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_VdwLJ_GeomW3P1_VF_avx_128_fma_single
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_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;
72 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
74 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
75 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
76 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
77 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
78 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
79 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
80 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
83 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
86 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
87 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
88 __m128 dummy_mask,cutoff_mask;
89 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
90 __m128 one = _mm_set1_ps(1.0);
91 __m128 two = _mm_set1_ps(2.0);
97 jindex = nlist->jindex;
99 shiftidx = nlist->shift;
101 shiftvec = fr->shift_vec[0];
102 fshift = fr->fshift[0];
103 facel = _mm_set1_ps(fr->epsfac);
104 charge = mdatoms->chargeA;
105 krf = _mm_set1_ps(fr->ic->k_rf);
106 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
107 crf = _mm_set1_ps(fr->ic->c_rf);
108 nvdwtype = fr->ntype;
110 vdwtype = mdatoms->typeA;
112 /* Setup water-specific parameters */
113 inr = nlist->iinr[0];
114 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
115 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
116 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
117 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
119 /* Avoid stupid compiler warnings */
120 jnrA = jnrB = jnrC = jnrD = 0;
129 for(iidx=0;iidx<4*DIM;iidx++)
134 /* Start outer loop over neighborlists */
135 for(iidx=0; iidx<nri; iidx++)
137 /* Load shift vector for this list */
138 i_shift_offset = DIM*shiftidx[iidx];
140 /* Load limits for loop over neighbors */
141 j_index_start = jindex[iidx];
142 j_index_end = jindex[iidx+1];
144 /* Get outer coordinate index */
146 i_coord_offset = DIM*inr;
148 /* Load i particle coords and add shift vector */
149 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
150 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
152 fix0 = _mm_setzero_ps();
153 fiy0 = _mm_setzero_ps();
154 fiz0 = _mm_setzero_ps();
155 fix1 = _mm_setzero_ps();
156 fiy1 = _mm_setzero_ps();
157 fiz1 = _mm_setzero_ps();
158 fix2 = _mm_setzero_ps();
159 fiy2 = _mm_setzero_ps();
160 fiz2 = _mm_setzero_ps();
162 /* Reset potential sums */
163 velecsum = _mm_setzero_ps();
164 vvdwsum = _mm_setzero_ps();
166 /* Start inner kernel loop */
167 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
170 /* Get j neighbor index, and coordinate index */
175 j_coord_offsetA = DIM*jnrA;
176 j_coord_offsetB = DIM*jnrB;
177 j_coord_offsetC = DIM*jnrC;
178 j_coord_offsetD = DIM*jnrD;
180 /* load j atom coordinates */
181 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
182 x+j_coord_offsetC,x+j_coord_offsetD,
185 /* Calculate displacement vector */
186 dx00 = _mm_sub_ps(ix0,jx0);
187 dy00 = _mm_sub_ps(iy0,jy0);
188 dz00 = _mm_sub_ps(iz0,jz0);
189 dx10 = _mm_sub_ps(ix1,jx0);
190 dy10 = _mm_sub_ps(iy1,jy0);
191 dz10 = _mm_sub_ps(iz1,jz0);
192 dx20 = _mm_sub_ps(ix2,jx0);
193 dy20 = _mm_sub_ps(iy2,jy0);
194 dz20 = _mm_sub_ps(iz2,jz0);
196 /* Calculate squared distance and things based on it */
197 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
198 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
199 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
201 rinv00 = gmx_mm_invsqrt_ps(rsq00);
202 rinv10 = gmx_mm_invsqrt_ps(rsq10);
203 rinv20 = gmx_mm_invsqrt_ps(rsq20);
205 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
206 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
207 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
209 /* Load parameters for j particles */
