2 * Note: this file was generated by the Gromacs avx_256_double kernel generator.
4 * This source code is part of
8 * Copyright (c) 2001-2012, The GROMACS Development Team
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
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
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.
28 #include "../nb_kernel.h"
29 #include "types/simple.h"
33 #include "gmx_math_x86_avx_256_double.h"
34 #include "kernelutil_x86_avx_256_double.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_avx_256_double
38 * Electrostatics interaction: ReactionField
39 * VdW interaction: LennardJones
40 * Geometry: Water3-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecRF_VdwLJ_GeomW3P1_VF_avx_256_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)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
62 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
63 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
64 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
66 real *shiftvec,*fshift,*x,*f;
67 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
69 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
70 real * vdwioffsetptr0;
71 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
72 real * vdwioffsetptr1;
73 __m256d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
74 real * vdwioffsetptr2;
75 __m256d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
76 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
77 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
78 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
79 __m256d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
80 __m256d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
81 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
84 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
87 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
88 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
89 __m256d dummy_mask,cutoff_mask;
90 __m128 tmpmask0,tmpmask1;
91 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
92 __m256d one = _mm256_set1_pd(1.0);
93 __m256d two = _mm256_set1_pd(2.0);
99 jindex = nlist->jindex;
101 shiftidx = nlist->shift;
103 shiftvec = fr->shift_vec[0];
104 fshift = fr->fshift[0];
105 facel = _mm256_set1_pd(fr->epsfac);
106 charge = mdatoms->chargeA;
107 krf = _mm256_set1_pd(fr->ic->k_rf);
108 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
109 crf = _mm256_set1_pd(fr->ic->c_rf);
110 nvdwtype = fr->ntype;
112 vdwtype = mdatoms->typeA;
114 /* Setup water-specific parameters */
115 inr = nlist->iinr[0];
116 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
117 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
118 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
119 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
121 /* Avoid stupid compiler warnings */
122 jnrA = jnrB = jnrC = jnrD = 0;
131 for(iidx=0;iidx<4*DIM;iidx++)
136 /* Start outer loop over neighborlists */
137 for(iidx=0; iidx<nri; iidx++)
139 /* Load shift vector for this list */
140 i_shift_offset = DIM*shiftidx[iidx];
142 /* Load limits for loop over neighbors */
143 j_index_start = jindex[iidx];
144 j_index_end = jindex[iidx+1];
146 /* Get outer coordinate index */
148 i_coord_offset = DIM*inr;
150 /* Load i particle coords and add shift vector */
151 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
152 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
154 fix0 = _mm256_setzero_pd();
155 fiy0 = _mm256_setzero_pd();
156 fiz0 = _mm256_setzero_pd();
157 fix1 = _mm256_setzero_pd();
158 fiy1 = _mm256_setzero_pd();
159 fiz1 = _mm256_setzero_pd();
160 fix2 = _mm256_setzero_pd();
161 fiy2 = _mm256_setzero_pd();
162 fiz2 = _mm256_setzero_pd();
164 /* Reset potential sums */
165 velecsum = _mm256_setzero_pd();
166 vvdwsum = _mm256_setzero_pd();
168 /* Start inner kernel loop */
169 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
172 /* Get j neighbor index, and coordinate index */
177 j_coord_offsetA = DIM*jnrA;
178 j_coord_offsetB = DIM*jnrB;
179 j_coord_offsetC = DIM*jnrC;
180 j_coord_offsetD = DIM*jnrD;
182 /* load j atom coordinates */
183 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
184 x+j_coord_offsetC,x+j_coord_offsetD,
187 /* Calculate displacement vector */
188 dx00 = _mm256_sub_pd(ix0,jx0);
189 dy00 = _mm256_sub_pd(iy0,jy0);
190 dz00 = _mm256_sub_pd(iz0,jz0);
191 dx10 = _mm256_sub_pd(ix1,jx0);
192 dy10 = _mm256_sub_pd(iy1,jy0);
193 dz10 = _mm256_sub_pd(iz1,jz0);
194 dx20 = _mm256_sub_pd(ix2,jx0);
195 dy20 = _mm256_sub_pd(iy2,jy0);
196 dz20 = _mm256_sub_pd(iz2,jz0);
198 /* Calculate squared distance and things based on it */
