoff = _mm256_set1_pd(rcutoff_scalar);
841 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
843 sh_vdw_invrcut6 = _mm256_set1_pd(fr->ic->sh_invrc6);
844 rvdw = _mm256_set1_pd(fr->rvdw);
846 /* Avoid stupid compiler warnings */
847 jnrA = jnrB = jnrC = jnrD = 0;
856 for(iidx=0;iidx<4*DIM;iidx++)
861 /* Start outer loop over neighborlists */
862 for(iidx=0; iidx<nri; iidx++)
864 /* Load shift vector for this list */
865 i_shift_offset = DIM*shiftidx[iidx];
867 /* Load limits for loop over neighbors */
868 j_index_start = jindex[iidx];
869 j_index_end = jindex[iidx+1];
871 /* Get outer coordinate index */
873 i_coord_offset = DIM*inr;
875 /* Load i particle coords and add shift vector */
876 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
877 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
879 fix0 = _mm256_setzero_pd();
880 fiy0 = _mm256_setzero_pd();
881 fiz0 = _mm256_setzero_pd();
882 fix1 = _mm256_setzero_pd();
883 fiy1 = _mm256_setzero_pd();
884 fiz1 = _mm256_setzero_pd();
885 fix2 = _mm256_setzero_pd();
886 fiy2 = _mm256_setzero_pd();
887 fiz2 = _mm256_setzero_pd();
889 /* Start inner kernel loop */
890 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
893 /* Get j neighbor index, and coordinate index */
898 j_coord_offsetA = DIM*jnrA;
899 j_coord_offsetB = DIM*jnrB;
900 j_coord_offsetC = DIM*jnrC;
901 j_coord_offsetD = DIM*jnrD;
903 /* load j atom coordinates */
904 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
905 x+j_coord_offsetC,x+j_coord_offsetD,
908 /* Calculate displacement vector */
909 dx00 = _mm256_sub_pd(ix0,jx0);
910 dy00 = _mm256_sub_pd(iy0,jy0);
911 dz00 = _mm256_sub_pd(iz0,jz0);
912 dx10 = _mm256_sub_pd(ix1,jx0);
913 dy10 = _mm256_sub_pd(iy1,jy0);
914 dz10 = _mm256_sub_pd(iz1,jz0);
915 dx20 = _mm256_sub_pd(ix2,jx0);
916 dy20 = _mm256_sub_pd(iy2,jy0);
917 dz20 = _mm256_sub_pd(iz2,jz0);
919 /* Calculate squared distance and things based on it */
920 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
921 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
922 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
924 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
925 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
926 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
928 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
929 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
930 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
932 /* Load parameters for j particles */
933 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
934 charge+jnrC+0,charge+jnrD+0);
935 vdwjidx0A = 2*vdwtype[jnrA+0];
936 vdwjidx0B = 2*vdwtype[jnrB+0];
937 vdwjidx0C = 2*vdwtype[jnrC+0];
938 vdwjidx0D = 2*vdwtype[jnrD+0];
940 fjx0 = _mm256_setzero_pd();
941 fjy0 = _mm256_setzero_pd();
942 fjz0 = _mm256_setzero_pd();
944 /**************************
945 * CALCULATE INTERACTIONS *
946 **************************/
948 if (gmx_mm256_any_lt(rsq00,rcutoff2))
951 r00 = _mm256_mul_pd(rsq00,rinv00);
953 /* Compute parameters for interactions between i and j atoms */
954 qq00 = _mm256_mul_pd(iq0,jq0);
955 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
956 vdwioffsetptr0+vdwjidx0B,
957 vdwioffsetptr0+vdwjidx0C,
958 vdwioffsetptr0+vdwjidx0D,
961 /* EWALD ELECTROSTATICS */
963 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
964 ewrt = _mm256_mul_pd(r00,ewtabscale);
965 ewitab = _mm256_cvttpd_epi32(ewrt);
966 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
967 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
968 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
970 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
971 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
