2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2017, 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.
44 #include "../nb_kernel.h"
45 #include "gromacs/gmxlib/nrnb.h"
47 #include "kernelutil_x86_avx_256_double.h"
50 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_double
51 * Electrostatics interaction: GeneralizedBorn
52 * VdW interaction: LennardJones
53 * Geometry: Particle-Particle
54 * Calculate force/pot: PotentialAndForce
57 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_double
58 (t_nblist * gmx_restrict nlist,
59 rvec * gmx_restrict xx,
60 rvec * gmx_restrict ff,
61 struct t_forcerec * gmx_restrict fr,
62 t_mdatoms * gmx_restrict mdatoms,
63 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
64 t_nrnb * gmx_restrict nrnb)
66 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
67 * just 0 for non-waters.
68 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
69 * jnr indices corresponding to data put in the four positions in the SIMD register.
71 int i_shift_offset,i_coord_offset,outeriter,inneriter;
72 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
73 int jnrA,jnrB,jnrC,jnrD;
74 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
75 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
76 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
77 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
79 real *shiftvec,*fshift,*x,*f;
80 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
82 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
83 real * vdwioffsetptr0;
84 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
85 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
86 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
87 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
88 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
91 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
92 __m256d minushalf = _mm256_set1_pd(-0.5);
93 real *invsqrta,*dvda,*gbtab;
95 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
98 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
99 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
101 __m128i ifour = _mm_set1_epi32(4);
102 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
104 __m256d dummy_mask,cutoff_mask;
105 __m128 tmpmask0,tmpmask1;
106 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
107 __m256d one = _mm256_set1_pd(1.0);
108 __m256d two = _mm256_set1_pd(2.0);
114 jindex = nlist->jindex;
116 shiftidx = nlist->shift;
118 shiftvec = fr->shift_vec[0];
119 fshift = fr->fshift[0];
120 facel = _mm256_set1_pd(fr->ic->epsfac);
121 charge = mdatoms->chargeA;
122 nvdwtype = fr->ntype;
124 vdwtype = mdatoms->typeA;
126 invsqrta = fr->invsqrta;
128 gbtabscale = _mm256_set1_pd(fr->gbtab->scale);
129 gbtab = fr->gbtab->data;
130 gbinvepsdiff = _mm256_set1_pd((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
132 /* Avoid stupid compiler warnings */
133 jnrA = jnrB = jnrC = jnrD = 0;
142 for(iidx=0;iidx<4*DIM;iidx++)
147 /* Start outer loop over neighborlists */
148 for(iidx=0; iidx<nri; iidx++)
150 /* Load shift vector for this list */
151 i_shift_offset = DIM*shiftidx[iidx];
153 /* Load limits for loop over neighbors */
154 j_index_start = jindex[iidx];
155 j_index_end = jindex[iidx+1];
157 /* Get outer coordinate index */
159 i_coord_offset = DIM*inr;
161 /* Load i particle coords and add shift vector */
162 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
164 fix0 = _mm256_setzero_pd();
165 fiy0 = _mm256_setzero_pd();
166 fiz0 = _mm256_setzero_pd();
168 /* Load parameters for i particles */
169 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
170 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
171 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
173 /* Reset potential sums */
174 velecsum = _mm256_setzero_pd();
175 vgbsum = _mm256_setzero_pd();
176 vvdwsum = _mm256_setzero_pd();
177 dvdasum = _mm256_setzero_pd();
179 /* Start inner kernel loop */
180 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
183 /* Get j neighbor index, and coordinate index */
188 j_coord_offsetA = DIM*jnrA;
189 j_coord_offsetB = DIM*jnrB;
190 j_coord_offsetC = DIM*jnrC;
191 j_coord_offsetD = DIM*jnrD;
193 /* load j atom coordinates */
194 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
195 x+j_coord_offsetC,x+j_coord_offsetD,
198 /* Calculate displacement vector */
199 dx00 = _mm256_sub_pd(ix0,jx0);
200 dy00 = _mm256_sub_pd(iy0,jy0);
201 dz00 = _mm256_sub_pd(iz0,jz0);
203 /* Calculate squared distance and things based on it */
204 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
206 rinv00 = avx256_invsqrt_d(rsq00);
208 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
210 /* Load parameters for j particles */
211 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
212 charge+jnrC+0,charge+jnrD+0);
213 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
214 invsqrta+jnrC+0,invsqrta+jnrD+0);
215 vdwjidx0A = 2*vdwtype[jnrA+0];
216 vdwjidx0B = 2*vdwtype[jnrB+0];
217 vdwjidx0C = 2*vdwtype[jnrC+0];
218 vdwjidx0D = 2*vdwtype[jnrD+0];
220 /**************************
221 * CALCULATE INTERACTIONS *
222 **************************/
224 r00 = _mm256_mul_pd(rsq00,rinv00);
226 /* Compute parameters for interactions between i and j atoms */
227 qq00 = _mm256_mul_pd(iq0,jq0);
228 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
229 vdwioffsetptr0+vdwjidx0B,
230 vdwioffsetptr0+vdwjidx0C,
231 vdwioffsetptr0+vdwjidx0D,
234 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
235 isaprod = _mm256_mul_pd(isai0,isaj0);
236 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
237 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
239 /* Calculate generalized born table index - this is a separate table from the normal one,
240 * but we use the same procedure by multiplying r with scale and truncating to integer.
