2 * Note: this file was generated by the Gromacs avx_128_fma_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_128_fma_double.h"
34 #include "kernelutil_x86_avx_128_fma_double.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecCSTab_VdwNone_GeomW3P1_VF_avx_128_fma_double
38 * Electrostatics interaction: CubicSplineTable
39 * VdW interaction: None
40 * Geometry: Water3-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecCSTab_VdwNone_GeomW3P1_VF_avx_128_fma_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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
61 int j_coord_offsetA,j_coord_offsetB;
62 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
64 real *shiftvec,*fshift,*x,*f;
65 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
67 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
69 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
71 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
72 int vdwjidx0A,vdwjidx0B;
73 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
74 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
75 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
76 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
77 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
80 __m128i ifour = _mm_set1_epi32(4);
81 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
83 __m128d dummy_mask,cutoff_mask;
84 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
85 __m128d one = _mm_set1_pd(1.0);
86 __m128d two = _mm_set1_pd(2.0);
92 jindex = nlist->jindex;
94 shiftidx = nlist->shift;
96 shiftvec = fr->shift_vec[0];
97 fshift = fr->fshift[0];
98 facel = _mm_set1_pd(fr->epsfac);
99 charge = mdatoms->chargeA;
101 vftab = kernel_data->table_elec->data;
102 vftabscale = _mm_set1_pd(kernel_data->table_elec->scale);
104 /* Setup water-specific parameters */
105 inr = nlist->iinr[0];
106 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
107 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
108 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
110 /* Avoid stupid compiler warnings */
118 /* Start outer loop over neighborlists */
119 for(iidx=0; iidx<nri; iidx++)
121 /* Load shift vector for this list */
122 i_shift_offset = DIM*shiftidx[iidx];
124 /* Load limits for loop over neighbors */
125 j_index_start = jindex[iidx];
126 j_index_end = jindex[iidx+1];
128 /* Get outer coordinate index */
130 i_coord_offset = DIM*inr;
132 /* Load i particle coords and add shift vector */
133 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
134 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
136 fix0 = _mm_setzero_pd();
137 fiy0 = _mm_setzero_pd();
138 fiz0 = _mm_setzero_pd();
139 fix1 = _mm_setzero_pd();
140 fiy1 = _mm_setzero_pd();
141 fiz1 = _mm_setzero_pd();
142 fix2 = _mm_setzero_pd();
143 fiy2 = _mm_setzero_pd();
144 fiz2 = _mm_setzero_pd();
146 /* Reset potential sums */
147 velecsum = _mm_setzero_pd();
149 /* Start inner kernel loop */
150 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
153 /* Get j neighbor index, and coordinate index */
156 j_coord_offsetA = DIM*jnrA;
157 j_coord_offsetB = DIM*jnrB;
159 /* load j atom coordinates */
160 gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
163 /* Calculate displacement vector */
164 dx00 = _mm_sub_pd(ix0,jx0);
165 dy00 = _mm_sub_pd(iy0,jy0);
166 dz00 = _mm_sub_pd(iz0,jz0);
167 dx10 = _mm_sub_pd(ix1,jx0);
168 dy10 = _mm_sub_pd(iy1,jy0);
169 dz10 = _mm_sub_pd(iz1,jz0);
170 dx20 = _mm_sub_pd(ix2,jx0);
171 dy20 = _mm_sub_pd(iy2,jy0);
172 dz20 = _mm_sub_pd(iz2,jz0);
174 /* Calculate squared distance and things based on it */
175 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
176 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
177 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
179 rinv00 = gmx_mm_invsqrt_pd(rsq00);
180 rinv10 = gmx_mm_invsqrt_pd(rsq10);
181 rinv20 = gmx_mm_invsqrt_pd(rsq20);
183 /* Load parameters for j particles */
184 jq0 = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
186 fjx0 = _mm_setzero_pd();
187 fjy0 = _mm_setzero_pd();
188 fjz0 = _mm_setzero_pd();
190 /**************************
191 * CALCULATE INTERACTIONS *
192 **************************/
