2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013, 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.
44 #include "../nb_kernel.h"
45 #include "types/simple.h"
49 #include "gromacs/simd/math_x86_avx_256_double.h"
50 #include "kernelutil_x86_avx_256_double.h"
53 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_256_double
54 * Electrostatics interaction: GeneralizedBorn
55 * VdW interaction: None
56 * Geometry: Particle-Particle
57 * Calculate force/pot: PotentialAndForce
60 nb_kernel_ElecGB_VdwNone_GeomP1P1_VF_avx_256_double
61 (t_nblist * gmx_restrict nlist,
62 rvec * gmx_restrict xx,
63 rvec * gmx_restrict ff,
64 t_forcerec * gmx_restrict fr,
65 t_mdatoms * gmx_restrict mdatoms,
66 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
67 t_nrnb * gmx_restrict nrnb)
69 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
70 * just 0 for non-waters.
71 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
72 * jnr indices corresponding to data put in the four positions in the SIMD register.
74 int i_shift_offset,i_coord_offset,outeriter,inneriter;
75 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
76 int jnrA,jnrB,jnrC,jnrD;
77 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
78 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
79 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
80 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
82 real *shiftvec,*fshift,*x,*f;
83 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
85 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
86 real * vdwioffsetptr0;
87 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
88 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
89 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
90 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
91 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
94 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
95 __m256d minushalf = _mm256_set1_pd(-0.5);
96 real *invsqrta,*dvda,*gbtab;
98 __m128i ifour = _mm_set1_epi32(4);
99 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
101 __m256d dummy_mask,cutoff_mask;
102 __m128 tmpmask0,tmpmask1;
103 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
104 __m256d one = _mm256_set1_pd(1.0);
105 __m256d two = _mm256_set1_pd(2.0);
111 jindex = nlist->jindex;
113 shiftidx = nlist->shift;
115 shiftvec = fr->shift_vec[0];
116 fshift = fr->fshift[0];
117 facel = _mm256_set1_pd(fr->epsfac);
118 charge = mdatoms->chargeA;
120 invsqrta = fr->invsqrta;
122 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
123 gbtab = fr->gbtab.data;
124 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
126 /* Avoid stupid compiler warnings */
127 jnrA = jnrB = jnrC = jnrD = 0;
136 for(iidx=0;iidx<4*DIM;iidx++)
141 /* Start outer loop over neighborlists */
142 for(iidx=0; iidx<nri; iidx++)
144 /* Load shift vector for this list */
145 i_shift_offset = DIM*shiftidx[iidx];
147 /* Load limits for loop over neighbors */
148 j_index_start = jindex[iidx];
149 j_index_end = jindex[iidx+1];
151 /* Get outer coordinate index */
153 i_coord_offset = DIM*inr;
155 /* Load i particle coords and add shift vector */
156 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
158 fix0 = _mm256_setzero_pd();
159 fiy0 = _mm256_setzero_pd();
160 fiz0 = _mm256_setzero_pd();
162 /* Load parameters for i particles */
163 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
164 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
166 /* Reset potential sums */
167 velecsum = _mm256_setzero_pd();
168 vgbsum = _mm256_setzero_pd();
169 dvdasum = _mm256_setzero_pd();
171 /* Start inner kernel loop */
172 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
175 /* Get j neighbor index, and coordinate index */
180 j_coord_offsetA = DIM*jnrA;
181 j_coord_offsetB = DIM*jnrB;
182 j_coord_offsetC = DIM*jnrC;
183 j_coord_offsetD = DIM*jnrD;
185 /* load j atom coordinates */
186 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
187 x+j_coord_offsetC,x+j_coord_offsetD,
190 /* Calculate displacement vector */
191 dx00 = _mm256_sub_pd(ix0,jx0);
192 dy00 = _mm256_sub_pd(iy0,jy0);
193 dz00 = _mm256_sub_pd(iz0,jz0);
195 /* Calculate squared distance and things based on it */
196 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
198 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
200 /* Load parameters for j particles */
201 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
202 charge+jnrC+0,charge+jnrD+0);
203 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
204 invsqrta+jnrC+0,invsqrta+jnrD+0);
206 /**************************
207 * CALCULATE INTERACTIONS *
208 **************************/
210 r00 = _mm256_mul_pd(rsq00,rinv00);
212 /* Compute parameters for interactions between i and j atoms */
213 qq00 = _mm256_mul_pd(iq0,jq0);
215 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
216 isaprod = _mm256_mul_pd(isai0,isaj0);
217 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
218 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
220 /* Calculate generalized born table index - this is a separate table from the normal one,
221 * but we use the same procedure by multiplying r with scale and truncating to integer.