242 rt = _mm256_mul_pd(r00,gbscale);
243 gbitab = _mm256_cvttpd_epi32(rt);
244 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
245 gbitab = _mm_slli_epi32(gbitab,2);
246 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
247 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
248 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
249 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
250 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
251 Heps = _mm256_mul_pd(gbeps,H);
252 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
253 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
254 vgb = _mm256_mul_pd(gbqqfactor,VV);
256 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
257 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
258 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
259 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
264 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
265 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
266 velec = _mm256_mul_pd(qq00,rinv00);
267 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
269 /* LENNARD-JONES DISPERSION/REPULSION */
271 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
272 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
273 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
274 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
275 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
277 /* Update potential sum for this i atom from the interaction with this j atom. */
278 velecsum = _mm256_add_pd(velecsum,velec);
279 vgbsum = _mm256_add_pd(vgbsum,vgb);
280 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
282 fscal = _mm256_add_pd(felec,fvdw);
284 /* Calculate temporary vectorial force */
285 tx = _mm256_mul_pd(fscal,dx00);
286 ty = _mm256_mul_pd(fscal,dy00);
287 tz = _mm256_mul_pd(fscal,dz00);
289 /* Update vectorial force */
290 fix0 = _mm256_add_pd(fix0,tx);
291 fiy0 = _mm256_add_pd(fiy0,ty);
292 fiz0 = _mm256_add_pd(fiz0,tz);
294 fjptrA = f+j_coord_offsetA;
295 fjptrB = f+j_coord_offsetB;
296 fjptrC = f+j_coord_offsetC;
297 fjptrD = f+j_coord_offsetD;
298 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
300 /* Inner loop uses 70 flops */
306 /* Get j neighbor index, and coordinate index */
307 jnrlistA = jjnr[jidx];
308 jnrlistB = jjnr[jidx+1];
309 jnrlistC = jjnr[jidx+2];
310 jnrlistD = jjnr[jidx+3];
311 /* Sign of each element will be negative for non-real atoms.
312 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
313 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
315 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
317 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
318 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
319 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
321 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
322 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
323 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
324 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
325 j_coord_offsetA = DIM*jnrA;
326 j_coord_offsetB = DIM*jnrB;
327 j_coord_offsetC = DIM*jnrC;
328 j_coord_offsetD = DIM*jnrD;
330 /* load j atom coordinates */
331 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
332 x+j_coord_offsetC,x+j_coord_offsetD,
335 /* Calculate displacement vector */
336 dx00 = _mm256_sub_pd(ix0,jx0);
337 dy00 = _mm256_sub_pd(iy0,jy0);
338 dz00 = _mm256_sub_pd(iz0,jz0);
340 /* Calculate squared distance and things based on it */
341 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
343 rinv00 = avx256_invsqrt_d(rsq00);
345 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
347 /* Load parameters for j particles */
348 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
349 charge+jnrC+0,charge+jnrD+0);
350 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
351 invsqrta+jnrC+0,invsqrta+jnrD+0);
352 vdwjidx0A = 2*vdwtype[jnrA+0];
353 vdwjidx0B = 2*vdwtype[jnrB+0];
354 vdwjidx0C = 2*vdwtype[jnrC+0];
355 vdwjidx0D = 2*vdwtype[jnrD+0];
357 /**************************
358 * CALCULATE INTERACTIONS *
359 **************************/
361 r00 = _mm256_mul_pd(rsq00,rinv00);
362 r00 = _mm256_andnot_pd(dummy_mask,r00);
364 /* Compute parameters for interactions between i and j atoms */
365 qq00 = _mm256_mul_pd(iq0,jq0);
366 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
367 vdwioffsetptr0+vdwjidx0B,
368 vdwioffsetptr0+vdwjidx0C,
369 vdwioffsetptr0+vdwjidx0D,
372 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
373 isaprod = _mm256_mul_pd(isai0,isaj0);
374 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
375 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
377 /* Calculate generalized born table index - this is a separate table from the normal one,
378 * but we use the same procedure by multiplying r with scale and truncating to integer.