973 /* LENNARD-JONES DISPERSION/REPULSION */
975 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
976 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
978 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
980 fscal = _mm256_add_pd(felec,fvdw);
982 fscal = _mm256_and_pd(fscal,cutoff_mask);
984 /* Calculate temporary vectorial force */
985 tx = _mm256_mul_pd(fscal,dx00);
986 ty = _mm256_mul_pd(fscal,dy00);
987 tz = _mm256_mul_pd(fscal,dz00);
989 /* Update vectorial force */
990 fix0 = _mm256_add_pd(fix0,tx);
991 fiy0 = _mm256_add_pd(fiy0,ty);
992 fiz0 = _mm256_add_pd(fiz0,tz);
994 fjx0 = _mm256_add_pd(fjx0,tx);
995 fjy0 = _mm256_add_pd(fjy0,ty);
996 fjz0 = _mm256_add_pd(fjz0,tz);
1000 /**************************
1001 * CALCULATE INTERACTIONS *
1002 **************************/
1004 if (gmx_mm256_any_lt(rsq10,rcutoff2))
1007 r10 = _mm256_mul_pd(rsq10,rinv10);
1009 /* Compute parameters for interactions between i and j atoms */
1010 qq10 = _mm256_mul_pd(iq1,jq0);
1012 /* EWALD ELECTROSTATICS */
1014 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1015 ewrt = _mm256_mul_pd(r10,ewtabscale);
1016 ewitab = _mm256_cvttpd_epi32(ewrt);
1017 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1018 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1019 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1021 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1022 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
1024 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
1028 fscal = _mm256_and_pd(fscal,cutoff_mask);
1030 /* Calculate temporary vectorial force */
1031 tx = _mm256_mul_pd(fscal,dx10);
1032 ty = _mm256_mul_pd(fscal,dy10);
1033 tz = _mm256_mul_pd(fscal,dz10);
1035 /* Update vectorial force */
1036 fix1 = _mm256_add_pd(fix1,tx);
1037 fiy1 = _mm256_add_pd(fiy1,ty);
1038 fiz1 = _mm256_add_pd(fiz1,tz);
1040 fjx0 = _mm256_add_pd(fjx0,tx);
1041 fjy0 = _mm256_add_pd(fjy0,ty);
1042 fjz0 = _mm256_add_pd(fjz0,tz);
1046 /**************************
1047 * CALCULATE INTERACTIONS *
1048 **************************/
1050 if (gmx_mm256_any_lt(rsq20,rcutoff2))
1053 r20 = _mm256_mul_pd(rsq20,rinv20);
1055 /* Compute parameters for interactions between i and j atoms */
1056 qq20 = _mm256_mul_pd(iq2,jq0);
1058 /* EWALD ELECTROSTATICS */
1060 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1061 ewrt = _mm256_mul_pd(r20,ewtabscale);
1062 ewitab = _mm256_cvttpd_epi32(ewrt);
1063 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1064 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1065 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1067 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1068 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
1070 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
1074 fscal = _mm256_and_pd(fscal,cutoff_mask);
1076 /* Calculate temporary vectorial force */
1077 tx = _mm256_mul_pd(fscal,dx20);
1078 ty = _mm256_mul_pd(fscal,dy20);
1079 tz = _mm256_mul_pd(fscal,dz20);
1081 /* Update vectorial force */
1082 fix2 = _mm256_add_pd(fix2,tx);
1083 fiy2 = _mm256_add_pd(fiy2,ty);
1084 fiz2 = _mm256_add_pd(fiz2,tz);
1086 fjx0 = _mm256_add_pd(fjx0,tx);
1087 fjy0 = _mm256_add_pd(fjy0,ty);
1088 fjz0 = _mm256_add_pd(fjz0,tz);
1092 fjptrA = f+j_coord_offsetA;
1093 fjptrB = f+j_coord_offsetB;
1094 fjptrC = f+j_coord_offsetC;
1095 fjptrD = f+j_coord_offsetD;
1097 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1099 /* Inner loop uses 127 flops */
1102 if(jidx<j_index_end)
1105 /* Get j neighbor index, and coordinate index */
1106 jnrlistA = jjnr[jidx];
1107 jnrlistB = jjnr[jidx+1];
1108 jnrlistC = jjnr[jidx+2];
1109 jnrlistD = jjnr[jidx+3];
1110 /* Sign of each element will be negative for non-real atoms.
