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 "types/simple.h"
44 #include "gromacs/math/vec.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_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_double
52 * Electrostatics interaction: GeneralizedBorn
53 * VdW interaction: LennardJones
54 * Geometry: Particle-Particle
55 * Calculate force/pot: PotentialAndForce
58 nb_kernel_ElecGB_VdwLJ_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 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
86 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
87 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
88 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
89 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
92 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
93 __m256d minushalf = _mm256_set1_pd(-0.5);
94 real *invsqrta,*dvda,*gbtab;
96 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
99 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
100 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
102 __m128i ifour = _mm_set1_epi32(4);
103 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
105 __m256d dummy_mask,cutoff_mask;
106 __m128 tmpmask0,tmpmask1;
107 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
108 __m256d one = _mm256_set1_pd(1.0);
109 __m256d two = _mm256_set1_pd(2.0);
115 jindex = nlist->jindex;
117 shiftidx = nlist->shift;
119 shiftvec = fr->shift_vec[0];
120 fshift = fr->fshift[0];
121 facel = _mm256_set1_pd(fr->epsfac);
122 charge = mdatoms->chargeA;
123 nvdwtype = fr->ntype;
125 vdwtype = mdatoms->typeA;
127 invsqrta = fr->invsqrta;
129 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
130 gbtab = fr->gbtab.data;
131 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
133 /* Avoid stupid compiler warnings */
134 jnrA = jnrB = jnrC = jnrD = 0;
143 for(iidx=0;iidx<4*DIM;iidx++)
148 /* Start outer loop over neighborlists */
149 for(iidx=0; iidx<nri; iidx++)
151 /* Load shift vector for this list */
152 i_shift_offset = DIM*shiftidx[iidx];
154 /* Load limits for loop over neighbors */
155 j_index_start = jindex[iidx];
156 j_index_end = jindex[iidx+1];
158 /* Get outer coordinate index */
160 i_coord_offset = DIM*inr;
162 /* Load i particle coords and add shift vector */
163 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
165 fix0 = _mm256_setzero_pd();
166 fiy0 = _mm256_setzero_pd();
167 fiz0 = _mm256_setzero_pd();
169 /* Load parameters for i particles */
170 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
171 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
172 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
174 /* Reset potential sums */
175 velecsum = _mm256_setzero_pd();
176 vgbsum = _mm256_setzero_pd();
177 vvdwsum = _mm256_setzero_pd();
178 dvdasum = _mm256_setzero_pd();
180 /* Start inner kernel loop */
181 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
184 /* Get j neighbor index, and coordinate index */
189 j_coord_offsetA = DIM*jnrA;
190 j_coord_offsetB = DIM*jnrB;
191 j_coord_offsetC = DIM*jnrC;
192 j_coord_offsetD = DIM*jnrD;
194 /* load j atom coordinates */
195 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
196 x+j_coord_offsetC,x+j_coord_offsetD,
199 /* Calculate displacement vector */
200 dx00 = _mm256_sub_pd(ix0,jx0);
201 dy00 = _mm256_sub_pd(iy0,jy0);
202 dz00 = _mm256_sub_pd(iz0,jz0);
204 /* Calculate squared distance and things based on it */
205 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
207 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
209 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
211 /* Load parameters for j particles */
212 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
213 charge+jnrC+0,charge+jnrD+0);
214 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
215 invsqrta+jnrC+0,invsqrta+jnrD+0);
216 vdwjidx0A = 2*vdwtype[jnrA+0];
217 vdwjidx0B = 2*vdwtype[jnrB+0];
218 vdwjidx0C = 2*vdwtype[jnrC+0];
219 vdwjidx0D = 2*vdwtype[jnrD+0];
221 /**************************
222 * CALCULATE INTERACTIONS *
223 **************************/
225 r00 = _mm256_mul_pd(rsq00,rinv00);
227 /* Compute parameters for interactions between i and j atoms */
228 qq00 = _mm256_mul_pd(iq0,jq0);
229 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
230 vdwioffsetptr0+vdwjidx0B,
231 vdwioffsetptr0+vdwjidx0C,
232 vdwioffsetptr0+vdwjidx0D,
235 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
236 isaprod = _mm256_mul_pd(isai0,isaj0);
237 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
238 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
240 /* Calculate generalized born table index - this is a separate table from the normal one,
241 * but we use the same procedure by multiplying r with scale and truncating to integer.
