2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 * Note: this file was generated by the GROMACS avx_256_double kernel generator.
42 #include "../nb_kernel.h"
43 #include "types/simple.h"
44 #include "gromacs/math/vec.h"
47 #include "gromacs/simd/math_x86_avx_256_double.h"
48 #include "kernelutil_x86_avx_256_double.h"
51 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_256_double
52 * Electrostatics interaction: GeneralizedBorn
53 * VdW interaction: None
54 * Geometry: Particle-Particle
55 * Calculate force/pot: PotentialAndForce
58 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_256_double
59 (t_nblist * gmx_restrict nlist,
60 rvec * gmx_restrict xx,
61 rvec * gmx_restrict ff,
62 t_forcerec * gmx_restrict fr,
63 t_mdatoms * gmx_restrict mdatoms,
64 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
65 t_nrnb * gmx_restrict nrnb)
67 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
68 * just 0 for non-waters.
69 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
70 * jnr indices corresponding to data put in the four positions in the SIMD register.
72 int i_shift_offset,i_coord_offset,outeriter,inneriter;
73 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
74 int jnrA,jnrB,jnrC,jnrD;
75 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
76 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
77 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
78 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
80 real *shiftvec,*fshift,*x,*f;
81 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
83 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
84 real * vdwioffsetptr0;
85 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
86 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
87 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
88 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
89 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
92 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
93 __m256d minushalf = _mm256_set1_pd(-0.5);
94 real *invsqrta,*dvda,*gbtab;
96 __m128i ifour = _mm_set1_epi32(4);
97 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
99 __m256d dummy_mask,cutoff_mask;
100 __m128 tmpmask0,tmpmask1;
101 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
102 __m256d one = _mm256_set1_pd(1.0);
103 __m256d two = _mm256_set1_pd(2.0);
109 jindex = nlist->jindex;
111 shiftidx = nlist->shift;
113 shiftvec = fr->shift_vec[0];
114 fshift = fr->fshift[0];
115 facel = _mm256_set1_pd(fr->epsfac);
116 charge = mdatoms->chargeA;
118 invsqrta = fr->invsqrta;
120 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
121 gbtab = fr->gbtab.data;
122 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
124 /* Avoid stupid compiler warnings */
125 jnrA = jnrB = jnrC = jnrD = 0;
134 for(iidx=0;iidx<4*DIM;iidx++)
139 /* Start outer loop over neighborlists */
140 for(iidx=0; iidx<nri; iidx++)
142 /* Load shift vector for this list */
143 i_shift_offset = DIM*shiftidx[iidx];
145 /* Load limits for loop over neighbors */
146 j_index_start = jindex[iidx];
147 j_index_end = jindex[iidx+1];
149 /* Get outer coordinate index */
151 i_coord_offset = DIM*inr;
153 /* Load i particle coords and add shift vector */
154 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
156 fix0 = _mm256_setzero_pd();
157 fiy0 = _mm256_setzero_pd();
158 fiz0 = _mm256_setzero_pd();
160 /* Load parameters for i particles */
161 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
162 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
164 /* Reset potential sums */
165 velecsum = _mm256_setzero_pd();
166 vgbsum = _mm256_setzero_pd();
167 dvdasum = _mm256_setzero_pd();
169 /* Start inner kernel loop */
170 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
173 /* Get j neighbor index, and coordinate index */
178 j_coord_offsetA = DIM*jnrA;
179 j_coord_offsetB = DIM*jnrB;
180 j_coord_offsetC = DIM*jnrC;
181 j_coord_offsetD = DIM*jnrD;
183 /* load j atom coordinates */
184 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
185 x+j_coord_offsetC,x+j_coord_offsetD,
188 /* Calculate displacement vector */
189 dx00 = _mm256_sub_pd(ix0,jx0);
190 dy00 = _mm256_sub_pd(iy0,jy0);
191 dz00 = _mm256_sub_pd(iz0,jz0);
193 /* Calculate squared distance and things based on it */
194 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
196 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
198 /* Load parameters for j particles */
199 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
200 charge+jnrC+0,charge+jnrD+0);
201 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
202 invsqrta+jnrC+0,invsqrta+jnrD+0);
204 /**************************
205 * CALCULATE INTERACTIONS *
206 **************************/
208 r00 = _mm256_mul_pd(rsq00,rinv00);
210 /* Compute parameters for interactions between i and j atoms */
211 qq00 = _mm256_mul_pd(iq0,jq0);
213 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
214 isaprod = _mm256_mul_pd(isai0,isaj0);
215 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
216 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
218 /* Calculate generalized born table index - this is a separate table from the normal one,
219 * but we use the same procedure by multiplying r with scale and truncating to integer.
