2 * Note: this file was generated by the Gromacs avx_256_double kernel generator.
4 * This source code is part of
8 * Copyright (c) 2001-2012, The GROMACS Development Team
10 * Gromacs is a library for molecular simulation and trajectory analysis,
11 * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
12 * a full list of developers and information, check out http://www.gromacs.org
14 * This program is free software; you can redistribute it and/or modify it under
15 * the terms of the GNU Lesser General Public License as published by the Free
16 * Software Foundation; either version 2 of the License, or (at your option) any
19 * To help fund GROMACS development, we humbly ask that you cite
20 * the papers people have written on it - you can find them on the website.
28 #include "../nb_kernel.h"
29 #include "types/simple.h"
33 #include "gmx_math_x86_avx_256_double.h"
34 #include "kernelutil_x86_avx_256_double.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecEwSh_VdwNone_GeomW3P1_VF_avx_256_double
38 * Electrostatics interaction: Ewald
39 * VdW interaction: None
40 * Geometry: Water3-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEwSh_VdwNone_GeomW3P1_VF_avx_256_double
45 (t_nblist * gmx_restrict nlist,
46 rvec * gmx_restrict xx,
47 rvec * gmx_restrict ff,
48 t_forcerec * gmx_restrict fr,
49 t_mdatoms * gmx_restrict mdatoms,
50 nb_kernel_data_t * gmx_restrict kernel_data,
51 t_nrnb * gmx_restrict nrnb)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
62 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
63 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
64 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
66 real *shiftvec,*fshift,*x,*f;
67 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
69 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
70 real * vdwioffsetptr0;
71 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
72 real * vdwioffsetptr1;
73 __m256d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
74 real * vdwioffsetptr2;
75 __m256d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
76 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
77 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
78 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
79 __m256d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
80 __m256d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
81 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
84 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
85 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
87 __m256d dummy_mask,cutoff_mask;
88 __m128 tmpmask0,tmpmask1;
89 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
90 __m256d one = _mm256_set1_pd(1.0);
91 __m256d two = _mm256_set1_pd(2.0);
97 jindex = nlist->jindex;
99 shiftidx = nlist->shift;
101 shiftvec = fr->shift_vec[0];
102 fshift = fr->fshift[0];
103 facel = _mm256_set1_pd(fr->epsfac);
104 charge = mdatoms->chargeA;
106 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
107 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
108 beta2 = _mm256_mul_pd(beta,beta);
109 beta3 = _mm256_mul_pd(beta,beta2);
111 ewtab = fr->ic->tabq_coul_FDV0;
112 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
113 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
115 /* Setup water-specific parameters */
116 inr = nlist->iinr[0];
117 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
118 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
119 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
121 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
122 rcutoff_scalar = fr->rcoulomb;
123 rcutoff = _mm256_set1_pd(rcutoff_scalar);
124 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
126 /* Avoid stupid compiler warnings */
127 jnrA = jnrB = jnrC = jnrD = 0;
136 for(iidx=0;iidx<4*DIM;iidx++)
141 /* Start outer loop over neighborlists */
142 for(iidx=0; iidx<nri; iidx++)
144 /* Load shift vector for this list */
145 i_shift_offset = DIM*shiftidx[iidx];
147 /* Load limits for loop over neighbors */
148 j_index_start = jindex[iidx];
149 j_index_end = jindex[iidx+1];
151 /* Get outer coordinate index */
153 i_coord_offset = DIM*inr;
155 /* Load i particle coords and add shift vector */
156 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
157 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
159 fix0 = _mm256_setzero_pd();
160 fiy0 = _mm256_setzero_pd();
161 fiz0 = _mm256_setzero_pd();
162 fix1 = _mm256_setzero_pd();
163 fiy1 = _mm256_setzero_pd();
164 fiz1 = _mm256_setzero_pd();
165 fix2 = _mm256_setzero_pd();
166 fiy2 = _mm256_setzero_pd();
167 fiz2 = _mm256_setzero_pd();
169 /* Reset potential sums */
170 velecsum = _mm256_setzero_pd();