243 rt = _mm256_mul_pd(r00,gbscale);
244 gbitab = _mm256_cvttpd_epi32(rt);
245 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
246 gbitab = _mm_slli_epi32(gbitab,2);
247 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
248 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
249 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
250 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
251 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
252 Heps = _mm256_mul_pd(gbeps,H);
253 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
254 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
255 vgb = _mm256_mul_pd(gbqqfactor,VV);
257 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
258 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
259 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
260 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
265 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
266 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
267 velec = _mm256_mul_pd(qq00,rinv00);
268 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
270 /* LENNARD-JONES DISPERSION/REPULSION */
272 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
273 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
274 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
275 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
276 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
278 /* Update potential sum for this i atom from the interaction with this j atom. */
279 velecsum = _mm256_add_pd(velecsum,velec);
280 vgbsum = _mm256_add_pd(vgbsum,vgb);
281 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
283 fscal = _mm256_add_pd(felec,fvdw);
285 /* Calculate temporary vectorial force */
286 tx = _mm256_mul_pd(fscal,dx00);
287 ty = _mm256_mul_pd(fscal,dy00);
288 tz = _mm256_mul_pd(fscal,dz00);
290 /* Update vectorial force */
291 fix0 = _mm256_add_pd(fix0,tx);
292 fiy0 = _mm256_add_pd(fiy0,ty);
293 fiz0 = _mm256_add_pd(fiz0,tz);
295 fjptrA = f+j_coord_offsetA;
296 fjptrB = f+j_coord_offsetB;
297 fjptrC = f+j_coord_offsetC;
298 fjptrD = f+j_coord_offsetD;
299 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
301 /* Inner loop uses 70 flops */
307 /* Get j neighbor index, and coordinate index */
308 jnrlistA = jjnr[jidx];
309 jnrlistB = jjnr[jidx+1];
310 jnrlistC = jjnr[jidx+2];
311 jnrlistD = jjnr[jidx+3];
312 /* Sign of each element will be negative for non-real atoms.
313 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
314 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
316 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
318 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
319 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
320 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
322 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
323 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
324 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
325 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
326 j_coord_offsetA = DIM*jnrA;
327 j_coord_offsetB = DIM*jnrB;
328 j_coord_offsetC = DIM*jnrC;
329 j_coord_offsetD = DIM*jnrD;
331 /* load j atom coordinates */
332 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
333 x+j_coord_offsetC,x+j_coord_offsetD,
336 /* Calculate displacement vector */
337 dx00 = _mm256_sub_pd(ix0,jx0);
338 dy00 = _mm256_sub_pd(iy0,jy0);
339 dz00 = _mm256_sub_pd(iz0,jz0);
341 /* Calculate squared distance and things based on it */
342 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
344 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
346 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
348 /* Load parameters for j particles */
349 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
350 charge+jnrC+0,charge+jnrD+0);
351 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
352 invsqrta+jnrC+0,invsqrta+jnrD+0);
353 vdwjidx0A = 2*vdwtype[jnrA+0];
354 vdwjidx0B = 2*vdwtype[jnrB+0];
355 vdwjidx0C = 2*vdwtype[jnrC+0];
356 vdwjidx0D = 2*vdwtype[jnrD+0];
358 /**************************
359 * CALCULATE INTERACTIONS *
360 **************************/
362 r00 = _mm256_mul_pd(rsq00,rinv00);
363 r00 = _mm256_andnot_pd(dummy_mask,r00);
365 /* Compute parameters for interactions between i and j atoms */
366 qq00 = _mm256_mul_pd(iq0,jq0);
367 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
368 vdwioffsetptr0+vdwjidx0B,
369 vdwioffsetptr0+vdwjidx0C,
370 vdwioffsetptr0+vdwjidx0D,
373 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
374 isaprod = _mm256_mul_pd(isai0,isaj0);
375 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
376 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
378 /* Calculate generalized born table index - this is a separate table from the normal one,
379 * but we use the same procedure by multiplying r with scale and truncating to integer.