221 rt = _mm256_mul_pd(r00,gbscale);
222 gbitab = _mm256_cvttpd_epi32(rt);
223 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
224 gbitab = _mm_slli_epi32(gbitab,2);
225 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
226 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
227 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
228 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
229 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
230 Heps = _mm256_mul_pd(gbeps,H);
231 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
232 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
233 vgb = _mm256_mul_pd(gbqqfactor,VV);
235 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
236 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
237 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
238 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
243 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
244 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
245 velec = _mm256_mul_pd(qq00,rinv00);
246 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
248 /* Update potential sum for this i atom from the interaction with this j atom. */
249 velecsum = _mm256_add_pd(velecsum,velec);
250 vgbsum = _mm256_add_pd(vgbsum,vgb);
254 /* Calculate temporary vectorial force */
255 tx = _mm256_mul_pd(fscal,dx00);
256 ty = _mm256_mul_pd(fscal,dy00);
257 tz = _mm256_mul_pd(fscal,dz00);
259 /* Update vectorial force */
260 fix0 = _mm256_add_pd(fix0,tx);
261 fiy0 = _mm256_add_pd(fiy0,ty);
262 fiz0 = _mm256_add_pd(fiz0,tz);
264 fjptrA = f+j_coord_offsetA;
265 fjptrB = f+j_coord_offsetB;
266 fjptrC = f+j_coord_offsetC;
267 fjptrD = f+j_coord_offsetD;
268 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
270 /* Inner loop uses 57 flops */
276 /* Get j neighbor index, and coordinate index */
277 jnrlistA = jjnr[jidx];
278 jnrlistB = jjnr[jidx+1];
279 jnrlistC = jjnr[jidx+2];
280 jnrlistD = jjnr[jidx+3];
281 /* Sign of each element will be negative for non-real atoms.
282 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
283 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
285 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
287 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
288 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
289 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
291 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
292 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
293 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
294 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
295 j_coord_offsetA = DIM*jnrA;
296 j_coord_offsetB = DIM*jnrB;
297 j_coord_offsetC = DIM*jnrC;
298 j_coord_offsetD = DIM*jnrD;
300 /* load j atom coordinates */
301 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
302 x+j_coord_offsetC,x+j_coord_offsetD,
305 /* Calculate displacement vector */
306 dx00 = _mm256_sub_pd(ix0,jx0);
307 dy00 = _mm256_sub_pd(iy0,jy0);
308 dz00 = _mm256_sub_pd(iz0,jz0);
310 /* Calculate squared distance and things based on it */
311 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
313 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
315 /* Load parameters for j particles */
316 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
317 charge+jnrC+0,charge+jnrD+0);
318 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
319 invsqrta+jnrC+0,invsqrta+jnrD+0);
321 /**************************
322 * CALCULATE INTERACTIONS *
323 **************************/
325 r00 = _mm256_mul_pd(rsq00,rinv00);
326 r00 = _mm256_andnot_pd(dummy_mask,r00);
328 /* Compute parameters for interactions between i and j atoms */
329 qq00 = _mm256_mul_pd(iq0,jq0);
331 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
332 isaprod = _mm256_mul_pd(isai0,isaj0);
333 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
334 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
336 /* Calculate generalized born table index - this is a separate table from the normal one,
337 * but we use the same procedure by multiplying r with scale and truncating to integer.
339 rt = _mm256_mul_pd(r00,gbscale);
340 gbitab = _mm256_cvttpd_epi32(rt);
341 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
342 gbitab = _mm_slli_epi32(gbitab,2);
343 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
344 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
345 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
346 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
347 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
348 Heps = _mm256_mul_pd(gbeps,H);
349 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
350 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
351 vgb = _mm256_mul_pd(gbqqfactor,VV);
353 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
354 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
355 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
356 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
357 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
358 /* 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. */
359 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
360 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
361 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
362 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
363 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
364 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
365 velec = _mm256_mul_pd(qq00,rinv00);
366 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
368 /* Update potential sum for this i atom from the interaction with this j atom. */
369 velec = _mm256_andnot_pd(dummy_mask,velec);
370 velecsum = _mm256_add_pd(velecsum,velec);
371 vgb = _mm256_andnot_pd(dummy_mask,vgb);
372 vgbsum = _mm256_add_pd(vgbsum,vgb);
376 fscal = _mm256_andnot_pd(dummy_mask,fscal);
378 /* Calculate temporary vectorial force */
379 tx = _mm256_mul_pd(fscal,dx00);
380 ty = _mm256_mul_pd(fscal,dy00);
381 tz = _mm256_mul_pd(fscal,dz00);
383 /* Update vectorial force */
384 fix0 = _mm256_add_pd(fix0,tx);
385 fiy0 = _mm256_add_pd(fiy0,ty);
386 fiz0 = _mm256_add_pd(fiz0,tz);
388 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
389 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
390 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
391 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
392 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
394 /* Inner loop uses 58 flops */
397 /* End of innermost loop */
399 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
400 f+i_coord_offset,fshift+i_shift_offset);
403 /* Update potential energies */
404 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
405 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
406 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
407 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
409 /* Increment number of inner iterations */
410 inneriter += j_index_end - j_index_start;
412 /* Outer loop uses 9 flops */
415 /* Increment number of outer iterations */
418 /* Update outer/inner flops */
420 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*58);
423 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
424 * Electrostatics interaction: GeneralizedBorn
425 * VdW interaction: None
426 * Geometry: Particle-Particle
427 * Calculate force/pot: Force
430 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
431 (t_nblist * gmx_restrict nlist,
432 rvec * gmx_restrict xx,
433 rvec * gmx_restrict ff,
434 t_forcerec * gmx_restrict fr,
435 t_mdatoms * gmx_restrict mdatoms,
436 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
437 t_nrnb * gmx_restrict nrnb)
439 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
440 * just 0 for non-waters.