380 rt = _mm256_mul_pd(r00,gbscale);
381 gbitab = _mm256_cvttpd_epi32(rt);
382 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
383 gbitab = _mm_slli_epi32(gbitab,2);
384 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
385 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
386 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
387 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
388 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
389 Heps = _mm256_mul_pd(gbeps,H);
390 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
391 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
392 vgb = _mm256_mul_pd(gbqqfactor,VV);
394 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
395 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
396 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
397 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
398 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
399 /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
400 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
401 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
402 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
403 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
404 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
405 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
406 velec = _mm256_mul_pd(qq00,rinv00);
407 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
409 /* LENNARD-JONES DISPERSION/REPULSION */
411 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
412 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
413 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
414 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
415 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
417 /* Update potential sum for this i atom from the interaction with this j atom. */
418 velec = _mm256_andnot_pd(dummy_mask,velec);
419 velecsum = _mm256_add_pd(velecsum,velec);
420 vgb = _mm256_andnot_pd(dummy_mask,vgb);
421 vgbsum = _mm256_add_pd(vgbsum,vgb);
422 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
423 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
425 fscal = _mm256_add_pd(felec,fvdw);
427 fscal = _mm256_andnot_pd(dummy_mask,fscal);
429 /* Calculate temporary vectorial force */
430 tx = _mm256_mul_pd(fscal,dx00);
431 ty = _mm256_mul_pd(fscal,dy00);
432 tz = _mm256_mul_pd(fscal,dz00);
434 /* Update vectorial force */
435 fix0 = _mm256_add_pd(fix0,tx);
436 fiy0 = _mm256_add_pd(fiy0,ty);
437 fiz0 = _mm256_add_pd(fiz0,tz);
439 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
440 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
441 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
442 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
443 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
445 /* Inner loop uses 71 flops */
448 /* End of innermost loop */
450 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
451 f+i_coord_offset,fshift+i_shift_offset);
454 /* Update potential energies */
455 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
456 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
457 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
458 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
459 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
461 /* Increment number of inner iterations */
462 inneriter += j_index_end - j_index_start;
464 /* Outer loop uses 10 flops */
467 /* Increment number of outer iterations */
470 /* Update outer/inner flops */
472 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*71);
475 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
476 * Electrostatics interaction: GeneralizedBorn
477 * VdW interaction: LennardJones
478 * Geometry: Particle-Particle
479 * Calculate force/pot: Force
482 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
483 (t_nblist * gmx_restrict nlist,
484 rvec * gmx_restrict xx,
485 rvec * gmx_restrict ff,
486 struct t_forcerec * gmx_restrict fr,
487 t_mdatoms * gmx_restrict mdatoms,
488 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
489 t_nrnb * gmx_restrict nrnb)
491 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
492 * just 0 for non-waters.
493 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
494 * jnr indices corresponding to data put in the four positions in the SIMD register.