381 rt = _mm256_mul_pd(r00,gbscale);
382 gbitab = _mm256_cvttpd_epi32(rt);
383 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
384 gbitab = _mm_slli_epi32(gbitab,2);
385 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
386 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
387 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
388 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
389 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
390 Heps = _mm256_mul_pd(gbeps,H);
391 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
392 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
393 vgb = _mm256_mul_pd(gbqqfactor,VV);
395 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
396 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
397 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
398 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
399 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
400 /* 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. */
401 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
402 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
403 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
404 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
405 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
406 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
407 velec = _mm256_mul_pd(qq00,rinv00);
408 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
410 /* LENNARD-JONES DISPERSION/REPULSION */
412 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
413 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
414 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
415 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
416 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
418 /* Update potential sum for this i atom from the interaction with this j atom. */
419 velec = _mm256_andnot_pd(dummy_mask,velec);
420 velecsum = _mm256_add_pd(velecsum,velec);
421 vgb = _mm256_andnot_pd(dummy_mask,vgb);
422 vgbsum = _mm256_add_pd(vgbsum,vgb);
423 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
424 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
426 fscal = _mm256_add_pd(felec,fvdw);
428 fscal = _mm256_andnot_pd(dummy_mask,fscal);
430 /* Calculate temporary vectorial force */
431 tx = _mm256_mul_pd(fscal,dx00);
432 ty = _mm256_mul_pd(fscal,dy00);
433 tz = _mm256_mul_pd(fscal,dz00);
435 /* Update vectorial force */
436 fix0 = _mm256_add_pd(fix0,tx);
437 fiy0 = _mm256_add_pd(fiy0,ty);
438 fiz0 = _mm256_add_pd(fiz0,tz);
440 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
441 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
442 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
443 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
444 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
446 /* Inner loop uses 71 flops */
449 /* End of innermost loop */
451 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
452 f+i_coord_offset,fshift+i_shift_offset);
455 /* Update potential energies */
456 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
457 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
458 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
459 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
460 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
462 /* Increment number of inner iterations */
463 inneriter += j_index_end - j_index_start;
465 /* Outer loop uses 10 flops */
468 /* Increment number of outer iterations */
471 /* Update outer/inner flops */
473 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*71);
476 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
477 * Electrostatics interaction: GeneralizedBorn
478 * VdW interaction: LennardJones
479 * Geometry: Particle-Particle
480 * Calculate force/pot: Force
483 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
484 (t_nblist * gmx_restrict nlist,
485 rvec * gmx_restrict xx,
486 rvec * gmx_restrict ff,
487 t_forcerec * gmx_restrict fr,
488 t_mdatoms * gmx_restrict mdatoms,
489 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
490 t_nrnb * gmx_restrict nrnb)
492 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
493 * just 0 for non-waters.
494 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
495 * jnr indices corresponding to data put in the four positions in the SIMD register.
497 int i_shift_offset,i_coord_offset,outeriter,inneriter;
498 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
499 int jnrA,jnrB,jnrC,jnrD;
500 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
501 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
502 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
503 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
505 real *shiftvec,*fshift,*x,*f;
506 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
508 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
509 real * vdwioffsetptr0;
510 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
511 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
512 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
513 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
514 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
517 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
518 __m256d minushalf = _mm256_set1_pd(-0.5);
519 real *invsqrta,*dvda,*gbtab;
521 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
524 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
525 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
527 __m128i ifour = _mm_set1_epi32(4);