199 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
200 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
201 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
203 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
204 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
205 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
207 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
208 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
209 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
211 /* Load parameters for j particles */
212 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
213 charge+jnrC+0,charge+jnrD+0);
214 vdwjidx0A = 2*vdwtype[jnrA+0];
215 vdwjidx0B = 2*vdwtype[jnrB+0];
216 vdwjidx0C = 2*vdwtype[jnrC+0];
217 vdwjidx0D = 2*vdwtype[jnrD+0];
219 fjx0 = _mm256_setzero_pd();
220 fjy0 = _mm256_setzero_pd();
221 fjz0 = _mm256_setzero_pd();
223 /**************************
224 * CALCULATE INTERACTIONS *
225 **************************/
227 /* Compute parameters for interactions between i and j atoms */
228 qq00 = _mm256_mul_pd(iq0,jq0);
229 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
230 vdwioffsetptr0+vdwjidx0B,
231 vdwioffsetptr0+vdwjidx0C,
232 vdwioffsetptr0+vdwjidx0D,
235 /* REACTION-FIELD ELECTROSTATICS */
236 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
237 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
239 /* LENNARD-JONES DISPERSION/REPULSION */
241 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
242 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
243 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
244 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
245 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
247 /* Update potential sum for this i atom from the interaction with this j atom. */
248 velecsum = _mm256_add_pd(velecsum,velec);
249 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
251 fscal = _mm256_add_pd(felec,fvdw);
253 /* Calculate temporary vectorial force */
254 tx = _mm256_mul_pd(fscal,dx00);
255 ty = _mm256_mul_pd(fscal,dy00);
256 tz = _mm256_mul_pd(fscal,dz00);
258 /* Update vectorial force */
259 fix0 = _mm256_add_pd(fix0,tx);
260 fiy0 = _mm256_add_pd(fiy0,ty);
261 fiz0 = _mm256_add_pd(fiz0,tz);
263 fjx0 = _mm256_add_pd(fjx0,tx);
264 fjy0 = _mm256_add_pd(fjy0,ty);
265 fjz0 = _mm256_add_pd(fjz0,tz);
267 /**************************
268 * CALCULATE INTERACTIONS *
269 **************************/
271 /* Compute parameters for interactions between i and j atoms */
272 qq10 = _mm256_mul_pd(iq1,jq0);
274 /* REACTION-FIELD ELECTROSTATICS */
275 velec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_add_pd(rinv10,_mm256_mul_pd(krf,rsq10)),crf));
276 felec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
278 /* Update potential sum for this i atom from the interaction with this j atom. */
279 velecsum = _mm256_add_pd(velecsum,velec);
283 /* Calculate temporary vectorial force */
284 tx = _mm256_mul_pd(fscal,dx10);
285 ty = _mm256_mul_pd(fscal,dy10);
286 tz = _mm256_mul_pd(fscal,dz10);
288 /* Update vectorial force */
289 fix1 = _mm256_add_pd(fix1,tx);
290 fiy1 = _mm256_add_pd(fiy1,ty);
291 fiz1 = _mm256_add_pd(fiz1,tz);
293 fjx0 = _mm256_add_pd(fjx0,tx);
294 fjy0 = _mm256_add_pd(fjy0,ty);
295 fjz0 = _mm256_add_pd(fjz0,tz);
297 /**************************
298 * CALCULATE INTERACTIONS *
299 **************************/
301 /* Compute parameters for interactions between i and j atoms */
302 qq20 = _mm256_mul_pd(iq2,jq0);
304 /* REACTION-FIELD ELECTROSTATICS */
305 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_add_pd(rinv20,_mm256_mul_pd(krf,rsq20)),crf));
306 felec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
308 /* Update potential sum for this i atom from the interaction with this j atom. */
309 velecsum = _mm256_add_pd(velecsum,velec);
313 /* Calculate temporary vectorial force */
314 tx = _mm256_mul_pd(fscal,dx20);
315 ty = _mm256_mul_pd(fscal,dy20);
316 tz = _mm256_mul_pd(fscal,dz20);
318 /* Update vectorial force */
319 fix2 = _mm256_add_pd(fix2,tx);
320 fiy2 = _mm256_add_pd(fiy2,ty);
321 fiz2 = _mm256_add_pd(fiz2,tz);
323 fjx0 = _mm256_add_pd(fjx0,tx);
324 fjy0 = _mm256_add_pd(fjy0,ty);
325 fjz0 = _mm256_add_pd(fjz0,tz);
327 fjptrA = f+j_coord_offsetA;
328 fjptrB = f+j_coord_offsetB;
329 fjptrC = f+j_coord_offsetC;
330 fjptrD = f+j_coord_offsetD;
332 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
334 /* Inner loop uses 111 flops */
340 /* Get j neighbor index, and coordinate index */
341 jnrlistA = jjnr[jidx];
342 jnrlistB = jjnr[jidx+1];
343 jnrlistC = jjnr[jidx+2];
344 jnrlistD = jjnr[jidx+3];
345 /* Sign of each element will be negative for non-real atoms.