172 /* Start inner kernel loop */
173 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
176 /* Get j neighbor index, and coordinate index */
181 j_coord_offsetA = DIM*jnrA;
182 j_coord_offsetB = DIM*jnrB;
183 j_coord_offsetC = DIM*jnrC;
184 j_coord_offsetD = DIM*jnrD;
186 /* load j atom coordinates */
187 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
188 x+j_coord_offsetC,x+j_coord_offsetD,
191 /* Calculate displacement vector */
192 dx00 = _mm256_sub_pd(ix0,jx0);
193 dy00 = _mm256_sub_pd(iy0,jy0);
194 dz00 = _mm256_sub_pd(iz0,jz0);
195 dx10 = _mm256_sub_pd(ix1,jx0);
196 dy10 = _mm256_sub_pd(iy1,jy0);
197 dz10 = _mm256_sub_pd(iz1,jz0);
198 dx20 = _mm256_sub_pd(ix2,jx0);
199 dy20 = _mm256_sub_pd(iy2,jy0);
200 dz20 = _mm256_sub_pd(iz2,jz0);
202 /* Calculate squared distance and things based on it */
203 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
204 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
205 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
207 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
208 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
209 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
211 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
212 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
213 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
215 /* Load parameters for j particles */
216 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
217 charge+jnrC+0,charge+jnrD+0);
219 fjx0 = _mm256_setzero_pd();
220 fjy0 = _mm256_setzero_pd();
221 fjz0 = _mm256_setzero_pd();
223 /**************************
224 * CALCULATE INTERACTIONS *
225 **************************/
227 if (gmx_mm256_any_lt(rsq00,rcutoff2))
230 r00 = _mm256_mul_pd(rsq00,rinv00);
232 /* Compute parameters for interactions between i and j atoms */
233 qq00 = _mm256_mul_pd(iq0,jq0);
235 /* EWALD ELECTROSTATICS */
237 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
238 ewrt = _mm256_mul_pd(r00,ewtabscale);
239 ewitab = _mm256_cvttpd_epi32(ewrt);
240 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
241 ewitab = _mm_slli_epi32(ewitab,2);
242 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
243 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
244 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
245 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
246 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
247 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
248 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
249 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
250 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
252 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
254 /* Update potential sum for this i atom from the interaction with this j atom. */
255 velec = _mm256_and_pd(velec,cutoff_mask);
256 velecsum = _mm256_add_pd(velecsum,velec);
260 fscal = _mm256_and_pd(fscal,cutoff_mask);
262 /* Calculate temporary vectorial force */
263 tx = _mm256_mul_pd(fscal,dx00);
264 ty = _mm256_mul_pd(fscal,dy00);
265 tz = _mm256_mul_pd(fscal,dz00);
267 /* Update vectorial force */
268 fix0 = _mm256_add_pd(fix0,tx);
269 fiy0 = _mm256_add_pd(fiy0,ty);
270 fiz0 = _mm256_add_pd(fiz0,tz);
272 fjx0 = _mm256_add_pd(fjx0,tx);
273 fjy0 = _mm256_add_pd(fjy0,ty);
274 fjz0 = _mm256_add_pd(fjz0,tz);
278 /**************************
279 * CALCULATE INTERACTIONS *
280 **************************/
282 if (gmx_mm256_any_lt(rsq10,rcutoff2))
285 r10 = _mm256_mul_pd(rsq10,rinv10);
287 /* Compute parameters for interactions between i and j atoms */
288 qq10 = _mm256_mul_pd(iq1,jq0);
290 /* EWALD ELECTROSTATICS */
292 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
293 ewrt = _mm256_mul_pd(r10,ewtabscale);
294 ewitab = _mm256_cvttpd_epi32(ewrt);
295 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
296 ewitab = _mm_slli_epi32(ewitab,2);
297 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
298 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
299 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
300 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
301 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
302 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
303 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
304 velec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_sub_pd(rinv10,sh_ewald),velec));
305 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
307 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
309 /* Update potential sum for this i atom from the interaction with this j atom. */
310 velec = _mm256_and_pd(velec,cutoff_mask);
311 velecsum = _mm256_add_pd(velecsum,velec);
