2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 * Note: this file was generated by the GROMACS avx_256_double kernel generator.
42 #include "../nb_kernel.h"
43 #include "gromacs/legacyheaders/types/simple.h"
44 #include "gromacs/math/vec.h"
45 #include "gromacs/legacyheaders/nrnb.h"
47 #include "gromacs/simd/math_x86_avx_256_double.h"
48 #include "kernelutil_x86_avx_256_double.h"
51 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwLJEw_GeomP1P1_VF_avx_256_double
52 * Electrostatics interaction: Ewald
53 * VdW interaction: LJEwald
54 * Geometry: Particle-Particle
55 * Calculate force/pot: PotentialAndForce
58 nb_kernel_ElecEw_VdwLJEw_GeomP1P1_VF_avx_256_double
59 (t_nblist * gmx_restrict nlist,
60 rvec * gmx_restrict xx,
61 rvec * gmx_restrict ff,
62 t_forcerec * gmx_restrict fr,
63 t_mdatoms * gmx_restrict mdatoms,
64 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
65 t_nrnb * gmx_restrict nrnb)
67 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
68 * just 0 for non-waters.
69 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
70 * jnr indices corresponding to data put in the four positions in the SIMD register.
72 int i_shift_offset,i_coord_offset,outeriter,inneriter;
73 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
74 int jnrA,jnrB,jnrC,jnrD;
75 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
76 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
77 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
78 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
80 real *shiftvec,*fshift,*x,*f;
81 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
83 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
84 real * vdwioffsetptr0;
85 real * vdwgridioffsetptr0;
86 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
87 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
88 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
89 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
90 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
93 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
96 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
97 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
100 __m256d ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
101 __m256d one_half = _mm256_set1_pd(0.5);
102 __m256d minus_one = _mm256_set1_pd(-1.0);
104 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
105 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
107 __m256d dummy_mask,cutoff_mask;
108 __m128 tmpmask0,tmpmask1;
109 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
110 __m256d one = _mm256_set1_pd(1.0);
111 __m256d two = _mm256_set1_pd(2.0);
117 jindex = nlist->jindex;
119 shiftidx = nlist->shift;
121 shiftvec = fr->shift_vec[0];
122 fshift = fr->fshift[0];
123 facel = _mm256_set1_pd(fr->epsfac);
124 charge = mdatoms->chargeA;
125 nvdwtype = fr->ntype;
127 vdwtype = mdatoms->typeA;
128 vdwgridparam = fr->ljpme_c6grid;
129 sh_lj_ewald = _mm256_set1_pd(fr->ic->sh_lj_ewald);
130 ewclj = _mm256_set1_pd(fr->ewaldcoeff_lj);
131 ewclj2 = _mm256_mul_pd(minus_one,_mm256_mul_pd(ewclj,ewclj));
133 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
134 beta = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
135 beta2 = _mm256_mul_pd(beta,beta);
136 beta3 = _mm256_mul_pd(beta,beta2);
138 ewtab = fr->ic->tabq_coul_FDV0;
139 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
140 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
142 /* Avoid stupid compiler warnings */
143 jnrA = jnrB = jnrC = jnrD = 0;
152 for(iidx=0;iidx<4*DIM;iidx++)
157 /* Start outer loop over neighborlists */
158 for(iidx=0; iidx<nri; iidx++)
160 /* Load shift vector for this list */
161 i_shift_offset = DIM*shiftidx[iidx];
163 /* Load limits for loop over neighbors */
164 j_index_start = jindex[iidx];
165 j_index_end = jindex[iidx+1];
167 /* Get outer coordinate index */
169 i_coord_offset = DIM*inr;
171 /* Load i particle coords and add shift vector */
172 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
174 fix0 = _mm256_setzero_pd();
175 fiy0 = _mm256_setzero_pd();
176 fiz0 = _mm256_setzero_pd();
178 /* Load parameters for i particles */
179 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
180 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
181 vdwgridioffsetptr0 = vdwgridparam+2*nvdwtype*vdwtype[inr+0];
