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 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
385 /* 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. */
386 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
387 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
388 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
389 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
390 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
391 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
392 velec = _mm256_mul_pd(qq00,rinv00);
393 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
395 /* LENNARD-JONES DISPERSION/REPULSION */
397 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
398 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
399 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
400 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
401 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
403 /* Update potential sum for this i atom from the interaction with this j atom. */
404 velec = _mm256_andnot_pd(dummy_mask,velec);
405 velecsum = _mm256_add_pd(velecsum,velec);
406 vgb = _mm256_andnot_pd(dummy_mask,vgb);
407 vgbsum = _mm256_add_pd(vgbsum,vgb);
408 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
409 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
411 fscal = _mm256_add_pd(felec,fvdw);
413 fscal = _mm256_andnot_pd(dummy_mask,fscal);
415 /* Calculate temporary vectorial force */
416 tx = _mm256_mul_pd(fscal,dx00);
417 ty = _mm256_mul_pd(fscal,dy00);
418 tz = _mm256_mul_pd(fscal,dz00);
420 /* Update vectorial force */
421 fix0 = _mm256_add_pd(fix0,tx);
422 fiy0 = _mm256_add_pd(fiy0,ty);
423 fiz0 = _mm256_add_pd(fiz0,tz);
425 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
426 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
427 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
428 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
429 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
431 /* Inner loop uses 71 flops */
434 /* End of innermost loop */
436 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
437 f+i_coord_offset,fshift+i_shift_offset);
440 /* Update potential energies */
441 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
442 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
443 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
444 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
445 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
447 /* Increment number of inner iterations */
448 inneriter += j_index_end - j_index_start;
450 /* Outer loop uses 10 flops */
453 /* Increment number of outer iterations */
456 /* Update outer/inner flops */
458 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*71);
461 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
462 * Electrostatics interaction: GeneralizedBorn
463 * VdW interaction: LennardJones
464 * Geometry: Particle-Particle
465 * Calculate force/pot: Force
468 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_avx_256_double
469 (t_nblist * gmx_restrict nlist,
470 rvec * gmx_restrict xx,
471 rvec * gmx_restrict ff,
472 t_forcerec * gmx_restrict fr,
473 t_mdatoms * gmx_restrict mdatoms,
474 nb_kernel_data_t * gmx_restrict kernel_data,
475 t_nrnb * gmx_restrict nrnb)
477 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
478 * just 0 for non-waters.
479 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
480 * jnr indices corresponding to data put in the four positions in the SIMD register.
482 int i_shift_offset,i_coord_offset,outeriter,inneriter;
483 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
484 int jnrA,jnrB,jnrC,jnrD;
485 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
486 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
487 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
488 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
490 real *shiftvec,*fshift,*x,*f;
491 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
493 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
494 real * vdwioffsetptr0;
495 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
496 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
497 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
498 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
499 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
502 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
503 __m256d minushalf = _mm256_set1_pd(-0.5);
504 real *invsqrta,*dvda,*gbtab;
506 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
509 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
510 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
512 __m128i ifour = _mm_set1_epi32(4);
513 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
515 __m256d dummy_mask,cutoff_mask;
516 __m128 tmpmask0,tmpmask1;
517 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
518 __m256d one = _mm256_set1_pd(1.0);
519 __m256d two = _mm256_set1_pd(2.0);
525 jindex = nlist->jindex;
527 shiftidx = nlist->shift;
529 shiftvec = fr->shift_vec[0];
530 fshift = fr->fshift[0];
531 facel = _mm256_set1_pd(fr->epsfac);
