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 dvdatmp = _mm256_andnot_ps(dummy_mask,dvdatmp);
343 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
344 /* 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. */
345 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
346 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
347 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
348 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
349 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
350 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
351 velec = _mm256_mul_pd(qq00,rinv00);
352 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
354 /* Update potential sum for this i atom from the interaction with this j atom. */
355 velec = _mm256_andnot_pd(dummy_mask,velec);
356 velecsum = _mm256_add_pd(velecsum,velec);
357 vgb = _mm256_andnot_pd(dummy_mask,vgb);
358 vgbsum = _mm256_add_pd(vgbsum,vgb);
362 fscal = _mm256_andnot_pd(dummy_mask,fscal);
364 /* Calculate temporary vectorial force */
365 tx = _mm256_mul_pd(fscal,dx00);
366 ty = _mm256_mul_pd(fscal,dy00);
367 tz = _mm256_mul_pd(fscal,dz00);
369 /* Update vectorial force */
370 fix0 = _mm256_add_pd(fix0,tx);
371 fiy0 = _mm256_add_pd(fiy0,ty);
372 fiz0 = _mm256_add_pd(fiz0,tz);
374 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
375 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
376 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
377 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
378 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
380 /* Inner loop uses 58 flops */
383 /* End of innermost loop */
385 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
386 f+i_coord_offset,fshift+i_shift_offset);
389 /* Update potential energies */
390 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
391 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
392 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
393 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
395 /* Increment number of inner iterations */
396 inneriter += j_index_end - j_index_start;
398 /* Outer loop uses 9 flops */
401 /* Increment number of outer iterations */
404 /* Update outer/inner flops */
406 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*58);
409 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
410 * Electrostatics interaction: GeneralizedBorn
411 * VdW interaction: None
412 * Geometry: Particle-Particle
413 * Calculate force/pot: Force
416 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
417 (t_nblist * gmx_restrict nlist,
418 rvec * gmx_restrict xx,
419 rvec * gmx_restrict ff,
420 t_forcerec * gmx_restrict fr,
421 t_mdatoms * gmx_restrict mdatoms,
422 nb_kernel_data_t * gmx_restrict kernel_data,
423 t_nrnb * gmx_restrict nrnb)
425 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
426 * just 0 for non-waters.
427 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
428 * jnr indices corresponding to data put in the four positions in the SIMD register.
430 int i_shift_offset,i_coord_offset,outeriter,inneriter;
431 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
432 int jnrA,jnrB,jnrC,jnrD;
433 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
434 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
435 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
436 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
438 real *shiftvec,*fshift,*x,*f;
439 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
441 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
442 real * vdwioffsetptr0;
443 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
444 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
445 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
446 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
447 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
450 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
451 __m256d minushalf = _mm256_set1_pd(-0.5);
452 real *invsqrta,*dvda,*gbtab;
454 __m128i ifour = _mm_set1_epi32(4);
455 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
457 __m256d dummy_mask,cutoff_mask;
458 __m128 tmpmask0,tmpmask1;
459 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
460 __m256d one = _mm256_set1_pd(1.0);
461 __m256d two = _mm256_set1_pd(2.0);
467 jindex = nlist->jindex;
469 shiftidx = nlist->shift;
471 shiftvec = fr->shift_vec[0];
472 fshift = fr->fshift[0];
473 facel = _mm256_set1_pd(fr->epsfac);
474 charge = mdatoms->chargeA;
476 invsqrta = fr->invsqrta;
478 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
479 gbtab = fr->gbtab.data;
480 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
482 /* Avoid stupid compiler warnings */
483 jnrA = jnrB = jnrC = jnrD = 0;
492 for(iidx=0;iidx<4*DIM;iidx++)
497 /* Start outer loop over neighborlists */
498 for(iidx=0; iidx<nri; iidx++)
500 /* Load shift vector for this list */
501 i_shift_offset = DIM*shiftidx[iidx];
503 /* Load limits for loop over neighbors */
504 j_index_start = jindex[iidx];
505 j_index_end = jindex[iidx+1];
507 /* Get outer coordinate index */
509 i_coord_offset = DIM*inr;
511 /* Load i particle coords and add shift vector */
512 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
514 fix0 = _mm256_setzero_pd();
515 fiy0 = _mm256_setzero_pd();
516 fiz0 = _mm256_setzero_pd();
518 /* Load parameters for i particles */
519 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
520 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
522 dvdasum = _mm256_setzero_pd();
524 /* Start inner kernel loop */
525 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
528 /* Get j neighbor index, and coordinate index */
533 j_coord_offsetA = DIM*jnrA;
534 j_coord_offsetB = DIM*jnrB;
535 j_coord_offsetC = DIM*jnrC;
536 j_coord_offsetD = DIM*jnrD;
538 /* load j atom coordinates */
539 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
540 x+j_coord_offsetC,x+j_coord_offsetD,
543 /* Calculate displacement vector */
544 dx00 = _mm256_sub_pd(ix0,jx0);
545 dy00 = _mm256_sub_pd(iy0,jy0);
546 dz00 = _mm256_sub_pd(iz0,jz0);
548 /* Calculate squared distance and things based on it */
549 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
551 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
553 /* Load parameters for j particles */
554 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
555 charge+jnrC+0,charge+jnrD+0);
556 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
557 invsqrta+jnrC+0,invsqrta+jnrD+0);
559 /**************************
560 * CALCULATE INTERACTIONS *
561 **************************/
563 r00 = _mm256_mul_pd(rsq00,rinv00);
565 /* Compute parameters for interactions between i and j atoms */
566 qq00 = _mm256_mul_pd(iq0,jq0);
568 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
569 isaprod = _mm256_mul_pd(isai0,isaj0);
570 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
571 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
573 /* Calculate generalized born table index - this is a separate table from the normal one,
574 * but we use the same procedure by multiplying r with scale and truncating to integer.