315 fscal = _mm256_and_pd(fscal,cutoff_mask);
317 /* Calculate temporary vectorial force */
318 tx = _mm256_mul_pd(fscal,dx10);
319 ty = _mm256_mul_pd(fscal,dy10);
320 tz = _mm256_mul_pd(fscal,dz10);
322 /* Update vectorial force */
323 fix1 = _mm256_add_pd(fix1,tx);
324 fiy1 = _mm256_add_pd(fiy1,ty);
325 fiz1 = _mm256_add_pd(fiz1,tz);
327 fjx0 = _mm256_add_pd(fjx0,tx);
328 fjy0 = _mm256_add_pd(fjy0,ty);
329 fjz0 = _mm256_add_pd(fjz0,tz);
333 /**************************
334 * CALCULATE INTERACTIONS *
335 **************************/
337 if (gmx_mm256_any_lt(rsq20,rcutoff2))
340 r20 = _mm256_mul_pd(rsq20,rinv20);
342 /* Compute parameters for interactions between i and j atoms */
343 qq20 = _mm256_mul_pd(iq2,jq0);
345 /* EWALD ELECTROSTATICS */
347 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
348 ewrt = _mm256_mul_pd(r20,ewtabscale);
349 ewitab = _mm256_cvttpd_epi32(ewrt);
350 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
351 ewitab = _mm_slli_epi32(ewitab,2);
352 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
353 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
354 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
355 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
356 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
357 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
358 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
359 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_sub_pd(rinv20,sh_ewald),velec));
360 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
362 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
364 /* Update potential sum for this i atom from the interaction with this j atom. */
365 velec = _mm256_and_pd(velec,cutoff_mask);
366 velecsum = _mm256_add_pd(velecsum,velec);
370 fscal = _mm256_and_pd(fscal,cutoff_mask);
372 /* Calculate temporary vectorial force */
373 tx = _mm256_mul_pd(fscal,dx20);
374 ty = _mm256_mul_pd(fscal,dy20);
375 tz = _mm256_mul_pd(fscal,dz20);
377 /* Update vectorial force */
378 fix2 = _mm256_add_pd(fix2,tx);
379 fiy2 = _mm256_add_pd(fiy2,ty);
380 fiz2 = _mm256_add_pd(fiz2,tz);
382 fjx0 = _mm256_add_pd(fjx0,tx);
383 fjy0 = _mm256_add_pd(fjy0,ty);
384 fjz0 = _mm256_add_pd(fjz0,tz);
388 fjptrA = f+j_coord_offsetA;
389 fjptrB = f+j_coord_offsetB;
390 fjptrC = f+j_coord_offsetC;
391 fjptrD = f+j_coord_offsetD;
393 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
395 /* Inner loop uses 141 flops */
401 /* Get j neighbor index, and coordinate index */
402 jnrlistA = jjnr[jidx];
403 jnrlistB = jjnr[jidx+1];
404 jnrlistC = jjnr[jidx+2];
405 jnrlistD = jjnr[jidx+3];
406 /* Sign of each element will be negative for non-real atoms.
407 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
408 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
410 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
412 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
413 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
414 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
416 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
417 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
418 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
419 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
420 j_coord_offsetA = DIM*jnrA;
421 j_coord_offsetB = DIM*jnrB;
422 j_coord_offsetC = DIM*jnrC;
423 j_coord_offsetD = DIM*jnrD;
425 /* load j atom coordinates */
426 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
427 x+j_coord_offsetC,x+j_coord_offsetD,
430 /* Calculate displacement vector */
431 dx00 = _mm256_sub_pd(ix0,jx0);
432 dy00 = _mm256_sub_pd(iy0,jy0);
433 dz00 = _mm256_sub_pd(iz0,jz0);
434 dx10 = _mm256_sub_pd(ix1,jx0);
435 dy10 = _mm256_sub_pd(iy1,jy0);
436 dz10 = _mm256_sub_pd(iz1,jz0);
437 dx20 = _mm256_sub_pd(ix2,jx0);
438 dy20 = _mm256_sub_pd(iy2,jy0);
439 dz20 = _mm256_sub_pd(iz2,jz0);
441 /* Calculate squared distance and things based on it */
442 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
443 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
444 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
446 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
447 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
448 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
450 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
451 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
452 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
454 /* Load parameters for j particles */
455 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
456 charge+jnrC+0,charge+jnrD+0);
458 fjx0 = _mm256_setzero_pd();
459 fjy0 = _mm256_setzero_pd();
460 fjz0 = _mm256_setzero_pd();
