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_ElecRF_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_ElecRF_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 /* Avoid stupid compiler warnings */
98 jnrA = jnrB = jnrC = jnrD = 0;
107 for(iidx=0;iidx<4*DIM;iidx++)
112 /* Start outer loop over neighborlists */
113 for(iidx=0; iidx<nri; iidx++)
115 /* Load shift vector for this list */
116 i_shift_offset = DIM*shiftidx[iidx];
118 /* Load limits for loop over neighbors */
119 j_index_start = jindex[iidx];
120 j_index_end = jindex[iidx+1];
122 /* Get outer coordinate index */
124 i_coord_offset = DIM*inr;
126 /* Load i particle coords and add shift vector */
127 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
129 fix0 = _mm_setzero_ps();
130 fiy0 = _mm_setzero_ps();
131 fiz0 = _mm_setzero_ps();
133 /* Load parameters for i particles */
134 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
136 /* Reset potential sums */
137 velecsum = _mm_setzero_ps();
139 /* Start inner kernel loop */
140 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
143 /* Get j neighbor index, and coordinate index */
148 j_coord_offsetA = DIM*jnrA;
149 j_coord_offsetB = DIM*jnrB;
150 j_coord_offsetC = DIM*jnrC;
151 j_coord_offsetD = DIM*jnrD;
153 /* load j atom coordinates */
154 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
155 x+j_coord_offsetC,x+j_coord_offsetD,
158 /* Calculate displacement vector */
159 dx00 = _mm_sub_ps(ix0,jx0);
160 dy00 = _mm_sub_ps(iy0,jy0);
161 dz00 = _mm_sub_ps(iz0,jz0);
163 /* Calculate squared distance and things based on it */
164 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
166 rinv00 = gmx_mm_invsqrt_ps(rsq00);
168 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
170 /* Load parameters for j particles */
171 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
172 charge+jnrC+0,charge+jnrD+0);
174 /**************************
175 * CALCULATE INTERACTIONS *
176 **************************/
178 /* Compute parameters for interactions between i and j atoms */
179 qq00 = _mm_mul_ps(iq0,jq0);
181 /* REACTION-FIELD ELECTROSTATICS */
182 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
183 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
185 /* Update potential sum for this i atom from the interaction with this j atom. */
186 velecsum = _mm_add_ps(velecsum,velec);
190 /* Update vectorial force */
191 fix0 = _mm_macc_ps(dx00,fscal,fix0);
192 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
193 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
195 fjptrA = f+j_coord_offsetA;
196 fjptrB = f+j_coord_offsetB;
197 fjptrC = f+j_coord_offsetC;
198 fjptrD = f+j_coord_offsetD;
199 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
200 _mm_mul_ps(dx00,fscal),
201 _mm_mul_ps(dy00,fscal),
202 _mm_mul_ps(dz00,fscal));
204 /* Inner loop uses 35 flops */
210 /* Get j neighbor index, and coordinate index */
211 jnrlistA = jjnr[jidx];
212 jnrlistB = jjnr[jidx+1];
213 jnrlistC = jjnr[jidx+2];
214 jnrlistD = jjnr[jidx+3];
215 /* Sign of each element will be negative for non-real atoms.