532 charge = mdatoms->chargeA;
533 nvdwtype = fr->ntype;
535 vdwtype = mdatoms->typeA;
537 invsqrta = fr->invsqrta;
539 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
540 gbtab = fr->gbtab.data;
541 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
543 /* Avoid stupid compiler warnings */
544 jnrA = jnrB = jnrC = jnrD = 0;
553 for(iidx=0;iidx<4*DIM;iidx++)
558 /* Start outer loop over neighborlists */
559 for(iidx=0; iidx<nri; iidx++)
561 /* Load shift vector for this list */
562 i_shift_offset = DIM*shiftidx[iidx];
564 /* Load limits for loop over neighbors */
565 j_index_start = jindex[iidx];
566 j_index_end = jindex[iidx+1];
568 /* Get outer coordinate index */
570 i_coord_offset = DIM*inr;
572 /* Load i particle coords and add shift vector */
573 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
575 fix0 = _mm256_setzero_pd();
576 fiy0 = _mm256_setzero_pd();
577 fiz0 = _mm256_setzero_pd();
579 /* Load parameters for i particles */
580 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
581 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
582 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
584 dvdasum = _mm256_setzero_pd();
586 /* Start inner kernel loop */
587 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
590 /* Get j neighbor index, and coordinate index */
595 j_coord_offsetA = DIM*jnrA;
596 j_coord_offsetB = DIM*jnrB;
597 j_coord_offsetC = DIM*jnrC;
598 j_coord_offsetD = DIM*jnrD;
600 /* load j atom coordinates */
601 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
602 x+j_coord_offsetC,x+j_coord_offsetD,
605 /* Calculate displacement vector */
606 dx00 = _mm256_sub_pd(ix0,jx0);
607 dy00 = _mm256_sub_pd(iy0,jy0);
608 dz00 = _mm256_sub_pd(iz0,jz0);
610 /* Calculate squared distance and things based on it */
611 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
613 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
615 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
617 /* Load parameters for j particles */
618 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
619 charge+jnrC+0,charge+jnrD+0);
620 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
621 invsqrta+jnrC+0,invsqrta+jnrD+0);
622 vdwjidx0A = 2*vdwtype[jnrA+0];
623 vdwjidx0B = 2*vdwtype[jnrB+0];
624 vdwjidx0C = 2*vdwtype[jnrC+0];
625 vdwjidx0D = 2*vdwtype[jnrD+0];
627 /**************************
628 * CALCULATE INTERACTIONS *
629 **************************/
631 r00 = _mm256_mul_pd(rsq00,rinv00);
633 /* Compute parameters for interactions between i and j atoms */
634 qq00 = _mm256_mul_pd(iq0,jq0);
635 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
636 vdwioffsetptr0+vdwjidx0B,
637 vdwioffsetptr0+vdwjidx0C,
638 vdwioffsetptr0+vdwjidx0D,
641 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
642 isaprod = _mm256_mul_pd(isai0,isaj0);
643 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
644 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
646 /* Calculate generalized born table index - this is a separate table from the normal one,
647 * but we use the same procedure by multiplying r with scale and truncating to integer.
649 rt = _mm256_mul_pd(r00,gbscale);
650 gbitab = _mm256_cvttpd_epi32(rt);
651 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
652 gbitab = _mm_slli_epi32(gbitab,2);
653 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
654 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
655 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
656 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
657 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
658 Heps = _mm256_mul_pd(gbeps,H);
659 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
660 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
661 vgb = _mm256_mul_pd(gbqqfactor,VV);
663 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
664 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
665 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
666 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
671 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
672 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
673 velec = _mm256_mul_pd(qq00,rinv00);
674 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
676 /* LENNARD-JONES DISPERSION/REPULSION */
678 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
679 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
681 fscal = _mm256_add_pd(felec,fvdw);
683 /* Calculate temporary vectorial force */
684 tx = _mm256_mul_pd(fscal,dx00);
685 ty = _mm256_mul_pd(fscal,dy00);
686 tz = _mm256_mul_pd(fscal,dz00);
688 /* Update vectorial force */
689 fix0 = _mm256_add_pd(fix0,tx);
690 fiy0 = _mm256_add_pd(fiy0,ty);
691 fiz0 = _mm256_add_pd(fiz0,tz);
693 fjptrA = f+j_coord_offsetA;
694 fjptrB = f+j_coord_offsetB;
695 fjptrC = f+j_coord_offsetC;
696 fjptrD = f+j_coord_offsetD;
697 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
699 /* Inner loop uses 63 flops */
705 /* Get j neighbor index, and coordinate index */
706 jnrlistA = jjnr[jidx];
707 jnrlistB = jjnr[jidx+1];
708 jnrlistC = jjnr[jidx+2];
709 jnrlistD = jjnr[jidx+3];
710 /* Sign of each element will be negative for non-real atoms.