441 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
442 * jnr indices corresponding to data put in the four positions in the SIMD register.
444 int i_shift_offset,i_coord_offset,outeriter,inneriter;
445 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
446 int jnrA,jnrB,jnrC,jnrD;
447 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
448 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
449 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
450 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
452 real *shiftvec,*fshift,*x,*f;
453 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
455 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
456 real * vdwioffsetptr0;
457 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
458 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
459 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
460 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
461 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
464 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
465 __m256d minushalf = _mm256_set1_pd(-0.5);
466 real *invsqrta,*dvda,*gbtab;
468 __m128i ifour = _mm_set1_epi32(4);
469 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
471 __m256d dummy_mask,cutoff_mask;
472 __m128 tmpmask0,tmpmask1;
473 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
474 __m256d one = _mm256_set1_pd(1.0);
475 __m256d two = _mm256_set1_pd(2.0);
481 jindex = nlist->jindex;
483 shiftidx = nlist->shift;
485 shiftvec = fr->shift_vec[0];
486 fshift = fr->fshift[0];
487 facel = _mm256_set1_pd(fr->epsfac);
488 charge = mdatoms->chargeA;
490 invsqrta = fr->invsqrta;
492 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
493 gbtab = fr->gbtab.data;
494 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
496 /* Avoid stupid compiler warnings */
497 jnrA = jnrB = jnrC = jnrD = 0;
506 for(iidx=0;iidx<4*DIM;iidx++)
511 /* Start outer loop over neighborlists */
512 for(iidx=0; iidx<nri; iidx++)
514 /* Load shift vector for this list */
515 i_shift_offset = DIM*shiftidx[iidx];
517 /* Load limits for loop over neighbors */
518 j_index_start = jindex[iidx];
519 j_index_end = jindex[iidx+1];
521 /* Get outer coordinate index */
523 i_coord_offset = DIM*inr;
525 /* Load i particle coords and add shift vector */
526 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
528 fix0 = _mm256_setzero_pd();
529 fiy0 = _mm256_setzero_pd();
530 fiz0 = _mm256_setzero_pd();
532 /* Load parameters for i particles */
533 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
534 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
536 dvdasum = _mm256_setzero_pd();
538 /* Start inner kernel loop */
539 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
542 /* Get j neighbor index, and coordinate index */
547 j_coord_offsetA = DIM*jnrA;
548 j_coord_offsetB = DIM*jnrB;
549 j_coord_offsetC = DIM*jnrC;
550 j_coord_offsetD = DIM*jnrD;
552 /* load j atom coordinates */
553 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
554 x+j_coord_offsetC,x+j_coord_offsetD,
557 /* Calculate displacement vector */
558 dx00 = _mm256_sub_pd(ix0,jx0);
559 dy00 = _mm256_sub_pd(iy0,jy0);
560 dz00 = _mm256_sub_pd(iz0,jz0);
562 /* Calculate squared distance and things based on it */
563 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
565 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
567 /* Load parameters for j particles */
568 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
569 charge+jnrC+0,charge+jnrD+0);
570 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
571 invsqrta+jnrC+0,invsqrta+jnrD+0);
573 /**************************
574 * CALCULATE INTERACTIONS *
575 **************************/
577 r00 = _mm256_mul_pd(rsq00,rinv00);
579 /* Compute parameters for interactions between i and j atoms */
580 qq00 = _mm256_mul_pd(iq0,jq0);
582 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
583 isaprod = _mm256_mul_pd(isai0,isaj0);
584 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
585 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
587 /* Calculate generalized born table index - this is a separate table from the normal one,
588 * but we use the same procedure by multiplying r with scale and truncating to integer.