210 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
211 charge+jnrC+0,charge+jnrD+0);
212 vdwjidx0A = 2*vdwtype[jnrA+0];
213 vdwjidx0B = 2*vdwtype[jnrB+0];
214 vdwjidx0C = 2*vdwtype[jnrC+0];
215 vdwjidx0D = 2*vdwtype[jnrD+0];
217 fjx0 = _mm_setzero_ps();
218 fjy0 = _mm_setzero_ps();
219 fjz0 = _mm_setzero_ps();
221 /**************************
222 * CALCULATE INTERACTIONS *
223 **************************/
225 /* Compute parameters for interactions between i and j atoms */
226 qq00 = _mm_mul_ps(iq0,jq0);
227 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
228 vdwparam+vdwioffset0+vdwjidx0B,
229 vdwparam+vdwioffset0+vdwjidx0C,
230 vdwparam+vdwioffset0+vdwjidx0D,
233 /* REACTION-FIELD ELECTROSTATICS */
234 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
235 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
237 /* LENNARD-JONES DISPERSION/REPULSION */
239 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
240 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
241 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
242 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
243 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
245 /* Update potential sum for this i atom from the interaction with this j atom. */
246 velecsum = _mm_add_ps(velecsum,velec);
247 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
249 fscal = _mm_add_ps(felec,fvdw);
251 /* Update vectorial force */
252 fix0 = _mm_macc_ps(dx00,fscal,fix0);
253 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
254 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
256 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
257 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
258 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
260 /**************************
261 * CALCULATE INTERACTIONS *
262 **************************/
264 /* Compute parameters for interactions between i and j atoms */
265 qq10 = _mm_mul_ps(iq1,jq0);
267 /* REACTION-FIELD ELECTROSTATICS */
268 velec = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
269 felec = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
271 /* Update potential sum for this i atom from the interaction with this j atom. */
272 velecsum = _mm_add_ps(velecsum,velec);
276 /* Update vectorial force */
277 fix1 = _mm_macc_ps(dx10,fscal,fix1);
278 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
279 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
281 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
282 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
283 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
285 /**************************
286 * CALCULATE INTERACTIONS *
287 **************************/
289 /* Compute parameters for interactions between i and j atoms */
290 qq20 = _mm_mul_ps(iq2,jq0);
292 /* REACTION-FIELD ELECTROSTATICS */
293 velec = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
294 felec = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
296 /* Update potential sum for this i atom from the interaction with this j atom. */
297 velecsum = _mm_add_ps(velecsum,velec);
301 /* Update vectorial force */
302 fix2 = _mm_macc_ps(dx20,fscal,fix2);
303 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
304 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
306 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
307 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
308 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
310 fjptrA = f+j_coord_offsetA;
311 fjptrB = f+j_coord_offsetB;
312 fjptrC = f+j_coord_offsetC;
313 fjptrD = f+j_coord_offsetD;
315 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
317 /* Inner loop uses 117 flops */
323 /* Get j neighbor index, and coordinate index */
324 jnrlistA = jjnr[jidx];
325 jnrlistB = jjnr[jidx+1];
326 jnrlistC = jjnr[jidx+2];
327 jnrlistD = jjnr[jidx+3];
328 /* Sign of each element will be negative for non-real atoms.