496 int i_shift_offset,i_coord_offset,outeriter,inneriter;
497 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
498 int jnrA,jnrB,jnrC,jnrD;
499 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
500 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
501 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
502 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
504 real *shiftvec,*fshift,*x,*f;
505 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
507 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
508 real * vdwioffsetptr0;
509 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
510 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
511 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
512 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
513 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
516 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
517 __m256d minushalf = _mm256_set1_pd(-0.5);
518 real *invsqrta,*dvda,*gbtab;
520 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
523 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
524 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
526 __m128i ifour = _mm_set1_epi32(4);
527 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
529 __m256d dummy_mask,cutoff_mask;
530 __m128 tmpmask0,tmpmask1;
531 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
532 __m256d one = _mm256_set1_pd(1.0);
533 __m256d two = _mm256_set1_pd(2.0);
539 jindex = nlist->jindex;
541 shiftidx = nlist->shift;
543 shiftvec = fr->shift_vec[0];
544 fshift = fr->fshift[0];
545 facel = _mm256_set1_pd(fr->ic->epsfac);
546 charge = mdatoms->chargeA;
547 nvdwtype = fr->ntype;
549 vdwtype = mdatoms->typeA;
551 invsqrta = fr->invsqrta;
553 gbtabscale = _mm256_set1_pd(fr->gbtab->scale);
554 gbtab = fr->gbtab->data;
555 gbinvepsdiff = _mm256_set1_pd((1.0/fr->ic->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
557 /* Avoid stupid compiler warnings */
558 jnrA = jnrB = jnrC = jnrD = 0;
567 for(iidx=0;iidx<4*DIM;iidx++)
572 /* Start outer loop over neighborlists */
573 for(iidx=0; iidx<nri; iidx++)
575 /* Load shift vector for this list */
576 i_shift_offset = DIM*shiftidx[iidx];
578 /* Load limits for loop over neighbors */
579 j_index_start = jindex[iidx];
580 j_index_end = jindex[iidx+1];
582 /* Get outer coordinate index */
584 i_coord_offset = DIM*inr;
586 /* Load i particle coords and add shift vector */
587 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
589 fix0 = _mm256_setzero_pd();
590 fiy0 = _mm256_setzero_pd();
591 fiz0 = _mm256_setzero_pd();
593 /* Load parameters for i particles */
594 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
595 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
596 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
598 dvdasum = _mm256_setzero_pd();
600 /* Start inner kernel loop */
601 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
604 /* Get j neighbor index, and coordinate index */
609 j_coord_offsetA = DIM*jnrA;
610 j_coord_offsetB = DIM*jnrB;
611 j_coord_offsetC = DIM*jnrC;
612 j_coord_offsetD = DIM*jnrD;
614 /* load j atom coordinates */
615 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
616 x+j_coord_offsetC,x+j_coord_offsetD,
619 /* Calculate displacement vector */
620 dx00 = _mm256_sub_pd(ix0,jx0);
621 dy00 = _mm256_sub_pd(iy0,jy0);
622 dz00 = _mm256_sub_pd(iz0,jz0);
624 /* Calculate squared distance and things based on it */
625 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
627 rinv00 = avx256_invsqrt_d(rsq00);
629 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
631 /* Load parameters for j particles */
632 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
633 charge+jnrC+0,charge+jnrD+0);
634 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
635 invsqrta+jnrC+0,invsqrta+jnrD+0);
636 vdwjidx0A = 2*vdwtype[jnrA+0];
637 vdwjidx0B = 2*vdwtype[jnrB+0];
638 vdwjidx0C = 2*vdwtype[jnrC+0];
639 vdwjidx0D = 2*vdwtype[jnrD+0];
641 /**************************
642 * CALCULATE INTERACTIONS *
643 **************************/
645 r00 = _mm256_mul_pd(rsq00,rinv00);
647 /* Compute parameters for interactions between i and j atoms */
648 qq00 = _mm256_mul_pd(iq0,jq0);
649 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
650 vdwioffsetptr0+vdwjidx0B,
651 vdwioffsetptr0+vdwjidx0C,
652 vdwioffsetptr0+vdwjidx0D,
655 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
656 isaprod = _mm256_mul_pd(isai0,isaj0);
657 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
658 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
660 /* Calculate generalized born table index - this is a separate table from the normal one,
661 * but we use the same procedure by multiplying r with scale and truncating to integer.
663 rt = _mm256_mul_pd(r00,gbscale);
664 gbitab = _mm256_cvttpd_epi32(rt);
665 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
666 gbitab = _mm_slli_epi32(gbitab,2);
667 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
668 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
669 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
670 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
671 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
672 Heps = _mm256_mul_pd(gbeps,H);
673 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
674 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
675 vgb = _mm256_mul_pd(gbqqfactor,VV);
677 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
678 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
679 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
680 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
685 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
686 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
687 velec = _mm256_mul_pd(qq00,rinv00);
688 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
690 /* LENNARD-JONES DISPERSION/REPULSION */
692 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
693 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
695 fscal = _mm256_add_pd(felec,fvdw);
697 /* Calculate temporary vectorial force */
698 tx = _mm256_mul_pd(fscal,dx00);
699 ty = _mm256_mul_pd(fscal,dy00);
700 tz = _mm256_mul_pd(fscal,dz00);
702 /* Update vectorial force */
703 fix0 = _mm256_add_pd(fix0,tx);
704 fiy0 = _mm256_add_pd(fiy0,ty);
705 fiz0 = _mm256_add_pd(fiz0,tz);
707 fjptrA = f+j_coord_offsetA;
708 fjptrB = f+j_coord_offsetB;
709 fjptrC = f+j_coord_offsetC;
710 fjptrD = f+j_coord_offsetD;
711 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
713 /* Inner loop uses 63 flops */
719 /* Get j neighbor index, and coordinate index */
720 jnrlistA = jjnr[jidx];
721 jnrlistB = jjnr[jidx+1];
722 jnrlistC = jjnr[jidx+2];
723 jnrlistD = jjnr[jidx+3];
724 /* Sign of each element will be negative for non-real atoms.
