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_VdwNone_GeomP1P1_VF_avx_256_double
38 * Electrostatics interaction: GeneralizedBorn
39 * VdW interaction: None
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecGB_VdwNone_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 __m128i ifour = _mm_set1_epi32(4);
83 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
85 __m256d dummy_mask,cutoff_mask;
86 __m128 tmpmask0,tmpmask1;
87 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
88 __m256d one = _mm256_set1_pd(1.0);
89 __m256d two = _mm256_set1_pd(2.0);
95 jindex = nlist->jindex;
97 shiftidx = nlist->shift;
99 shiftvec = fr->shift_vec[0];
100 fshift = fr->fshift[0];
101 facel = _mm256_set1_pd(fr->epsfac);
102 charge = mdatoms->chargeA;
104 invsqrta = fr->invsqrta;
106 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
107 gbtab = fr->gbtab.data;
108 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
110 /* Avoid stupid compiler warnings */
111 jnrA = jnrB = jnrC = jnrD = 0;
120 for(iidx=0;iidx<4*DIM;iidx++)
125 /* Start outer loop over neighborlists */
126 for(iidx=0; iidx<nri; iidx++)
128 /* Load shift vector for this list */
129 i_shift_offset = DIM*shiftidx[iidx];
131 /* Load limits for loop over neighbors */
132 j_index_start = jindex[iidx];
133 j_index_end = jindex[iidx+1];
135 /* Get outer coordinate index */
137 i_coord_offset = DIM*inr;
139 /* Load i particle coords and add shift vector */
140 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
142 fix0 = _mm256_setzero_pd();
143 fiy0 = _mm256_setzero_pd();
144 fiz0 = _mm256_setzero_pd();
146 /* Load parameters for i particles */
147 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
148 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
150 /* Reset potential sums */
151 velecsum = _mm256_setzero_pd();
152 vgbsum = _mm256_setzero_pd();
153 dvdasum = _mm256_setzero_pd();
155 /* Start inner kernel loop */
156 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
159 /* Get j neighbor index, and coordinate index */
164 j_coord_offsetA = DIM*jnrA;
165 j_coord_offsetB = DIM*jnrB;
166 j_coord_offsetC = DIM*jnrC;
167 j_coord_offsetD = DIM*jnrD;
169 /* load j atom coordinates */
170 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
171 x+j_coord_offsetC,x+j_coord_offsetD,
174 /* Calculate displacement vector */
175 dx00 = _mm256_sub_pd(ix0,jx0);
176 dy00 = _mm256_sub_pd(iy0,jy0);
177 dz00 = _mm256_sub_pd(iz0,jz0);
179 /* Calculate squared distance and things based on it */
180 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
182 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
184 /* Load parameters for j particles */
185 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
186 charge+jnrC+0,charge+jnrD+0);
187 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
188 invsqrta+jnrC+0,invsqrta+jnrD+0);
190 /**************************
191 * CALCULATE INTERACTIONS *
192 **************************/
194 r00 = _mm256_mul_pd(rsq00,rinv00);
196 /* Compute parameters for interactions between i and j atoms */
197 qq00 = _mm256_mul_pd(iq0,jq0);
199 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
200 isaprod = _mm256_mul_pd(isai0,isaj0);
201 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
202 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
204 /* Calculate generalized born table index - this is a separate table from the normal one,
205 * but we use the same procedure by multiplying r with scale and truncating to integer.
207 rt = _mm256_mul_pd(r00,gbscale);
208 gbitab = _mm256_cvttpd_epi32(rt);
209 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
210 gbitab = _mm_slli_epi32(gbitab,2);
211 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
212 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
213 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
214 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
215 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
216 Heps = _mm256_mul_pd(gbeps,H);
217 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
218 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
219 vgb = _mm256_mul_pd(gbqqfactor,VV);
221 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
222 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
223 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
224 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
229 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
230 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
231 velec = _mm256_mul_pd(qq00,rinv00);
232 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
234 /* Update potential sum for this i atom from the interaction with this j atom. */
235 velecsum = _mm256_add_pd(velecsum,velec);
236 vgbsum = _mm256_add_pd(vgbsum,vgb);
240 /* Calculate temporary vectorial force */
241 tx = _mm256_mul_pd(fscal,dx00);
242 ty = _mm256_mul_pd(fscal,dy00);
243 tz = _mm256_mul_pd(fscal,dz00);
245 /* Update vectorial force */
246 fix0 = _mm256_add_pd(fix0,tx);
247 fiy0 = _mm256_add_pd(fiy0,ty);
248 fiz0 = _mm256_add_pd(fiz0,tz);
250 fjptrA = f+j_coord_offsetA;
251 fjptrB = f+j_coord_offsetB;
252 fjptrC = f+j_coord_offsetC;
253 fjptrD = f+j_coord_offsetD;
254 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
256 /* Inner loop uses 57 flops */
262 /* Get j neighbor index, and coordinate index */
263 jnrlistA = jjnr[jidx];
264 jnrlistB = jjnr[jidx+1];
265 jnrlistC = jjnr[jidx+2];
266 jnrlistD = jjnr[jidx+3];
267 /* Sign of each element will be negative for non-real atoms.
