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_VdwNone_GeomP1P1_VF_avx_256_double
38 * Electrostatics interaction: ReactionField
39 * VdW interaction: None
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecRF_VdwNone_GeomP1P1_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 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
73 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
74 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
75 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
77 __m256d dummy_mask,cutoff_mask;
78 __m128 tmpmask0,tmpmask1;
79 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
80 __m256d one = _mm256_set1_pd(1.0);
81 __m256d two = _mm256_set1_pd(2.0);
87 jindex = nlist->jindex;
89 shiftidx = nlist->shift;
91 shiftvec = fr->shift_vec[0];
92 fshift = fr->fshift[0];
93 facel = _mm256_set1_pd(fr->epsfac);
94 charge = mdatoms->chargeA;
95 krf = _mm256_set1_pd(fr->ic->k_rf);
96 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
97 crf = _mm256_set1_pd(fr->ic->c_rf);
99 /* Avoid stupid compiler warnings */
100 jnrA = jnrB = jnrC = jnrD = 0;
109 for(iidx=0;iidx<4*DIM;iidx++)
114 /* Start outer loop over neighborlists */
115 for(iidx=0; iidx<nri; iidx++)
117 /* Load shift vector for this list */
118 i_shift_offset = DIM*shiftidx[iidx];
120 /* Load limits for loop over neighbors */
121 j_index_start = jindex[iidx];
122 j_index_end = jindex[iidx+1];
124 /* Get outer coordinate index */
126 i_coord_offset = DIM*inr;
128 /* Load i particle coords and add shift vector */
129 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
131 fix0 = _mm256_setzero_pd();
132 fiy0 = _mm256_setzero_pd();
133 fiz0 = _mm256_setzero_pd();
135 /* Load parameters for i particles */
136 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
138 /* Reset potential sums */
139 velecsum = _mm256_setzero_pd();
141 /* Start inner kernel loop */
142 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
145 /* Get j neighbor index, and coordinate index */
150 j_coord_offsetA = DIM*jnrA;
151 j_coord_offsetB = DIM*jnrB;
152 j_coord_offsetC = DIM*jnrC;
153 j_coord_offsetD = DIM*jnrD;
155 /* load j atom coordinates */
156 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
157 x+j_coord_offsetC,x+j_coord_offsetD,
160 /* Calculate displacement vector */
161 dx00 = _mm256_sub_pd(ix0,jx0);
162 dy00 = _mm256_sub_pd(iy0,jy0);
163 dz00 = _mm256_sub_pd(iz0,jz0);
165 /* Calculate squared distance and things based on it */
166 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
168 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
170 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
172 /* Load parameters for j particles */
173 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
174 charge+jnrC+0,charge+jnrD+0);
176 /**************************
177 * CALCULATE INTERACTIONS *
178 **************************/
180 /* Compute parameters for interactions between i and j atoms */
181 qq00 = _mm256_mul_pd(iq0,jq0);
183 /* REACTION-FIELD ELECTROSTATICS */
184 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
185 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
187 /* Update potential sum for this i atom from the interaction with this j atom. */
188 velecsum = _mm256_add_pd(velecsum,velec);
192 /* Calculate temporary vectorial force */
193 tx = _mm256_mul_pd(fscal,dx00);
194 ty = _mm256_mul_pd(fscal,dy00);
195 tz = _mm256_mul_pd(fscal,dz00);
197 /* Update vectorial force */
198 fix0 = _mm256_add_pd(fix0,tx);
199 fiy0 = _mm256_add_pd(fiy0,ty);
200 fiz0 = _mm256_add_pd(fiz0,tz);
202 fjptrA = f+j_coord_offsetA;
203 fjptrB = f+j_coord_offsetB;
204 fjptrC = f+j_coord_offsetC;
205 fjptrD = f+j_coord_offsetD;
206 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
208 /* Inner loop uses 32 flops */
214 /* Get j neighbor index, and coordinate index */
215 jnrlistA = jjnr[jidx];
216 jnrlistB = jjnr[jidx+1];
217 jnrlistC = jjnr[jidx+2];
218 jnrlistD = jjnr[jidx+3];
219 /* Sign of each element will be negative for non-real atoms.