462 /**************************
463 * CALCULATE INTERACTIONS *
464 **************************/
466 if (gmx_mm256_any_lt(rsq00,rcutoff2))
469 r00 = _mm256_mul_pd(rsq00,rinv00);
470 r00 = _mm256_andnot_pd(dummy_mask,r00);
472 /* Compute parameters for interactions between i and j atoms */
473 qq00 = _mm256_mul_pd(iq0,jq0);
475 /* EWALD ELECTROSTATICS */
477 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
478 ewrt = _mm256_mul_pd(r00,ewtabscale);
479 ewitab = _mm256_cvttpd_epi32(ewrt);
480 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
481 ewitab = _mm_slli_epi32(ewitab,2);
482 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
483 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
484 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
485 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
486 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
487 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
488 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
489 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
490 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
492 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
494 /* Update potential sum for this i atom from the interaction with this j atom. */
495 velec = _mm256_and_pd(velec,cutoff_mask);
496 velec = _mm256_andnot_pd(dummy_mask,velec);
497 velecsum = _mm256_add_pd(velecsum,velec);
501 fscal = _mm256_and_pd(fscal,cutoff_mask);
503 fscal = _mm256_andnot_pd(dummy_mask,fscal);
505 /* Calculate temporary vectorial force */
506 tx = _mm256_mul_pd(fscal,dx00);
507 ty = _mm256_mul_pd(fscal,dy00);
508 tz = _mm256_mul_pd(fscal,dz00);
510 /* Update vectorial force */
511 fix0 = _mm256_add_pd(fix0,tx);
512 fiy0 = _mm256_add_pd(fiy0,ty);
513 fiz0 = _mm256_add_pd(fiz0,tz);
515 fjx0 = _mm256_add_pd(fjx0,tx);
516 fjy0 = _mm256_add_pd(fjy0,ty);
517 fjz0 = _mm256_add_pd(fjz0,tz);
521 /**************************
522 * CALCULATE INTERACTIONS *
523 **************************/
525 if (gmx_mm256_any_lt(rsq10,rcutoff2))
528 r10 = _mm256_mul_pd(rsq10,rinv10);
529 r10 = _mm256_andnot_pd(dummy_mask,r10);
531 /* Compute parameters for interactions between i and j atoms */
532 qq10 = _mm256_mul_pd(iq1,jq0);
534 /* EWALD ELECTROSTATICS */
536 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
537 ewrt = _mm256_mul_pd(r10,ewtabscale);
538 ewitab = _mm256_cvttpd_epi32(ewrt);
539 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
540 ewitab = _mm_slli_epi32(ewitab,2);
541 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
542 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
543 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
544 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
545 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
546 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
547 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
548 velec = _mm256_mul_pd(qq10,_mm256_sub_pd(_mm256_sub_pd(rinv10,sh_ewald),velec));
549 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
551 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
553 /* Update potential sum for this i atom from the interaction with this j atom. */
554 velec = _mm256_and_pd(velec,cutoff_mask);
555 velec = _mm256_andnot_pd(dummy_mask,velec);
556 velecsum = _mm256_add_pd(velecsum,velec);
560 fscal = _mm256_and_pd(fscal,cutoff_mask);
562 fscal = _mm256_andnot_pd(dummy_mask,fscal);
564 /* Calculate temporary vectorial force */
565 tx = _mm256_mul_pd(fscal,dx10);
566 ty = _mm256_mul_pd(fscal,dy10);
567 tz = _mm256_mul_pd(fscal,dz10);
569 /* Update vectorial force */
570 fix1 = _mm256_add_pd(fix1,tx);
571 fiy1 = _mm256_add_pd(fiy1,ty);
572 fiz1 = _mm256_add_pd(fiz1,tz);
574 fjx0 = _mm256_add_pd(fjx0,tx);
575 fjy0 = _mm256_add_pd(fjy0,ty);
576 fjz0 = _mm256_add_pd(fjz0,tz);
580 /**************************
581 * CALCULATE INTERACTIONS *
582 **************************/
584 if (gmx_mm256_any_lt(rsq20,rcutoff2))
587 r20 = _mm256_mul_pd(rsq20,rinv20);
588 r20 = _mm256_andnot_pd(dummy_mask,r20);
590 /* Compute parameters for interactions between i and j atoms */
591 qq20 = _mm256_mul_pd(iq2,jq0);
593 /* EWALD ELECTROSTATICS */
595 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
596 ewrt = _mm256_mul_pd(r20,ewtabscale);
597 ewitab = _mm256_cvttpd_epi32(ewrt);
598 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
599 ewitab = _mm_slli_epi32(ewitab,2);
600 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
601 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
602 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
603 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
604 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