268 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
269 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
271 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
273 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
274 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
275 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
277 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
278 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
279 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
280 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
281 j_coord_offsetA = DIM*jnrA;
282 j_coord_offsetB = DIM*jnrB;
283 j_coord_offsetC = DIM*jnrC;
284 j_coord_offsetD = DIM*jnrD;
286 /* load j atom coordinates */
287 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
288 x+j_coord_offsetC,x+j_coord_offsetD,
291 /* Calculate displacement vector */
292 dx00 = _mm256_sub_pd(ix0,jx0);
293 dy00 = _mm256_sub_pd(iy0,jy0);
294 dz00 = _mm256_sub_pd(iz0,jz0);
296 /* Calculate squared distance and things based on it */
297 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
299 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
301 /* Load parameters for j particles */
302 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
303 charge+jnrC+0,charge+jnrD+0);
304 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
305 invsqrta+jnrC+0,invsqrta+jnrD+0);
307 /**************************
308 * CALCULATE INTERACTIONS *
309 **************************/
311 r00 = _mm256_mul_pd(rsq00,rinv00);
312 r00 = _mm256_andnot_pd(dummy_mask,r00);
314 /* Compute parameters for interactions between i and j atoms */
315 qq00 = _mm256_mul_pd(iq0,jq0);
317 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
318 isaprod = _mm256_mul_pd(isai0,isaj0);
319 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
320 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
322 /* Calculate generalized born table index - this is a separate table from the normal one,
323 * but we use the same procedure by multiplying r with scale and truncating to integer.
325 rt = _mm256_mul_pd(r00,gbscale);
326 gbitab = _mm256_cvttpd_epi32(rt);
327 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
328 gbitab = _mm_slli_epi32(gbitab,2);
329 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
330 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
331 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
332 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
333 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
334 Heps = _mm256_mul_pd(gbeps,H);
335 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
336 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
337 vgb = _mm256_mul_pd(gbqqfactor,VV);
339 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
340 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
341 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
342 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
343 /* 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. */
344 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
345 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
346 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
347 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
348 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
349 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
350 velec = _mm256_mul_pd(qq00,rinv00);
351 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
353 /* Update potential sum for this i atom from the interaction with this j atom. */
354 velec = _mm256_andnot_pd(dummy_mask,velec);
355 velecsum = _mm256_add_pd(velecsum,velec);
356 vgb = _mm256_andnot_pd(dummy_mask,vgb);
357 vgbsum = _mm256_add_pd(vgbsum,vgb);
361 fscal = _mm256_andnot_pd(dummy_mask,fscal);
363 /* Calculate temporary vectorial force */
364 tx = _mm256_mul_pd(fscal,dx00);
365 ty = _mm256_mul_pd(fscal,dy00);
366 tz = _mm256_mul_pd(fscal,dz00);
368 /* Update vectorial force */
369 fix0 = _mm256_add_pd(fix0,tx);
370 fiy0 = _mm256_add_pd(fiy0,ty);
371 fiz0 = _mm256_add_pd(fiz0,tz);
373 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
374 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
375 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
376 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
377 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
379 /* Inner loop uses 58 flops */
382 /* End of innermost loop */
384 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
385 f+i_coord_offset,fshift+i_shift_offset);
388 /* Update potential energies */
389 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
390 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
391 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
392 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
394 /* Increment number of inner iterations */
395 inneriter += j_index_end - j_index_start;
397 /* Outer loop uses 9 flops */
400 /* Increment number of outer iterations */
403 /* Update outer/inner flops */
405 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*58);
408 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
409 * Electrostatics interaction: GeneralizedBorn
410 * VdW interaction: None
411 * Geometry: Particle-Particle
412 * Calculate force/pot: Force
415 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
416 (t_nblist * gmx_restrict nlist,
417 rvec * gmx_restrict xx,
418 rvec * gmx_restrict ff,
419 t_forcerec * gmx_restrict fr,
420 t_mdatoms * gmx_restrict mdatoms,
421 nb_kernel_data_t * gmx_restrict kernel_data,
422 t_nrnb * gmx_restrict nrnb)
424 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
425 * just 0 for non-waters.
