2 * Note: this file was generated by the Gromacs avx_128_fma_single 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_128_fma_single.h"
34 #include "kernelutil_x86_avx_128_fma_single.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwNone_GeomP1P1_VF_avx_128_fma_single
38 * Electrostatics interaction: ReactionField
39 * VdW interaction: None
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecRFCut_VdwNone_GeomP1P1_VF_avx_128_fma_single
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_128, 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 j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
63 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
65 real *shiftvec,*fshift,*x,*f;
66 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
68 __m128 fscal,rcutoff,rcutoff2,jidxall;
70 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
71 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
72 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
73 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
74 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
76 __m128 dummy_mask,cutoff_mask;
77 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
78 __m128 one = _mm_set1_ps(1.0);
79 __m128 two = _mm_set1_ps(2.0);
85 jindex = nlist->jindex;
87 shiftidx = nlist->shift;
89 shiftvec = fr->shift_vec[0];
90 fshift = fr->fshift[0];
91 facel = _mm_set1_ps(fr->epsfac);
92 charge = mdatoms->chargeA;
93 krf = _mm_set1_ps(fr->ic->k_rf);
94 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
95 crf = _mm_set1_ps(fr->ic->c_rf);
97 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
98 rcutoff_scalar = fr->rcoulomb;
99 rcutoff = _mm_set1_ps(rcutoff_scalar);
100 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
102 /* Avoid stupid compiler warnings */
103 jnrA = jnrB = jnrC = jnrD = 0;
112 for(iidx=0;iidx<4*DIM;iidx++)
117 /* Start outer loop over neighborlists */
118 for(iidx=0; iidx<nri; iidx++)
120 /* Load shift vector for this list */
121 i_shift_offset = DIM*shiftidx[iidx];
123 /* Load limits for loop over neighbors */
124 j_index_start = jindex[iidx];
125 j_index_end = jindex[iidx+1];
127 /* Get outer coordinate index */
129 i_coord_offset = DIM*inr;
131 /* Load i particle coords and add shift vector */
132 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
134 fix0 = _mm_setzero_ps();
135 fiy0 = _mm_setzero_ps();
136 fiz0 = _mm_setzero_ps();
138 /* Load parameters for i particles */
139 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
141 /* Reset potential sums */
142 velecsum = _mm_setzero_ps();
144 /* Start inner kernel loop */
145 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
148 /* Get j neighbor index, and coordinate index */
153 j_coord_offsetA = DIM*jnrA;
154 j_coord_offsetB = DIM*jnrB;
155 j_coord_offsetC = DIM*jnrC;
156 j_coord_offsetD = DIM*jnrD;
158 /* load j atom coordinates */
159 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
160 x+j_coord_offsetC,x+j_coord_offsetD,
163 /* Calculate displacement vector */
164 dx00 = _mm_sub_ps(ix0,jx0);
165 dy00 = _mm_sub_ps(iy0,jy0);
166 dz00 = _mm_sub_ps(iz0,jz0);
168 /* Calculate squared distance and things based on it */
169 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
171 rinv00 = gmx_mm_invsqrt_ps(rsq00);
173 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
175 /* Load parameters for j particles */
176 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
177 charge+jnrC+0,charge+jnrD+0);
179 /**************************
180 * CALCULATE INTERACTIONS *
181 **************************/
183 if (gmx_mm_any_lt(rsq00,rcutoff2))
186 /* Compute parameters for interactions between i and j atoms */
187 qq00 = _mm_mul_ps(iq0,jq0);
189 /* REACTION-FIELD ELECTROSTATICS */
190 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
191 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
193 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
195 /* Update potential sum for this i atom from the interaction with this j atom. */
196 velec = _mm_and_ps(velec,cutoff_mask);
197 velecsum = _mm_add_ps(velecsum,velec);
201 fscal = _mm_and_ps(fscal,cutoff_mask);
203 /* Update vectorial force */
204 fix0 = _mm_macc_ps(dx00,fscal,fix0);
205 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
206 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
208 fjptrA = f+j_coord_offsetA;
209 fjptrB = f+j_coord_offsetB;
210 fjptrC = f+j_coord_offsetC;
211 fjptrD = f+j_coord_offsetD;
212 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
213 _mm_mul_ps(dx00,fscal),
214 _mm_mul_ps(dy00,fscal),
215 _mm_mul_ps(dz00,fscal));
219 /* Inner loop uses 39 flops */
225 /* Get j neighbor index, and coordinate index */
226 jnrlistA = jjnr[jidx];
227 jnrlistB = jjnr[jidx+1];
228 jnrlistC = jjnr[jidx+2];
229 jnrlistD = jjnr[jidx+3];
230 /* Sign of each element will be negative for non-real atoms.
