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_ElecEw_VdwLJ_GeomW3P1_VF_avx_256_double
38 * Electrostatics interaction: Ewald
39 * VdW interaction: LennardJones
40 * Geometry: Water3-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEw_VdwLJ_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 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
87 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
88 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
90 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
91 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
93 __m256d dummy_mask,cutoff_mask;
94 __m128 tmpmask0,tmpmask1;
95 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
96 __m256d one = _mm256_set1_pd(1.0);
97 __m256d two = _mm256_set1_pd(2.0);
103 jindex = nlist->jindex;
105 shiftidx = nlist->shift;
107 shiftvec = fr->shift_vec[0];
108 fshift = fr->fshift[0];
109 facel = _mm256_set1_pd(fr->epsfac);
110 charge = mdatoms->chargeA;
111 nvdwtype = fr->ntype;
113 vdwtype = mdatoms->typeA;
115 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
116 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
117 beta2 = _mm256_mul_pd(beta,beta);
118 beta3 = _mm256_mul_pd(beta,beta2);
120 ewtab = fr->ic->tabq_coul_FDV0;
121 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
122 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
124 /* Setup water-specific parameters */
125 inr = nlist->iinr[0];
126 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
127 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
128 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
129 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
131 /* Avoid stupid compiler warnings */
132 jnrA = jnrB = jnrC = jnrD = 0;
141 for(iidx=0;iidx<4*DIM;iidx++)
146 /* Start outer loop over neighborlists */
147 for(iidx=0; iidx<nri; iidx++)
149 /* Load shift vector for this list */
150 i_shift_offset = DIM*shiftidx[iidx];
152 /* Load limits for loop over neighbors */
153 j_index_start = jindex[iidx];
154 j_index_end = jindex[iidx+1];
156 /* Get outer coordinate index */
158 i_coord_offset = DIM*inr;
160 /* Load i particle coords and add shift vector */
161 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
162 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
164 fix0 = _mm256_setzero_pd();
165 fiy0 = _mm256_setzero_pd();
166 fiz0 = _mm256_setzero_pd();
167 fix1 = _mm256_setzero_pd();
168 fiy1 = _mm256_setzero_pd();
169 fiz1 = _mm256_setzero_pd();
170 fix2 = _mm256_setzero_pd();
171 fiy2 = _mm256_setzero_pd();
172 fiz2 = _mm256_setzero_pd();
174 /* Reset potential sums */
175 velecsum = _mm256_setzero_pd();
176 vvdwsum = _mm256_setzero_pd();
178 /* Start inner kernel loop */
179 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
182 /* Get j neighbor index, and coordinate index */
187 j_coord_offsetA = DIM*jnrA;
188 j_coord_offsetB = DIM*jnrB;
189 j_coord_offsetC = DIM*jnrC;
190 j_coord_offsetD = DIM*jnrD;
192 /* load j atom coordinates */
193 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
194 x+j_coord_offsetC,x+j_coord_offsetD,
197 /* Calculate displacement vector */
198 dx00 = _mm256_sub_pd(ix0,jx0);
199 dy00 = _mm256_sub_pd(iy0,jy0);
200 dz00 = _mm256_sub_pd(iz0,jz0);
201 dx10 = _mm256_sub_pd(ix1,jx0);
202 dy10 = _mm256_sub_pd(iy1,jy0);
203 dz10 = _mm256_sub_pd(iz1,jz0);
204 dx20 = _mm256_sub_pd(ix2,jx0);
205 dy20 = _mm256_sub_pd(iy2,jy0);
206 dz20 = _mm256_sub_pd(iz2,jz0);
208 /* Calculate squared distance and things based on it */
209 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
210 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
211 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
213 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
214 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
215 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
217 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
218 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
219 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
221 /* Load parameters for j particles */
222 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
223 charge+jnrC+0,charge+jnrD+0);
224 vdwjidx0A = 2*vdwtype[jnrA+0];
225 vdwjidx0B = 2*vdwtype[jnrB+0];
226 vdwjidx0C = 2*vdwtype[jnrC+0];
227 vdwjidx0D = 2*vdwtype[jnrD+0];
229 fjx0 = _mm256_setzero_pd();
230 fjy0 = _mm256_setzero_pd();
231 fjz0 = _mm256_setzero_pd();
233 /**************************
234 * CALCULATE INTERACTIONS *
235 **************************/
237 r00 = _mm256_mul_pd(rsq00,rinv00);
239 /* Compute parameters for interactions between i and j atoms */
240 qq00 = _mm256_mul_pd(iq0,jq0);
241 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
242 vdwioffsetptr0+vdwjidx0B,
243 vdwioffsetptr0+vdwjidx0C,
244 vdwioffsetptr0+vdwjidx0D,
247 /* EWALD ELECTROSTATICS */
249 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
250 ewrt = _mm256_mul_pd(r00,ewtabscale);
251 ewitab = _mm256_cvttpd_epi32(ewrt);
252 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
253 ewitab = _mm_slli_epi32(ewitab,2);