346 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
347 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
349 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
351 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
352 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
353 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
355 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
356 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
357 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
358 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
359 j_coord_offsetA = DIM*jnrA;
360 j_coord_offsetB = DIM*jnrB;
361 j_coord_offsetC = DIM*jnrC;
362 j_coord_offsetD = DIM*jnrD;
364 /* load j atom coordinates */
365 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
366 x+j_coord_offsetC,x+j_coord_offsetD,
369 /* Calculate displacement vector */
370 dx00 = _mm256_sub_pd(ix0,jx0);
371 dy00 = _mm256_sub_pd(iy0,jy0);
372 dz00 = _mm256_sub_pd(iz0,jz0);
373 dx10 = _mm256_sub_pd(ix1,jx0);
374 dy10 = _mm256_sub_pd(iy1,jy0);
375 dz10 = _mm256_sub_pd(iz1,jz0);
376 dx20 = _mm256_sub_pd(ix2,jx0);
377 dy20 = _mm256_sub_pd(iy2,jy0);
378 dz20 = _mm256_sub_pd(iz2,jz0);
380 /* Calculate squared distance and things based on it */
381 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
382 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
383 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
385 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
386 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
387 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
389 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
390 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
391 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
393 /* Load parameters for j particles */
394 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
395 charge+jnrC+0,charge+jnrD+0);
396 vdwjidx0A = 2*vdwtype[jnrA+0];
397 vdwjidx0B = 2*vdwtype[jnrB+0];
398 vdwjidx0C = 2*vdwtype[jnrC+0];
399 vdwjidx0D = 2*vdwtype[jnrD+0];
401 fjx0 = _mm256_setzero_pd();
402 fjy0 = _mm256_setzero_pd();
403 fjz0 = _mm256_setzero_pd();
405 /**************************
406 * CALCULATE INTERACTIONS *
407 **************************/
409 /* Compute parameters for interactions between i and j atoms */
410 qq00 = _mm256_mul_pd(iq0,jq0);
411 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
412 vdwioffsetptr0+vdwjidx0B,
413 vdwioffsetptr0+vdwjidx0C,
414 vdwioffsetptr0+vdwjidx0D,
417 /* REACTION-FIELD ELECTROSTATICS */
418 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
419 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
421 /* LENNARD-JONES DISPERSION/REPULSION */
423 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
424 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
425 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
426 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
427 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
429 /* Update potential sum for this i atom from the interaction with this j atom. */
430 velec = _mm256_andnot_pd(dummy_mask,velec);
431 velecsum = _mm256_add_pd(velecsum,velec);
432 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
433 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
435 fscal = _mm256_add_pd(felec,fvdw);
437 fscal = _mm256_andnot_pd(dummy_mask,fscal);
439 /* Calculate temporary vectorial force */
440 tx = _mm256_mul_pd(fscal,dx00);
441 ty = _mm256_mul_pd(fscal,dy00);
442 tz = _mm256_mul_pd(fscal,dz00);
444 /* Update vectorial force */
445 fix0 = _mm256_add_pd(fix0,tx);
446 fiy0 = _mm256_add_pd(fiy0,ty);
447 fiz0 = _mm256_add_pd(fiz0,tz);
449 fjx0 = _mm256_add_pd(fjx0,tx);
450 fjy0 = _mm256_add_pd(fjy0,ty);