194 r00 = _mm_mul_pd(rsq00,rinv00);
196 /* Compute parameters for interactions between i and j atoms */
197 qq00 = _mm_mul_pd(iq0,jq0);
199 /* Calculate table index by multiplying r with table scale and truncate to integer */
200 rt = _mm_mul_pd(r00,vftabscale);
201 vfitab = _mm_cvttpd_epi32(rt);
203 vfeps = _mm_frcz_pd(rt);
205 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
207 twovfeps = _mm_add_pd(vfeps,vfeps);
208 vfitab = _mm_slli_epi32(vfitab,2);
210 /* CUBIC SPLINE TABLE ELECTROSTATICS */
211 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
212 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
213 GMX_MM_TRANSPOSE2_PD(Y,F);
214 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
215 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
216 GMX_MM_TRANSPOSE2_PD(G,H);
217 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
218 VV = _mm_macc_pd(vfeps,Fp,Y);
219 velec = _mm_mul_pd(qq00,VV);
220 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
221 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq00,FF),_mm_mul_pd(vftabscale,rinv00)));
223 /* Update potential sum for this i atom from the interaction with this j atom. */
224 velecsum = _mm_add_pd(velecsum,velec);
228 /* Update vectorial force */
229 fix0 = _mm_macc_pd(dx00,fscal,fix0);
230 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
231 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
233 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
234 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
235 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
237 /**************************
238 * CALCULATE INTERACTIONS *
239 **************************/
241 r10 = _mm_mul_pd(rsq10,rinv10);
243 /* Compute parameters for interactions between i and j atoms */
244 qq10 = _mm_mul_pd(iq1,jq0);
246 /* Calculate table index by multiplying r with table scale and truncate to integer */
247 rt = _mm_mul_pd(r10,vftabscale);
248 vfitab = _mm_cvttpd_epi32(rt);
250 vfeps = _mm_frcz_pd(rt);
252 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
254 twovfeps = _mm_add_pd(vfeps,vfeps);
255 vfitab = _mm_slli_epi32(vfitab,2);
257 /* CUBIC SPLINE TABLE ELECTROSTATICS */
258 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
259 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
260 GMX_MM_TRANSPOSE2_PD(Y,F);
261 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
262 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
263 GMX_MM_TRANSPOSE2_PD(G,H);
264 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
265 VV = _mm_macc_pd(vfeps,Fp,Y);
266 velec = _mm_mul_pd(qq10,VV);
267 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
268 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq10,FF),_mm_mul_pd(vftabscale,rinv10)));
270 /* Update potential sum for this i atom from the interaction with this j atom. */
271 velecsum = _mm_add_pd(velecsum,velec);
275 /* Update vectorial force */
276 fix1 = _mm_macc_pd(dx10,fscal,fix1);
277 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
278 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
280 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
281 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
282 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
284 /**************************
285 * CALCULATE INTERACTIONS *
286 **************************/
288 r20 = _mm_mul_pd(rsq20,rinv20);
290 /* Compute parameters for interactions between i and j atoms */
291 qq20 = _mm_mul_pd(iq2,jq0);
293 /* Calculate table index by multiplying r with table scale and truncate to integer */
294 rt = _mm_mul_pd(r20,vftabscale);
295 vfitab = _mm_cvttpd_epi32(rt);
297 vfeps = _mm_frcz_pd(rt);
299 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
301 twovfeps = _mm_add_pd(vfeps,vfeps);
302 vfitab = _mm_slli_epi32(vfitab,2);
304 /* CUBIC SPLINE TABLE ELECTROSTATICS */
305 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
306 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
307 GMX_MM_TRANSPOSE2_PD(Y,F);
308 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
309 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
310 GMX_MM_TRANSPOSE2_PD(G,H);
311 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