426 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
427 * jnr indices corresponding to data put in the four positions in the SIMD register.
429 int i_shift_offset,i_coord_offset,outeriter,inneriter;
430 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
431 int jnrA,jnrB,jnrC,jnrD;
432 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
433 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
434 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
435 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
437 real *shiftvec,*fshift,*x,*f;
438 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
440 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
441 real * vdwioffsetptr0;
442 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
443 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
444 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
445 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
446 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
449 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
450 __m256d minushalf = _mm256_set1_pd(-0.5);
451 real *invsqrta,*dvda,*gbtab;
453 __m128i ifour = _mm_set1_epi32(4);
454 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
456 __m256d dummy_mask,cutoff_mask;
457 __m128 tmpmask0,tmpmask1;
458 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
459 __m256d one = _mm256_set1_pd(1.0);
460 __m256d two = _mm256_set1_pd(2.0);
466 jindex = nlist->jindex;
468 shiftidx = nlist->shift;
470 shiftvec = fr->shift_vec[0];
471 fshift = fr->fshift[0];
472 facel = _mm256_set1_pd(fr->epsfac);
473 charge = mdatoms->chargeA;
475 invsqrta = fr->invsqrta;
477 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
478 gbtab = fr->gbtab.data;
479 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
481 /* Avoid stupid compiler warnings */
482 jnrA = jnrB = jnrC = jnrD = 0;
491 for(iidx=0;iidx<4*DIM;iidx++)
496 /* Start outer loop over neighborlists */
497 for(iidx=0; iidx<nri; iidx++)
499 /* Load shift vector for this list */
500 i_shift_offset = DIM*shiftidx[iidx];
502 /* Load limits for loop over neighbors */
503 j_index_start = jindex[iidx];
504 j_index_end = jindex[iidx+1];
506 /* Get outer coordinate index */
508 i_coord_offset = DIM*inr;
510 /* Load i particle coords and add shift vector */
511 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
513 fix0 = _mm256_setzero_pd();
514 fiy0 = _mm256_setzero_pd();
515 fiz0 = _mm256_setzero_pd();
517 /* Load parameters for i particles */
518 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
519 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
521 dvdasum = _mm256_setzero_pd();
523 /* Start inner kernel loop */
524 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
527 /* Get j neighbor index, and coordinate index */
532 j_coord_offsetA = DIM*jnrA;
533 j_coord_offsetB = DIM*jnrB;
534 j_coord_offsetC = DIM*jnrC;
535 j_coord_offsetD = DIM*jnrD;
537 /* load j atom coordinates */
538 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
539 x+j_coord_offsetC,x+j_coord_offsetD,
542 /* Calculate displacement vector */
543 dx00 = _mm256_sub_pd(ix0,jx0);
544 dy00 = _mm256_sub_pd(iy0,jy0);
545 dz00 = _mm256_sub_pd(iz0,jz0);
547 /* Calculate squared distance and things based on it */
548 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
550 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
552 /* Load parameters for j particles */
553 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
554 charge+jnrC+0,charge+jnrD+0);
555 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
556 invsqrta+jnrC+0,invsqrta+jnrD+0);
558 /**************************
559 * CALCULATE INTERACTIONS *
560 **************************/
562 r00 = _mm256_mul_pd(rsq00,rinv00);
564 /* Compute parameters for interactions between i and j atoms */
565 qq00 = _mm256_mul_pd(iq0,jq0);
567 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
568 isaprod = _mm256_mul_pd(isai0,isaj0);
569 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
570 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
572 /* Calculate generalized born table index - this is a separate table from the normal one,
573 * but we use the same procedure by multiplying r with scale and truncating to integer.