605 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
606 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
607 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(_mm256_sub_pd(rinv20,sh_ewald),velec));
608 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
610 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
612 /* Update potential sum for this i atom from the interaction with this j atom. */
613 velec = _mm256_and_pd(velec,cutoff_mask);
614 velec = _mm256_andnot_pd(dummy_mask,velec);
615 velecsum = _mm256_add_pd(velecsum,velec);
619 fscal = _mm256_and_pd(fscal,cutoff_mask);
621 fscal = _mm256_andnot_pd(dummy_mask,fscal);
623 /* Calculate temporary vectorial force */
624 tx = _mm256_mul_pd(fscal,dx20);
625 ty = _mm256_mul_pd(fscal,dy20);
626 tz = _mm256_mul_pd(fscal,dz20);
628 /* Update vectorial force */
629 fix2 = _mm256_add_pd(fix2,tx);
630 fiy2 = _mm256_add_pd(fiy2,ty);
631 fiz2 = _mm256_add_pd(fiz2,tz);
633 fjx0 = _mm256_add_pd(fjx0,tx);
634 fjy0 = _mm256_add_pd(fjy0,ty);
635 fjz0 = _mm256_add_pd(fjz0,tz);
639 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
640 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
641 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
642 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
644 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
646 /* Inner loop uses 144 flops */
649 /* End of innermost loop */
651 gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
652 f+i_coord_offset,fshift+i_shift_offset);
655 /* Update potential energies */
656 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
658 /* Increment number of inner iterations */
659 inneriter += j_index_end - j_index_start;
661 /* Outer loop uses 19 flops */
664 /* Increment number of outer iterations */
667 /* Update outer/inner flops */
669 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_VF,outeriter*19 + inneriter*144);
672 * Gromacs nonbonded kernel: nb_kernel_ElecEwSh_VdwNone_GeomW3P1_F_avx_256_double
673 * Electrostatics interaction: Ewald
674 * VdW interaction: None
675 * Geometry: Water3-Particle
676 * Calculate force/pot: Force
679 nb_kernel_ElecEwSh_VdwNone_GeomW3P1_F_avx_256_double
680 (t_nblist * gmx_restrict nlist,
681 rvec * gmx_restrict xx,
682 rvec * gmx_restrict ff,
683 t_forcerec * gmx_restrict fr,
684 t_mdatoms * gmx_restrict mdatoms,
685 nb_kernel_data_t * gmx_restrict kernel_data,
686 t_nrnb * gmx_restrict nrnb)
688 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
689 * just 0 for non-waters.
690 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
691 * jnr indices corresponding to data put in the four positions in the SIMD register.
693 int i_shift_offset,i_coord_offset,outeriter,inneriter;
694 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
695 int jnrA,jnrB,jnrC,jnrD;
696 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
697 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
698 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
699 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
701 real *shiftvec,*fshift,*x,*f;
702 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
704 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
705 real * vdwioffsetptr0;
706 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
707 real * vdwioffsetptr1;
708 __m256d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
709 real * vdwioffsetptr2;
710 __m256d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
711 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
712 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
713 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
714 __m256d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
715 __m256d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
716 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
719 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
720 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
722 __m256d dummy_mask,cutoff_mask;
723 __m128 tmpmask0,tmpmask1;
724 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
725 __m256d one = _mm256_set1_pd(1.0);
726 __m256d two = _mm256_set1_pd(2.0);
732 jindex = nlist->jindex;
734 shiftidx = nlist->shift;
736 shiftvec = fr->shift_vec[0];
737 fshift = fr->fshift[0];
738 facel = _mm256_set1_pd(fr->epsfac);
739 charge = mdatoms->chargeA;
741 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
742 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
743 beta2 = _mm256_mul_pd(beta,beta);
744 beta3 = _mm256_mul_pd(beta,beta2);
746 ewtab = fr->ic->tabq_coul_F;
747 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
748 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