329 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
330 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
332 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
333 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
334 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
335 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
336 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
337 j_coord_offsetA = DIM*jnrA;
338 j_coord_offsetB = DIM*jnrB;
339 j_coord_offsetC = DIM*jnrC;
340 j_coord_offsetD = DIM*jnrD;
342 /* load j atom coordinates */
343 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
344 x+j_coord_offsetC,x+j_coord_offsetD,
347 /* Calculate displacement vector */
348 dx00 = _mm_sub_ps(ix0,jx0);
349 dy00 = _mm_sub_ps(iy0,jy0);
350 dz00 = _mm_sub_ps(iz0,jz0);
351 dx10 = _mm_sub_ps(ix1,jx0);
352 dy10 = _mm_sub_ps(iy1,jy0);
353 dz10 = _mm_sub_ps(iz1,jz0);
354 dx20 = _mm_sub_ps(ix2,jx0);
355 dy20 = _mm_sub_ps(iy2,jy0);
356 dz20 = _mm_sub_ps(iz2,jz0);
358 /* Calculate squared distance and things based on it */
359 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
360 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
361 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
363 rinv00 = gmx_mm_invsqrt_ps(rsq00);
364 rinv10 = gmx_mm_invsqrt_ps(rsq10);
365 rinv20 = gmx_mm_invsqrt_ps(rsq20);
367 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
368 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
369 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
371 /* Load parameters for j particles */
372 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
373 charge+jnrC+0,charge+jnrD+0);
374 vdwjidx0A = 2*vdwtype[jnrA+0];
375 vdwjidx0B = 2*vdwtype[jnrB+0];
376 vdwjidx0C = 2*vdwtype[jnrC+0];
377 vdwjidx0D = 2*vdwtype[jnrD+0];
379 fjx0 = _mm_setzero_ps();
380 fjy0 = _mm_setzero_ps();
381 fjz0 = _mm_setzero_ps();
383 /**************************
384 * CALCULATE INTERACTIONS *
385 **************************/
387 /* Compute parameters for interactions between i and j atoms */
388 qq00 = _mm_mul_ps(iq0,jq0);
389 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
390 vdwparam+vdwioffset0+vdwjidx0B,
391 vdwparam+vdwioffset0+vdwjidx0C,
392 vdwparam+vdwioffset0+vdwjidx0D,
395 /* REACTION-FIELD ELECTROSTATICS */
396 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
397 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
399 /* LENNARD-JONES DISPERSION/REPULSION */
401 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
402 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
403 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
404 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
405 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
407 /* Update potential sum for this i atom from the interaction with this j atom. */
408 velec = _mm_andnot_ps(dummy_mask,velec);
409 velecsum = _mm_add_ps(velecsum,velec);
410 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
411 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
413 fscal = _mm_add_ps(felec,fvdw);
415 fscal = _mm_andnot_ps(dummy_mask,fscal);
417 /* Update vectorial force */
418 fix0 = _mm_macc_ps(dx00,fscal,fix0);
419 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
420 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
422 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
423 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
424 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
426 /**************************
427 * CALCULATE INTERACTIONS *
428 **************************/
430 /* Compute parameters for interactions between i and j atoms */
431 qq10 = _mm_mul_ps(iq1,jq0);
433 /* REACTION-FIELD ELECTROSTATICS */
434 velec = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
435 felec = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
437 /* Update potential sum for this i atom from the interaction with this j atom. */
438 velec = _mm_andnot_ps(dummy_mask,velec);
439 velecsum = _mm_add_ps(velecsum,velec);
443 fscal = _mm_andnot_ps(dummy_mask,fscal);
445 /* Update vectorial force */
446 fix1 = _mm_macc_ps(dx10,fscal,fix1);
447 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
448 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
450 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
451 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
452 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
454 /**************************
455 * CALCULATE INTERACTIONS *
456 **************************/
458 /* Compute parameters for interactions between i and j atoms */
459 qq20 = _mm_mul_ps(iq2,jq0);
461 /* REACTION-FIELD ELECTROSTATICS */
462 velec = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
463 felec = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
465 /* Update potential sum for this i atom from the interaction with this j atom. */
466 velec = _mm_andnot_ps(dummy_mask,velec);
467 velecsum = _mm_add_ps(velecsum,velec);
471 fscal = _mm_andnot_ps(dummy_mask,fscal);
473 /* Update vectorial force */
474 fix2 = _mm_macc_ps(dx20,fscal,fix2);
475 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
476 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
478 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
479 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
480 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
482 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
483 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
484 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
485 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
487 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
489 /* Inner loop uses 117 flops */
492 /* End of innermost loop */
494 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
495 f+i_coord_offset,fshift+i_shift_offset);
498 /* Update potential energies */
499 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
500 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
502 /* Increment number of inner iterations */
503 inneriter += j_index_end - j_index_start;
505 /* Outer loop uses 20 flops */
508 /* Increment number of outer iterations */
511 /* Update outer/inner flops */
513 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*117);
516 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_avx_128_fma_single
517 * Electrostatics interaction: ReactionField
518 * VdW interaction: LennardJones
519 * Geometry: Water3-Particle
520 * Calculate force/pot: Force
523 nb_kernel_ElecRF_VdwLJ_GeomW3P1_F_avx_128_fma_single
524 (t_nblist * gmx_restrict nlist,
525 rvec * gmx_restrict xx,
526 rvec * gmx_restrict ff,
527 t_forcerec * gmx_restrict fr,
528 t_mdatoms * gmx_restrict mdatoms,
529 nb_kernel_data_t * gmx_restrict kernel_data,
530 t_nrnb * gmx_restrict nrnb)
532 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
533 * just 0 for non-waters.