231 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
232 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
234 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
235 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
236 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
237 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
238 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
239 j_coord_offsetA = DIM*jnrA;
240 j_coord_offsetB = DIM*jnrB;
241 j_coord_offsetC = DIM*jnrC;
242 j_coord_offsetD = DIM*jnrD;
244 /* load j atom coordinates */
245 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
246 x+j_coord_offsetC,x+j_coord_offsetD,
249 /* Calculate displacement vector */
250 dx00 = _mm_sub_ps(ix0,jx0);
251 dy00 = _mm_sub_ps(iy0,jy0);
252 dz00 = _mm_sub_ps(iz0,jz0);
254 /* Calculate squared distance and things based on it */
255 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
257 rinv00 = gmx_mm_invsqrt_ps(rsq00);
259 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
261 /* Load parameters for j particles */
262 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
263 charge+jnrC+0,charge+jnrD+0);
265 /**************************
266 * CALCULATE INTERACTIONS *
267 **************************/
269 if (gmx_mm_any_lt(rsq00,rcutoff2))
272 /* Compute parameters for interactions between i and j atoms */
273 qq00 = _mm_mul_ps(iq0,jq0);
275 /* REACTION-FIELD ELECTROSTATICS */
276 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
277 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
279 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
281 /* Update potential sum for this i atom from the interaction with this j atom. */
282 velec = _mm_and_ps(velec,cutoff_mask);
283 velec = _mm_andnot_ps(dummy_mask,velec);
284 velecsum = _mm_add_ps(velecsum,velec);
288 fscal = _mm_and_ps(fscal,cutoff_mask);
290 fscal = _mm_andnot_ps(dummy_mask,fscal);
292 /* Update vectorial force */
293 fix0 = _mm_macc_ps(dx00,fscal,fix0);
294 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
295 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
297 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
298 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
299 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
300 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
301 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
302 _mm_mul_ps(dx00,fscal),
303 _mm_mul_ps(dy00,fscal),
304 _mm_mul_ps(dz00,fscal));
308 /* Inner loop uses 39 flops */
311 /* End of innermost loop */
313 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
314 f+i_coord_offset,fshift+i_shift_offset);
317 /* Update potential energies */
318 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
320 /* Increment number of inner iterations */
321 inneriter += j_index_end - j_index_start;
323 /* Outer loop uses 8 flops */
326 /* Increment number of outer iterations */
329 /* Update outer/inner flops */
331 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*39);
334 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwNone_GeomP1P1_F_avx_128_fma_single
335 * Electrostatics interaction: ReactionField
336 * VdW interaction: None
337 * Geometry: Particle-Particle
338 * Calculate force/pot: Force
341 nb_kernel_ElecRFCut_VdwNone_GeomP1P1_F_avx_128_fma_single
342 (t_nblist * gmx_restrict nlist,
343 rvec * gmx_restrict xx,
344 rvec * gmx_restrict ff,
345 t_forcerec * gmx_restrict fr,
346 t_mdatoms * gmx_restrict mdatoms,
347 nb_kernel_data_t * gmx_restrict kernel_data,
348 t_nrnb * gmx_restrict nrnb)
350 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
351 * just 0 for non-waters.
352 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
353 * jnr indices corresponding to data put in the four positions in the SIMD register.
355 int i_shift_offset,i_coord_offset,outeriter,inneriter;
356 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
357 int jnrA,jnrB,jnrC,jnrD;
358 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
359 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
360 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
362 real *shiftvec,*fshift,*x,*f;
363 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
365 __m128 fscal,rcutoff,rcutoff2,jidxall;
367 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
368 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
369 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
370 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
371 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
373 __m128 dummy_mask,cutoff_mask;
374 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
375 __m128 one = _mm_set1_ps(1.0);
376 __m128 two = _mm_set1_ps(2.0);
382 jindex = nlist->jindex;
384 shiftidx = nlist->shift;
386 shiftvec = fr->shift_vec[0];
387 fshift = fr->fshift[0];
388 facel = _mm_set1_ps(fr->epsfac);
389 charge = mdatoms->chargeA;
390 krf = _mm_set1_ps(fr->ic->k_rf);
391 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
392 crf = _mm_set1_ps(fr->ic->c_rf);
394 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
395 rcutoff_scalar = fr->rcoulomb;
396 rcutoff = _mm_set1_ps(rcutoff_scalar);
397 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
399 /* Avoid stupid compiler warnings */
400 jnrA = jnrB = jnrC = jnrD = 0;
409 for(iidx=0;iidx<4*DIM;iidx++)