254 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
255 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
256 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
257 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
258 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
259 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
260 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
261 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
262 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
264 /* LENNARD-JONES DISPERSION/REPULSION */
266 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
267 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
268 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
269 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
270 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
272 /* Update potential sum for this i atom from the interaction with this j atom. */
273 velecsum = _mm256_add_pd(velecsum,velec);
274 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
276 fscal = _mm256_add_pd(felec,fvdw);
278 /* Calculate temporary vectorial force */
279 tx = _mm256_mul_pd(fscal,dx00);
280 ty = _mm256_mul_pd(fscal,dy00);
281 tz = _mm256_mul_pd(fscal,dz00);
283 /* Update vectorial force */
284 fix0 = _mm256_add_pd(fix0,tx);
285 fiy0 = _mm256_add_pd(fiy0,ty);
286 fiz0 = _mm256_add_pd(fiz0,tz);
288 fjx0 = _mm256_add_pd(fjx0,tx);
289 fjy0 = _mm256_add_pd(fjy0,ty);
290 fjz0 = _mm256_add_pd(fjz0,tz);
292 /**************************
293 * CALCULATE INTERACTIONS *
294 **************************/
296 r10 = _mm256_mul_pd(rsq10,rinv10);
298 /* Compute parameters for interactions between i and j atoms */
299 qq10 = _mm256_mul_pd(iq1,jq0);
301 /* EWALD ELECTROSTATICS */
303 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
304 ewrt = _mm256_mul_pd(r10,ewtabscale);
305 ewitab = _mm256_cvttpd_epi32(ewrt);
306 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
307 ewitab = _mm_slli_epi32(ewitab,2);
308 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
309 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
310 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
311 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
312 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
313 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
314 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
315 velec = _mm256_mul_pd(qq10,_mm256_sub_pd(rinv10,velec));
316 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
318 /* Update potential sum for this i atom from the interaction with this j atom. */
319 velecsum = _mm256_add_pd(velecsum,velec);
323 /* Calculate temporary vectorial force */
324 tx = _mm256_mul_pd(fscal,dx10);
325 ty = _mm256_mul_pd(fscal,dy10);
326 tz = _mm256_mul_pd(fscal,dz10);
328 /* Update vectorial force */
329 fix1 = _mm256_add_pd(fix1,tx);
330 fiy1 = _mm256_add_pd(fiy1,ty);
331 fiz1 = _mm256_add_pd(fiz1,tz);
333 fjx0 = _mm256_add_pd(fjx0,tx);
334 fjy0 = _mm256_add_pd(fjy0,ty);
335 fjz0 = _mm256_add_pd(fjz0,tz);
337 /**************************
338 * CALCULATE INTERACTIONS *
339 **************************/
341 r20 = _mm256_mul_pd(rsq20,rinv20);
343 /* Compute parameters for interactions between i and j atoms */
344 qq20 = _mm256_mul_pd(iq2,jq0);
346 /* EWALD ELECTROSTATICS */
348 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
349 ewrt = _mm256_mul_pd(r20,ewtabscale);
350 ewitab = _mm256_cvttpd_epi32(ewrt);
351 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
352 ewitab = _mm_slli_epi32(ewitab,2);
353 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
354 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
355 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
356 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
357 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
358 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
359 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
360 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(rinv20,velec));
361 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
363 /* Update potential sum for this i atom from the interaction with this j atom. */
364 velecsum = _mm256_add_pd(velecsum,velec);
368 /* Calculate temporary vectorial force */
369 tx = _mm256_mul_pd(fscal,dx20);
370 ty = _mm256_mul_pd(fscal,dy20);
371 tz = _mm256_mul_pd(fscal,dz20);
373 /* Update vectorial force */
374 fix2 = _mm256_add_pd(fix2,tx);
375 fiy2 = _mm256_add_pd(fiy2,ty);
376 fiz2 = _mm256_add_pd(fiz2,tz);
378 fjx0 = _mm256_add_pd(fjx0,tx);
379 fjy0 = _mm256_add_pd(fjy0,ty);
380 fjz0 = _mm256_add_pd(fjz0,tz);
382 fjptrA = f+j_coord_offsetA;
383 fjptrB = f+j_coord_offsetB;
384 fjptrC = f+j_coord_offsetC;
385 fjptrD = f+j_coord_offsetD;
387 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
389 /* Inner loop uses 138 flops */
395 /* Get j neighbor index, and coordinate index */
396 jnrlistA = jjnr[jidx];
397 jnrlistB = jjnr[jidx+1];
398 jnrlistC = jjnr[jidx+2];
399 jnrlistD = jjnr[jidx+3];
400 /* Sign of each element will be negative for non-real atoms.