534 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
535 * jnr indices corresponding to data put in the four positions in the SIMD register.
537 int i_shift_offset,i_coord_offset,outeriter,inneriter;
538 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
539 int jnrA,jnrB,jnrC,jnrD;
540 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
541 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
542 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
544 real *shiftvec,*fshift,*x,*f;
545 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
547 __m128 fscal,rcutoff,rcutoff2,jidxall;
549 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
551 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
553 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
554 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
555 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
556 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
557 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
558 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
559 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
562 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
565 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
566 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
567 __m128 dummy_mask,cutoff_mask;
568 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
569 __m128 one = _mm_set1_ps(1.0);
570 __m128 two = _mm_set1_ps(2.0);
576 jindex = nlist->jindex;
578 shiftidx = nlist->shift;
580 shiftvec = fr->shift_vec[0];
581 fshift = fr->fshift[0];
582 facel = _mm_set1_ps(fr->epsfac);
583 charge = mdatoms->chargeA;
584 krf = _mm_set1_ps(fr->ic->k_rf);
585 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
586 crf = _mm_set1_ps(fr->ic->c_rf);
587 nvdwtype = fr->ntype;
589 vdwtype = mdatoms->typeA;
591 /* Setup water-specific parameters */
592 inr = nlist->iinr[0];
593 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
594 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
595 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
596 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
598 /* Avoid stupid compiler warnings */
599 jnrA = jnrB = jnrC = jnrD = 0;
608 for(iidx=0;iidx<4*DIM;iidx++)
613 /* Start outer loop over neighborlists */
614 for(iidx=0; iidx<nri; iidx++)
616 /* Load shift vector for this list */
617 i_shift_offset = DIM*shiftidx[iidx];
619 /* Load limits for loop over neighbors */
620 j_index_start = jindex[iidx];
621 j_index_end = jindex[iidx+1];
623 /* Get outer coordinate index */
625 i_coord_offset = DIM*inr;
627 /* Load i particle coords and add shift vector */
628 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
629 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
631 fix0 = _mm_setzero_ps();
632 fiy0 = _mm_setzero_ps();
633 fiz0 = _mm_setzero_ps();
634 fix1 = _mm_setzero_ps();
635 fiy1 = _mm_setzero_ps();
636 fiz1 = _mm_setzero_ps();
637 fix2 = _mm_setzero_ps();
638 fiy2 = _mm_setzero_ps();
639 fiz2 = _mm_setzero_ps();
641 /* Start inner kernel loop */
642 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
645 /* Get j neighbor index, and coordinate index */
650 j_coord_offsetA = DIM*jnrA;
651 j_coord_offsetB = DIM*jnrB;
652 j_coord_offsetC = DIM*jnrC;
653 j_coord_offsetD = DIM*jnrD;
655 /* load j atom coordinates */
656 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
657 x+j_coord_offsetC,x+j_coord_offsetD,
660 /* Calculate displacement vector */
661 dx00 = _mm_sub_ps(ix0,jx0);
662 dy00 = _mm_sub_ps(iy0,jy0);
663 dz00 = _mm_sub_ps(iz0,jz0);
664 dx10 = _mm_sub_ps(ix1,jx0);
665 dy10 = _mm_sub_ps(iy1,jy0);
666 dz10 = _mm_sub_ps(iz1,jz0);
667 dx20 = _mm_sub_ps(ix2,jx0);
668 dy20 = _mm_sub_ps(iy2,jy0);