414 /* Start outer loop over neighborlists */
415 for(iidx=0; iidx<nri; iidx++)
417 /* Load shift vector for this list */
418 i_shift_offset = DIM*shiftidx[iidx];
420 /* Load limits for loop over neighbors */
421 j_index_start = jindex[iidx];
422 j_index_end = jindex[iidx+1];
424 /* Get outer coordinate index */
426 i_coord_offset = DIM*inr;
428 /* Load i particle coords and add shift vector */
429 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
431 fix0 = _mm_setzero_ps();
432 fiy0 = _mm_setzero_ps();
433 fiz0 = _mm_setzero_ps();
435 /* Load parameters for i particles */
436 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
438 /* Start inner kernel loop */
439 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
442 /* Get j neighbor index, and coordinate index */
447 j_coord_offsetA = DIM*jnrA;
448 j_coord_offsetB = DIM*jnrB;
449 j_coord_offsetC = DIM*jnrC;
450 j_coord_offsetD = DIM*jnrD;
452 /* load j atom coordinates */
453 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
454 x+j_coord_offsetC,x+j_coord_offsetD,
457 /* Calculate displacement vector */
458 dx00 = _mm_sub_ps(ix0,jx0);
459 dy00 = _mm_sub_ps(iy0,jy0);
460 dz00 = _mm_sub_ps(iz0,jz0);
462 /* Calculate squared distance and things based on it */
463 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
465 rinv00 = gmx_mm_invsqrt_ps(rsq00);
467 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
469 /* Load parameters for j particles */
470 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
471 charge+jnrC+0,charge+jnrD+0);
473 /**************************
474 * CALCULATE INTERACTIONS *
475 **************************/
477 if (gmx_mm_any_lt(rsq00,rcutoff2))
480 /* Compute parameters for interactions between i and j atoms */
481 qq00 = _mm_mul_ps(iq0,jq0);
483 /* REACTION-FIELD ELECTROSTATICS */
484 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
486 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
490 fscal = _mm_and_ps(fscal,cutoff_mask);
492 /* Update vectorial force */
493 fix0 = _mm_macc_ps(dx00,fscal,fix0);
494 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
495 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
497 fjptrA = f+j_coord_offsetA;
498 fjptrB = f+j_coord_offsetB;
499 fjptrC = f+j_coord_offsetC;
500 fjptrD = f+j_coord_offsetD;
501 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
502 _mm_mul_ps(dx00,fscal),
503 _mm_mul_ps(dy00,fscal),
504 _mm_mul_ps(dz00,fscal));
508 /* Inner loop uses 33 flops */
514 /* Get j neighbor index, and coordinate index */
515 jnrlistA = jjnr[jidx];
516 jnrlistB = jjnr[jidx+1];
517 jnrlistC = jjnr[jidx+2];
518 jnrlistD = jjnr[jidx+3];
519 /* Sign of each element will be negative for non-real atoms.
520 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
521 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
523 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
524 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
525 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
526 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
527 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
528 j_coord_offsetA = DIM*jnrA;
529 j_coord_offsetB = DIM*jnrB;
530 j_coord_offsetC = DIM*jnrC;
531 j_coord_offsetD = DIM*jnrD;
533 /* load j atom coordinates */
534 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
535 x+j_coord_offsetC,x+j_coord_offsetD,
538 /* Calculate displacement vector */
539 dx00 = _mm_sub_ps(ix0,jx0);
540 dy00 = _mm_sub_ps(iy0,jy0);
541 dz00 = _mm_sub_ps(iz0,jz0);
543 /* Calculate squared distance and things based on it */
544 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
546 rinv00 = gmx_mm_invsqrt_ps(rsq00);
548 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
550 /* Load parameters for j particles */
551 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
552 charge+jnrC+0,charge+jnrD+0);
554 /**************************
555 * CALCULATE INTERACTIONS *
556 **************************/
558 if (gmx_mm_any_lt(rsq00,rcutoff2))
561 /* Compute parameters for interactions between i and j atoms */
562 qq00 = _mm_mul_ps(iq0,jq0);
564 /* REACTION-FIELD ELECTROSTATICS */
565 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
567 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
571 fscal = _mm_and_ps(fscal,cutoff_mask);
573 fscal = _mm_andnot_ps(dummy_mask,fscal);
575 /* Update vectorial force */
576 fix0 = _mm_macc_ps(dx00,fscal,fix0);
577 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
578 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
580 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
581 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
582 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
583 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
584 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
585 _mm_mul_ps(dx00,fscal),
586 _mm_mul_ps(dy00,fscal),
587 _mm_mul_ps(dz00,fscal));
591 /* Inner loop uses 33 flops */
594 /* End of innermost loop */
596 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
597 f+i_coord_offset,fshift+i_shift_offset);
599 /* Increment number of inner iterations */
600 inneriter += j_index_end - j_index_start;
602 /* Outer loop uses 7 flops */
605 /* Increment number of outer iterations */
608 /* Update outer/inner flops */
610 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*33);