216 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
217 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
219 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
220 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
221 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
222 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
223 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
224 j_coord_offsetA = DIM*jnrA;
225 j_coord_offsetB = DIM*jnrB;
226 j_coord_offsetC = DIM*jnrC;
227 j_coord_offsetD = DIM*jnrD;
229 /* load j atom coordinates */
230 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
231 x+j_coord_offsetC,x+j_coord_offsetD,
234 /* Calculate displacement vector */
235 dx00 = _mm_sub_ps(ix0,jx0);
236 dy00 = _mm_sub_ps(iy0,jy0);
237 dz00 = _mm_sub_ps(iz0,jz0);
239 /* Calculate squared distance and things based on it */
240 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
242 rinv00 = gmx_mm_invsqrt_ps(rsq00);
244 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
246 /* Load parameters for j particles */
247 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
248 charge+jnrC+0,charge+jnrD+0);
250 /**************************
251 * CALCULATE INTERACTIONS *
252 **************************/
254 /* Compute parameters for interactions between i and j atoms */
255 qq00 = _mm_mul_ps(iq0,jq0);
257 /* REACTION-FIELD ELECTROSTATICS */
258 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
259 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
261 /* Update potential sum for this i atom from the interaction with this j atom. */
262 velec = _mm_andnot_ps(dummy_mask,velec);
263 velecsum = _mm_add_ps(velecsum,velec);
267 fscal = _mm_andnot_ps(dummy_mask,fscal);
269 /* Update vectorial force */
270 fix0 = _mm_macc_ps(dx00,fscal,fix0);
271 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
272 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
274 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
275 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
276 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
277 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
278 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
279 _mm_mul_ps(dx00,fscal),
280 _mm_mul_ps(dy00,fscal),
281 _mm_mul_ps(dz00,fscal));
283 /* Inner loop uses 35 flops */
286 /* End of innermost loop */
288 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
289 f+i_coord_offset,fshift+i_shift_offset);
292 /* Update potential energies */
293 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
295 /* Increment number of inner iterations */
296 inneriter += j_index_end - j_index_start;
298 /* Outer loop uses 8 flops */
301 /* Increment number of outer iterations */
304 /* Update outer/inner flops */
306 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*35);
309 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwNone_GeomP1P1_F_avx_128_fma_single
310 * Electrostatics interaction: ReactionField
311 * VdW interaction: None
312 * Geometry: Particle-Particle
313 * Calculate force/pot: Force
316 nb_kernel_ElecRF_VdwNone_GeomP1P1_F_avx_128_fma_single
317 (t_nblist * gmx_restrict nlist,
318 rvec * gmx_restrict xx,
319 rvec * gmx_restrict ff,
320 t_forcerec * gmx_restrict fr,
321 t_mdatoms * gmx_restrict mdatoms,
322 nb_kernel_data_t * gmx_restrict kernel_data,
323 t_nrnb * gmx_restrict nrnb)
325 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
326 * just 0 for non-waters.
327 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
328 * jnr indices corresponding to data put in the four positions in the SIMD register.
330 int i_shift_offset,i_coord_offset,outeriter,inneriter;
331 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
332 int jnrA,jnrB,jnrC,jnrD;
333 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
334 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
335 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
337 real *shiftvec,*fshift,*x,*f;
338 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
340 __m128 fscal,rcutoff,rcutoff2,jidxall;
342 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
343 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
344 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
345 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
346 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
348 __m128 dummy_mask,cutoff_mask;
349 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
350 __m128 one = _mm_set1_ps(1.0);
351 __m128 two = _mm_set1_ps(2.0);
357 jindex = nlist->jindex;
359 shiftidx = nlist->shift;
361 shiftvec = fr->shift_vec[0];
362 fshift = fr->fshift[0];
363 facel = _mm_set1_ps(fr->epsfac);
364 charge = mdatoms->chargeA;
365 krf = _mm_set1_ps(fr->ic->k_rf);
366 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
367 crf = _mm_set1_ps(fr->ic->c_rf);
369 /* Avoid stupid compiler warnings */
370 jnrA = jnrB = jnrC = jnrD = 0;
379 for(iidx=0;iidx<4*DIM;iidx++)
384 /* Start outer loop over neighborlists */