1111 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1112 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
1114 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1116 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
1117 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
1118 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
1120 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1121 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1122 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1123 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1124 j_coord_offsetA = DIM*jnrA;
1125 j_coord_offsetB = DIM*jnrB;
1126 j_coord_offsetC = DIM*jnrC;
1127 j_coord_offsetD = DIM*jnrD;
1129 /* load j atom coordinates */
1130 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1131 x+j_coord_offsetC,x+j_coord_offsetD,
1134 /* Calculate displacement vector */
1135 dx00 = _mm256_sub_pd(ix0,jx0);
1136 dy00 = _mm256_sub_pd(iy0,jy0);
1137 dz00 = _mm256_sub_pd(iz0,jz0);
1138 dx10 = _mm256_sub_pd(ix1,jx0);
1139 dy10 = _mm256_sub_pd(iy1,jy0);
1140 dz10 = _mm256_sub_pd(iz1,jz0);
1141 dx20 = _mm256_sub_pd(ix2,jx0);
1142 dy20 = _mm256_sub_pd(iy2,jy0);
1143 dz20 = _mm256_sub_pd(iz2,jz0);
1145 /* Calculate squared distance and things based on it */
1146 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
1147 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
1148 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
1150 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
1151 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
1152 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
1154 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
1155 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
1156 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
1158 /* Load parameters for j particles */
1159 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
1160 charge+jnrC+0,charge+jnrD+0);
1161 vdwjidx0A = 2*vdwtype[jnrA+0];
1162 vdwjidx0B = 2*vdwtype[jnrB+0];
1163 vdwjidx0C = 2*vdwtype[jnrC+0];
1164 vdwjidx0D = 2*vdwtype[jnrD+0];
1166 fjx0 = _mm256_setzero_pd();
1167 fjy0 = _mm256_setzero_pd();
1168 fjz0 = _mm256_setzero_pd();
1170 /**************************
1171 * CALCULATE INTERACTIONS *
1172 **************************/
1174 if (gmx_mm256_any_lt(rsq00,rcutoff2))
1177 r00 = _mm256_mul_pd(rsq00,rinv00);
1178 r00 = _mm256_andnot_pd(dummy_mask,r00);
1180 /* Compute parameters for interactions between i and j atoms */
1181 qq00 = _mm256_mul_pd(iq0,jq0);
1182 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
1183 vdwioffsetptr0+vdwjidx0B,
1184 vdwioffsetptr0+vdwjidx0C,
1185 vdwioffsetptr0+vdwjidx0D,
1188 /* EWALD ELECTROSTATICS */
1190 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1191 ewrt = _mm256_mul_pd(r00,ewtabscale);
1192 ewitab = _mm256_cvttpd_epi32(ewrt);
1193 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1194 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1195 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1197 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1198 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
1200 /* LENNARD-JONES DISPERSION/REPULSION */
1202 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1203 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
1205 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
1207 fscal = _mm256_add_pd(felec,fvdw);
1209 fscal = _mm256_and_pd(fscal,cutoff_mask);
1211 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1213 /* Calculate temporary vectorial force */
1214 tx = _mm256_mul_pd(fscal,dx00);
1215 ty = _mm256_mul_pd(fscal,dy00);
1216 tz = _mm256_mul_pd(fscal,dz00);
1218 /* Update vectorial force */