575 rt = _mm256_mul_pd(r00,gbscale);
576 gbitab = _mm256_cvttpd_epi32(rt);
577 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
578 gbitab = _mm_slli_epi32(gbitab,2);
579 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
580 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
581 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
582 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
583 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
584 Heps = _mm256_mul_pd(gbeps,H);
585 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
586 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
587 vgb = _mm256_mul_pd(gbqqfactor,VV);
589 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
590 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
591 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
592 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
597 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
598 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
599 velec = _mm256_mul_pd(qq00,rinv00);
600 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
604 /* Calculate temporary vectorial force */
605 tx = _mm256_mul_pd(fscal,dx00);
606 ty = _mm256_mul_pd(fscal,dy00);
607 tz = _mm256_mul_pd(fscal,dz00);
609 /* Update vectorial force */
610 fix0 = _mm256_add_pd(fix0,tx);
611 fiy0 = _mm256_add_pd(fiy0,ty);
612 fiz0 = _mm256_add_pd(fiz0,tz);
614 fjptrA = f+j_coord_offsetA;
615 fjptrB = f+j_coord_offsetB;
616 fjptrC = f+j_coord_offsetC;
617 fjptrD = f+j_coord_offsetD;
618 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
620 /* Inner loop uses 55 flops */
626 /* Get j neighbor index, and coordinate index */
627 jnrlistA = jjnr[jidx];
628 jnrlistB = jjnr[jidx+1];
629 jnrlistC = jjnr[jidx+2];
630 jnrlistD = jjnr[jidx+3];
631 /* Sign of each element will be negative for non-real atoms.
632 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
633 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
635 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
637 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
638 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
639 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
641 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
642 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
643 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
644 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
645 j_coord_offsetA = DIM*jnrA;
646 j_coord_offsetB = DIM*jnrB;
647 j_coord_offsetC = DIM*jnrC;
648 j_coord_offsetD = DIM*jnrD;
650 /* load j atom coordinates */
651 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
652 x+j_coord_offsetC,x+j_coord_offsetD,
655 /* Calculate displacement vector */
656 dx00 = _mm256_sub_pd(ix0,jx0);
657 dy00 = _mm256_sub_pd(iy0,jy0);
658 dz00 = _mm256_sub_pd(iz0,jz0);
660 /* Calculate squared distance and things based on it */
661 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
663 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
665 /* Load parameters for j particles */
666 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
667 charge+jnrC+0,charge+jnrD+0);
668 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
669 invsqrta+jnrC+0,invsqrta+jnrD+0);
671 /**************************
672 * CALCULATE INTERACTIONS *
673 **************************/
675 r00 = _mm256_mul_pd(rsq00,rinv00);
676 r00 = _mm256_andnot_pd(dummy_mask,r00);
678 /* Compute parameters for interactions between i and j atoms */
679 qq00 = _mm256_mul_pd(iq0,jq0);
681 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
682 isaprod = _mm256_mul_pd(isai0,isaj0);
683 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
684 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
686 /* Calculate generalized born table index - this is a separate table from the normal one,
687 * but we use the same procedure by multiplying r with scale and truncating to integer.
689 rt = _mm256_mul_pd(r00,gbscale);
690 gbitab = _mm256_cvttpd_epi32(rt);
691 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
692 gbitab = _mm_slli_epi32(gbitab,2);
693 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
694 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
695 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
696 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
697 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
698 Heps = _mm256_mul_pd(gbeps,H);
699 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
700 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
701 vgb = _mm256_mul_pd(gbqqfactor,VV);
703 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
704 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
705 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
706 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
707 /* 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. */
708 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
709 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
710 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
711 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
712 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
713 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
714 velec = _mm256_mul_pd(qq00,rinv00);
715 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
719 fscal = _mm256_andnot_pd(dummy_mask,fscal);
721 /* Calculate temporary vectorial force */
722 tx = _mm256_mul_pd(fscal,dx00);
723 ty = _mm256_mul_pd(fscal,dy00);
724 tz = _mm256_mul_pd(fscal,dz00);
726 /* Update vectorial force */
727 fix0 = _mm256_add_pd(fix0,tx);
728 fiy0 = _mm256_add_pd(fiy0,ty);
729 fiz0 = _mm256_add_pd(fiz0,tz);
731 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
732 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
733 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
734 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
735 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
737 /* Inner loop uses 56 flops */
740 /* End of innermost loop */
742 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
743 f+i_coord_offset,fshift+i_shift_offset);
745 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
746 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
748 /* Increment number of inner iterations */
749 inneriter += j_index_end - j_index_start;
751 /* Outer loop uses 7 flops */
754 /* Increment number of outer iterations */
757 /* Update outer/inner flops */
759 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*56);