2 * Note: this file was generated by the Gromacs avx_256_double kernel generator.
4 * This source code is part of
8 * Copyright (c) 2001-2012, The GROMACS Development Team
10 * Gromacs is a library for molecular simulation and trajectory analysis,
11 * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
12 * a full list of developers and information, check out http://www.gromacs.org
14 * This program is free software; you can redistribute it and/or modify it under
15 * the terms of the GNU Lesser General Public License as published by the Free
16 * Software Foundation; either version 2 of the License, or (at your option) any
19 * To help fund GROMACS development, we humbly ask that you cite
20 * the papers people have written on it - you can find them on the website.
28 #include "../nb_kernel.h"
29 #include "types/simple.h"
33 #include "gmx_math_x86_avx_256_double.h"
34 #include "kernelutil_x86_avx_256_double.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_double
38 * Electrostatics interaction: GeneralizedBorn
39 * VdW interaction: LennardJones
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_avx_256_double
45 (t_nblist * gmx_restrict nlist,
46 rvec * gmx_restrict xx,
47 rvec * gmx_restrict ff,
48 t_forcerec * gmx_restrict fr,
49 t_mdatoms * gmx_restrict mdatoms,
50 nb_kernel_data_t * gmx_restrict kernel_data,
51 t_nrnb * gmx_restrict nrnb)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
62 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
63 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
64 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
66 real *shiftvec,*fshift,*x,*f;
67 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
69 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
70 real * vdwioffsetptr0;
71 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
72 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
73 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
74 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
75 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
78 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
79 __m256d minushalf = _mm256_set1_pd(-0.5);
80 real *invsqrta,*dvda,*gbtab;
82 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
85 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
86 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
88 __m128i ifour = _mm_set1_epi32(4);
89 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
91 __m256d dummy_mask,cutoff_mask;
92 __m128 tmpmask0,tmpmask1;
93 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
94 __m256d one = _mm256_set1_pd(1.0);
95 __m256d two = _mm256_set1_pd(2.0);
101 jindex = nlist->jindex;
103 shiftidx = nlist->shift;
105 shiftvec = fr->shift_vec[0];
106 fshift = fr->fshift[0];
107 facel = _mm256_set1_pd(fr->epsfac);
108 charge = mdatoms->chargeA;
109 nvdwtype = fr->ntype;
111 vdwtype = mdatoms->typeA;
113 invsqrta = fr->invsqrta;
115 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
116 gbtab = fr->gbtab.data;
117 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
119 /* Avoid stupid compiler warnings */
120 jnrA = jnrB = jnrC = jnrD = 0;
129 for(iidx=0;iidx<4*DIM;iidx++)
134 /* Start outer loop over neighborlists */
135 for(iidx=0; iidx<nri; iidx++)
137 /* Load shift vector for this list */
138 i_shift_offset = DIM*shiftidx[iidx];
140 /* Load limits for loop over neighbors */
141 j_index_start = jindex[iidx];
142 j_index_end = jindex[iidx+1];
144 /* Get outer coordinate index */
146 i_coord_offset = DIM*inr;
148 /* Load i particle coords and add shift vector */
149 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
151 fix0 = _mm256_setzero_pd();
152 fiy0 = _mm256_setzero_pd();
153 fiz0 = _mm256_setzero_pd();
155 /* Load parameters for i particles */
156 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
157 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
158 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
160 /* Reset potential sums */
161 velecsum = _mm256_setzero_pd();
162 vgbsum = _mm256_setzero_pd();
163 vvdwsum = _mm256_setzero_pd();
164 dvdasum = _mm256_setzero_pd();
166 /* Start inner kernel loop */
167 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
170 /* Get j neighbor index, and coordinate index */