401 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
402 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
404 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
406 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
407 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
408 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
410 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
411 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
412 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
413 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
414 j_coord_offsetA = DIM*jnrA;
415 j_coord_offsetB = DIM*jnrB;
416 j_coord_offsetC = DIM*jnrC;
417 j_coord_offsetD = DIM*jnrD;
419 /* load j atom coordinates */
420 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
421 x+j_coord_offsetC,x+j_coord_offsetD,
424 /* Calculate displacement vector */
425 dx00 = _mm256_sub_pd(ix0,jx0);
426 dy00 = _mm256_sub_pd(iy0,jy0);
427 dz00 = _mm256_sub_pd(iz0,jz0);
428 dx10 = _mm256_sub_pd(ix1,jx0);
429 dy10 = _mm256_sub_pd(iy1,jy0);
430 dz10 = _mm256_sub_pd(iz1,jz0);
431 dx20 = _mm256_sub_pd(ix2,jx0);
432 dy20 = _mm256_sub_pd(iy2,jy0);
433 dz20 = _mm256_sub_pd(iz2,jz0);
435 /* Calculate squared distance and things based on it */
436 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
437 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
438 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
440 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
441 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
442 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
444 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
445 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
446 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
448 /* Load parameters for j particles */
449 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
450 charge+jnrC+0,charge+jnrD+0);
451 vdwjidx0A = 2*vdwtype[jnrA+0];
452 vdwjidx0B = 2*vdwtype[jnrB+0];
453 vdwjidx0C = 2*vdwtype[jnrC+0];
454 vdwjidx0D = 2*vdwtype[jnrD+0];
456 fjx0 = _mm256_setzero_pd();
457 fjy0 = _mm256_setzero_pd();
458 fjz0 = _mm256_setzero_pd();
460 /**************************
461 * CALCULATE INTERACTIONS *
462 **************************/
464 r00 = _mm256_mul_pd(rsq00,rinv00);
465 r00 = _mm256_andnot_pd(dummy_mask,r00);
467 /* Compute parameters for interactions between i and j atoms */
468 qq00 = _mm256_mul_pd(iq0,jq0);
469 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
470 vdwioffsetptr0+vdwjidx0B,
471 vdwioffsetptr0+vdwjidx0C,
472 vdwioffsetptr0+vdwjidx0D,
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(rinv00,velec));
490 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
492 /* LENNARD-JONES DISPERSION/REPULSION */
494 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
495 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
496 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
497 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
498 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
500 /* Update potential sum for this i atom from the interaction with this j atom. */
501 velec = _mm256_andnot_pd(dummy_mask,velec);
502 velecsum = _mm256_add_pd(velecsum,velec);
503 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
504 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
506 fscal = _mm256_add_pd(felec,fvdw);
508 fscal = _mm256_andnot_pd(dummy_mask,fscal);
510 /* Calculate temporary vectorial force */
511 tx = _mm256_mul_pd(fscal,dx00);
512 ty = _mm256_mul_pd(fscal,dy00);
513 tz = _mm256_mul_pd(fscal,dz00);
515 /* Update vectorial force */
516 fix0 = _mm256_add_pd(fix0,tx);
517 fiy0 = _mm256_add_pd(fiy0,ty);
518 fiz0 = _mm256_add_pd(fiz0,tz);
520 fjx0 = _mm256_add_pd(fjx0,tx);
521 fjy0 = _mm256_add_pd(fjy0,ty);
522 fjz0 = _mm256_add_pd(fjz0,tz);
524 /**************************
525 * CALCULATE INTERACTIONS *
526 **************************/