312 VV = _mm_macc_pd(vfeps,Fp,Y);
313 velec = _mm_mul_pd(qq20,VV);
314 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
315 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq20,FF),_mm_mul_pd(vftabscale,rinv20)));
317 /* Update potential sum for this i atom from the interaction with this j atom. */
318 velecsum = _mm_add_pd(velecsum,velec);
322 /* Update vectorial force */
323 fix2 = _mm_macc_pd(dx20,fscal,fix2);
324 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
325 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
327 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
328 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
329 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
331 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
333 /* Inner loop uses 141 flops */
340 j_coord_offsetA = DIM*jnrA;
342 /* load j atom coordinates */
343 gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
346 /* Calculate displacement vector */
347 dx00 = _mm_sub_pd(ix0,jx0);
348 dy00 = _mm_sub_pd(iy0,jy0);
349 dz00 = _mm_sub_pd(iz0,jz0);
350 dx10 = _mm_sub_pd(ix1,jx0);
351 dy10 = _mm_sub_pd(iy1,jy0);
352 dz10 = _mm_sub_pd(iz1,jz0);
353 dx20 = _mm_sub_pd(ix2,jx0);
354 dy20 = _mm_sub_pd(iy2,jy0);
355 dz20 = _mm_sub_pd(iz2,jz0);
357 /* Calculate squared distance and things based on it */
358 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
359 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
360 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
362 rinv00 = gmx_mm_invsqrt_pd(rsq00);
363 rinv10 = gmx_mm_invsqrt_pd(rsq10);
364 rinv20 = gmx_mm_invsqrt_pd(rsq20);
366 /* Load parameters for j particles */
367 jq0 = _mm_load_sd(charge+jnrA+0);
369 fjx0 = _mm_setzero_pd();
370 fjy0 = _mm_setzero_pd();
371 fjz0 = _mm_setzero_pd();
373 /**************************
374 * CALCULATE INTERACTIONS *
375 **************************/
377 r00 = _mm_mul_pd(rsq00,rinv00);
379 /* Compute parameters for interactions between i and j atoms */
380 qq00 = _mm_mul_pd(iq0,jq0);
382 /* Calculate table index by multiplying r with table scale and truncate to integer */
383 rt = _mm_mul_pd(r00,vftabscale);
384 vfitab = _mm_cvttpd_epi32(rt);
386 vfeps = _mm_frcz_pd(rt);
388 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
390 twovfeps = _mm_add_pd(vfeps,vfeps);
391 vfitab = _mm_slli_epi32(vfitab,2);
393 /* CUBIC SPLINE TABLE ELECTROSTATICS */
394 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
395 F = _mm_setzero_pd();
396 GMX_MM_TRANSPOSE2_PD(Y,F);
397 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
398 H = _mm_setzero_pd();
399 GMX_MM_TRANSPOSE2_PD(G,H);
400 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
401 VV = _mm_macc_pd(vfeps,Fp,Y);
402 velec = _mm_mul_pd(qq00,VV);
403 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
404 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq00,FF),_mm_mul_pd(vftabscale,rinv00)));
406 /* Update potential sum for this i atom from the interaction with this j atom. */
407 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
408 velecsum = _mm_add_pd(velecsum,velec);
412 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
414 /* Update vectorial force */
415 fix0 = _mm_macc_pd(dx00,fscal,fix0);
416 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
417 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
419 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
420 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
421 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
423 /**************************
424 * CALCULATE INTERACTIONS *
425 **************************/
427 r10 = _mm_mul_pd(rsq10,rinv10);
429 /* Compute parameters for interactions between i and j atoms */
430 qq10 = _mm_mul_pd(iq1,jq0);
432 /* Calculate table index by multiplying r with table scale and truncate to integer */
433 rt = _mm_mul_pd(r10,vftabscale);
434 vfitab = _mm_cvttpd_epi32(rt);
436 vfeps = _mm_frcz_pd(rt);
438 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
440 twovfeps = _mm_add_pd(vfeps,vfeps);
441 vfitab = _mm_slli_epi32(vfitab,2);
443 /* CUBIC SPLINE TABLE ELECTROSTATICS */
444 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