175 j_coord_offsetA = DIM*jnrA;
176 j_coord_offsetB = DIM*jnrB;
177 j_coord_offsetC = DIM*jnrC;
178 j_coord_offsetD = DIM*jnrD;
180 /* load j atom coordinates */
181 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
182 x+j_coord_offsetC,x+j_coord_offsetD,
185 /* Calculate displacement vector */
186 dx00 = _mm256_sub_pd(ix0,jx0);
187 dy00 = _mm256_sub_pd(iy0,jy0);
188 dz00 = _mm256_sub_pd(iz0,jz0);
190 /* Calculate squared distance and things based on it */
191 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
193 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
195 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
197 /* Load parameters for j particles */
198 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
199 charge+jnrC+0,charge+jnrD+0);
200 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
201 invsqrta+jnrC+0,invsqrta+jnrD+0);
202 vdwjidx0A = 2*vdwtype[jnrA+0];
203 vdwjidx0B = 2*vdwtype[jnrB+0];
204 vdwjidx0C = 2*vdwtype[jnrC+0];
205 vdwjidx0D = 2*vdwtype[jnrD+0];
207 /**************************
208 * CALCULATE INTERACTIONS *
209 **************************/
211 r00 = _mm256_mul_pd(rsq00,rinv00);
213 /* Compute parameters for interactions between i and j atoms */
214 qq00 = _mm256_mul_pd(iq0,jq0);
215 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
216 vdwioffsetptr0+vdwjidx0B,
217 vdwioffsetptr0+vdwjidx0C,
218 vdwioffsetptr0+vdwjidx0D,
221 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
222 isaprod = _mm256_mul_pd(isai0,isaj0);
223 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
224 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
226 /* Calculate generalized born table index - this is a separate table from the normal one,
227 * but we use the same procedure by multiplying r with scale and truncating to integer.
229 rt = _mm256_mul_pd(r00,gbscale);
230 gbitab = _mm256_cvttpd_epi32(rt);
231 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
232 gbitab = _mm_slli_epi32(gbitab,2);
233 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
234 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
235 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
236 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
237 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
238 Heps = _mm256_mul_pd(gbeps,H);
239 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
240 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
241 vgb = _mm256_mul_pd(gbqqfactor,VV);
243 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
244 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
245 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
246 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
251 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
252 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
253 velec = _mm256_mul_pd(qq00,rinv00);
254 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
256 /* LENNARD-JONES DISPERSION/REPULSION */
258 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
259 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
260 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
261 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
262 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
264 /* Update potential sum for this i atom from the interaction with this j atom. */
265 velecsum = _mm256_add_pd(velecsum,velec);
266 vgbsum = _mm256_add_pd(vgbsum,vgb);
267 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
269 fscal = _mm256_add_pd(felec,fvdw);
271 /* Calculate temporary vectorial force */
272 tx = _mm256_mul_pd(fscal,dx00);
273 ty = _mm256_mul_pd(fscal,dy00);
274 tz = _mm256_mul_pd(fscal,dz00);
276 /* Update vectorial force */
277 fix0 = _mm256_add_pd(fix0,tx);
278 fiy0 = _mm256_add_pd(fiy0,ty);
279 fiz0 = _mm256_add_pd(fiz0,tz);
281 fjptrA = f+j_coord_offsetA;
282 fjptrB = f+j_coord_offsetB;
283 fjptrC = f+j_coord_offsetC;
284 fjptrD = f+j_coord_offsetD;
285 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
287 /* Inner loop uses 70 flops */
293 /* Get j neighbor index, and coordinate index */
294 jnrlistA = jjnr[jidx];
295 jnrlistB = jjnr[jidx+1];
296 jnrlistC = jjnr[jidx+2];
297 jnrlistD = jjnr[jidx+3];
298 /* Sign of each element will be negative for non-real atoms.