451 fjz0 = _mm256_add_pd(fjz0,tz);
453 /**************************
454 * CALCULATE INTERACTIONS *
455 **************************/
457 /* Compute parameters for interactions between i and j atoms */
458 qq10 = _mm256_mul_pd(iq1,jq0);
460 /* REACTION-FIELD ELECTROSTATICS */
461 velec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_add_pd(rinv10,_mm256_mul_pd(krf,rsq10)),crf));
462 felec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
464 /* Update potential sum for this i atom from the interaction with this j atom. */
465 velec = _mm256_andnot_pd(dummy_mask,velec);
466 velecsum = _mm256_add_pd(velecsum,velec);
470 fscal = _mm256_andnot_pd(dummy_mask,fscal);
472 /* Calculate temporary vectorial force */
473 tx = _mm256_mul_pd(fscal,dx10);
474 ty = _mm256_mul_pd(fscal,dy10);
475 tz = _mm256_mul_pd(fscal,dz10);
477 /* Update vectorial force */
478 fix1 = _mm256_add_pd(fix1,tx);
479 fiy1 = _mm256_add_pd(fiy1,ty);
480 fiz1 = _mm256_add_pd(fiz1,tz);
482 fjx0 = _mm256_add_pd(fjx0,tx);
483 fjy0 = _mm256_add_pd(fjy0,ty);
484 fjz0 = _mm256_add_pd(fjz0,tz);
486 /**************************
487 * CALCULATE INTERACTIONS *
488 **************************/
490 /* Compute parameters for interactions between i and j atoms */
491 qq20 = _mm256_mul_pd(iq2,jq0);
493 /* REACTION-FIELD ELECTROSTATICS */
494 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_add_pd(rinv20,_mm256_mul_pd(krf,rsq20)),crf));
495 felec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
497 /* Update potential sum for this i atom from the interaction with this j atom. */
498 velec = _mm256_andnot_pd(dummy_mask,velec);
499 velecsum = _mm256_add_pd(velecsum,velec);
503 fscal = _mm256_andnot_pd(dummy_mask,fscal);
505 /* Calculate temporary vectorial force */
506 tx = _mm256_mul_pd(fscal,dx20);
507 ty = _mm256_mul_pd(fscal,dy20);
508 tz = _mm256_mul_pd(fscal,dz20);
510 /* Update vectorial force */
511 fix2 = _mm256_add_pd(fix2,tx);
512 fiy2 = _mm256_add_pd(fiy2,ty);
513 fiz2 = _mm256_add_pd(fiz2,tz);
515 fjx0 = _mm256_add_pd(fjx0,tx);
516 fjy0 = _mm256_add_pd(fjy0,ty);
517 fjz0 = _mm256_add_pd(fjz0,tz);
519 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
520 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
521 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
522 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
524 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
526 /* Inner loop uses 111 flops */
529 /* End of innermost loop */
531 gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
532 f+i_coord_offset,fshift+i_shift_offset);
535 /* Update potential energies */
536 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
537 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
539 /* Increment number of inner iterations */
540 inneriter += j_index_end - j_index_start;
542 /* Outer loop uses 20 flops */
545 /* Increment number of outer iterations */
548 /* Update outer/inner flops */
550 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*111);
553 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_avx_256_double
554 * Electrostatics interaction: ReactionField
555 * VdW interaction: LennardJones
556 * Geometry: Water3-Particle
557 * Calculate force/pot: Force
560 nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_avx_256_double
561 (t_nblist * gmx_restrict nlist,
562 rvec * gmx_restrict xx,
563 rvec * gmx_restrict ff,
564 t_forcerec * gmx_restrict fr,
565 t_mdatoms * gmx_restrict mdatoms,
566 nb_kernel_data_t * gmx_restrict kernel_data,
567 t_nrnb * gmx_restrict nrnb)
569 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
570 * just 0 for non-waters.
571 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
572 * jnr indices corresponding to data put in the four positions in the SIMD register.