750 /* Setup water-specific parameters */
751 inr = nlist->iinr[0];
752 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
753 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
754 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
756 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
757 rcutoff_scalar = fr->rcoulomb;
758 rcutoff = _mm256_set1_pd(rcutoff_scalar);
759 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
761 /* Avoid stupid compiler warnings */
762 jnrA = jnrB = jnrC = jnrD = 0;
771 for(iidx=0;iidx<4*DIM;iidx++)
776 /* Start outer loop over neighborlists */
777 for(iidx=0; iidx<nri; iidx++)
779 /* Load shift vector for this list */
780 i_shift_offset = DIM*shiftidx[iidx];
782 /* Load limits for loop over neighbors */
783 j_index_start = jindex[iidx];
784 j_index_end = jindex[iidx+1];
786 /* Get outer coordinate index */
788 i_coord_offset = DIM*inr;
790 /* Load i particle coords and add shift vector */
791 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
792 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
794 fix0 = _mm256_setzero_pd();
795 fiy0 = _mm256_setzero_pd();
796 fiz0 = _mm256_setzero_pd();
797 fix1 = _mm256_setzero_pd();
798 fiy1 = _mm256_setzero_pd();
799 fiz1 = _mm256_setzero_pd();
800 fix2 = _mm256_setzero_pd();
801 fiy2 = _mm256_setzero_pd();
802 fiz2 = _mm256_setzero_pd();
804 /* Start inner kernel loop */
805 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
808 /* Get j neighbor index, and coordinate index */
813 j_coord_offsetA = DIM*jnrA;
814 j_coord_offsetB = DIM*jnrB;
815 j_coord_offsetC = DIM*jnrC;
816 j_coord_offsetD = DIM*jnrD;
818 /* load j atom coordinates */
819 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
820 x+j_coord_offsetC,x+j_coord_offsetD,
823 /* Calculate displacement vector */
824 dx00 = _mm256_sub_pd(ix0,jx0);
825 dy00 = _mm256_sub_pd(iy0,jy0);
826 dz00 = _mm256_sub_pd(iz0,jz0);
827 dx10 = _mm256_sub_pd(ix1,jx0);
828 dy10 = _mm256_sub_pd(iy1,jy0);
829 dz10 = _mm256_sub_pd(iz1,jz0);
830 dx20 = _mm256_sub_pd(ix2,jx0);
831 dy20 = _mm256_sub_pd(iy2,jy0);
832 dz20 = _mm256_sub_pd(iz2,jz0);
834 /* Calculate squared distance and things based on it */
835 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
836 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
837 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
839 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
840 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
841 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
843 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
844 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
845 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
847 /* Load parameters for j particles */
848 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
849 charge+jnrC+0,charge+jnrD+0);
851 fjx0 = _mm256_setzero_pd();
852 fjy0 = _mm256_setzero_pd();
853 fjz0 = _mm256_setzero_pd();
855 /**************************
856 * CALCULATE INTERACTIONS *
857 **************************/
859 if (gmx_mm256_any_lt(rsq00,rcutoff2))
862 r00 = _mm256_mul_pd(rsq00,rinv00);
864 /* Compute parameters for interactions between i and j atoms */
865 qq00 = _mm256_mul_pd(iq0,jq0);
867 /* EWALD ELECTROSTATICS */
869 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
870 ewrt = _mm256_mul_pd(r00,ewtabscale);
871 ewitab = _mm256_cvttpd_epi32(ewrt);
872 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
873 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
874 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
876 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
877 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
879 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
883 fscal = _mm256_and_pd(fscal,cutoff_mask);
885 /* Calculate temporary vectorial force */
886 tx = _mm256_mul_pd(fscal,dx00);
887 ty = _mm256_mul_pd(fscal,dy00);
888 tz = _mm256_mul_pd(fscal,dz00);
890 /* Update vectorial force */
891 fix0 = _mm256_add_pd(fix0,tx);
892 fiy0 = _mm256_add_pd(fiy0,ty);
893 fiz0 = _mm256_add_pd(fiz0,tz);
895 fjx0 = _mm256_add_pd(fjx0,tx);
896 fjy0 = _mm256_add_pd(fjy0,ty);
897 fjz0 = _mm256_add_pd(fjz0,tz);
901 /**************************
902 * CALCULATE INTERACTIONS *
903 **************************/
905 if (gmx_mm256_any_lt(rsq10,rcutoff2))
908 r10 = _mm256_mul_pd(rsq10,rinv10);
910 /* Compute parameters for interactions between i and j atoms */
911 qq10 = _mm256_mul_pd(iq1,jq0);
913 /* EWALD ELECTROSTATICS */
915 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
916 ewrt = _mm256_mul_pd(r10,ewtabscale);
917 ewitab = _mm256_cvttpd_epi32(ewrt);
918 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