725 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
726 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
728 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
730 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
731 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
732 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
734 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
735 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
736 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
737 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
738 j_coord_offsetA = DIM*jnrA;
739 j_coord_offsetB = DIM*jnrB;
740 j_coord_offsetC = DIM*jnrC;
741 j_coord_offsetD = DIM*jnrD;
743 /* load j atom coordinates */
744 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
745 x+j_coord_offsetC,x+j_coord_offsetD,
748 /* Calculate displacement vector */
749 dx00 = _mm256_sub_pd(ix0,jx0);
750 dy00 = _mm256_sub_pd(iy0,jy0);
751 dz00 = _mm256_sub_pd(iz0,jz0);
753 /* Calculate squared distance and things based on it */
754 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
756 rinv00 = avx256_invsqrt_d(rsq00);
758 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
760 /* Load parameters for j particles */
761 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
762 charge+jnrC+0,charge+jnrD+0);
763 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
764 invsqrta+jnrC+0,invsqrta+jnrD+0);
765 vdwjidx0A = 2*vdwtype[jnrA+0];
766 vdwjidx0B = 2*vdwtype[jnrB+0];
767 vdwjidx0C = 2*vdwtype[jnrC+0];
768 vdwjidx0D = 2*vdwtype[jnrD+0];
770 /**************************
771 * CALCULATE INTERACTIONS *
772 **************************/
774 r00 = _mm256_mul_pd(rsq00,rinv00);
775 r00 = _mm256_andnot_pd(dummy_mask,r00);
777 /* Compute parameters for interactions between i and j atoms */
778 qq00 = _mm256_mul_pd(iq0,jq0);
779 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
780 vdwioffsetptr0+vdwjidx0B,
781 vdwioffsetptr0+vdwjidx0C,
782 vdwioffsetptr0+vdwjidx0D,
785 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
786 isaprod = _mm256_mul_pd(isai0,isaj0);
787 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
788 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
790 /* Calculate generalized born table index - this is a separate table from the normal one,
791 * but we use the same procedure by multiplying r with scale and truncating to integer.
793 rt = _mm256_mul_pd(r00,gbscale);
794 gbitab = _mm256_cvttpd_epi32(rt);
795 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
796 gbitab = _mm_slli_epi32(gbitab,2);
797 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
798 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
799 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
800 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
801 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
802 Heps = _mm256_mul_pd(gbeps,H);
803 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
804 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
805 vgb = _mm256_mul_pd(gbqqfactor,VV);
807 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
808 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
809 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
810 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
811 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
812 /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
813 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
814 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
815 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
816 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
817 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
818 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
819 velec = _mm256_mul_pd(qq00,rinv00);
820 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
822 /* LENNARD-JONES DISPERSION/REPULSION */
824 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
825 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
827 fscal = _mm256_add_pd(felec,fvdw);
829 fscal = _mm256_andnot_pd(dummy_mask,fscal);
831 /* Calculate temporary vectorial force */
832 tx = _mm256_mul_pd(fscal,dx00);
833 ty = _mm256_mul_pd(fscal,dy00);
834 tz = _mm256_mul_pd(fscal,dz00);
836 /* Update vectorial force */
837 fix0 = _mm256_add_pd(fix0,tx);
838 fiy0 = _mm256_add_pd(fiy0,ty);
839 fiz0 = _mm256_add_pd(fiz0,tz);
841 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
842 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
843 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
844 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
845 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
847 /* Inner loop uses 64 flops */
850 /* End of innermost loop */
852 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
853 f+i_coord_offset,fshift+i_shift_offset);
855 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
856 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
858 /* Increment number of inner iterations */
859 inneriter += j_index_end - j_index_start;
861 /* Outer loop uses 7 flops */
864 /* Increment number of outer iterations */
867 /* Update outer/inner flops */
869 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*64);