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(rinv10,velec));
549 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
551 /* Update potential sum for this i atom from the interaction with this j atom. */
552 velec = _mm256_andnot_pd(dummy_mask,velec);
553 velecsum = _mm256_add_pd(velecsum,velec);
557 fscal = _mm256_andnot_pd(dummy_mask,fscal);
559 /* Calculate temporary vectorial force */
560 tx = _mm256_mul_pd(fscal,dx10);
561 ty = _mm256_mul_pd(fscal,dy10);
562 tz = _mm256_mul_pd(fscal,dz10);
564 /* Update vectorial force */
565 fix1 = _mm256_add_pd(fix1,tx);
566 fiy1 = _mm256_add_pd(fiy1,ty);
567 fiz1 = _mm256_add_pd(fiz1,tz);
569 fjx0 = _mm256_add_pd(fjx0,tx);
570 fjy0 = _mm256_add_pd(fjy0,ty);
571 fjz0 = _mm256_add_pd(fjz0,tz);
573 /**************************
574 * CALCULATE INTERACTIONS *
575 **************************/
577 r20 = _mm256_mul_pd(rsq20,rinv20);
578 r20 = _mm256_andnot_pd(dummy_mask,r20);
580 /* Compute parameters for interactions between i and j atoms */
581 qq20 = _mm256_mul_pd(iq2,jq0);
583 /* EWALD ELECTROSTATICS */
585 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
586 ewrt = _mm256_mul_pd(r20,ewtabscale);
587 ewitab = _mm256_cvttpd_epi32(ewrt);
588 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
589 ewitab = _mm_slli_epi32(ewitab,2);
590 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
591 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
592 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
593 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
594 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
595 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
596 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
597 velec = _mm256_mul_pd(qq20,_mm256_sub_pd(rinv20,velec));
598 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
600 /* Update potential sum for this i atom from the interaction with this j atom. */
601 velec = _mm256_andnot_pd(dummy_mask,velec);
602 velecsum = _mm256_add_pd(velecsum,velec);
606 fscal = _mm256_andnot_pd(dummy_mask,fscal);
608 /* Calculate temporary vectorial force */
609 tx = _mm256_mul_pd(fscal,dx20);
610 ty = _mm256_mul_pd(fscal,dy20);
611 tz = _mm256_mul_pd(fscal,dz20);
613 /* Update vectorial force */
614 fix2 = _mm256_add_pd(fix2,tx);
615 fiy2 = _mm256_add_pd(fiy2,ty);
616 fiz2 = _mm256_add_pd(fiz2,tz);
618 fjx0 = _mm256_add_pd(fjx0,tx);
619 fjy0 = _mm256_add_pd(fjy0,ty);
620 fjz0 = _mm256_add_pd(fjz0,tz);
622 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
623 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
624 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
625 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
627 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
629 /* Inner loop uses 141 flops */
632 /* End of innermost loop */
634 gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
635 f+i_coord_offset,fshift+i_shift_offset);
638 /* Update potential energies */
639 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
640 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
642 /* Increment number of inner iterations */
643 inneriter += j_index_end - j_index_start;
645 /* Outer loop uses 20 flops */
648 /* Increment number of outer iterations */
651 /* Update outer/inner flops */
653 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*141);
656 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwLJ_GeomW3P1_F_avx_256_double
657 * Electrostatics interaction: Ewald
658 * VdW interaction: LennardJones
659 * Geometry: Water3-Particle
660 * Calculate force/pot: Force
663 nb_kernel_ElecEw_VdwLJ_GeomW3P1_F_avx_256_double
664 (t_nblist * gmx_restrict nlist,
665 rvec * gmx_restrict xx,
666 rvec * gmx_restrict ff,
667 t_forcerec * gmx_restrict fr,
668 t_mdatoms * gmx_restrict mdatoms,
669 nb_kernel_data_t * gmx_restrict kernel_data,
670 t_nrnb * gmx_restrict nrnb)
672 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
673 * just 0 for non-waters.
