2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013, 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 "types/simple.h"
49 #include "gromacs/simd/math_x86_avx_256_double.h"
50 #include "kernelutil_x86_avx_256_double.h"
53 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_double
54 * Electrostatics interaction: GeneralizedBorn
55 * VdW interaction: LennardJones
56 * Geometry: Particle-Particle
57 * Calculate force/pot: PotentialAndForce
60 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_double
61 (t_nblist * gmx_restrict nlist,
62 rvec * gmx_restrict xx,
63 rvec * gmx_restrict ff,
64 t_forcerec * gmx_restrict fr,
65 t_mdatoms * gmx_restrict mdatoms,
66 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
67 t_nrnb * gmx_restrict nrnb)
69 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
70 * just 0 for non-waters.
71 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
72 * jnr indices corresponding to data put in the four positions in the SIMD register.
74 int i_shift_offset,i_coord_offset,outeriter,inneriter;
75 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
76 int jnrA,jnrB,jnrC,jnrD;
77 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
78 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
79 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
80 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
82 real *shiftvec,*fshift,*x,*f;
83 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
85 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
86 real * vdwioffsetptr0;
87 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
88 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
89 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
90 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
91 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
94 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
95 __m256d minushalf = _mm256_set1_pd(-0.5);
96 real *invsqrta,*dvda,*gbtab;
98 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
101 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
102 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
104 __m128i ifour = _mm_set1_epi32(4);
105 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
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;
129 invsqrta = fr->invsqrta;
131 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
132 gbtab = fr->gbtab.data;
133 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
135 /* Avoid stupid compiler warnings */
136 jnrA = jnrB = jnrC = jnrD = 0;
145 for(iidx=0;iidx<4*DIM;iidx++)
150 /* Start outer loop over neighborlists */
151 for(iidx=0; iidx<nri; iidx++)
153 /* Load shift vector for this list */
154 i_shift_offset = DIM*shiftidx[iidx];
156 /* Load limits for loop over neighbors */
157 j_index_start = jindex[iidx];
158 j_index_end = jindex[iidx+1];
160 /* Get outer coordinate index */
162 i_coord_offset = DIM*inr;
164 /* Load i particle coords and add shift vector */
165 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
167 fix0 = _mm256_setzero_pd();
168 fiy0 = _mm256_setzero_pd();
169 fiz0 = _mm256_setzero_pd();
171 /* Load parameters for i particles */
172 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
173 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
174 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
176 /* Reset potential sums */
177 velecsum = _mm256_setzero_pd();
178 vgbsum = _mm256_setzero_pd();
179 vvdwsum = _mm256_setzero_pd();
180 dvdasum = _mm256_setzero_pd();
182 /* Start inner kernel loop */
183 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
186 /* Get j neighbor index, and coordinate index */
191 j_coord_offsetA = DIM*jnrA;
192 j_coord_offsetB = DIM*jnrB;
193 j_coord_offsetC = DIM*jnrC;
194 j_coord_offsetD = DIM*jnrD;
196 /* load j atom coordinates */
197 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
198 x+j_coord_offsetC,x+j_coord_offsetD,
201 /* Calculate displacement vector */
202 dx00 = _mm256_sub_pd(ix0,jx0);
203 dy00 = _mm256_sub_pd(iy0,jy0);
204 dz00 = _mm256_sub_pd(iz0,jz0);
206 /* Calculate squared distance and things based on it */
207 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
209 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
211 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
213 /* Load parameters for j particles */
214 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
215 charge+jnrC+0,charge+jnrD+0);
216 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
217 invsqrta+jnrC+0,invsqrta+jnrD+0);
218 vdwjidx0A = 2*vdwtype[jnrA+0];
219 vdwjidx0B = 2*vdwtype[jnrB+0];
220 vdwjidx0C = 2*vdwtype[jnrC+0];
221 vdwjidx0D = 2*vdwtype[jnrD+0];
223 /**************************
224 * CALCULATE INTERACTIONS *
225 **************************/
227 r00 = _mm256_mul_pd(rsq00,rinv00);
229 /* Compute parameters for interactions between i and j atoms */
230 qq00 = _mm256_mul_pd(iq0,jq0);
231 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
232 vdwioffsetptr0+vdwjidx0B,
233 vdwioffsetptr0+vdwjidx0C,
234 vdwioffsetptr0+vdwjidx0D,
237 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
238 isaprod = _mm256_mul_pd(isai0,isaj0);
239 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
240 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
242 /* Calculate generalized born table index - this is a separate table from the normal one,
243 * but we use the same procedure by multiplying r with scale and truncating to integer.