711 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
712 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
714 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
716 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
717 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
718 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
720 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
721 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
722 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
723 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
724 j_coord_offsetA = DIM*jnrA;
725 j_coord_offsetB = DIM*jnrB;
726 j_coord_offsetC = DIM*jnrC;
727 j_coord_offsetD = DIM*jnrD;
729 /* load j atom coordinates */
730 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
731 x+j_coord_offsetC,x+j_coord_offsetD,
734 /* Calculate displacement vector */
735 dx00 = _mm256_sub_pd(ix0,jx0);
736 dy00 = _mm256_sub_pd(iy0,jy0);
737 dz00 = _mm256_sub_pd(iz0,jz0);
739 /* Calculate squared distance and things based on it */
740 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
742 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
744 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
746 /* Load parameters for j particles */
747 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
748 charge+jnrC+0,charge+jnrD+0);
749 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
750 invsqrta+jnrC+0,invsqrta+jnrD+0);
751 vdwjidx0A = 2*vdwtype[jnrA+0];
752 vdwjidx0B = 2*vdwtype[jnrB+0];
753 vdwjidx0C = 2*vdwtype[jnrC+0];
754 vdwjidx0D = 2*vdwtype[jnrD+0];
756 /**************************
757 * CALCULATE INTERACTIONS *
758 **************************/
760 r00 = _mm256_mul_pd(rsq00,rinv00);
761 r00 = _mm256_andnot_pd(dummy_mask,r00);
763 /* Compute parameters for interactions between i and j atoms */
764 qq00 = _mm256_mul_pd(iq0,jq0);
765 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
766 vdwioffsetptr0+vdwjidx0B,
767 vdwioffsetptr0+vdwjidx0C,
768 vdwioffsetptr0+vdwjidx0D,
771 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
772 isaprod = _mm256_mul_pd(isai0,isaj0);
773 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
774 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
776 /* Calculate generalized born table index - this is a separate table from the normal one,
777 * but we use the same procedure by multiplying r with scale and truncating to integer.
779 rt = _mm256_mul_pd(r00,gbscale);
780 gbitab = _mm256_cvttpd_epi32(rt);
781 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
782 gbitab = _mm_slli_epi32(gbitab,2);
783 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
784 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
785 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
786 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
787 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
788 Heps = _mm256_mul_pd(gbeps,H);
789 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
790 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
791 vgb = _mm256_mul_pd(gbqqfactor,VV);
793 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
794 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
795 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
796 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
797 /* 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. */
798 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
799 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
800 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
801 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
802 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
803 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
804 velec = _mm256_mul_pd(qq00,rinv00);
805 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
807 /* LENNARD-JONES DISPERSION/REPULSION */
809 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
810 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
812 fscal = _mm256_add_pd(felec,fvdw);
814 fscal = _mm256_andnot_pd(dummy_mask,fscal);
816 /* Calculate temporary vectorial force */
817 tx = _mm256_mul_pd(fscal,dx00);
818 ty = _mm256_mul_pd(fscal,dy00);
819 tz = _mm256_mul_pd(fscal,dz00);
821 /* Update vectorial force */
822 fix0 = _mm256_add_pd(fix0,tx);
823 fiy0 = _mm256_add_pd(fiy0,ty);
824 fiz0 = _mm256_add_pd(fiz0,tz);
826 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
827 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
828 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
829 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
830 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
832 /* Inner loop uses 64 flops */
835 /* End of innermost loop */
837 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
838 f+i_coord_offset,fshift+i_shift_offset);
840 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
841 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
843 /* Increment number of inner iterations */
844 inneriter += j_index_end - j_index_start;
846 /* Outer loop uses 7 flops */
849 /* Increment number of outer iterations */
852 /* Update outer/inner flops */
854 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*64);