528 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
530 __m256d dummy_mask,cutoff_mask;
531 __m128 tmpmask0,tmpmask1;
532 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
533 __m256d one = _mm256_set1_pd(1.0);
534 __m256d two = _mm256_set1_pd(2.0);
540 jindex = nlist->jindex;
542 shiftidx = nlist->shift;
544 shiftvec = fr->shift_vec[0];
545 fshift = fr->fshift[0];
546 facel = _mm256_set1_pd(fr->epsfac);
547 charge = mdatoms->chargeA;
548 nvdwtype = fr->ntype;
550 vdwtype = mdatoms->typeA;
552 invsqrta = fr->invsqrta;
554 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
555 gbtab = fr->gbtab.data;
556 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
558 /* Avoid stupid compiler warnings */
559 jnrA = jnrB = jnrC = jnrD = 0;
568 for(iidx=0;iidx<4*DIM;iidx++)
573 /* Start outer loop over neighborlists */
574 for(iidx=0; iidx<nri; iidx++)
576 /* Load shift vector for this list */
577 i_shift_offset = DIM*shiftidx[iidx];
579 /* Load limits for loop over neighbors */
580 j_index_start = jindex[iidx];
581 j_index_end = jindex[iidx+1];
583 /* Get outer coordinate index */
585 i_coord_offset = DIM*inr;
587 /* Load i particle coords and add shift vector */
588 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
590 fix0 = _mm256_setzero_pd();
591 fiy0 = _mm256_setzero_pd();
592 fiz0 = _mm256_setzero_pd();
594 /* Load parameters for i particles */
595 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
596 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
597 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
599 dvdasum = _mm256_setzero_pd();
601 /* Start inner kernel loop */
602 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
605 /* Get j neighbor index, and coordinate index */
610 j_coord_offsetA = DIM*jnrA;
611 j_coord_offsetB = DIM*jnrB;
612 j_coord_offsetC = DIM*jnrC;
613 j_coord_offsetD = DIM*jnrD;
615 /* load j atom coordinates */
616 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
617 x+j_coord_offsetC,x+j_coord_offsetD,
620 /* Calculate displacement vector */
621 dx00 = _mm256_sub_pd(ix0,jx0);
622 dy00 = _mm256_sub_pd(iy0,jy0);
623 dz00 = _mm256_sub_pd(iz0,jz0);
625 /* Calculate squared distance and things based on it */
626 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
628 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
630 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
632 /* Load parameters for j particles */
633 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
634 charge+jnrC+0,charge+jnrD+0);
635 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
636 invsqrta+jnrC+0,invsqrta+jnrD+0);
637 vdwjidx0A = 2*vdwtype[jnrA+0];
638 vdwjidx0B = 2*vdwtype[jnrB+0];
639 vdwjidx0C = 2*vdwtype[jnrC+0];
640 vdwjidx0D = 2*vdwtype[jnrD+0];
642 /**************************
643 * CALCULATE INTERACTIONS *
644 **************************/
646 r00 = _mm256_mul_pd(rsq00,rinv00);
648 /* Compute parameters for interactions between i and j atoms */
649 qq00 = _mm256_mul_pd(iq0,jq0);
650 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
651 vdwioffsetptr0+vdwjidx0B,
652 vdwioffsetptr0+vdwjidx0C,
653 vdwioffsetptr0+vdwjidx0D,
656 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
657 isaprod = _mm256_mul_pd(isai0,isaj0);
658 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
659 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
661 /* Calculate generalized born table index - this is a separate table from the normal one,
662 * but we use the same procedure by multiplying r with scale and truncating to integer.
664 rt = _mm256_mul_pd(r00,gbscale);
665 gbitab = _mm256_cvttpd_epi32(rt);
666 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
667 gbitab = _mm_slli_epi32(gbitab,2);
668 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
669 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
670 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
671 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
672 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
673 Heps = _mm256_mul_pd(gbeps,H);
674 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
675 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
676 vgb = _mm256_mul_pd(gbqqfactor,VV);
678 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
679 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
680 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
681 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
686 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
687 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
688 velec = _mm256_mul_pd(qq00,rinv00);
689 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
691 /* LENNARD-JONES DISPERSION/REPULSION */
693 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
694 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
696 fscal = _mm256_add_pd(felec,fvdw);
698 /* Calculate temporary vectorial force */
699 tx = _mm256_mul_pd(fscal,dx00);
700 ty = _mm256_mul_pd(fscal,dy00);
701 tz = _mm256_mul_pd(fscal,dz00);
703 /* Update vectorial force */
704 fix0 = _mm256_add_pd(fix0,tx);
705 fiy0 = _mm256_add_pd(fiy0,ty);
706 fiz0 = _mm256_add_pd(fiz0,tz);
708 fjptrA = f+j_coord_offsetA;
709 fjptrB = f+j_coord_offsetB;
710 fjptrC = f+j_coord_offsetC;
711 fjptrD = f+j_coord_offsetD;
712 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
714 /* Inner loop uses 63 flops */
720 /* Get j neighbor index, and coordinate index */
721 jnrlistA = jjnr[jidx];
722 jnrlistB = jjnr[jidx+1];
723 jnrlistC = jjnr[jidx+2];
724 jnrlistD = jjnr[jidx+3];
725 /* Sign of each element will be negative for non-real atoms.