183 /* Reset potential sums */
184 velecsum = _mm256_setzero_pd();
185 vvdwsum = _mm256_setzero_pd();
187 /* Start inner kernel loop */
188 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
191 /* Get j neighbor index, and coordinate index */
196 j_coord_offsetA = DIM*jnrA;
197 j_coord_offsetB = DIM*jnrB;
198 j_coord_offsetC = DIM*jnrC;
199 j_coord_offsetD = DIM*jnrD;
201 /* load j atom coordinates */
202 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
203 x+j_coord_offsetC,x+j_coord_offsetD,
206 /* Calculate displacement vector */
207 dx00 = _mm256_sub_pd(ix0,jx0);
208 dy00 = _mm256_sub_pd(iy0,jy0);
209 dz00 = _mm256_sub_pd(iz0,jz0);
211 /* Calculate squared distance and things based on it */
212 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
214 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
216 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
218 /* Load parameters for j particles */
219 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
220 charge+jnrC+0,charge+jnrD+0);
221 vdwjidx0A = 2*vdwtype[jnrA+0];
222 vdwjidx0B = 2*vdwtype[jnrB+0];
223 vdwjidx0C = 2*vdwtype[jnrC+0];
224 vdwjidx0D = 2*vdwtype[jnrD+0];
226 /**************************
227 * CALCULATE INTERACTIONS *
228 **************************/
230 r00 = _mm256_mul_pd(rsq00,rinv00);
232 /* Compute parameters for interactions between i and j atoms */
233 qq00 = _mm256_mul_pd(iq0,jq0);
234 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
235 vdwioffsetptr0+vdwjidx0B,
236 vdwioffsetptr0+vdwjidx0C,
237 vdwioffsetptr0+vdwjidx0D,
240 c6grid_00 = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
241 vdwgridioffsetptr0+vdwjidx0B,
242 vdwgridioffsetptr0+vdwjidx0C,
243 vdwgridioffsetptr0+vdwjidx0D);
245 /* EWALD ELECTROSTATICS */
247 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
248 ewrt = _mm256_mul_pd(r00,ewtabscale);
249 ewitab = _mm256_cvttpd_epi32(ewrt);
250 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
251 ewitab = _mm_slli_epi32(ewitab,2);
252 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
253 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
254 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
255 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
256 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
257 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
258 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
259 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
260 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
262 /* Analytical LJ-PME */
263 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
264 ewcljrsq = _mm256_mul_pd(ewclj2,rsq00);
265 ewclj6 = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
266 exponent = gmx_simd_exp_d(ewcljrsq);
267 /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
268 poly = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
269 /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
270 vvdw6 = _mm256_mul_pd(_mm256_sub_pd(c6_00,_mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly))),rinvsix);
271 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
272 vvdw = _mm256_sub_pd(_mm256_mul_pd(vvdw12,one_twelfth),_mm256_mul_pd(vvdw6,one_sixth));
273 /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
274 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,_mm256_sub_pd(vvdw6,_mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6)))),rinvsq00);
276 /* Update potential sum for this i atom from the interaction with this j atom. */
277 velecsum = _mm256_add_pd(velecsum,velec);
278 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
280 fscal = _mm256_add_pd(felec,fvdw);
282 /* Calculate temporary vectorial force */
283 tx = _mm256_mul_pd(fscal,dx00);
284 ty = _mm256_mul_pd(fscal,dy00);
285 tz = _mm256_mul_pd(fscal,dz00);
287 /* Update vectorial force */
288 fix0 = _mm256_add_pd(fix0,tx);
289 fiy0 = _mm256_add_pd(fiy0,ty);
290 fiz0 = _mm256_add_pd(fiz0,tz);
292 fjptrA = f+j_coord_offsetA;
293 fjptrB = f+j_coord_offsetB;
294 fjptrC = f+j_coord_offsetC;
295 fjptrD = f+j_coord_offsetD;
296 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
298 /* Inner loop uses 72 flops */
304 /* Get j neighbor index, and coordinate index */
305 jnrlistA = jjnr[jidx];
306 jnrlistB = jjnr[jidx+1];
307 jnrlistC = jjnr[jidx+2];
308 jnrlistD = jjnr[jidx+3];
309 /* Sign of each element will be negative for non-real atoms.