245 rt = _mm256_mul_pd(r00,gbscale);
246 gbitab = _mm256_cvttpd_epi32(rt);
247 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
248 gbitab = _mm_slli_epi32(gbitab,2);
249 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
250 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
251 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
252 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
253 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
254 Heps = _mm256_mul_pd(gbeps,H);
255 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
256 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
257 vgb = _mm256_mul_pd(gbqqfactor,VV);
259 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
260 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
261 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
262 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
267 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
268 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
269 velec = _mm256_mul_pd(qq00,rinv00);
270 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
272 /* LENNARD-JONES DISPERSION/REPULSION */
274 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
275 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
276 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
277 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
278 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
280 /* Update potential sum for this i atom from the interaction with this j atom. */
281 velecsum = _mm256_add_pd(velecsum,velec);
282 vgbsum = _mm256_add_pd(vgbsum,vgb);
283 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
285 fscal = _mm256_add_pd(felec,fvdw);
287 /* Calculate temporary vectorial force */
288 tx = _mm256_mul_pd(fscal,dx00);
289 ty = _mm256_mul_pd(fscal,dy00);
290 tz = _mm256_mul_pd(fscal,dz00);
292 /* Update vectorial force */
293 fix0 = _mm256_add_pd(fix0,tx);
294 fiy0 = _mm256_add_pd(fiy0,ty);
295 fiz0 = _mm256_add_pd(fiz0,tz);
297 fjptrA = f+j_coord_offsetA;
298 fjptrB = f+j_coord_offsetB;
299 fjptrC = f+j_coord_offsetC;
300 fjptrD = f+j_coord_offsetD;
301 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
303 /* Inner loop uses 70 flops */
309 /* Get j neighbor index, and coordinate index */
310 jnrlistA = jjnr[jidx];
311 jnrlistB = jjnr[jidx+1];
312 jnrlistC = jjnr[jidx+2];
313 jnrlistD = jjnr[jidx+3];
314 /* Sign of each element will be negative for non-real atoms.
315 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
316 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
318 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
320 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
321 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
322 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
324 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
325 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
326 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
327 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
328 j_coord_offsetA = DIM*jnrA;
329 j_coord_offsetB = DIM*jnrB;
330 j_coord_offsetC = DIM*jnrC;
331 j_coord_offsetD = DIM*jnrD;
333 /* load j atom coordinates */
334 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
335 x+j_coord_offsetC,x+j_coord_offsetD,
338 /* Calculate displacement vector */
339 dx00 = _mm256_sub_pd(ix0,jx0);
340 dy00 = _mm256_sub_pd(iy0,jy0);
341 dz00 = _mm256_sub_pd(iz0,jz0);
343 /* Calculate squared distance and things based on it */
344 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
346 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
348 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
350 /* Load parameters for j particles */
351 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
352 charge+jnrC+0,charge+jnrD+0);
353 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
354 invsqrta+jnrC+0,invsqrta+jnrD+0);
355 vdwjidx0A = 2*vdwtype[jnrA+0];
356 vdwjidx0B = 2*vdwtype[jnrB+0];
357 vdwjidx0C = 2*vdwtype[jnrC+0];
358 vdwjidx0D = 2*vdwtype[jnrD+0];
360 /**************************
361 * CALCULATE INTERACTIONS *
362 **************************/
364 r00 = _mm256_mul_pd(rsq00,rinv00);
365 r00 = _mm256_andnot_pd(dummy_mask,r00);
367 /* Compute parameters for interactions between i and j atoms */
368 qq00 = _mm256_mul_pd(iq0,jq0);
369 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
370 vdwioffsetptr0+vdwjidx0B,
371 vdwioffsetptr0+vdwjidx0C,
372 vdwioffsetptr0+vdwjidx0D,
375 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
376 isaprod = _mm256_mul_pd(isai0,isaj0);
377 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
378 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
380 /* Calculate generalized born table index - this is a separate table from the normal one,
381 * but we use the same procedure by multiplying r with scale and truncating to integer.