574 int i_shift_offset,i_coord_offset,outeriter,inneriter;
575 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
576 int jnrA,jnrB,jnrC,jnrD;
577 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
578 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
579 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
580 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
582 real *shiftvec,*fshift,*x,*f;
583 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
585 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
586 real * vdwioffsetptr0;
587 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
588 real * vdwioffsetptr1;
589 __m256d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
590 real * vdwioffsetptr2;
591 __m256d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
592 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
593 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
594 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
595 __m256d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
596 __m256d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
597 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
600 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
603 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
604 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
605 __m256d dummy_mask,cutoff_mask;
606 __m128 tmpmask0,tmpmask1;
607 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
608 __m256d one = _mm256_set1_pd(1.0);
609 __m256d two = _mm256_set1_pd(2.0);
615 jindex = nlist->jindex;
617 shiftidx = nlist->shift;
619 shiftvec = fr->shift_vec[0];
620 fshift = fr->fshift[0];
621 facel = _mm256_set1_pd(fr->epsfac);
622 charge = mdatoms->chargeA;
623 krf = _mm256_set1_pd(fr->ic->k_rf);
624 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
625 crf = _mm256_set1_pd(fr->ic->c_rf);
626 nvdwtype = fr->ntype;
628 vdwtype = mdatoms->typeA;
630 /* Setup water-specific parameters */
631 inr = nlist->iinr[0];
632 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
633 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
634 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
635 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
637 /* Avoid stupid compiler warnings */
638 jnrA = jnrB = jnrC = jnrD = 0;
647 for(iidx=0;iidx<4*DIM;iidx++)
652 /* Start outer loop over neighborlists */
653 for(iidx=0; iidx<nri; iidx++)
655 /* Load shift vector for this list */
656 i_shift_offset = DIM*shiftidx[iidx];
658 /* Load limits for loop over neighbors */
659 j_index_start = jindex[iidx];
660 j_index_end = jindex[iidx+1];
662 /* Get outer coordinate index */
664 i_coord_offset = DIM*inr;
666 /* Load i particle coords and add shift vector */
667 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
668 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
670 fix0 = _mm256_setzero_pd();
671 fiy0 = _mm256_setzero_pd();
672 fiz0 = _mm256_setzero_pd();
673 fix1 = _mm256_setzero_pd();
674 fiy1 = _mm256_setzero_pd();
675 fiz1 = _mm256_setzero_pd();
676 fix2 = _mm256_setzero_pd();
677 fiy2 = _mm256_setzero_pd();
678 fiz2 = _mm256_setzero_pd();
680 /* Start inner kernel loop */
681 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
684 /* Get j neighbor index, and coordinate index */
689 j_coord_offsetA = DIM*jnrA;
690 j_coord_offsetB = DIM*jnrB;
691 j_coord_offsetC = DIM*jnrC;
692 j_coord_offsetD = DIM*jnrD;
694 /* load j atom coordinates */
695 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
696 x+j_coord_offsetC,x+j_coord_offsetD,
699 /* Calculate displacement vector */
700 dx00 = _mm256_sub_pd(ix0,jx0);
701 dy00 = _mm256_sub_pd(iy0,jy0);
702 dz00 = _mm256_sub_pd(iz0,jz0);
703 dx10 = _mm256_sub_pd(ix1,jx0);
704 dy10 = _mm256_sub_pd(iy1,jy0);
705 dz10 = _mm256_sub_pd(iz1,jz0);
706 dx20 = _mm256_sub_pd(ix2,jx0);
707 dy20 = _mm256_sub_pd(iy2,jy0);
708 dz20 = _mm256_sub_pd(iz2,jz0);
710 /* Calculate squared distance and things based on it */
711 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
712 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
713 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
715 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
716 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
717 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
719 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
720 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