576 rt = _mm256_mul_pd(r00,gbscale);
577 gbitab = _mm256_cvttpd_epi32(rt);
578 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
579 gbitab = _mm_slli_epi32(gbitab,2);
580 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
581 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
582 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
583 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
584 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
585 Heps = _mm256_mul_pd(gbeps,H);
586 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
587 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
588 vgb = _mm256_mul_pd(gbqqfactor,VV);
590 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
591 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
592 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
593 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
598 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
599 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
600 velec = _mm256_mul_pd(qq00,rinv00);
601 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
605 /* Calculate temporary vectorial force */
606 tx = _mm256_mul_pd(fscal,dx00);
607 ty = _mm256_mul_pd(fscal,dy00);
608 tz = _mm256_mul_pd(fscal,dz00);
610 /* Update vectorial force */
611 fix0 = _mm256_add_pd(fix0,tx);
612 fiy0 = _mm256_add_pd(fiy0,ty);
613 fiz0 = _mm256_add_pd(fiz0,tz);
615 fjptrA = f+j_coord_offsetA;
616 fjptrB = f+j_coord_offsetB;
617 fjptrC = f+j_coord_offsetC;
618 fjptrD = f+j_coord_offsetD;
619 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
621 /* Inner loop uses 55 flops */
627 /* Get j neighbor index, and coordinate index */
628 jnrlistA = jjnr[jidx];
629 jnrlistB = jjnr[jidx+1];
630 jnrlistC = jjnr[jidx+2];
631 jnrlistD = jjnr[jidx+3];
632 /* Sign of each element will be negative for non-real atoms.
633 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
634 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
636 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
638 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
639 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
640 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
642 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
643 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
644 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
645 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
646 j_coord_offsetA = DIM*jnrA;
647 j_coord_offsetB = DIM*jnrB;
648 j_coord_offsetC = DIM*jnrC;
649 j_coord_offsetD = DIM*jnrD;
651 /* load j atom coordinates */
652 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
653 x+j_coord_offsetC,x+j_coord_offsetD,
656 /* Calculate displacement vector */
657 dx00 = _mm256_sub_pd(ix0,jx0);
658 dy00 = _mm256_sub_pd(iy0,jy0);
659 dz00 = _mm256_sub_pd(iz0,jz0);
661 /* Calculate squared distance and things based on it */
662 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
664 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
666 /* Load parameters for j particles */
667 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
668 charge+jnrC+0,charge+jnrD+0);
669 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
670 invsqrta+jnrC+0,invsqrta+jnrD+0);
672 /**************************
673 * CALCULATE INTERACTIONS *
674 **************************/
676 r00 = _mm256_mul_pd(rsq00,rinv00);
677 r00 = _mm256_andnot_pd(dummy_mask,r00);
679 /* Compute parameters for interactions between i and j atoms */
680 qq00 = _mm256_mul_pd(iq0,jq0);
682 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
683 isaprod = _mm256_mul_pd(isai0,isaj0);
684 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
685 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
687 /* Calculate generalized born table index - this is a separate table from the normal one,
688 * but we use the same procedure by multiplying r with scale and truncating to integer.
690 rt = _mm256_mul_pd(r00,gbscale);
691 gbitab = _mm256_cvttpd_epi32(rt);
692 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
693 gbitab = _mm_slli_epi32(gbitab,2);
694 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
695 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
696 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
697 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
698 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
699 Heps = _mm256_mul_pd(gbeps,H);
700 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
701 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
702 vgb = _mm256_mul_pd(gbqqfactor,VV);
704 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
705 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
706 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
707 dvdatmp = _mm256_andnot_ps(dummy_mask,dvdatmp);
708 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
709 /* 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. */
710 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
711 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
712 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
713 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
714 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
715 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
716 velec = _mm256_mul_pd(qq00,rinv00);
717 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
721 fscal = _mm256_andnot_pd(dummy_mask,fscal);
723 /* Calculate temporary vectorial force */
724 tx = _mm256_mul_pd(fscal,dx00);
725 ty = _mm256_mul_pd(fscal,dy00);
726 tz = _mm256_mul_pd(fscal,dz00);
728 /* Update vectorial force */
729 fix0 = _mm256_add_pd(fix0,tx);
730 fiy0 = _mm256_add_pd(fiy0,ty);
731 fiz0 = _mm256_add_pd(fiz0,tz);
733 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
734 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
735 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
736 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
737 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
739 /* Inner loop uses 56 flops */
742 /* End of innermost loop */
744 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
745 f+i_coord_offset,fshift+i_shift_offset);
747 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
748 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
750 /* Increment number of inner iterations */
751 inneriter += j_index_end - j_index_start;
753 /* Outer loop uses 7 flops */
756 /* Increment number of outer iterations */
759 /* Update outer/inner flops */
761 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*56);