445 F = _mm_setzero_pd();
446 GMX_MM_TRANSPOSE2_PD(Y,F);
447 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
448 H = _mm_setzero_pd();
449 GMX_MM_TRANSPOSE2_PD(G,H);
450 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
451 VV = _mm_macc_pd(vfeps,Fp,Y);
452 velec = _mm_mul_pd(qq10,VV);
453 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
454 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq10,FF),_mm_mul_pd(vftabscale,rinv10)));
456 /* Update potential sum for this i atom from the interaction with this j atom. */
457 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
458 velecsum = _mm_add_pd(velecsum,velec);
462 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
464 /* Update vectorial force */
465 fix1 = _mm_macc_pd(dx10,fscal,fix1);
466 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
467 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
469 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
470 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
471 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
473 /**************************
474 * CALCULATE INTERACTIONS *
475 **************************/
477 r20 = _mm_mul_pd(rsq20,rinv20);
479 /* Compute parameters for interactions between i and j atoms */
480 qq20 = _mm_mul_pd(iq2,jq0);
482 /* Calculate table index by multiplying r with table scale and truncate to integer */
483 rt = _mm_mul_pd(r20,vftabscale);
484 vfitab = _mm_cvttpd_epi32(rt);
486 vfeps = _mm_frcz_pd(rt);
488 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
490 twovfeps = _mm_add_pd(vfeps,vfeps);
491 vfitab = _mm_slli_epi32(vfitab,2);
493 /* CUBIC SPLINE TABLE ELECTROSTATICS */
494 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
495 F = _mm_setzero_pd();
496 GMX_MM_TRANSPOSE2_PD(Y,F);
497 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
498 H = _mm_setzero_pd();
499 GMX_MM_TRANSPOSE2_PD(G,H);
500 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
501 VV = _mm_macc_pd(vfeps,Fp,Y);
502 velec = _mm_mul_pd(qq20,VV);
503 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
504 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq20,FF),_mm_mul_pd(vftabscale,rinv20)));
506 /* Update potential sum for this i atom from the interaction with this j atom. */
507 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
508 velecsum = _mm_add_pd(velecsum,velec);
512 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
514 /* Update vectorial force */
515 fix2 = _mm_macc_pd(dx20,fscal,fix2);
516 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
517 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
519 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
520 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
521 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
523 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
525 /* Inner loop uses 141 flops */
528 /* End of innermost loop */
530 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
531 f+i_coord_offset,fshift+i_shift_offset);
534 /* Update potential energies */
535 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
537 /* Increment number of inner iterations */
538 inneriter += j_index_end - j_index_start;
540 /* Outer loop uses 19 flops */
543 /* Increment number of outer iterations */
546 /* Update outer/inner flops */
548 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_VF,outeriter*19 + inneriter*141);
551 * Gromacs nonbonded kernel: nb_kernel_ElecCSTab_VdwNone_GeomW3P1_F_avx_128_fma_double
552 * Electrostatics interaction: CubicSplineTable
553 * VdW interaction: None
554 * Geometry: Water3-Particle
555 * Calculate force/pot: Force
558 nb_kernel_ElecCSTab_VdwNone_GeomW3P1_F_avx_128_fma_double
559 (t_nblist * gmx_restrict nlist,
560 rvec * gmx_restrict xx,
561 rvec * gmx_restrict ff,
562 t_forcerec * gmx_restrict fr,
563 t_mdatoms * gmx_restrict mdatoms,
564 nb_kernel_data_t * gmx_restrict kernel_data,
565 t_nrnb * gmx_restrict nrnb)
567 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
568 * just 0 for non-waters.
569 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
570 * jnr indices corresponding to data put in the four positions in the SIMD register.