385 for(iidx=0; iidx<nri; iidx++)
387 /* Load shift vector for this list */
388 i_shift_offset = DIM*shiftidx[iidx];
390 /* Load limits for loop over neighbors */
391 j_index_start = jindex[iidx];
392 j_index_end = jindex[iidx+1];
394 /* Get outer coordinate index */
396 i_coord_offset = DIM*inr;
398 /* Load i particle coords and add shift vector */
399 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
401 fix0 = _mm_setzero_ps();
402 fiy0 = _mm_setzero_ps();
403 fiz0 = _mm_setzero_ps();
405 /* Load parameters for i particles */
406 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
408 /* Start inner kernel loop */
409 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
412 /* Get j neighbor index, and coordinate index */
417 j_coord_offsetA = DIM*jnrA;
418 j_coord_offsetB = DIM*jnrB;
419 j_coord_offsetC = DIM*jnrC;
420 j_coord_offsetD = DIM*jnrD;
422 /* load j atom coordinates */
423 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
424 x+j_coord_offsetC,x+j_coord_offsetD,
427 /* Calculate displacement vector */
428 dx00 = _mm_sub_ps(ix0,jx0);
429 dy00 = _mm_sub_ps(iy0,jy0);
430 dz00 = _mm_sub_ps(iz0,jz0);
432 /* Calculate squared distance and things based on it */
433 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
435 rinv00 = gmx_mm_invsqrt_ps(rsq00);
437 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
439 /* Load parameters for j particles */
440 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
441 charge+jnrC+0,charge+jnrD+0);
443 /**************************
444 * CALCULATE INTERACTIONS *
445 **************************/
447 /* Compute parameters for interactions between i and j atoms */
448 qq00 = _mm_mul_ps(iq0,jq0);
450 /* REACTION-FIELD ELECTROSTATICS */
451 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
455 /* Update vectorial force */
456 fix0 = _mm_macc_ps(dx00,fscal,fix0);
457 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
458 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
460 fjptrA = f+j_coord_offsetA;
461 fjptrB = f+j_coord_offsetB;
462 fjptrC = f+j_coord_offsetC;
463 fjptrD = f+j_coord_offsetD;
464 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
465 _mm_mul_ps(dx00,fscal),
466 _mm_mul_ps(dy00,fscal),
467 _mm_mul_ps(dz00,fscal));
469 /* Inner loop uses 30 flops */
475 /* Get j neighbor index, and coordinate index */
476 jnrlistA = jjnr[jidx];
477 jnrlistB = jjnr[jidx+1];
478 jnrlistC = jjnr[jidx+2];
479 jnrlistD = jjnr[jidx+3];
480 /* Sign of each element will be negative for non-real atoms.
481 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
482 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
484 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
485 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
486 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
487 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
488 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
489 j_coord_offsetA = DIM*jnrA;
490 j_coord_offsetB = DIM*jnrB;
491 j_coord_offsetC = DIM*jnrC;
492 j_coord_offsetD = DIM*jnrD;
494 /* load j atom coordinates */
495 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
496 x+j_coord_offsetC,x+j_coord_offsetD,
499 /* Calculate displacement vector */
500 dx00 = _mm_sub_ps(ix0,jx0);
501 dy00 = _mm_sub_ps(iy0,jy0);
502 dz00 = _mm_sub_ps(iz0,jz0);
504 /* Calculate squared distance and things based on it */
505 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
507 rinv00 = gmx_mm_invsqrt_ps(rsq00);
509 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
511 /* Load parameters for j particles */
512 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
513 charge+jnrC+0,charge+jnrD+0);
515 /**************************
516 * CALCULATE INTERACTIONS *
517 **************************/
519 /* Compute parameters for interactions between i and j atoms */
520 qq00 = _mm_mul_ps(iq0,jq0);
522 /* REACTION-FIELD ELECTROSTATICS */
523 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
527 fscal = _mm_andnot_ps(dummy_mask,fscal);
529 /* Update vectorial force */
530 fix0 = _mm_macc_ps(dx00,fscal,fix0);
531 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
532 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
534 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
535 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
536 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
537 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
538 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
539 _mm_mul_ps(dx00,fscal),
540 _mm_mul_ps(dy00,fscal),
541 _mm_mul_ps(dz00,fscal));
543 /* Inner loop uses 30 flops */
546 /* End of innermost loop */
548 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
549 f+i_coord_offset,fshift+i_shift_offset);
551 /* Increment number of inner iterations */
552 inneriter += j_index_end - j_index_start;
554 /* Outer loop uses 7 flops */
557 /* Increment number of outer iterations */
560 /* Update outer/inner flops */
562 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*30);