919 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
920 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
922 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
923 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
925 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
929 fscal = _mm256_and_pd(fscal,cutoff_mask);
931 /* Calculate temporary vectorial force */
932 tx = _mm256_mul_pd(fscal,dx10);
933 ty = _mm256_mul_pd(fscal,dy10);
934 tz = _mm256_mul_pd(fscal,dz10);
936 /* Update vectorial force */
937 fix1 = _mm256_add_pd(fix1,tx);
938 fiy1 = _mm256_add_pd(fiy1,ty);
939 fiz1 = _mm256_add_pd(fiz1,tz);
941 fjx0 = _mm256_add_pd(fjx0,tx);
942 fjy0 = _mm256_add_pd(fjy0,ty);
943 fjz0 = _mm256_add_pd(fjz0,tz);
947 /**************************
948 * CALCULATE INTERACTIONS *
949 **************************/
951 if (gmx_mm256_any_lt(rsq20,rcutoff2))
954 r20 = _mm256_mul_pd(rsq20,rinv20);
956 /* Compute parameters for interactions between i and j atoms */
957 qq20 = _mm256_mul_pd(iq2,jq0);
959 /* EWALD ELECTROSTATICS */
961 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
962 ewrt = _mm256_mul_pd(r20,ewtabscale);
963 ewitab = _mm256_cvttpd_epi32(ewrt);
964 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
965 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
966 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
968 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
969 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
971 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
975 fscal = _mm256_and_pd(fscal,cutoff_mask);
977 /* Calculate temporary vectorial force */
978 tx = _mm256_mul_pd(fscal,dx20);
979 ty = _mm256_mul_pd(fscal,dy20);
980 tz = _mm256_mul_pd(fscal,dz20);
982 /* Update vectorial force */
983 fix2 = _mm256_add_pd(fix2,tx);
984 fiy2 = _mm256_add_pd(fiy2,ty);
985 fiz2 = _mm256_add_pd(fiz2,tz);
987 fjx0 = _mm256_add_pd(fjx0,tx);
988 fjy0 = _mm256_add_pd(fjy0,ty);
989 fjz0 = _mm256_add_pd(fjz0,tz);
993 fjptrA = f+j_coord_offsetA;
994 fjptrB = f+j_coord_offsetB;
995 fjptrC = f+j_coord_offsetC;
996 fjptrD = f+j_coord_offsetD;
998 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1000 /* Inner loop uses 120 flops */
1003 if(jidx<j_index_end)
1006 /* Get j neighbor index, and coordinate index */
1007 jnrlistA = jjnr[jidx];
1008 jnrlistB = jjnr[jidx+1];
1009 jnrlistC = jjnr[jidx+2];
1010 jnrlistD = jjnr[jidx+3];
1011 /* Sign of each element will be negative for non-real atoms.
1012 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1013 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
1015 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1017 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
1018 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
1019 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
1021 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1022 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1023 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1024 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1025 j_coord_offsetA = DIM*jnrA;
1026 j_coord_offsetB = DIM*jnrB;
1027 j_coord_offsetC = DIM*jnrC;
1028 j_coord_offsetD = DIM*jnrD;
1030 /* load j atom coordinates */
1031 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1032 x+j_coord_offsetC,x+j_coord_offsetD,
1035 /* Calculate displacement vector */
1036 dx00 = _mm256_sub_pd(ix0,jx0);
1037 dy00 = _mm256_sub_pd(iy0,jy0);
1038 dz00 = _mm256_sub_pd(iz0,jz0);
1039 dx10 = _mm256_sub_pd(ix1,jx0);
1040 dy10 = _mm256_sub_pd(iy1,jy0);
1041 dz10 = _mm256_sub_pd(iz1,jz0);
1042 dx20 = _mm256_sub_pd(ix2,jx0);
1043 dy20 = _mm256_sub_pd(iy2,jy0);
1044 dz20 = _mm256_sub_pd(iz2,jz0);
1046 /* Calculate squared distance and things based on it */
1047 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
1048 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
1049 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
1051 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
1052 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
1053 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
1055 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
1056 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
1057 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
1059 /* Load parameters for j particles */
1060 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
1061 charge+jnrC+0,charge+jnrD+0);
1063 fjx0 = _mm256_setzero_pd();
1064 fjy0 = _mm256_setzero_pd();
1065 fjz0 = _mm256_setzero_pd();
1067 /**************************
1068 * CALCULATE INTERACTIONS *
1069 **************************/
1071 if (gmx_mm256_any_lt(rsq00,rcutoff2))
1074 r00 = _mm256_mul_pd(rsq00,rinv00);