572 int i_shift_offset,i_coord_offset,outeriter,inneriter;
573 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
575 int j_coord_offsetA,j_coord_offsetB;
576 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
578 real *shiftvec,*fshift,*x,*f;
579 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
581 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
583 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
585 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
586 int vdwjidx0A,vdwjidx0B;
587 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
588 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
589 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
590 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
591 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
594 __m128i ifour = _mm_set1_epi32(4);
595 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
597 __m128d dummy_mask,cutoff_mask;
598 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
599 __m128d one = _mm_set1_pd(1.0);
600 __m128d two = _mm_set1_pd(2.0);
606 jindex = nlist->jindex;
608 shiftidx = nlist->shift;
610 shiftvec = fr->shift_vec[0];
611 fshift = fr->fshift[0];
612 facel = _mm_set1_pd(fr->epsfac);
613 charge = mdatoms->chargeA;
615 vftab = kernel_data->table_elec->data;
616 vftabscale = _mm_set1_pd(kernel_data->table_elec->scale);
618 /* Setup water-specific parameters */
619 inr = nlist->iinr[0];
620 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
621 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
622 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
624 /* Avoid stupid compiler warnings */
632 /* Start outer loop over neighborlists */
633 for(iidx=0; iidx<nri; iidx++)
635 /* Load shift vector for this list */
636 i_shift_offset = DIM*shiftidx[iidx];
638 /* Load limits for loop over neighbors */
639 j_index_start = jindex[iidx];
640 j_index_end = jindex[iidx+1];
642 /* Get outer coordinate index */
644 i_coord_offset = DIM*inr;
646 /* Load i particle coords and add shift vector */
647 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
648 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
650 fix0 = _mm_setzero_pd();
651 fiy0 = _mm_setzero_pd();
652 fiz0 = _mm_setzero_pd();
653 fix1 = _mm_setzero_pd();
654 fiy1 = _mm_setzero_pd();
655 fiz1 = _mm_setzero_pd();
656 fix2 = _mm_setzero_pd();
657 fiy2 = _mm_setzero_pd();
658 fiz2 = _mm_setzero_pd();
660 /* Start inner kernel loop */
661 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
664 /* Get j neighbor index, and coordinate index */
667 j_coord_offsetA = DIM*jnrA;
668 j_coord_offsetB = DIM*jnrB;
670 /* load j atom coordinates */
671 gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
674 /* Calculate displacement vector */
675 dx00 = _mm_sub_pd(ix0,jx0);
676 dy00 = _mm_sub_pd(iy0,jy0);
677 dz00 = _mm_sub_pd(iz0,jz0);
678 dx10 = _mm_sub_pd(ix1,jx0);
679 dy10 = _mm_sub_pd(iy1,jy0);
680 dz10 = _mm_sub_pd(iz1,jz0);
681 dx20 = _mm_sub_pd(ix2,jx0);
682 dy20 = _mm_sub_pd(iy2,jy0);
683 dz20 = _mm_sub_pd(iz2,jz0);
685 /* Calculate squared distance and things based on it */
686 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
687 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
688 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
690 rinv00 = gmx_mm_invsqrt_pd(rsq00);
691 rinv10 = gmx_mm_invsqrt_pd(rsq10);
692 rinv20 = gmx_mm_invsqrt_pd(rsq20);
694 /* Load parameters for j particles */
695 jq0 = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
697 fjx0 = _mm_setzero_pd();
698 fjy0 = _mm_setzero_pd();
699 fjz0 = _mm_setzero_pd();
701 /**************************
702 * CALCULATE INTERACTIONS *
703 **************************/
705 r00 = _mm_mul_pd(rsq00,rinv00);
707 /* Compute parameters for interactions between i and j atoms */
708 qq00 = _mm_mul_pd(iq0,jq0);
710 /* Calculate table index by multiplying r with table scale and truncate to integer */
711 rt = _mm_mul_pd(r00,vftabscale);
712 vfitab = _mm_cvttpd_epi32(rt);
714 vfeps = _mm_frcz_pd(rt);
716 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
718 twovfeps = _mm_add_pd(vfeps,vfeps);
719 vfitab = _mm_slli_epi32(vfitab,2);
721 /* CUBIC SPLINE TABLE ELECTROSTATICS */
722 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
723 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
724 GMX_MM_TRANSPOSE2_PD(Y,F);
725 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
726 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
727 GMX_MM_TRANSPOSE2_PD(G,H);
728 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
729 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
730 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq00,FF),_mm_mul_pd(vftabscale,rinv00)));