220 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
221 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
223 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
225 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
226 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
227 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
229 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
230 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
231 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
232 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
233 j_coord_offsetA = DIM*jnrA;
234 j_coord_offsetB = DIM*jnrB;
235 j_coord_offsetC = DIM*jnrC;
236 j_coord_offsetD = DIM*jnrD;
238 /* load j atom coordinates */
239 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
240 x+j_coord_offsetC,x+j_coord_offsetD,
243 /* Calculate displacement vector */
244 dx00 = _mm256_sub_pd(ix0,jx0);
245 dy00 = _mm256_sub_pd(iy0,jy0);
246 dz00 = _mm256_sub_pd(iz0,jz0);
248 /* Calculate squared distance and things based on it */
249 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
251 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
253 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
255 /* Load parameters for j particles */
256 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
257 charge+jnrC+0,charge+jnrD+0);
259 /**************************
260 * CALCULATE INTERACTIONS *
261 **************************/
263 /* Compute parameters for interactions between i and j atoms */
264 qq00 = _mm256_mul_pd(iq0,jq0);
266 /* REACTION-FIELD ELECTROSTATICS */
267 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
268 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
270 /* Update potential sum for this i atom from the interaction with this j atom. */
271 velec = _mm256_andnot_pd(dummy_mask,velec);
272 velecsum = _mm256_add_pd(velecsum,velec);
276 fscal = _mm256_andnot_pd(dummy_mask,fscal);
278 /* Calculate temporary vectorial force */
279 tx = _mm256_mul_pd(fscal,dx00);
280 ty = _mm256_mul_pd(fscal,dy00);
281 tz = _mm256_mul_pd(fscal,dz00);
283 /* Update vectorial force */
284 fix0 = _mm256_add_pd(fix0,tx);
285 fiy0 = _mm256_add_pd(fiy0,ty);
286 fiz0 = _mm256_add_pd(fiz0,tz);
288 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
289 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
290 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
291 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
292 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
294 /* Inner loop uses 32 flops */
297 /* End of innermost loop */
299 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
300 f+i_coord_offset,fshift+i_shift_offset);
303 /* Update potential energies */
304 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
306 /* Increment number of inner iterations */
307 inneriter += j_index_end - j_index_start;
309 /* Outer loop uses 8 flops */
312 /* Increment number of outer iterations */
315 /* Update outer/inner flops */
317 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*32);
320 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwNone_GeomP1P1_F_avx_256_double
321 * Electrostatics interaction: ReactionField
322 * VdW interaction: None
323 * Geometry: Particle-Particle
324 * Calculate force/pot: Force
327 nb_kernel_ElecRF_VdwNone_GeomP1P1_F_avx_256_double
328 (t_nblist * gmx_restrict nlist,
329 rvec * gmx_restrict xx,
330 rvec * gmx_restrict ff,
331 t_forcerec * gmx_restrict fr,
332 t_mdatoms * gmx_restrict mdatoms,
333 nb_kernel_data_t * gmx_restrict kernel_data,
334 t_nrnb * gmx_restrict nrnb)
336 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
337 * just 0 for non-waters.
338 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
339 * jnr indices corresponding to data put in the four positions in the SIMD register.
341 int i_shift_offset,i_coord_offset,outeriter,inneriter;
342 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
343 int jnrA,jnrB,jnrC,jnrD;
344 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
345 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
346 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
347 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
349 real *shiftvec,*fshift,*x,*f;
350 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
352 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
353 real * vdwioffsetptr0;
354 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
355 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
356 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
357 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
358 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
360 __m256d dummy_mask,cutoff_mask;
361 __m128 tmpmask0,tmpmask1;
362 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
363 __m256d one = _mm256_set1_pd(1.0);
364 __m256d two = _mm256_set1_pd(2.0);
370 jindex = nlist->jindex;
372 shiftidx = nlist->shift;
374 shiftvec = fr->shift_vec[0];
375 fshift = fr->fshift[0];
376 facel = _mm256_set1_pd(fr->epsfac);
377 charge = mdatoms->chargeA;
378 krf = _mm256_set1_pd(fr->ic->k_rf);
379 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
380 crf = _mm256_set1_pd(fr->ic->c_rf);
382 /* Avoid stupid compiler warnings */
383 jnrA = jnrB = jnrC = jnrD = 0;
392 for(iidx=0;iidx<4*DIM;iidx++)