721 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
723 /* Load parameters for j particles */
724 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
725 charge+jnrC+0,charge+jnrD+0);
726 vdwjidx0A = 2*vdwtype[jnrA+0];
727 vdwjidx0B = 2*vdwtype[jnrB+0];
728 vdwjidx0C = 2*vdwtype[jnrC+0];
729 vdwjidx0D = 2*vdwtype[jnrD+0];
731 fjx0 = _mm256_setzero_pd();
732 fjy0 = _mm256_setzero_pd();
733 fjz0 = _mm256_setzero_pd();
735 /**************************
736 * CALCULATE INTERACTIONS *
737 **************************/
739 /* Compute parameters for interactions between i and j atoms */
740 qq00 = _mm256_mul_pd(iq0,jq0);
741 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
742 vdwioffsetptr0+vdwjidx0B,
743 vdwioffsetptr0+vdwjidx0C,
744 vdwioffsetptr0+vdwjidx0D,
747 /* REACTION-FIELD ELECTROSTATICS */
748 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
750 /* LENNARD-JONES DISPERSION/REPULSION */
752 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
753 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
755 fscal = _mm256_add_pd(felec,fvdw);
757 /* Calculate temporary vectorial force */
758 tx = _mm256_mul_pd(fscal,dx00);
759 ty = _mm256_mul_pd(fscal,dy00);
760 tz = _mm256_mul_pd(fscal,dz00);
762 /* Update vectorial force */
763 fix0 = _mm256_add_pd(fix0,tx);
764 fiy0 = _mm256_add_pd(fiy0,ty);
765 fiz0 = _mm256_add_pd(fiz0,tz);
767 fjx0 = _mm256_add_pd(fjx0,tx);
768 fjy0 = _mm256_add_pd(fjy0,ty);
769 fjz0 = _mm256_add_pd(fjz0,tz);
771 /**************************
772 * CALCULATE INTERACTIONS *
773 **************************/
775 /* Compute parameters for interactions between i and j atoms */
776 qq10 = _mm256_mul_pd(iq1,jq0);
778 /* REACTION-FIELD ELECTROSTATICS */
779 felec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
783 /* Calculate temporary vectorial force */
784 tx = _mm256_mul_pd(fscal,dx10);
785 ty = _mm256_mul_pd(fscal,dy10);
786 tz = _mm256_mul_pd(fscal,dz10);
788 /* Update vectorial force */
789 fix1 = _mm256_add_pd(fix1,tx);
790 fiy1 = _mm256_add_pd(fiy1,ty);
791 fiz1 = _mm256_add_pd(fiz1,tz);
793 fjx0 = _mm256_add_pd(fjx0,tx);
794 fjy0 = _mm256_add_pd(fjy0,ty);
795 fjz0 = _mm256_add_pd(fjz0,tz);
797 /**************************
798 * CALCULATE INTERACTIONS *
799 **************************/
801 /* Compute parameters for interactions between i and j atoms */
802 qq20 = _mm256_mul_pd(iq2,jq0);
804 /* REACTION-FIELD ELECTROSTATICS */
805 felec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
809 /* Calculate temporary vectorial force */
810 tx = _mm256_mul_pd(fscal,dx20);
811 ty = _mm256_mul_pd(fscal,dy20);
812 tz = _mm256_mul_pd(fscal,dz20);
814 /* Update vectorial force */
815 fix2 = _mm256_add_pd(fix2,tx);
816 fiy2 = _mm256_add_pd(fiy2,ty);
817 fiz2 = _mm256_add_pd(fiz2,tz);
819 fjx0 = _mm256_add_pd(fjx0,tx);
820 fjy0 = _mm256_add_pd(fjy0,ty);
821 fjz0 = _mm256_add_pd(fjz0,tz);
823 fjptrA = f+j_coord_offsetA;
824 fjptrB = f+j_coord_offsetB;
825 fjptrC = f+j_coord_offsetC;
826 fjptrD = f+j_coord_offsetD;
828 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
830 /* Inner loop uses 91 flops */
836 /* Get j neighbor index, and coordinate index */
837 jnrlistA = jjnr[jidx];
838 jnrlistB = jjnr[jidx+1];
839 jnrlistC = jjnr[jidx+2];
840 jnrlistD = jjnr[jidx+3];
841 /* Sign of each element will be negative for non-real atoms.
842 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
843 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
845 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
847 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
848 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
849 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
851 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
852 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
853 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
854 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
855 j_coord_offsetA = DIM*jnrA;
856 j_coord_offsetB = DIM*jnrB;
857 j_coord_offsetC = DIM*jnrC;
858 j_coord_offsetD = DIM*jnrD;
860 /* load j atom coordinates */
861 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
862 x+j_coord_offsetC,x+j_coord_offsetD,
865 /* Calculate displacement vector */
866 dx00 = _mm256_sub_pd(ix0,jx0);
867 dy00 = _mm256_sub_pd(iy0,jy0);
868 dz00 = _mm256_sub_pd(iz0,jz0);
869 dx10 = _mm256_sub_pd(ix1,jx0);
870 dy10 = _mm256_sub_pd(iy1,jy0);
871 dz10 = _mm256_sub_pd(iz1,jz0);
872 dx20 = _mm256_sub_pd(ix2,jx0);