734 /* Update vectorial force */
735 fix0 = _mm_macc_pd(dx00,fscal,fix0);
736 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
737 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
739 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
740 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
741 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
743 /**************************
744 * CALCULATE INTERACTIONS *
745 **************************/
747 r10 = _mm_mul_pd(rsq10,rinv10);
749 /* Compute parameters for interactions between i and j atoms */
750 qq10 = _mm_mul_pd(iq1,jq0);
752 /* Calculate table index by multiplying r with table scale and truncate to integer */
753 rt = _mm_mul_pd(r10,vftabscale);
754 vfitab = _mm_cvttpd_epi32(rt);
756 vfeps = _mm_frcz_pd(rt);
758 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
760 twovfeps = _mm_add_pd(vfeps,vfeps);
761 vfitab = _mm_slli_epi32(vfitab,2);
763 /* CUBIC SPLINE TABLE ELECTROSTATICS */
764 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
765 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
766 GMX_MM_TRANSPOSE2_PD(Y,F);
767 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
768 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
769 GMX_MM_TRANSPOSE2_PD(G,H);
770 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
771 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
772 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq10,FF),_mm_mul_pd(vftabscale,rinv10)));
776 /* Update vectorial force */
777 fix1 = _mm_macc_pd(dx10,fscal,fix1);
778 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
779 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
781 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
782 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
783 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
785 /**************************
786 * CALCULATE INTERACTIONS *
787 **************************/
789 r20 = _mm_mul_pd(rsq20,rinv20);
791 /* Compute parameters for interactions between i and j atoms */
792 qq20 = _mm_mul_pd(iq2,jq0);
794 /* Calculate table index by multiplying r with table scale and truncate to integer */
795 rt = _mm_mul_pd(r20,vftabscale);
796 vfitab = _mm_cvttpd_epi32(rt);
798 vfeps = _mm_frcz_pd(rt);
800 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
802 twovfeps = _mm_add_pd(vfeps,vfeps);
803 vfitab = _mm_slli_epi32(vfitab,2);
805 /* CUBIC SPLINE TABLE ELECTROSTATICS */
806 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
807 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
808 GMX_MM_TRANSPOSE2_PD(Y,F);
809 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
810 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
811 GMX_MM_TRANSPOSE2_PD(G,H);
812 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
813 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
814 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq20,FF),_mm_mul_pd(vftabscale,rinv20)));
818 /* Update vectorial force */
819 fix2 = _mm_macc_pd(dx20,fscal,fix2);
820 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
821 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
823 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
824 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
825 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
827 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
829 /* Inner loop uses 129 flops */
836 j_coord_offsetA = DIM*jnrA;
838 /* load j atom coordinates */
839 gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
842 /* Calculate displacement vector */
843 dx00 = _mm_sub_pd(ix0,jx0);
844 dy00 = _mm_sub_pd(iy0,jy0);
845 dz00 = _mm_sub_pd(iz0,jz0);
846 dx10 = _mm_sub_pd(ix1,jx0);
847 dy10 = _mm_sub_pd(iy1,jy0);
848 dz10 = _mm_sub_pd(iz1,jz0);
849 dx20 = _mm_sub_pd(ix2,jx0);
850 dy20 = _mm_sub_pd(iy2,jy0);
851 dz20 = _mm_sub_pd(iz2,jz0);
853 /* Calculate squared distance and things based on it */
854 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
855 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
856 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
858 rinv00 = gmx_mm_invsqrt_pd(rsq00);
859 rinv10 = gmx_mm_invsqrt_pd(rsq10);
860 rinv20 = gmx_mm_invsqrt_pd(rsq20);
862 /* Load parameters for j particles */
863 jq0 = _mm_load_sd(charge+jnrA+0);
865 fjx0 = _mm_setzero_pd();
866 fjy0 = _mm_setzero_pd();
867 fjz0 = _mm_setzero_pd();
869 /**************************
870 * CALCULATE INTERACTIONS *
871 **************************/
873 r00 = _mm_mul_pd(rsq00,rinv00);
875 /* Compute parameters for interactions between i and j atoms */
876 qq00 = _mm_mul_pd(iq0,jq0);
878 /* Calculate table index by multiplying r with table scale and truncate to integer */