726 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
727 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
729 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
731 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
732 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
733 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
735 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
736 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
737 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
738 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
739 j_coord_offsetA = DIM*jnrA;
740 j_coord_offsetB = DIM*jnrB;
741 j_coord_offsetC = DIM*jnrC;
742 j_coord_offsetD = DIM*jnrD;
744 /* load j atom coordinates */
745 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
746 x+j_coord_offsetC,x+j_coord_offsetD,
749 /* Calculate displacement vector */
750 dx00 = _mm256_sub_pd(ix0,jx0);
751 dy00 = _mm256_sub_pd(iy0,jy0);
752 dz00 = _mm256_sub_pd(iz0,jz0);
754 /* Calculate squared distance and things based on it */
755 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
757 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
759 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
761 /* Load parameters for j particles */
762 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
763 charge+jnrC+0,charge+jnrD+0);
764 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
765 invsqrta+jnrC+0,invsqrta+jnrD+0);
766 vdwjidx0A = 2*vdwtype[jnrA+0];
767 vdwjidx0B = 2*vdwtype[jnrB+0];
768 vdwjidx0C = 2*vdwtype[jnrC+0];
769 vdwjidx0D = 2*vdwtype[jnrD+0];
771 /**************************
772 * CALCULATE INTERACTIONS *
773 **************************/
775 r00 = _mm256_mul_pd(rsq00,rinv00);
776 r00 = _mm256_andnot_pd(dummy_mask,r00);
778 /* Compute parameters for interactions between i and j atoms */
779 qq00 = _mm256_mul_pd(iq0,jq0);
780 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
781 vdwioffsetptr0+vdwjidx0B,
782 vdwioffsetptr0+vdwjidx0C,
783 vdwioffsetptr0+vdwjidx0D,
786 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
787 isaprod = _mm256_mul_pd(isai0,isaj0);
788 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
789 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
791 /* Calculate generalized born table index - this is a separate table from the normal one,
792 * but we use the same procedure by multiplying r with scale and truncating to integer.
794 rt = _mm256_mul_pd(r00,gbscale);
795 gbitab = _mm256_cvttpd_epi32(rt);
796 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
797 gbitab = _mm_slli_epi32(gbitab,2);
798 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
799 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
800 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
801 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
802 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
803 Heps = _mm256_mul_pd(gbeps,H);
804 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
805 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
806 vgb = _mm256_mul_pd(gbqqfactor,VV);
808 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
809 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
810 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
811 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
812 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
813 /* 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. */
814 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
815 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
816 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
817 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
818 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
819 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
820 velec = _mm256_mul_pd(qq00,rinv00);
821 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
823 /* LENNARD-JONES DISPERSION/REPULSION */
825 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
826 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
828 fscal = _mm256_add_pd(felec,fvdw);
830 fscal = _mm256_andnot_pd(dummy_mask,fscal);
832 /* Calculate temporary vectorial force */
833 tx = _mm256_mul_pd(fscal,dx00);
834 ty = _mm256_mul_pd(fscal,dy00);
835 tz = _mm256_mul_pd(fscal,dz00);
837 /* Update vectorial force */
838 fix0 = _mm256_add_pd(fix0,tx);
839 fiy0 = _mm256_add_pd(fiy0,ty);
840 fiz0 = _mm256_add_pd(fiz0,tz);
842 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
843 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
844 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
845 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
846 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
848 /* Inner loop uses 64 flops */
851 /* End of innermost loop */
853 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
854 f+i_coord_offset,fshift+i_shift_offset);
856 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
857 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
859 /* Increment number of inner iterations */
860 inneriter += j_index_end - j_index_start;
862 /* Outer loop uses 7 flops */
865 /* Increment number of outer iterations */
868 /* Update outer/inner flops */
870 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*64);