674 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
675 * jnr indices corresponding to data put in the four positions in the SIMD register.
677 int i_shift_offset,i_coord_offset,outeriter,inneriter;
678 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
679 int jnrA,jnrB,jnrC,jnrD;
680 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
681 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
682 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
683 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
685 real *shiftvec,*fshift,*x,*f;
686 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
688 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
689 real * vdwioffsetptr0;
690 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
691 real * vdwioffsetptr1;
692 __m256d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
693 real * vdwioffsetptr2;
694 __m256d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
695 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
696 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
697 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
698 __m256d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
699 __m256d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
700 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
703 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
706 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
707 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
709 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
710 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
712 __m256d dummy_mask,cutoff_mask;
713 __m128 tmpmask0,tmpmask1;
714 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
715 __m256d one = _mm256_set1_pd(1.0);
716 __m256d two = _mm256_set1_pd(2.0);
722 jindex = nlist->jindex;
724 shiftidx = nlist->shift;
726 shiftvec = fr->shift_vec[0];
727 fshift = fr->fshift[0];
728 facel = _mm256_set1_pd(fr->epsfac);
729 charge = mdatoms->chargeA;
730 nvdwtype = fr->ntype;
732 vdwtype = mdatoms->typeA;
734 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
735 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
736 beta2 = _mm256_mul_pd(beta,beta);
737 beta3 = _mm256_mul_pd(beta,beta2);
739 ewtab = fr->ic->tabq_coul_F;
740 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
741 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
743 /* Setup water-specific parameters */
744 inr = nlist->iinr[0];
745 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
746 iq1 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+1]));
747 iq2 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+2]));
748 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
750 /* Avoid stupid compiler warnings */
751 jnrA = jnrB = jnrC = jnrD = 0;
760 for(iidx=0;iidx<4*DIM;iidx++)
765 /* Start outer loop over neighborlists */
766 for(iidx=0; iidx<nri; iidx++)
768 /* Load shift vector for this list */
769 i_shift_offset = DIM*shiftidx[iidx];
771 /* Load limits for loop over neighbors */
772 j_index_start = jindex[iidx];
773 j_index_end = jindex[iidx+1];
775 /* Get outer coordinate index */
777 i_coord_offset = DIM*inr;
779 /* Load i particle coords and add shift vector */
780 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
781 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
783 fix0 = _mm256_setzero_pd();
784 fiy0 = _mm256_setzero_pd();
785 fiz0 = _mm256_setzero_pd();
786 fix1 = _mm256_setzero_pd();
787 fiy1 = _mm256_setzero_pd();
788 fiz1 = _mm256_setzero_pd();
789 fix2 = _mm256_setzero_pd();
790 fiy2 = _mm256_setzero_pd();
791 fiz2 = _mm256_setzero_pd();
793 /* Start inner kernel loop */
794 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
797 /* Get j neighbor index, and coordinate index */
802 j_coord_offsetA = DIM*jnrA;
803 j_coord_offsetB = DIM*jnrB;
804 j_coord_offsetC = DIM*jnrC;
805 j_coord_offsetD = DIM*jnrD;
807 /* load j atom coordinates */
808 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
809 x+j_coord_offsetC,x+j_coord_offsetD,
812 /* Calculate displacement vector */
813 dx00 = _mm256_sub_pd(ix0,jx0);
814 dy00 = _mm256_sub_pd(iy0,jy0);
815 dz00 = _mm256_sub_pd(iz0,jz0);
816 dx10 = _mm256_sub_pd(ix1,jx0);
817 dy10 = _mm256_sub_pd(iy1,jy0);
818 dz10 = _mm256_sub_pd(iz1,jz0);
819 dx20 = _mm256_sub_pd(ix2,jx0);
820 dy20 = _mm256_sub_pd(iy2,jy0);
821 dz20 = _mm256_sub_pd(iz2,jz0);
823 /* Calculate squared distance and things based on it */
824 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
825 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
826 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
828 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
829 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