383 rt = _mm256_mul_pd(r00,gbscale);
384 gbitab = _mm256_cvttpd_epi32(rt);
385 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
386 gbitab = _mm_slli_epi32(gbitab,2);
387 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
388 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
389 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
390 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
391 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
392 Heps = _mm256_mul_pd(gbeps,H);
393 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
394 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
395 vgb = _mm256_mul_pd(gbqqfactor,VV);
397 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
398 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
399 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
400 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
401 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
402 /* 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. */
403 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
404 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
405 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
406 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
407 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
408 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
409 velec = _mm256_mul_pd(qq00,rinv00);
410 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
412 /* LENNARD-JONES DISPERSION/REPULSION */
414 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
415 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
416 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
417 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
418 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
420 /* Update potential sum for this i atom from the interaction with this j atom. */
421 velec = _mm256_andnot_pd(dummy_mask,velec);
422 velecsum = _mm256_add_pd(velecsum,velec);
423 vgb = _mm256_andnot_pd(dummy_mask,vgb);
424 vgbsum = _mm256_add_pd(vgbsum,vgb);
425 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
426 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
428 fscal = _mm256_add_pd(felec,fvdw);
430 fscal = _mm256_andnot_pd(dummy_mask,fscal);
432 /* Calculate temporary vectorial force */
433 tx = _mm256_mul_pd(fscal,dx00);
434 ty = _mm256_mul_pd(fscal,dy00);
435 tz = _mm256_mul_pd(fscal,dz00);
437 /* Update vectorial force */
438 fix0 = _mm256_add_pd(fix0,tx);
439 fiy0 = _mm256_add_pd(fiy0,ty);
440 fiz0 = _mm256_add_pd(fiz0,tz);
442 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
443 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
444 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
445 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
446 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
448 /* Inner loop uses 71 flops */
451 /* End of innermost loop */
453 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
454 f+i_coord_offset,fshift+i_shift_offset);
457 /* Update potential energies */
458 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
459 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
460 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
461 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
462 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
464 /* Increment number of inner iterations */
465 inneriter += j_index_end - j_index_start;
467 /* Outer loop uses 10 flops */
470 /* Increment number of outer iterations */
473 /* Update outer/inner flops */
475 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*71);
478 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
479 * Electrostatics interaction: GeneralizedBorn
480 * VdW interaction: LennardJones
481 * Geometry: Particle-Particle
482 * Calculate force/pot: Force
485 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
486 (t_nblist * gmx_restrict nlist,
487 rvec * gmx_restrict xx,
488 rvec * gmx_restrict ff,
489 t_forcerec * gmx_restrict fr,
490 t_mdatoms * gmx_restrict mdatoms,
491 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
492 t_nrnb * gmx_restrict nrnb)
494 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
495 * just 0 for non-waters.
496 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
497 * jnr indices corresponding to data put in the four positions in the SIMD register.