299 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
300 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
302 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
304 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
305 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
306 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
308 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
309 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
310 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
311 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
312 j_coord_offsetA = DIM*jnrA;
313 j_coord_offsetB = DIM*jnrB;
314 j_coord_offsetC = DIM*jnrC;
315 j_coord_offsetD = DIM*jnrD;
317 /* load j atom coordinates */
318 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
319 x+j_coord_offsetC,x+j_coord_offsetD,
322 /* Calculate displacement vector */
323 dx00 = _mm256_sub_pd(ix0,jx0);
324 dy00 = _mm256_sub_pd(iy0,jy0);
325 dz00 = _mm256_sub_pd(iz0,jz0);
327 /* Calculate squared distance and things based on it */
328 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
330 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
332 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
334 /* Load parameters for j particles */
335 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
336 charge+jnrC+0,charge+jnrD+0);
337 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
338 invsqrta+jnrC+0,invsqrta+jnrD+0);
339 vdwjidx0A = 2*vdwtype[jnrA+0];
340 vdwjidx0B = 2*vdwtype[jnrB+0];
341 vdwjidx0C = 2*vdwtype[jnrC+0];
342 vdwjidx0D = 2*vdwtype[jnrD+0];
344 /**************************
345 * CALCULATE INTERACTIONS *
346 **************************/
348 r00 = _mm256_mul_pd(rsq00,rinv00);
349 r00 = _mm256_andnot_pd(dummy_mask,r00);
351 /* Compute parameters for interactions between i and j atoms */
352 qq00 = _mm256_mul_pd(iq0,jq0);
353 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
354 vdwioffsetptr0+vdwjidx0B,
355 vdwioffsetptr0+vdwjidx0C,
356 vdwioffsetptr0+vdwjidx0D,
359 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
360 isaprod = _mm256_mul_pd(isai0,isaj0);
361 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
362 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
364 /* Calculate generalized born table index - this is a separate table from the normal one,
365 * but we use the same procedure by multiplying r with scale and truncating to integer.
367 rt = _mm256_mul_pd(r00,gbscale);
368 gbitab = _mm256_cvttpd_epi32(rt);
369 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
370 gbitab = _mm_slli_epi32(gbitab,2);
371 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
372 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
373 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
374 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
375 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
376 Heps = _mm256_mul_pd(gbeps,H);
377 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
378 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
379 vgb = _mm256_mul_pd(gbqqfactor,VV);
381 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
382 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
383 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
384 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
385 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
386 /* 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. */
387 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
388 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
389 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
390 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
391 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
392 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
393 velec = _mm256_mul_pd(qq00,rinv00);
394 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
396 /* LENNARD-JONES DISPERSION/REPULSION */
398 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
399 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
400 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
401 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
402 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),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 vgb = _mm256_andnot_pd(dummy_mask,vgb);
408 vgbsum = _mm256_add_pd(vgbsum,vgb);
409 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
410 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
412 fscal = _mm256_add_pd(felec,fvdw);
414 fscal = _mm256_andnot_pd(dummy_mask,fscal);
416 /* Calculate temporary vectorial force */
417 tx = _mm256_mul_pd(fscal,dx00);
418 ty = _mm256_mul_pd(fscal,dy00);
419 tz = _mm256_mul_pd(fscal,dz00);
421 /* Update vectorial force */
422 fix0 = _mm256_add_pd(fix0,tx);
423 fiy0 = _mm256_add_pd(fiy0,ty);
424 fiz0 = _mm256_add_pd(fiz0,tz);
426 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
427 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
428 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
429 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
430 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
432 /* Inner loop uses 71 flops */
435 /* End of innermost loop */
437 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
438 f+i_coord_offset,fshift+i_shift_offset);
441 /* Update potential energies */
442 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
443 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
444 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
445 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
446 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
448 /* Increment number of inner iterations */
449 inneriter += j_index_end - j_index_start;
451 /* Outer loop uses 10 flops */
454 /* Increment number of outer iterations */
457 /* Update outer/inner flops */
459 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*71);
462 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
463 * Electrostatics interaction: GeneralizedBorn
464 * VdW interaction: LennardJones
465 * Geometry: Particle-Particle
466 * Calculate force/pot: Force
469 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
470 (t_nblist * gmx_restrict nlist,
471 rvec * gmx_restrict xx,
472 rvec * gmx_restrict ff,
473 t_forcerec * gmx_restrict fr,
474 t_mdatoms * gmx_restrict mdatoms,
475 nb_kernel_data_t * gmx_restrict kernel_data,
476 t_nrnb * gmx_restrict nrnb)
478 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
479 * just 0 for non-waters.
480 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
481 * jnr indices corresponding to data put in the four positions in the SIMD register.