397 /* Start outer loop over neighborlists */
398 for(iidx=0; iidx<nri; iidx++)
400 /* Load shift vector for this list */
401 i_shift_offset = DIM*shiftidx[iidx];
403 /* Load limits for loop over neighbors */
404 j_index_start = jindex[iidx];
405 j_index_end = jindex[iidx+1];
407 /* Get outer coordinate index */
409 i_coord_offset = DIM*inr;
411 /* Load i particle coords and add shift vector */
412 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
414 fix0 = _mm256_setzero_pd();
415 fiy0 = _mm256_setzero_pd();
416 fiz0 = _mm256_setzero_pd();
418 /* Load parameters for i particles */
419 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
421 /* Start inner kernel loop */
422 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
425 /* Get j neighbor index, and coordinate index */
430 j_coord_offsetA = DIM*jnrA;
431 j_coord_offsetB = DIM*jnrB;
432 j_coord_offsetC = DIM*jnrC;
433 j_coord_offsetD = DIM*jnrD;
435 /* load j atom coordinates */
436 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
437 x+j_coord_offsetC,x+j_coord_offsetD,
440 /* Calculate displacement vector */
441 dx00 = _mm256_sub_pd(ix0,jx0);
442 dy00 = _mm256_sub_pd(iy0,jy0);
443 dz00 = _mm256_sub_pd(iz0,jz0);
445 /* Calculate squared distance and things based on it */
446 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
448 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
450 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
452 /* Load parameters for j particles */
453 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
454 charge+jnrC+0,charge+jnrD+0);
456 /**************************
457 * CALCULATE INTERACTIONS *
458 **************************/
460 /* Compute parameters for interactions between i and j atoms */
461 qq00 = _mm256_mul_pd(iq0,jq0);
463 /* REACTION-FIELD ELECTROSTATICS */
464 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
468 /* Calculate temporary vectorial force */
469 tx = _mm256_mul_pd(fscal,dx00);
470 ty = _mm256_mul_pd(fscal,dy00);
471 tz = _mm256_mul_pd(fscal,dz00);
473 /* Update vectorial force */
474 fix0 = _mm256_add_pd(fix0,tx);
475 fiy0 = _mm256_add_pd(fiy0,ty);
476 fiz0 = _mm256_add_pd(fiz0,tz);
478 fjptrA = f+j_coord_offsetA;
479 fjptrB = f+j_coord_offsetB;
480 fjptrC = f+j_coord_offsetC;
481 fjptrD = f+j_coord_offsetD;
482 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
484 /* Inner loop uses 27 flops */
490 /* Get j neighbor index, and coordinate index */
491 jnrlistA = jjnr[jidx];
492 jnrlistB = jjnr[jidx+1];
493 jnrlistC = jjnr[jidx+2];
494 jnrlistD = jjnr[jidx+3];
495 /* Sign of each element will be negative for non-real atoms.
496 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
497 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
499 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
501 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
502 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
503 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
505 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
506 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
507 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
508 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
509 j_coord_offsetA = DIM*jnrA;
510 j_coord_offsetB = DIM*jnrB;
511 j_coord_offsetC = DIM*jnrC;
512 j_coord_offsetD = DIM*jnrD;
514 /* load j atom coordinates */
515 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
516 x+j_coord_offsetC,x+j_coord_offsetD,
519 /* Calculate displacement vector */
520 dx00 = _mm256_sub_pd(ix0,jx0);
521 dy00 = _mm256_sub_pd(iy0,jy0);
522 dz00 = _mm256_sub_pd(iz0,jz0);
524 /* Calculate squared distance and things based on it */
525 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
527 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
529 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
531 /* Load parameters for j particles */
532 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
533 charge+jnrC+0,charge+jnrD+0);
535 /**************************
536 * CALCULATE INTERACTIONS *
537 **************************/
539 /* Compute parameters for interactions between i and j atoms */
540 qq00 = _mm256_mul_pd(iq0,jq0);
542 /* REACTION-FIELD ELECTROSTATICS */
543 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
547 fscal = _mm256_andnot_pd(dummy_mask,fscal);
549 /* Calculate temporary vectorial force */
550 tx = _mm256_mul_pd(fscal,dx00);
551 ty = _mm256_mul_pd(fscal,dy00);
552 tz = _mm256_mul_pd(fscal,dz00);
554 /* Update vectorial force */
555 fix0 = _mm256_add_pd(fix0,tx);
556 fiy0 = _mm256_add_pd(fiy0,ty);
557 fiz0 = _mm256_add_pd(fiz0,tz);
559 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
560 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
561 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
562 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
563 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
565 /* Inner loop uses 27 flops */
568 /* End of innermost loop */
570 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
571 f+i_coord_offset,fshift+i_shift_offset);
573 /* Increment number of inner iterations */
574 inneriter += j_index_end - j_index_start;
576 /* Outer loop uses 7 flops */
579 /* Increment number of outer iterations */
582 /* Update outer/inner flops */
584 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*27);