499 int i_shift_offset,i_coord_offset,outeriter,inneriter;
500 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
501 int jnrA,jnrB,jnrC,jnrD;
502 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
503 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
504 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
505 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
507 real *shiftvec,*fshift,*x,*f;
508 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
510 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
511 real * vdwioffsetptr0;
512 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
513 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
514 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
515 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
516 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
519 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
520 __m256d minushalf = _mm256_set1_pd(-0.5);
521 real *invsqrta,*dvda,*gbtab;
523 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
526 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
527 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
529 __m128i ifour = _mm_set1_epi32(4);
530 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
532 __m256d dummy_mask,cutoff_mask;
533 __m128 tmpmask0,tmpmask1;
534 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
535 __m256d one = _mm256_set1_pd(1.0);
536 __m256d two = _mm256_set1_pd(2.0);
542 jindex = nlist->jindex;
544 shiftidx = nlist->shift;
546 shiftvec = fr->shift_vec[0];
547 fshift = fr->fshift[0];
548 facel = _mm256_set1_pd(fr->epsfac);
549 charge = mdatoms->chargeA;
550 nvdwtype = fr->ntype;
552 vdwtype = mdatoms->typeA;
554 invsqrta = fr->invsqrta;
556 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
557 gbtab = fr->gbtab.data;
558 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
560 /* Avoid stupid compiler warnings */
561 jnrA = jnrB = jnrC = jnrD = 0;
570 for(iidx=0;iidx<4*DIM;iidx++)
575 /* Start outer loop over neighborlists */
576 for(iidx=0; iidx<nri; iidx++)
578 /* Load shift vector for this list */
579 i_shift_offset = DIM*shiftidx[iidx];
581 /* Load limits for loop over neighbors */
582 j_index_start = jindex[iidx];
583 j_index_end = jindex[iidx+1];
585 /* Get outer coordinate index */
587 i_coord_offset = DIM*inr;
589 /* Load i particle coords and add shift vector */
590 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
592 fix0 = _mm256_setzero_pd();
593 fiy0 = _mm256_setzero_pd();
594 fiz0 = _mm256_setzero_pd();
596 /* Load parameters for i particles */
597 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
598 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
599 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
601 dvdasum = _mm256_setzero_pd();
603 /* Start inner kernel loop */
604 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
607 /* Get j neighbor index, and coordinate index */
612 j_coord_offsetA = DIM*jnrA;
613 j_coord_offsetB = DIM*jnrB;
614 j_coord_offsetC = DIM*jnrC;
615 j_coord_offsetD = DIM*jnrD;
617 /* load j atom coordinates */
618 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
619 x+j_coord_offsetC,x+j_coord_offsetD,
622 /* Calculate displacement vector */
623 dx00 = _mm256_sub_pd(ix0,jx0);
624 dy00 = _mm256_sub_pd(iy0,jy0);
625 dz00 = _mm256_sub_pd(iz0,jz0);
627 /* Calculate squared distance and things based on it */
628 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
630 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
632 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
634 /* Load parameters for j particles */
635 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
636 charge+jnrC+0,charge+jnrD+0);
637 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
638 invsqrta+jnrC+0,invsqrta+jnrD+0);
639 vdwjidx0A = 2*vdwtype[jnrA+0];
640 vdwjidx0B = 2*vdwtype[jnrB+0];
641 vdwjidx0C = 2*vdwtype[jnrC+0];
642 vdwjidx0D = 2*vdwtype[jnrD+0];
644 /**************************
645 * CALCULATE INTERACTIONS *
646 **************************/
648 r00 = _mm256_mul_pd(rsq00,rinv00);
650 /* Compute parameters for interactions between i and j atoms */
651 qq00 = _mm256_mul_pd(iq0,jq0);
652 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
653 vdwioffsetptr0+vdwjidx0B,
654 vdwioffsetptr0+vdwjidx0C,
655 vdwioffsetptr0+vdwjidx0D,
658 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
659 isaprod = _mm256_mul_pd(isai0,isaj0);
660 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
661 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
663 /* Calculate generalized born table index - this is a separate table from the normal one,
664 * but we use the same procedure by multiplying r with scale and truncating to integer.
666 rt = _mm256_mul_pd(r00,gbscale);
667 gbitab = _mm256_cvttpd_epi32(rt);
668 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
669 gbitab = _mm_slli_epi32(gbitab,2);
670 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
671 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
672 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
673 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
674 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
675 Heps = _mm256_mul_pd(gbeps,H);
676 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
677 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
678 vgb = _mm256_mul_pd(gbqqfactor,VV);
680 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
681 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
682 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
683 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
688 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
689 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
690 velec = _mm256_mul_pd(qq00,rinv00);
691 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
693 /* LENNARD-JONES DISPERSION/REPULSION */
695 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
696 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
698 fscal = _mm256_add_pd(felec,fvdw);
700 /* Calculate temporary vectorial force */
701 tx = _mm256_mul_pd(fscal,dx00);
702 ty = _mm256_mul_pd(fscal,dy00);
703 tz = _mm256_mul_pd(fscal,dz00);
705 /* Update vectorial force */
706 fix0 = _mm256_add_pd(fix0,tx);
707 fiy0 = _mm256_add_pd(fiy0,ty);
708 fiz0 = _mm256_add_pd(fiz0,tz);
710 fjptrA = f+j_coord_offsetA;
711 fjptrB = f+j_coord_offsetB;
712 fjptrC = f+j_coord_offsetC;
713 fjptrD = f+j_coord_offsetD;
714 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
716 /* Inner loop uses 63 flops */
722 /* Get j neighbor index, and coordinate index */
723 jnrlistA = jjnr[jidx];
724 jnrlistB = jjnr[jidx+1];
725 jnrlistC = jjnr[jidx+2];
726 jnrlistD = jjnr[jidx+3];
727 /* Sign of each element will be negative for non-real atoms.