483 int i_shift_offset,i_coord_offset,outeriter,inneriter;
484 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
485 int jnrA,jnrB,jnrC,jnrD;
486 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
487 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
488 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
489 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
491 real *shiftvec,*fshift,*x,*f;
492 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
494 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
495 real * vdwioffsetptr0;
496 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
497 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
498 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
499 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
500 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
503 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
504 __m256d minushalf = _mm256_set1_pd(-0.5);
505 real *invsqrta,*dvda,*gbtab;
507 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
510 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
511 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
513 __m128i ifour = _mm_set1_epi32(4);
514 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
516 __m256d dummy_mask,cutoff_mask;
517 __m128 tmpmask0,tmpmask1;
518 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
519 __m256d one = _mm256_set1_pd(1.0);
520 __m256d two = _mm256_set1_pd(2.0);
526 jindex = nlist->jindex;
528 shiftidx = nlist->shift;
530 shiftvec = fr->shift_vec[0];
531 fshift = fr->fshift[0];
532 facel = _mm256_set1_pd(fr->epsfac);
533 charge = mdatoms->chargeA;
534 nvdwtype = fr->ntype;
536 vdwtype = mdatoms->typeA;
538 invsqrta = fr->invsqrta;
540 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
541 gbtab = fr->gbtab.data;
542 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
544 /* Avoid stupid compiler warnings */
545 jnrA = jnrB = jnrC = jnrD = 0;
554 for(iidx=0;iidx<4*DIM;iidx++)
559 /* Start outer loop over neighborlists */
560 for(iidx=0; iidx<nri; iidx++)
562 /* Load shift vector for this list */
563 i_shift_offset = DIM*shiftidx[iidx];
565 /* Load limits for loop over neighbors */
566 j_index_start = jindex[iidx];
567 j_index_end = jindex[iidx+1];
569 /* Get outer coordinate index */
571 i_coord_offset = DIM*inr;
573 /* Load i particle coords and add shift vector */
574 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
576 fix0 = _mm256_setzero_pd();
577 fiy0 = _mm256_setzero_pd();
578 fiz0 = _mm256_setzero_pd();
580 /* Load parameters for i particles */
581 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
582 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
583 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
585 dvdasum = _mm256_setzero_pd();
587 /* Start inner kernel loop */
588 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
591 /* Get j neighbor index, and coordinate index */
596 j_coord_offsetA = DIM*jnrA;
597 j_coord_offsetB = DIM*jnrB;
598 j_coord_offsetC = DIM*jnrC;
599 j_coord_offsetD = DIM*jnrD;
601 /* load j atom coordinates */
602 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
603 x+j_coord_offsetC,x+j_coord_offsetD,
606 /* Calculate displacement vector */
607 dx00 = _mm256_sub_pd(ix0,jx0);
608 dy00 = _mm256_sub_pd(iy0,jy0);
609 dz00 = _mm256_sub_pd(iz0,jz0);
611 /* Calculate squared distance and things based on it */
612 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
614 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
616 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
618 /* Load parameters for j particles */
619 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
620 charge+jnrC+0,charge+jnrD+0);
621 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
622 invsqrta+jnrC+0,invsqrta+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 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
643 isaprod = _mm256_mul_pd(isai0,isaj0);
644 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
645 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
647 /* Calculate generalized born table index - this is a separate table from the normal one,
648 * but we use the same procedure by multiplying r with scale and truncating to integer.
650 rt = _mm256_mul_pd(r00,gbscale);
651 gbitab = _mm256_cvttpd_epi32(rt);
652 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
653 gbitab = _mm_slli_epi32(gbitab,2);
654 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
655 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
656 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
657 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
658 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
659 Heps = _mm256_mul_pd(gbeps,H);
660 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
661 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
662 vgb = _mm256_mul_pd(gbqqfactor,VV);
664 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
665 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
666 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
667 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
672 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
673 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
674 velec = _mm256_mul_pd(qq00,rinv00);
675 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
677 /* LENNARD-JONES DISPERSION/REPULSION */
679 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
680 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
682 fscal = _mm256_add_pd(felec,fvdw);
684 /* Calculate temporary vectorial force */
685 tx = _mm256_mul_pd(fscal,dx00);
686 ty = _mm256_mul_pd(fscal,dy00);
687 tz = _mm256_mul_pd(fscal,dz00);
689 /* Update vectorial force */
690 fix0 = _mm256_add_pd(fix0,tx);
691 fiy0 = _mm256_add_pd(fiy0,ty);
692 fiz0 = _mm256_add_pd(fiz0,tz);
694 fjptrA = f+j_coord_offsetA;
695 fjptrB = f+j_coord_offsetB;
696 fjptrC = f+j_coord_offsetC;
697 fjptrD = f+j_coord_offsetD;
698 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
700 /* Inner loop uses 63 flops */
706 /* Get j neighbor index, and coordinate index */
707 jnrlistA = jjnr[jidx];
708 jnrlistB = jjnr[jidx+1];
709 jnrlistC = jjnr[jidx+2];
710 jnrlistD = jjnr[jidx+3];
711 /* Sign of each element will be negative for non-real atoms.