830 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
832 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
833 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
834 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
836 /* Load parameters for j particles */
837 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
838 charge+jnrC+0,charge+jnrD+0);
839 vdwjidx0A = 2*vdwtype[jnrA+0];
840 vdwjidx0B = 2*vdwtype[jnrB+0];
841 vdwjidx0C = 2*vdwtype[jnrC+0];
842 vdwjidx0D = 2*vdwtype[jnrD+0];
844 fjx0 = _mm256_setzero_pd();
845 fjy0 = _mm256_setzero_pd();
846 fjz0 = _mm256_setzero_pd();
848 /**************************
849 * CALCULATE INTERACTIONS *
850 **************************/
852 r00 = _mm256_mul_pd(rsq00,rinv00);
854 /* Compute parameters for interactions between i and j atoms */
855 qq00 = _mm256_mul_pd(iq0,jq0);
856 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
857 vdwioffsetptr0+vdwjidx0B,
858 vdwioffsetptr0+vdwjidx0C,
859 vdwioffsetptr0+vdwjidx0D,
862 /* EWALD ELECTROSTATICS */
864 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
865 ewrt = _mm256_mul_pd(r00,ewtabscale);
866 ewitab = _mm256_cvttpd_epi32(ewrt);
867 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
868 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
869 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
871 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
872 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
874 /* LENNARD-JONES DISPERSION/REPULSION */
876 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
877 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
879 fscal = _mm256_add_pd(felec,fvdw);
881 /* Calculate temporary vectorial force */
882 tx = _mm256_mul_pd(fscal,dx00);
883 ty = _mm256_mul_pd(fscal,dy00);
884 tz = _mm256_mul_pd(fscal,dz00);
886 /* Update vectorial force */
887 fix0 = _mm256_add_pd(fix0,tx);
888 fiy0 = _mm256_add_pd(fiy0,ty);
889 fiz0 = _mm256_add_pd(fiz0,tz);
891 fjx0 = _mm256_add_pd(fjx0,tx);
892 fjy0 = _mm256_add_pd(fjy0,ty);
893 fjz0 = _mm256_add_pd(fjz0,tz);
895 /**************************
896 * CALCULATE INTERACTIONS *
897 **************************/
899 r10 = _mm256_mul_pd(rsq10,rinv10);
901 /* Compute parameters for interactions between i and j atoms */
902 qq10 = _mm256_mul_pd(iq1,jq0);
904 /* EWALD ELECTROSTATICS */
906 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
907 ewrt = _mm256_mul_pd(r10,ewtabscale);
908 ewitab = _mm256_cvttpd_epi32(ewrt);
909 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
910 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
911 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
913 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
914 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
918 /* Calculate temporary vectorial force */
919 tx = _mm256_mul_pd(fscal,dx10);
920 ty = _mm256_mul_pd(fscal,dy10);
921 tz = _mm256_mul_pd(fscal,dz10);
923 /* Update vectorial force */
924 fix1 = _mm256_add_pd(fix1,tx);
925 fiy1 = _mm256_add_pd(fiy1,ty);
926 fiz1 = _mm256_add_pd(fiz1,tz);
928 fjx0 = _mm256_add_pd(fjx0,tx);
929 fjy0 = _mm256_add_pd(fjy0,ty);
930 fjz0 = _mm256_add_pd(fjz0,tz);
932 /**************************
933 * CALCULATE INTERACTIONS *
934 **************************/
936 r20 = _mm256_mul_pd(rsq20,rinv20);
938 /* Compute parameters for interactions between i and j atoms */
939 qq20 = _mm256_mul_pd(iq2,jq0);
941 /* EWALD ELECTROSTATICS */
943 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
944 ewrt = _mm256_mul_pd(r20,ewtabscale);
945 ewitab = _mm256_cvttpd_epi32(ewrt);
946 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
947 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
948 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
950 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
951 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
955 /* Calculate temporary vectorial force */
956 tx = _mm256_mul_pd(fscal,dx20);
957 ty = _mm256_mul_pd(fscal,dy20);
958 tz = _mm256_mul_pd(fscal,dz20);
960 /* Update vectorial force */
961 fix2 = _mm256_add_pd(fix2,tx);
962 fiy2 = _mm256_add_pd(fiy2,ty);
963 fiz2 = _mm256_add_pd(fiz2,tz);
965 fjx0 = _mm256_add_pd(fjx0,tx);
966 fjy0 = _mm256_add_pd(fjy0,ty);
967 fjz0 = _mm256_add_pd(fjz0,tz);
969 fjptrA = f+j_coord_offsetA;
970 fjptrB = f+j_coord_offsetB;
971 fjptrC = f+j_coord_offsetC;
972 fjptrD = f+j_coord_offsetD;
974 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
976 /* Inner loop uses 118 flops */
982 /* Get j neighbor index, and coordinate index */
983 jnrlistA = jjnr[jidx];
984 jnrlistB = jjnr[jidx+1];
985 jnrlistC = jjnr[jidx+2];
986 jnrlistD = jjnr[jidx+3];
987 /* Sign of each element will be negative for non-real atoms.