728 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
729 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
731 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
733 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
734 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
735 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
737 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
738 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
739 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
740 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
741 j_coord_offsetA = DIM*jnrA;
742 j_coord_offsetB = DIM*jnrB;
743 j_coord_offsetC = DIM*jnrC;
744 j_coord_offsetD = DIM*jnrD;
746 /* load j atom coordinates */
747 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
748 x+j_coord_offsetC,x+j_coord_offsetD,
751 /* Calculate displacement vector */
752 dx00 = _mm256_sub_pd(ix0,jx0);
753 dy00 = _mm256_sub_pd(iy0,jy0);
754 dz00 = _mm256_sub_pd(iz0,jz0);
756 /* Calculate squared distance and things based on it */
757 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
759 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
761 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
763 /* Load parameters for j particles */
764 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
765 charge+jnrC+0,charge+jnrD+0);
766 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
767 invsqrta+jnrC+0,invsqrta+jnrD+0);
768 vdwjidx0A = 2*vdwtype[jnrA+0];
769 vdwjidx0B = 2*vdwtype[jnrB+0];
770 vdwjidx0C = 2*vdwtype[jnrC+0];
771 vdwjidx0D = 2*vdwtype[jnrD+0];
773 /**************************
774 * CALCULATE INTERACTIONS *
775 **************************/
777 r00 = _mm256_mul_pd(rsq00,rinv00);
778 r00 = _mm256_andnot_pd(dummy_mask,r00);
780 /* Compute parameters for interactions between i and j atoms */
781 qq00 = _mm256_mul_pd(iq0,jq0);
782 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
783 vdwioffsetptr0+vdwjidx0B,
784 vdwioffsetptr0+vdwjidx0C,
785 vdwioffsetptr0+vdwjidx0D,
788 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
789 isaprod = _mm256_mul_pd(isai0,isaj0);
790 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
791 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
793 /* Calculate generalized born table index - this is a separate table from the normal one,
794 * but we use the same procedure by multiplying r with scale and truncating to integer.
796 rt = _mm256_mul_pd(r00,gbscale);
797 gbitab = _mm256_cvttpd_epi32(rt);
798 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
799 gbitab = _mm_slli_epi32(gbitab,2);
800 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
801 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
802 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
803 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
804 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
805 Heps = _mm256_mul_pd(gbeps,H);
806 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
807 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
808 vgb = _mm256_mul_pd(gbqqfactor,VV);
810 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
811 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
812 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
813 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
814 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
815 /* 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. */
816 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
817 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
818 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
819 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
820 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
821 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
822 velec = _mm256_mul_pd(qq00,rinv00);
823 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
825 /* LENNARD-JONES DISPERSION/REPULSION */
827 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
828 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
830 fscal = _mm256_add_pd(felec,fvdw);
832 fscal = _mm256_andnot_pd(dummy_mask,fscal);
834 /* Calculate temporary vectorial force */
835 tx = _mm256_mul_pd(fscal,dx00);
836 ty = _mm256_mul_pd(fscal,dy00);
837 tz = _mm256_mul_pd(fscal,dz00);
839 /* Update vectorial force */
840 fix0 = _mm256_add_pd(fix0,tx);
841 fiy0 = _mm256_add_pd(fiy0,ty);
842 fiz0 = _mm256_add_pd(fiz0,tz);
844 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
845 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
846 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
847 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
848 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
850 /* Inner loop uses 64 flops */
853 /* End of innermost loop */
855 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
856 f+i_coord_offset,fshift+i_shift_offset);
858 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
859 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
861 /* Increment number of inner iterations */
862 inneriter += j_index_end - j_index_start;
864 /* Outer loop uses 7 flops */
867 /* Increment number of outer iterations */
870 /* Update outer/inner flops */
872 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*64);