1219 fix0 = _mm256_add_pd(fix0,tx);
1220 fiy0 = _mm256_add_pd(fiy0,ty);
1221 fiz0 = _mm256_add_pd(fiz0,tz);
1223 fjx0 = _mm256_add_pd(fjx0,tx);
1224 fjy0 = _mm256_add_pd(fjy0,ty);
1225 fjz0 = _mm256_add_pd(fjz0,tz);
1229 /**************************
1230 * CALCULATE INTERACTIONS *
1231 **************************/
1233 if (gmx_mm256_any_lt(rsq10,rcutoff2))
1236 r10 = _mm256_mul_pd(rsq10,rinv10);
1237 r10 = _mm256_andnot_pd(dummy_mask,r10);
1239 /* Compute parameters for interactions between i and j atoms */
1240 qq10 = _mm256_mul_pd(iq1,jq0);
1242 /* EWALD ELECTROSTATICS */
1244 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1245 ewrt = _mm256_mul_pd(r10,ewtabscale);
1246 ewitab = _mm256_cvttpd_epi32(ewrt);
1247 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1248 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1249 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1251 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1252 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
1254 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
1258 fscal = _mm256_and_pd(fscal,cutoff_mask);
1260 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1262 /* Calculate temporary vectorial force */
1263 tx = _mm256_mul_pd(fscal,dx10);
1264 ty = _mm256_mul_pd(fscal,dy10);
1265 tz = _mm256_mul_pd(fscal,dz10);
1267 /* Update vectorial force */
1268 fix1 = _mm256_add_pd(fix1,tx);
1269 fiy1 = _mm256_add_pd(fiy1,ty);
1270 fiz1 = _mm256_add_pd(fiz1,tz);
1272 fjx0 = _mm256_add_pd(fjx0,tx);
1273 fjy0 = _mm256_add_pd(fjy0,ty);
1274 fjz0 = _mm256_add_pd(fjz0,tz);
1278 /**************************
1279 * CALCULATE INTERACTIONS *
1280 **************************/
1282 if (gmx_mm256_any_lt(rsq20,rcutoff2))
1285 r20 = _mm256_mul_pd(rsq20,rinv20);
1286 r20 = _mm256_andnot_pd(dummy_mask,r20);
1288 /* Compute parameters for interactions between i and j atoms */
1289 qq20 = _mm256_mul_pd(iq2,jq0);
1291 /* EWALD ELECTROSTATICS */
1293 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1294 ewrt = _mm256_mul_pd(r20,ewtabscale);
1295 ewitab = _mm256_cvttpd_epi32(ewrt);
1296 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1297 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1298 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1300 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1301 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
1303 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
1307 fscal = _mm256_and_pd(fscal,cutoff_mask);
1309 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1311 /* Calculate temporary vectorial force */
1312 tx = _mm256_mul_pd(fscal,dx20);
1313 ty = _mm256_mul_pd(fscal,dy20);
1314 tz = _mm256_mul_pd(fscal,dz20);
1316 /* Update vectorial force */
1317 fix2 = _mm256_add_pd(fix2,tx);
1318 fiy2 = _mm256_add_pd(fiy2,ty);
1319 fiz2 = _mm256_add_pd(fiz2,tz);
1321 fjx0 = _mm256_add_pd(fjx0,tx);
1322 fjy0 = _mm256_add_pd(fjy0,ty);
1323 fjz0 = _mm256_add_pd(fjz0,tz);
1327 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1328 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1329 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1330 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1332 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1334 /* Inner loop uses 130 flops */
1337 /* End of innermost loop */
1339 gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1340 f+i_coord_offset,fshift+i_shift_offset);
1342 /* Increment number of inner iterations */
1343 inneriter += j_index_end - j_index_start;
1345 /* Outer loop uses 18 flops */
1348 /* Increment number of outer iterations */
1351 /* Update outer/inner flops */
1353 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*130);