669 dz20 = _mm_sub_ps(iz2,jz0);
671 /* Calculate squared distance and things based on it */
672 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
673 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
674 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
676 rinv00 = gmx_mm_invsqrt_ps(rsq00);
677 rinv10 = gmx_mm_invsqrt_ps(rsq10);
678 rinv20 = gmx_mm_invsqrt_ps(rsq20);
680 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
681 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
682 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
684 /* Load parameters for j particles */
685 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
686 charge+jnrC+0,charge+jnrD+0);
687 vdwjidx0A = 2*vdwtype[jnrA+0];
688 vdwjidx0B = 2*vdwtype[jnrB+0];
689 vdwjidx0C = 2*vdwtype[jnrC+0];
690 vdwjidx0D = 2*vdwtype[jnrD+0];
692 fjx0 = _mm_setzero_ps();
693 fjy0 = _mm_setzero_ps();
694 fjz0 = _mm_setzero_ps();
696 /**************************
697 * CALCULATE INTERACTIONS *
698 **************************/
700 /* Compute parameters for interactions between i and j atoms */
701 qq00 = _mm_mul_ps(iq0,jq0);
702 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
703 vdwparam+vdwioffset0+vdwjidx0B,
704 vdwparam+vdwioffset0+vdwjidx0C,
705 vdwparam+vdwioffset0+vdwjidx0D,
708 /* REACTION-FIELD ELECTROSTATICS */
709 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
711 /* LENNARD-JONES DISPERSION/REPULSION */
713 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
714 fvdw = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
716 fscal = _mm_add_ps(felec,fvdw);
718 /* Update vectorial force */
719 fix0 = _mm_macc_ps(dx00,fscal,fix0);
720 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
721 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
723 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
724 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
725 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
727 /**************************
728 * CALCULATE INTERACTIONS *
729 **************************/
731 /* Compute parameters for interactions between i and j atoms */
732 qq10 = _mm_mul_ps(iq1,jq0);
734 /* REACTION-FIELD ELECTROSTATICS */
735 felec = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
739 /* Update vectorial force */
740 fix1 = _mm_macc_ps(dx10,fscal,fix1);
741 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
742 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
744 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
745 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
746 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
748 /**************************
749 * CALCULATE INTERACTIONS *
750 **************************/
752 /* Compute parameters for interactions between i and j atoms */
753 qq20 = _mm_mul_ps(iq2,jq0);
755 /* REACTION-FIELD ELECTROSTATICS */
756 felec = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
760 /* Update vectorial force */
761 fix2 = _mm_macc_ps(dx20,fscal,fix2);
762 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
763 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
765 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
766 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
767 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
769 fjptrA = f+j_coord_offsetA;
770 fjptrB = f+j_coord_offsetB;
771 fjptrC = f+j_coord_offsetC;
772 fjptrD = f+j_coord_offsetD;
774 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
776 /* Inner loop uses 97 flops */
782 /* Get j neighbor index, and coordinate index */
783 jnrlistA = jjnr[jidx];
784 jnrlistB = jjnr[jidx+1];
785 jnrlistC = jjnr[jidx+2];
786 jnrlistD = jjnr[jidx+3];
787 /* Sign of each element will be negative for non-real atoms.