310 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
311 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
313 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
315 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
316 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
317 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
319 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
320 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
321 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
322 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
323 j_coord_offsetA = DIM*jnrA;
324 j_coord_offsetB = DIM*jnrB;
325 j_coord_offsetC = DIM*jnrC;
326 j_coord_offsetD = DIM*jnrD;
328 /* load j atom coordinates */
329 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
330 x+j_coord_offsetC,x+j_coord_offsetD,
333 /* Calculate displacement vector */
334 dx00 = _mm256_sub_pd(ix0,jx0);
335 dy00 = _mm256_sub_pd(iy0,jy0);
336 dz00 = _mm256_sub_pd(iz0,jz0);
338 /* Calculate squared distance and things based on it */
339 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
341 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
343 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
345 /* Load parameters for j particles */
346 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
347 charge+jnrC+0,charge+jnrD+0);
348 vdwjidx0A = 2*vdwtype[jnrA+0];
349 vdwjidx0B = 2*vdwtype[jnrB+0];
350 vdwjidx0C = 2*vdwtype[jnrC+0];
351 vdwjidx0D = 2*vdwtype[jnrD+0];
353 /**************************
354 * CALCULATE INTERACTIONS *
355 **************************/
357 r00 = _mm256_mul_pd(rsq00,rinv00);
358 r00 = _mm256_andnot_pd(dummy_mask,r00);
360 /* Compute parameters for interactions between i and j atoms */
361 qq00 = _mm256_mul_pd(iq0,jq0);
362 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
363 vdwioffsetptr0+vdwjidx0B,
364 vdwioffsetptr0+vdwjidx0C,
365 vdwioffsetptr0+vdwjidx0D,
368 c6grid_00 = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
369 vdwgridioffsetptr0+vdwjidx0B,
370 vdwgridioffsetptr0+vdwjidx0C,
371 vdwgridioffsetptr0+vdwjidx0D);
373 /* EWALD ELECTROSTATICS */
375 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
376 ewrt = _mm256_mul_pd(r00,ewtabscale);
377 ewitab = _mm256_cvttpd_epi32(ewrt);
378 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
379 ewitab = _mm_slli_epi32(ewitab,2);
380 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
381 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
382 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
383 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
384 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
385 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
386 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
387 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
388 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
390 /* Analytical LJ-PME */
391 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
392 ewcljrsq = _mm256_mul_pd(ewclj2,rsq00);
393 ewclj6 = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
394 exponent = gmx_simd_exp_d(ewcljrsq);
395 /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
396 poly = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
397 /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
398 vvdw6 = _mm256_mul_pd(_mm256_sub_pd(c6_00,_mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly))),rinvsix);
399 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
400 vvdw = _mm256_sub_pd(_mm256_mul_pd(vvdw12,one_twelfth),_mm256_mul_pd(vvdw6,one_sixth));
401 /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
402 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,_mm256_sub_pd(vvdw6,_mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6)))),rinvsq00);
404 /* Update potential sum for this i atom from the interaction with this j atom. */
405 velec = _mm256_andnot_pd(dummy_mask,velec);
406 velecsum = _mm256_add_pd(velecsum,velec);
407 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
408 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
410 fscal = _mm256_add_pd(felec,fvdw);
412 fscal = _mm256_andnot_pd(dummy_mask,fscal);
414 /* Calculate temporary vectorial force */
415 tx = _mm256_mul_pd(fscal,dx00);
416 ty = _mm256_mul_pd(fscal,dy00);
417 tz = _mm256_mul_pd(fscal,dz00);
419 /* Update vectorial force */
420 fix0 = _mm256_add_pd(fix0,tx);
421 fiy0 = _mm256_add_pd(fiy0,ty);
422 fiz0 = _mm256_add_pd(fiz0,tz);
424 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
425 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
426 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
427 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
428 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
430 /* Inner loop uses 73 flops */
433 /* End of innermost loop */
435 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
436 f+i_coord_offset,fshift+i_shift_offset);
439 /* Update potential energies */
440 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
441 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
443 /* Increment number of inner iterations */
444 inneriter += j_index_end - j_index_start;
446 /* Outer loop uses 9 flops */
449 /* Increment number of outer iterations */
452 /* Update outer/inner flops */
454 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*73);
457 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwLJEw_GeomP1P1_F_avx_256_double
458 * Electrostatics interaction: Ewald
459 * VdW interaction: LJEwald
460 * Geometry: Particle-Particle
461 * Calculate force/pot: Force
464 nb_kernel_ElecEw_VdwLJEw_GeomP1P1_F_avx_256_double
465 (t_nblist * gmx_restrict nlist,
466 rvec * gmx_restrict xx,
467 rvec * gmx_restrict ff,
468 t_forcerec * gmx_restrict fr,
469 t_mdatoms * gmx_restrict mdatoms,
470 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
471 t_nrnb * gmx_restrict nrnb)
473 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
474 * just 0 for non-waters.