879 rt = _mm_mul_pd(r00,vftabscale);
880 vfitab = _mm_cvttpd_epi32(rt);
882 vfeps = _mm_frcz_pd(rt);
884 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
886 twovfeps = _mm_add_pd(vfeps,vfeps);
887 vfitab = _mm_slli_epi32(vfitab,2);
889 /* CUBIC SPLINE TABLE ELECTROSTATICS */
890 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
891 F = _mm_setzero_pd();
892 GMX_MM_TRANSPOSE2_PD(Y,F);
893 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
894 H = _mm_setzero_pd();
895 GMX_MM_TRANSPOSE2_PD(G,H);
896 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
897 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
898 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq00,FF),_mm_mul_pd(vftabscale,rinv00)));
902 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
904 /* Update vectorial force */
905 fix0 = _mm_macc_pd(dx00,fscal,fix0);
906 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
907 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
909 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
910 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
911 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
913 /**************************
914 * CALCULATE INTERACTIONS *
915 **************************/
917 r10 = _mm_mul_pd(rsq10,rinv10);
919 /* Compute parameters for interactions between i and j atoms */
920 qq10 = _mm_mul_pd(iq1,jq0);
922 /* Calculate table index by multiplying r with table scale and truncate to integer */
923 rt = _mm_mul_pd(r10,vftabscale);
924 vfitab = _mm_cvttpd_epi32(rt);
926 vfeps = _mm_frcz_pd(rt);
928 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
930 twovfeps = _mm_add_pd(vfeps,vfeps);
931 vfitab = _mm_slli_epi32(vfitab,2);
933 /* CUBIC SPLINE TABLE ELECTROSTATICS */
934 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
935 F = _mm_setzero_pd();
936 GMX_MM_TRANSPOSE2_PD(Y,F);
937 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
938 H = _mm_setzero_pd();
939 GMX_MM_TRANSPOSE2_PD(G,H);
940 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
941 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
942 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq10,FF),_mm_mul_pd(vftabscale,rinv10)));
946 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
948 /* Update vectorial force */
949 fix1 = _mm_macc_pd(dx10,fscal,fix1);
950 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
951 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
953 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
954 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
955 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
957 /**************************
958 * CALCULATE INTERACTIONS *
959 **************************/
961 r20 = _mm_mul_pd(rsq20,rinv20);
963 /* Compute parameters for interactions between i and j atoms */
964 qq20 = _mm_mul_pd(iq2,jq0);
966 /* Calculate table index by multiplying r with table scale and truncate to integer */
967 rt = _mm_mul_pd(r20,vftabscale);
968 vfitab = _mm_cvttpd_epi32(rt);
970 vfeps = _mm_frcz_pd(rt);
972 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
974 twovfeps = _mm_add_pd(vfeps,vfeps);
975 vfitab = _mm_slli_epi32(vfitab,2);
977 /* CUBIC SPLINE TABLE ELECTROSTATICS */
978 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
979 F = _mm_setzero_pd();
980 GMX_MM_TRANSPOSE2_PD(Y,F);
981 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
982 H = _mm_setzero_pd();
983 GMX_MM_TRANSPOSE2_PD(G,H);
984 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(vfeps,H,G),F);
985 FF = _mm_macc_pd(_mm_macc_pd(twovfeps,H,G),vfeps,Fp);
986 felec = _mm_xor_pd(signbit,_mm_mul_pd(_mm_mul_pd(qq20,FF),_mm_mul_pd(vftabscale,rinv20)));
990 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
992 /* Update vectorial force */
993 fix2 = _mm_macc_pd(dx20,fscal,fix2);
994 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
995 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
997 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
998 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
999 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
1001 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
1003 /* Inner loop uses 129 flops */
1006 /* End of innermost loop */
1008 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1009 f+i_coord_offset,fshift+i_shift_offset);
1011 /* Increment number of inner iterations */
1012 inneriter += j_index_end - j_index_start;
1014 /* Outer loop uses 18 flops */
1017 /* Increment number of outer iterations */
1020 /* Update outer/inner flops */
1022 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_F,outeriter*18 + inneriter*129);