988 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
989 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
991 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
993 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
994 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
995 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
997 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
998 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
999 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1000 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1001 j_coord_offsetA = DIM*jnrA;
1002 j_coord_offsetB = DIM*jnrB;
1003 j_coord_offsetC = DIM*jnrC;
1004 j_coord_offsetD = DIM*jnrD;
1006 /* load j atom coordinates */
1007 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1008 x+j_coord_offsetC,x+j_coord_offsetD,
1011 /* Calculate displacement vector */
1012 dx00 = _mm256_sub_pd(ix0,jx0);
1013 dy00 = _mm256_sub_pd(iy0,jy0);
1014 dz00 = _mm256_sub_pd(iz0,jz0);
1015 dx10 = _mm256_sub_pd(ix1,jx0);
1016 dy10 = _mm256_sub_pd(iy1,jy0);
1017 dz10 = _mm256_sub_pd(iz1,jz0);
1018 dx20 = _mm256_sub_pd(ix2,jx0);
1019 dy20 = _mm256_sub_pd(iy2,jy0);
1020 dz20 = _mm256_sub_pd(iz2,jz0);
1022 /* Calculate squared distance and things based on it */
1023 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
1024 rsq10 = gmx_mm256_calc_rsq_pd(dx10,dy10,dz10);
1025 rsq20 = gmx_mm256_calc_rsq_pd(dx20,dy20,dz20);
1027 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
1028 rinv10 = gmx_mm256_invsqrt_pd(rsq10);
1029 rinv20 = gmx_mm256_invsqrt_pd(rsq20);
1031 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
1032 rinvsq10 = _mm256_mul_pd(rinv10,rinv10);
1033 rinvsq20 = _mm256_mul_pd(rinv20,rinv20);
1035 /* Load parameters for j particles */
1036 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
1037 charge+jnrC+0,charge+jnrD+0);
1038 vdwjidx0A = 2*vdwtype[jnrA+0];
1039 vdwjidx0B = 2*vdwtype[jnrB+0];
1040 vdwjidx0C = 2*vdwtype[jnrC+0];
1041 vdwjidx0D = 2*vdwtype[jnrD+0];
1043 fjx0 = _mm256_setzero_pd();
1044 fjy0 = _mm256_setzero_pd();
1045 fjz0 = _mm256_setzero_pd();
1047 /**************************
1048 * CALCULATE INTERACTIONS *
1049 **************************/
1051 r00 = _mm256_mul_pd(rsq00,rinv00);
1052 r00 = _mm256_andnot_pd(dummy_mask,r00);
1054 /* Compute parameters for interactions between i and j atoms */
1055 qq00 = _mm256_mul_pd(iq0,jq0);
1056 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
1057 vdwioffsetptr0+vdwjidx0B,
1058 vdwioffsetptr0+vdwjidx0C,
1059 vdwioffsetptr0+vdwjidx0D,
1062 /* EWALD ELECTROSTATICS */
1064 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1065 ewrt = _mm256_mul_pd(r00,ewtabscale);
1066 ewitab = _mm256_cvttpd_epi32(ewrt);
1067 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1068 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1069 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1071 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1072 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
1074 /* LENNARD-JONES DISPERSION/REPULSION */
1076 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1077 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
1079 fscal = _mm256_add_pd(felec,fvdw);
1081 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1083 /* Calculate temporary vectorial force */
1084 tx = _mm256_mul_pd(fscal,dx00);
1085 ty = _mm256_mul_pd(fscal,dy00);
1086 tz = _mm256_mul_pd(fscal,dz00);
1088 /* Update vectorial force */