475 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
476 * jnr indices corresponding to data put in the four positions in the SIMD register.
478 int i_shift_offset,i_coord_offset,outeriter,inneriter;
479 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
480 int jnrA,jnrB,jnrC,jnrD;
481 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
482 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
483 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
484 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
486 real *shiftvec,*fshift,*x,*f;
487 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
489 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
490 real * vdwioffsetptr0;
491 real * vdwgridioffsetptr0;
492 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
493 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
494 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
495 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
496 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
499 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
502 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
503 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
506 __m256d ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
507 __m256d one_half = _mm256_set1_pd(0.5);
508 __m256d minus_one = _mm256_set1_pd(-1.0);
510 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
511 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
513 __m256d dummy_mask,cutoff_mask;
514 __m128 tmpmask0,tmpmask1;
515 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
516 __m256d one = _mm256_set1_pd(1.0);
517 __m256d two = _mm256_set1_pd(2.0);
523 jindex = nlist->jindex;
525 shiftidx = nlist->shift;
527 shiftvec = fr->shift_vec[0];
528 fshift = fr->fshift[0];
529 facel = _mm256_set1_pd(fr->epsfac);
530 charge = mdatoms->chargeA;
531 nvdwtype = fr->ntype;
533 vdwtype = mdatoms->typeA;
534 vdwgridparam = fr->ljpme_c6grid;
535 sh_lj_ewald = _mm256_set1_pd(fr->ic->sh_lj_ewald);
536 ewclj = _mm256_set1_pd(fr->ewaldcoeff_lj);
537 ewclj2 = _mm256_mul_pd(minus_one,_mm256_mul_pd(ewclj,ewclj));
539 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
540 beta = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
541 beta2 = _mm256_mul_pd(beta,beta);
542 beta3 = _mm256_mul_pd(beta,beta2);
544 ewtab = fr->ic->tabq_coul_F;
545 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
546 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
548 /* Avoid stupid compiler warnings */
549 jnrA = jnrB = jnrC = jnrD = 0;
558 for(iidx=0;iidx<4*DIM;iidx++)
563 /* Start outer loop over neighborlists */
564 for(iidx=0; iidx<nri; iidx++)
566 /* Load shift vector for this list */
567 i_shift_offset = DIM*shiftidx[iidx];
569 /* Load limits for loop over neighbors */
570 j_index_start = jindex[iidx];
571 j_index_end = jindex[iidx+1];
573 /* Get outer coordinate index */
575 i_coord_offset = DIM*inr;
577 /* Load i particle coords and add shift vector */
578 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
580 fix0 = _mm256_setzero_pd();
581 fiy0 = _mm256_setzero_pd();
582 fiz0 = _mm256_setzero_pd();
584 /* Load parameters for i particles */
585 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
586 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
587 vdwgridioffsetptr0 = vdwgridparam+2*nvdwtype*vdwtype[inr+0];
589 /* Start inner kernel loop */
590 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
593 /* Get j neighbor index, and coordinate index */