873 dy20 = _mm256_sub_pd(iy2,jy0);
874 dz20 = _mm256_sub_pd(iz2,jz0);
876 /* Calculate squared distance and things based on it */
877 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
878 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
879 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
881 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
882 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
883 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
885 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
886 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
887 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
889 /* Load parameters for j particles */
890 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
891 charge+jnrC+0,charge+jnrD+0);
892 vdwjidx0A = 2*vdwtype[jnrA+0];
893 vdwjidx0B = 2*vdwtype[jnrB+0];
894 vdwjidx0C = 2*vdwtype[jnrC+0];
895 vdwjidx0D = 2*vdwtype[jnrD+0];
897 fjx0 = _mm256_setzero_pd();
898 fjy0 = _mm256_setzero_pd();
899 fjz0 = _mm256_setzero_pd();
901 /**************************
902 * CALCULATE INTERACTIONS *
903 **************************/
905 /* Compute parameters for interactions between i and j atoms */
906 qq00 = _mm256_mul_pd(iq0,jq0);
907 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
908 vdwioffsetptr0+vdwjidx0B,
909 vdwioffsetptr0+vdwjidx0C,
910 vdwioffsetptr0+vdwjidx0D,
913 /* REACTION-FIELD ELECTROSTATICS */
914 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
916 /* LENNARD-JONES DISPERSION/REPULSION */
918 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
919 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
921 fscal = _mm256_add_pd(felec,fvdw);
923 fscal = _mm256_andnot_pd(dummy_mask,fscal);
925 /* Calculate temporary vectorial force */
926 tx = _mm256_mul_pd(fscal,dx00);
927 ty = _mm256_mul_pd(fscal,dy00);
928 tz = _mm256_mul_pd(fscal,dz00);
930 /* Update vectorial force */
931 fix0 = _mm256_add_pd(fix0,tx);
932 fiy0 = _mm256_add_pd(fiy0,ty);
933 fiz0 = _mm256_add_pd(fiz0,tz);
935 fjx0 = _mm256_add_pd(fjx0,tx);
936 fjy0 = _mm256_add_pd(fjy0,ty);
937 fjz0 = _mm256_add_pd(fjz0,tz);
939 /**************************
940 * CALCULATE INTERACTIONS *
941 **************************/
943 /* Compute parameters for interactions between i and j atoms */
944 qq10 = _mm256_mul_pd(iq1,jq0);
946 /* REACTION-FIELD ELECTROSTATICS */
947 felec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_mul_pd(rinv10,rinvsq10),krf2));
951 fscal = _mm256_andnot_pd(dummy_mask,fscal);
953 /* Calculate temporary vectorial force */
954 tx = _mm256_mul_pd(fscal,dx10);
955 ty = _mm256_mul_pd(fscal,dy10);
956 tz = _mm256_mul_pd(fscal,dz10);
958 /* Update vectorial force */
959 fix1 = _mm256_add_pd(fix1,tx);
960 fiy1 = _mm256_add_pd(fiy1,ty);
961 fiz1 = _mm256_add_pd(fiz1,tz);
963 fjx0 = _mm256_add_pd(fjx0,tx);
964 fjy0 = _mm256_add_pd(fjy0,ty);
965 fjz0 = _mm256_add_pd(fjz0,tz);
967 /**************************
968 * CALCULATE INTERACTIONS *
969 **************************/
971 /* Compute parameters for interactions between i and j atoms */
972 qq20 = _mm256_mul_pd(iq2,jq0);
974 /* REACTION-FIELD ELECTROSTATICS */
975 felec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_mul_pd(rinv20,rinvsq20),krf2));
979 fscal = _mm256_andnot_pd(dummy_mask,fscal);
981 /* Calculate temporary vectorial force */
982 tx = _mm256_mul_pd(fscal,dx20);
983 ty = _mm256_mul_pd(fscal,dy20);
984 tz = _mm256_mul_pd(fscal,dz20);
986 /* Update vectorial force */
987 fix2 = _mm256_add_pd(fix2,tx);
988 fiy2 = _mm256_add_pd(fiy2,ty);
989 fiz2 = _mm256_add_pd(fiz2,tz);
991 fjx0 = _mm256_add_pd(fjx0,tx);
992 fjy0 = _mm256_add_pd(fjy0,ty);
993 fjz0 = _mm256_add_pd(fjz0,tz);
995 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
996 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
997 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
998 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1000 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1002 /* Inner loop uses 91 flops */
1005 /* End of innermost loop */
1007 gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1008 f+i_coord_offset,fshift+i_shift_offset);
1010 /* Increment number of inner iterations */
1011 inneriter += j_index_end - j_index_start;
1013 /* Outer loop uses 18 flops */
1016 /* Increment number of outer iterations */
1019 /* Update outer/inner flops */
1021 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*91);