1089 fix0 = _mm256_add_pd(fix0,tx);
1090 fiy0 = _mm256_add_pd(fiy0,ty);
1091 fiz0 = _mm256_add_pd(fiz0,tz);
1093 fjx0 = _mm256_add_pd(fjx0,tx);
1094 fjy0 = _mm256_add_pd(fjy0,ty);
1095 fjz0 = _mm256_add_pd(fjz0,tz);
1097 /**************************
1098 * CALCULATE INTERACTIONS *
1099 **************************/
1101 r10 = _mm256_mul_pd(rsq10,rinv10);
1102 r10 = _mm256_andnot_pd(dummy_mask,r10);
1104 /* Compute parameters for interactions between i and j atoms */
1105 qq10 = _mm256_mul_pd(iq1,jq0);
1107 /* EWALD ELECTROSTATICS */
1109 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1110 ewrt = _mm256_mul_pd(r10,ewtabscale);
1111 ewitab = _mm256_cvttpd_epi32(ewrt);
1112 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1113 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1114 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1116 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1117 felec = _mm256_mul_pd(_mm256_mul_pd(qq10,rinv10),_mm256_sub_pd(rinvsq10,felec));
1121 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1123 /* Calculate temporary vectorial force */
1124 tx = _mm256_mul_pd(fscal,dx10);
1125 ty = _mm256_mul_pd(fscal,dy10);
1126 tz = _mm256_mul_pd(fscal,dz10);
1128 /* Update vectorial force */
1129 fix1 = _mm256_add_pd(fix1,tx);
1130 fiy1 = _mm256_add_pd(fiy1,ty);
1131 fiz1 = _mm256_add_pd(fiz1,tz);
1133 fjx0 = _mm256_add_pd(fjx0,tx);
1134 fjy0 = _mm256_add_pd(fjy0,ty);
1135 fjz0 = _mm256_add_pd(fjz0,tz);
1137 /**************************
1138 * CALCULATE INTERACTIONS *
1139 **************************/
1141 r20 = _mm256_mul_pd(rsq20,rinv20);
1142 r20 = _mm256_andnot_pd(dummy_mask,r20);
1144 /* Compute parameters for interactions between i and j atoms */
1145 qq20 = _mm256_mul_pd(iq2,jq0);
1147 /* EWALD ELECTROSTATICS */
1149 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1150 ewrt = _mm256_mul_pd(r20,ewtabscale);
1151 ewitab = _mm256_cvttpd_epi32(ewrt);
1152 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
1153 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
1154 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
1156 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
1157 felec = _mm256_mul_pd(_mm256_mul_pd(qq20,rinv20),_mm256_sub_pd(rinvsq20,felec));
1161 fscal = _mm256_andnot_pd(dummy_mask,fscal);
1163 /* Calculate temporary vectorial force */
1164 tx = _mm256_mul_pd(fscal,dx20);
1165 ty = _mm256_mul_pd(fscal,dy20);
1166 tz = _mm256_mul_pd(fscal,dz20);
1168 /* Update vectorial force */
1169 fix2 = _mm256_add_pd(fix2,tx);
1170 fiy2 = _mm256_add_pd(fiy2,ty);
1171 fiz2 = _mm256_add_pd(fiz2,tz);
1173 fjx0 = _mm256_add_pd(fjx0,tx);
1174 fjy0 = _mm256_add_pd(fjy0,ty);
1175 fjz0 = _mm256_add_pd(fjz0,tz);
1177 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1178 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1179 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1180 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1182 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1184 /* Inner loop uses 121 flops */
1187 /* End of innermost loop */
1189 gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1190 f+i_coord_offset,fshift+i_shift_offset);
1192 /* Increment number of inner iterations */
1193 inneriter += j_index_end - j_index_start;
1195 /* Outer loop uses 18 flops */
1198 /* Increment number of outer iterations */
1201 /* Update outer/inner flops */
1203 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*121);