598 j_coord_offsetA = DIM*jnrA;
599 j_coord_offsetB = DIM*jnrB;
600 j_coord_offsetC = DIM*jnrC;
601 j_coord_offsetD = DIM*jnrD;
603 /* load j atom coordinates */
604 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
605 x+j_coord_offsetC,x+j_coord_offsetD,
608 /* Calculate displacement vector */
609 dx00 = _mm256_sub_pd(ix0,jx0);
610 dy00 = _mm256_sub_pd(iy0,jy0);
611 dz00 = _mm256_sub_pd(iz0,jz0);
613 /* Calculate squared distance and things based on it */
614 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
616 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
618 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
620 /* Load parameters for j particles */
621 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
622 charge+jnrC+0,charge+jnrD+0);
623 vdwjidx0A = 2*vdwtype[jnrA+0];
624 vdwjidx0B = 2*vdwtype[jnrB+0];
625 vdwjidx0C = 2*vdwtype[jnrC+0];
626 vdwjidx0D = 2*vdwtype[jnrD+0];
628 /**************************
629 * CALCULATE INTERACTIONS *
630 **************************/
632 r00 = _mm256_mul_pd(rsq00,rinv00);
634 /* Compute parameters for interactions between i and j atoms */
635 qq00 = _mm256_mul_pd(iq0,jq0);
636 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
637 vdwioffsetptr0+vdwjidx0B,
638 vdwioffsetptr0+vdwjidx0C,
639 vdwioffsetptr0+vdwjidx0D,
642 c6grid_00 = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
643 vdwgridioffsetptr0+vdwjidx0B,
644 vdwgridioffsetptr0+vdwjidx0C,
645 vdwgridioffsetptr0+vdwjidx0D);
647 /* EWALD ELECTROSTATICS */
649 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
650 ewrt = _mm256_mul_pd(r00,ewtabscale);
651 ewitab = _mm256_cvttpd_epi32(ewrt);
652 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
653 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
654 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
656 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
657 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
659 /* Analytical LJ-PME */
660 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
661 ewcljrsq = _mm256_mul_pd(ewclj2,rsq00);
662 ewclj6 = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
663 exponent = gmx_simd_exp_d(ewcljrsq);
664 /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
665 poly = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
666 /* f6A = 6 * C6grid * (1 - poly) */
667 f6A = _mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly));
668 /* f6B = C6grid * exponent * beta^6 */
669 f6B = _mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6));
670 /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
671 fvdw = _mm256_mul_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),_mm256_sub_pd(c6_00,f6A)),rinvsix),f6B),rinvsq00);
673 fscal = _mm256_add_pd(felec,fvdw);
675 /* Calculate temporary vectorial force */
676 tx = _mm256_mul_pd(fscal,dx00);
677 ty = _mm256_mul_pd(fscal,dy00);
678 tz = _mm256_mul_pd(fscal,dz00);
680 /* Update vectorial force */
681 fix0 = _mm256_add_pd(fix0,tx);
682 fiy0 = _mm256_add_pd(fiy0,ty);
683 fiz0 = _mm256_add_pd(fiz0,tz);
685 fjptrA = f+j_coord_offsetA;
686 fjptrB = f+j_coord_offsetB;
687 fjptrC = f+j_coord_offsetC;
688 fjptrD = f+j_coord_offsetD;
689 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
691 /* Inner loop uses 59 flops */
697 /* Get j neighbor index, and coordinate index */
698 jnrlistA = jjnr[jidx];
699 jnrlistB = jjnr[jidx+1];
700 jnrlistC = jjnr[jidx+2];
701 jnrlistD = jjnr[jidx+3];
702 /* Sign of each element will be negative for non-real atoms.