590 rt = _mm256_mul_pd(r00,gbscale);
591 gbitab = _mm256_cvttpd_epi32(rt);
592 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
593 gbitab = _mm_slli_epi32(gbitab,2);
594 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
595 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
596 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
597 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
598 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
599 Heps = _mm256_mul_pd(gbeps,H);
600 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
601 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
602 vgb = _mm256_mul_pd(gbqqfactor,VV);
604 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
605 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
606 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
607 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
612 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
613 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
614 velec = _mm256_mul_pd(qq00,rinv00);
615 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
619 /* Calculate temporary vectorial force */
620 tx = _mm256_mul_pd(fscal,dx00);
621 ty = _mm256_mul_pd(fscal,dy00);
622 tz = _mm256_mul_pd(fscal,dz00);
624 /* Update vectorial force */
625 fix0 = _mm256_add_pd(fix0,tx);
626 fiy0 = _mm256_add_pd(fiy0,ty);
627 fiz0 = _mm256_add_pd(fiz0,tz);
629 fjptrA = f+j_coord_offsetA;
630 fjptrB = f+j_coord_offsetB;
631 fjptrC = f+j_coord_offsetC;
632 fjptrD = f+j_coord_offsetD;
633 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
635 /* Inner loop uses 55 flops */
641 /* Get j neighbor index, and coordinate index */
642 jnrlistA = jjnr[jidx];
643 jnrlistB = jjnr[jidx+1];
644 jnrlistC = jjnr[jidx+2];
645 jnrlistD = jjnr[jidx+3];
646 /* Sign of each element will be negative for non-real atoms.
647 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
648 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
650 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
652 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
653 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
654 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
656 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
657 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
658 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
659 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
660 j_coord_offsetA = DIM*jnrA;
661 j_coord_offsetB = DIM*jnrB;
662 j_coord_offsetC = DIM*jnrC;
663 j_coord_offsetD = DIM*jnrD;
665 /* load j atom coordinates */
666 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
667 x+j_coord_offsetC,x+j_coord_offsetD,
670 /* Calculate displacement vector */
671 dx00 = _mm256_sub_pd(ix0,jx0);
672 dy00 = _mm256_sub_pd(iy0,jy0);
673 dz00 = _mm256_sub_pd(iz0,jz0);
675 /* Calculate squared distance and things based on it */
676 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
678 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
680 /* Load parameters for j particles */
681 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
682 charge+jnrC+0,charge+jnrD+0);
683 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
684 invsqrta+jnrC+0,invsqrta+jnrD+0);
686 /**************************
687 * CALCULATE INTERACTIONS *
688 **************************/
690 r00 = _mm256_mul_pd(rsq00,rinv00);
691 r00 = _mm256_andnot_pd(dummy_mask,r00);
693 /* Compute parameters for interactions between i and j atoms */
694 qq00 = _mm256_mul_pd(iq0,jq0);
696 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
697 isaprod = _mm256_mul_pd(isai0,isaj0);
698 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
699 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
701 /* Calculate generalized born table index - this is a separate table from the normal one,
702 * but we use the same procedure by multiplying r with scale and truncating to integer.
704 rt = _mm256_mul_pd(r00,gbscale);
705 gbitab = _mm256_cvttpd_epi32(rt);
706 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
707 gbitab = _mm_slli_epi32(gbitab,2);
708 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
709 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
710 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
711 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
712 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
713 Heps = _mm256_mul_pd(gbeps,H);
714 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
715 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
716 vgb = _mm256_mul_pd(gbqqfactor,VV);
718 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
719 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
720 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
721 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
722 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
723 /* 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. */
724 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
725 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
726 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
727 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
728 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
729 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
730 velec = _mm256_mul_pd(qq00,rinv00);
731 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
735 fscal = _mm256_andnot_pd(dummy_mask,fscal);
737 /* Calculate temporary vectorial force */
738 tx = _mm256_mul_pd(fscal,dx00);
739 ty = _mm256_mul_pd(fscal,dy00);
740 tz = _mm256_mul_pd(fscal,dz00);
742 /* Update vectorial force */
743 fix0 = _mm256_add_pd(fix0,tx);
744 fiy0 = _mm256_add_pd(fiy0,ty);
745 fiz0 = _mm256_add_pd(fiz0,tz);
747 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
748 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
749 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
750 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
751 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
753 /* Inner loop uses 56 flops */
756 /* End of innermost loop */
758 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
759 f+i_coord_offset,fshift+i_shift_offset);
761 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
762 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
764 /* Increment number of inner iterations */
765 inneriter += j_index_end - j_index_start;
767 /* Outer loop uses 7 flops */
770 /* Increment number of outer iterations */
773 /* Update outer/inner flops */
775 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*56);