1075 r00 = _mm256_andnot_pd(dummy_mask,r00);
1077 /* Compute parameters for interactions between i and j atoms */
1078 qq00 = _mm256_mul_pd(iq0,jq0);
1080 /* EWALD ELECTROSTATICS */
1082 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1083 ewrt = _mm256_mul_pd(r00,ewtabscale);
1084 ewitab = _mm256_cvttpd_epi32(ewrt);
1085 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1086 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1087 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1089 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1090 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
1092 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
1096 fscal = _mm256_and_pd(fscal,cutoff_mask);
1098 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1100 /* Calculate temporary vectorial force */
1101 tx = _mm256_mul_pd(fscal,dx00);
1102 ty = _mm256_mul_pd(fscal,dy00);
1103 tz = _mm256_mul_pd(fscal,dz00);
1105 /* Update vectorial force */
1106 fix0 = _mm256_add_pd(fix0,tx);
1107 fiy0 = _mm256_add_pd(fiy0,ty);
1108 fiz0 = _mm256_add_pd(fiz0,tz);
1110 fjx0 = _mm256_add_pd(fjx0,tx);
1111 fjy0 = _mm256_add_pd(fjy0,ty);
1112 fjz0 = _mm256_add_pd(fjz0,tz);
1116 /**************************
1117 * CALCULATE INTERACTIONS *
1118 **************************/
1120 if (gmx_mm256_any_lt(rsq10,rcutoff2))
1123 r10 = _mm256_mul_pd(rsq10,rinv10);
1124 r10 = _mm256_andnot_pd(dummy_mask,r10);
1126 /* Compute parameters for interactions between i and j atoms */
1127 qq10 = _mm256_mul_pd(iq1,jq0);
1129 /* EWALD ELECTROSTATICS */
1131 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1132 ewrt = _mm256_mul_pd(r10,ewtabscale);
1133 ewitab = _mm256_cvttpd_epi32(ewrt);
1134 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1135 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1136 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1138 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1139 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
1141 cutoff_mask = _mm256_cmp_pd(rsq10,rcutoff2,_CMP_LT_OQ);
1145 fscal = _mm256_and_pd(fscal,cutoff_mask);
1147 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1149 /* Calculate temporary vectorial force */
1150 tx = _mm256_mul_pd(fscal,dx10);
1151 ty = _mm256_mul_pd(fscal,dy10);
1152 tz = _mm256_mul_pd(fscal,dz10);
1154 /* Update vectorial force */
1155 fix1 = _mm256_add_pd(fix1,tx);
1156 fiy1 = _mm256_add_pd(fiy1,ty);
1157 fiz1 = _mm256_add_pd(fiz1,tz);
1159 fjx0 = _mm256_add_pd(fjx0,tx);
1160 fjy0 = _mm256_add_pd(fjy0,ty);
1161 fjz0 = _mm256_add_pd(fjz0,tz);
1165 /**************************
1166 * CALCULATE INTERACTIONS *
1167 **************************/
1169 if (gmx_mm256_any_lt(rsq20,rcutoff2))
1172 r20 = _mm256_mul_pd(rsq20,rinv20);
1173 r20 = _mm256_andnot_pd(dummy_mask,r20);
1175 /* Compute parameters for interactions between i and j atoms */
1176 qq20 = _mm256_mul_pd(iq2,jq0);
1178 /* EWALD ELECTROSTATICS */
1180 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1181 ewrt = _mm256_mul_pd(r20,ewtabscale);
1182 ewitab = _mm256_cvttpd_epi32(ewrt);
1183 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1184 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1185 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1187 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1188 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
1190 cutoff_mask = _mm256_cmp_pd(rsq20,rcutoff2,_CMP_LT_OQ);
1194 fscal = _mm256_and_pd(fscal,cutoff_mask);
1196 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1198 /* Calculate temporary vectorial force */
1199 tx = _mm256_mul_pd(fscal,dx20);
1200 ty = _mm256_mul_pd(fscal,dy20);
1201 tz = _mm256_mul_pd(fscal,dz20);
1203 /* Update vectorial force */
1204 fix2 = _mm256_add_pd(fix2,tx);
1205 fiy2 = _mm256_add_pd(fiy2,ty);
1206 fiz2 = _mm256_add_pd(fiz2,tz);
1208 fjx0 = _mm256_add_pd(fjx0,tx);
1209 fjy0 = _mm256_add_pd(fjy0,ty);
1210 fjz0 = _mm256_add_pd(fjz0,tz);
1214 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1215 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1216 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1217 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1219 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1221 /* Inner loop uses 123 flops */
1224 /* End of innermost loop */
1226 gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1227 f+i_coord_offset,fshift+i_shift_offset);
1229 /* Increment number of inner iterations */
1230 inneriter += j_index_end - j_index_start;
1232 /* Outer loop uses 18 flops */
1235 /* Increment number of outer iterations */
1238 /* Update outer/inner flops */
1240 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_F,outeriter*18 + inneriter*123);