703 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
704 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
706 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
708 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
709 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
710 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
712 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
713 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
714 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
715 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
716 j_coord_offsetA = DIM*jnrA;
717 j_coord_offsetB = DIM*jnrB;
718 j_coord_offsetC = DIM*jnrC;
719 j_coord_offsetD = DIM*jnrD;
721 /* load j atom coordinates */
722 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
723 x+j_coord_offsetC,x+j_coord_offsetD,
726 /* Calculate displacement vector */
727 dx00 = _mm256_sub_pd(ix0,jx0);
728 dy00 = _mm256_sub_pd(iy0,jy0);
729 dz00 = _mm256_sub_pd(iz0,jz0);
731 /* Calculate squared distance and things based on it */
732 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
734 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
736 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
738 /* Load parameters for j particles */
739 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
740 charge+jnrC+0,charge+jnrD+0);
741 vdwjidx0A = 2*vdwtype[jnrA+0];
742 vdwjidx0B = 2*vdwtype[jnrB+0];
743 vdwjidx0C = 2*vdwtype[jnrC+0];
744 vdwjidx0D = 2*vdwtype[jnrD+0];
746 /**************************
747 * CALCULATE INTERACTIONS *
748 **************************/
750 r00 = _mm256_mul_pd(rsq00,rinv00);
751 r00 = _mm256_andnot_pd(dummy_mask,r00);
753 /* Compute parameters for interactions between i and j atoms */
754 qq00 = _mm256_mul_pd(iq0,jq0);
755 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
756 vdwioffsetptr0+vdwjidx0B,
757 vdwioffsetptr0+vdwjidx0C,
758 vdwioffsetptr0+vdwjidx0D,
761 c6grid_00 = gmx_mm256_load_4real_swizzle_pd(vdwgridioffsetptr0+vdwjidx0A,
762 vdwgridioffsetptr0+vdwjidx0B,
763 vdwgridioffsetptr0+vdwjidx0C,
764 vdwgridioffsetptr0+vdwjidx0D);
766 /* EWALD ELECTROSTATICS */
768 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
769 ewrt = _mm256_mul_pd(r00,ewtabscale);
770 ewitab = _mm256_cvttpd_epi32(ewrt);
771 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
772 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
773 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
775 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
776 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
778 /* Analytical LJ-PME */
779 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
780 ewcljrsq = _mm256_mul_pd(ewclj2,rsq00);
781 ewclj6 = _mm256_mul_pd(ewclj2,_mm256_mul_pd(ewclj2,ewclj2));
782 exponent = gmx_simd_exp_d(ewcljrsq);
783 /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
784 poly = _mm256_mul_pd(exponent,_mm256_add_pd(_mm256_sub_pd(one,ewcljrsq),_mm256_mul_pd(_mm256_mul_pd(ewcljrsq,ewcljrsq),one_half)));
785 /* f6A = 6 * C6grid * (1 - poly) */
786 f6A = _mm256_mul_pd(c6grid_00,_mm256_sub_pd(one,poly));
787 /* f6B = C6grid * exponent * beta^6 */
788 f6B = _mm256_mul_pd(_mm256_mul_pd(c6grid_00,one_sixth),_mm256_mul_pd(exponent,ewclj6));
789 /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
790 fvdw = _mm256_mul_pd(_mm256_add_pd(_mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),_mm256_sub_pd(c6_00,f6A)),rinvsix),f6B),rinvsq00);
792 fscal = _mm256_add_pd(felec,fvdw);
794 fscal = _mm256_andnot_pd(dummy_mask,fscal);
796 /* Calculate temporary vectorial force */
797 tx = _mm256_mul_pd(fscal,dx00);
798 ty = _mm256_mul_pd(fscal,dy00);
799 tz = _mm256_mul_pd(fscal,dz00);
801 /* Update vectorial force */
802 fix0 = _mm256_add_pd(fix0,tx);
803 fiy0 = _mm256_add_pd(fiy0,ty);
804 fiz0 = _mm256_add_pd(fiz0,tz);
806 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
807 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
808 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
809 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
810 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
812 /* Inner loop uses 60 flops */
815 /* End of innermost loop */
817 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
818 f+i_coord_offset,fshift+i_shift_offset);
820 /* Increment number of inner iterations */
821 inneriter += j_index_end - j_index_start;
823 /* Outer loop uses 7 flops */
826 /* Increment number of outer iterations */
829 /* Update outer/inner flops */
831 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*60);