788 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
789 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
791 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
792 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
793 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
794 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
795 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
796 j_coord_offsetA = DIM*jnrA;
797 j_coord_offsetB = DIM*jnrB;
798 j_coord_offsetC = DIM*jnrC;
799 j_coord_offsetD = DIM*jnrD;
801 /* load j atom coordinates */
802 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
803 x+j_coord_offsetC,x+j_coord_offsetD,
806 /* Calculate displacement vector */
807 dx00 = _mm_sub_ps(ix0,jx0);
808 dy00 = _mm_sub_ps(iy0,jy0);
809 dz00 = _mm_sub_ps(iz0,jz0);
810 dx10 = _mm_sub_ps(ix1,jx0);
811 dy10 = _mm_sub_ps(iy1,jy0);
812 dz10 = _mm_sub_ps(iz1,jz0);
813 dx20 = _mm_sub_ps(ix2,jx0);
814 dy20 = _mm_sub_ps(iy2,jy0);
815 dz20 = _mm_sub_ps(iz2,jz0);
817 /* Calculate squared distance and things based on it */
818 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
819 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
820 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
822 rinv00 = gmx_mm_invsqrt_ps(rsq00);
823 rinv10 = gmx_mm_invsqrt_ps(rsq10);
824 rinv20 = gmx_mm_invsqrt_ps(rsq20);
826 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
827 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
828 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
830 /* Load parameters for j particles */
831 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
832 charge+jnrC+0,charge+jnrD+0);
833 vdwjidx0A = 2*vdwtype[jnrA+0];
834 vdwjidx0B = 2*vdwtype[jnrB+0];
835 vdwjidx0C = 2*vdwtype[jnrC+0];
836 vdwjidx0D = 2*vdwtype[jnrD+0];
838 fjx0 = _mm_setzero_ps();
839 fjy0 = _mm_setzero_ps();
840 fjz0 = _mm_setzero_ps();
842 /**************************
843 * CALCULATE INTERACTIONS *
844 **************************/
846 /* Compute parameters for interactions between i and j atoms */
847 qq00 = _mm_mul_ps(iq0,jq0);
848 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
849 vdwparam+vdwioffset0+vdwjidx0B,
850 vdwparam+vdwioffset0+vdwjidx0C,
851 vdwparam+vdwioffset0+vdwjidx0D,
854 /* REACTION-FIELD ELECTROSTATICS */
855 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
857 /* LENNARD-JONES DISPERSION/REPULSION */
859 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
860 fvdw = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
862 fscal = _mm_add_ps(felec,fvdw);
864 fscal = _mm_andnot_ps(dummy_mask,fscal);
866 /* Update vectorial force */
867 fix0 = _mm_macc_ps(dx00,fscal,fix0);
868 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
869 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
871 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
872 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
873 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
875 /**************************
876 * CALCULATE INTERACTIONS *
877 **************************/
879 /* Compute parameters for interactions between i and j atoms */
880 qq10 = _mm_mul_ps(iq1,jq0);
882 /* REACTION-FIELD ELECTROSTATICS */
883 felec = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
887 fscal = _mm_andnot_ps(dummy_mask,fscal);
889 /* Update vectorial force */
890 fix1 = _mm_macc_ps(dx10,fscal,fix1);
891 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
892 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
894 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
895 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
896 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
898 /**************************
899 * CALCULATE INTERACTIONS *
900 **************************/
902 /* Compute parameters for interactions between i and j atoms */
903 qq20 = _mm_mul_ps(iq2,jq0);
905 /* REACTION-FIELD ELECTROSTATICS */
906 felec = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
910 fscal = _mm_andnot_ps(dummy_mask,fscal);
912 /* Update vectorial force */
913 fix2 = _mm_macc_ps(dx20,fscal,fix2);
914 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
915 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
917 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
918 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
919 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
921 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
922 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
923 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
924 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
926 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
928 /* Inner loop uses 97 flops */
931 /* End of innermost loop */
933 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
934 f+i_coord_offset,fshift+i_shift_offset);
936 /* Increment number of inner iterations */
937 inneriter += j_index_end - j_index_start;
939 /* Outer loop uses 18 flops */
942 /* Increment number of outer iterations */
945 /* Update outer/inner flops */
947 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*97);