712 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
713 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
715 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
717 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
718 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
719 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
721 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
722 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
723 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
724 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
725 j_coord_offsetA = DIM*jnrA;
726 j_coord_offsetB = DIM*jnrB;
727 j_coord_offsetC = DIM*jnrC;
728 j_coord_offsetD = DIM*jnrD;
730 /* load j atom coordinates */
731 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
732 x+j_coord_offsetC,x+j_coord_offsetD,
735 /* Calculate displacement vector */
736 dx00 = _mm256_sub_pd(ix0,jx0);
737 dy00 = _mm256_sub_pd(iy0,jy0);
738 dz00 = _mm256_sub_pd(iz0,jz0);
740 /* Calculate squared distance and things based on it */
741 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
743 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
745 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
747 /* Load parameters for j particles */
748 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
749 charge+jnrC+0,charge+jnrD+0);
750 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
751 invsqrta+jnrC+0,invsqrta+jnrD+0);
752 vdwjidx0A = 2*vdwtype[jnrA+0];
753 vdwjidx0B = 2*vdwtype[jnrB+0];
754 vdwjidx0C = 2*vdwtype[jnrC+0];
755 vdwjidx0D = 2*vdwtype[jnrD+0];
757 /**************************
758 * CALCULATE INTERACTIONS *
759 **************************/
761 r00 = _mm256_mul_pd(rsq00,rinv00);
762 r00 = _mm256_andnot_pd(dummy_mask,r00);
764 /* Compute parameters for interactions between i and j atoms */
765 qq00 = _mm256_mul_pd(iq0,jq0);
766 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
767 vdwioffsetptr0+vdwjidx0B,
768 vdwioffsetptr0+vdwjidx0C,
769 vdwioffsetptr0+vdwjidx0D,
772 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
773 isaprod = _mm256_mul_pd(isai0,isaj0);
774 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
775 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
777 /* Calculate generalized born table index - this is a separate table from the normal one,
778 * but we use the same procedure by multiplying r with scale and truncating to integer.
780 rt = _mm256_mul_pd(r00,gbscale);
781 gbitab = _mm256_cvttpd_epi32(rt);
782 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
783 gbitab = _mm_slli_epi32(gbitab,2);
784 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
785 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
786 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
787 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
788 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
789 Heps = _mm256_mul_pd(gbeps,H);
790 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
791 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
792 vgb = _mm256_mul_pd(gbqqfactor,VV);
794 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
795 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
796 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
797 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
798 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
799 /* 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. */
800 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
801 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
802 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
803 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
804 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
805 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
806 velec = _mm256_mul_pd(qq00,rinv00);
807 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
809 /* LENNARD-JONES DISPERSION/REPULSION */
811 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
812 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
814 fscal = _mm256_add_pd(felec,fvdw);
816 fscal = _mm256_andnot_pd(dummy_mask,fscal);
818 /* Calculate temporary vectorial force */
819 tx = _mm256_mul_pd(fscal,dx00);
820 ty = _mm256_mul_pd(fscal,dy00);
821 tz = _mm256_mul_pd(fscal,dz00);
823 /* Update vectorial force */
824 fix0 = _mm256_add_pd(fix0,tx);
825 fiy0 = _mm256_add_pd(fiy0,ty);
826 fiz0 = _mm256_add_pd(fiz0,tz);
828 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
829 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
830 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
831 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
832 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
834 /* Inner loop uses 64 flops */
837 /* End of innermost loop */
839 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
840 f+i_coord_offset,fshift+i_shift_offset);
842 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
843 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
845 /* Increment number of inner iterations */
846 inneriter += j_index_end - j_index_start;
848 /* Outer loop uses 7 flops */
851 /* Increment number of outer iterations */
854 /* Update outer/inner flops */
856 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*64);