223 rt = _mm256_mul_pd(r00,gbscale);
224 gbitab = _mm256_cvttpd_epi32(rt);
225 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
226 gbitab = _mm_slli_epi32(gbitab,2);
227 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
228 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
229 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
230 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
231 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
232 Heps = _mm256_mul_pd(gbeps,H);
233 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
234 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
235 vgb = _mm256_mul_pd(gbqqfactor,VV);
237 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
238 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
239 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
240 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
245 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
246 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
247 velec = _mm256_mul_pd(qq00,rinv00);
248 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
250 /* Update potential sum for this i atom from the interaction with this j atom. */
251 velecsum = _mm256_add_pd(velecsum,velec);
252 vgbsum = _mm256_add_pd(vgbsum,vgb);
256 /* Calculate temporary vectorial force */
257 tx = _mm256_mul_pd(fscal,dx00);
258 ty = _mm256_mul_pd(fscal,dy00);
259 tz = _mm256_mul_pd(fscal,dz00);
261 /* Update vectorial force */
262 fix0 = _mm256_add_pd(fix0,tx);
263 fiy0 = _mm256_add_pd(fiy0,ty);
264 fiz0 = _mm256_add_pd(fiz0,tz);
266 fjptrA = f+j_coord_offsetA;
267 fjptrB = f+j_coord_offsetB;
268 fjptrC = f+j_coord_offsetC;
269 fjptrD = f+j_coord_offsetD;
270 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
272 /* Inner loop uses 57 flops */
278 /* Get j neighbor index, and coordinate index */
279 jnrlistA = jjnr[jidx];
280 jnrlistB = jjnr[jidx+1];
281 jnrlistC = jjnr[jidx+2];
282 jnrlistD = jjnr[jidx+3];
283 /* Sign of each element will be negative for non-real atoms.
284 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
285 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
287 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
289 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
290 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
291 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
293 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
294 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
295 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
296 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
297 j_coord_offsetA = DIM*jnrA;
298 j_coord_offsetB = DIM*jnrB;
299 j_coord_offsetC = DIM*jnrC;
300 j_coord_offsetD = DIM*jnrD;
302 /* load j atom coordinates */
303 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
304 x+j_coord_offsetC,x+j_coord_offsetD,
307 /* Calculate displacement vector */
308 dx00 = _mm256_sub_pd(ix0,jx0);
309 dy00 = _mm256_sub_pd(iy0,jy0);
310 dz00 = _mm256_sub_pd(iz0,jz0);
312 /* Calculate squared distance and things based on it */
313 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
315 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
317 /* Load parameters for j particles */
318 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
319 charge+jnrC+0,charge+jnrD+0);
320 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
321 invsqrta+jnrC+0,invsqrta+jnrD+0);
323 /**************************
324 * CALCULATE INTERACTIONS *
325 **************************/
327 r00 = _mm256_mul_pd(rsq00,rinv00);
328 r00 = _mm256_andnot_pd(dummy_mask,r00);
330 /* Compute parameters for interactions between i and j atoms */
331 qq00 = _mm256_mul_pd(iq0,jq0);
333 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
334 isaprod = _mm256_mul_pd(isai0,isaj0);
335 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
336 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
338 /* Calculate generalized born table index - this is a separate table from the normal one,
339 * but we use the same procedure by multiplying r with scale and truncating to integer.
341 rt = _mm256_mul_pd(r00,gbscale);
342 gbitab = _mm256_cvttpd_epi32(rt);
343 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
344 gbitab = _mm_slli_epi32(gbitab,2);
345 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
346 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
347 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
348 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
349 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
350 Heps = _mm256_mul_pd(gbeps,H);
351 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
352 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
353 vgb = _mm256_mul_pd(gbqqfactor,VV);
355 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
356 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
357 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
358 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
359 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
360 /* 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. */
361 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
362 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
363 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
364 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
365 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
366 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
367 velec = _mm256_mul_pd(qq00,rinv00);
368 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
370 /* Update potential sum for this i atom from the interaction with this j atom. */
371 velec = _mm256_andnot_pd(dummy_mask,velec);
372 velecsum = _mm256_add_pd(velecsum,velec);
373 vgb = _mm256_andnot_pd(dummy_mask,vgb);
374 vgbsum = _mm256_add_pd(vgbsum,vgb);
378 fscal = _mm256_andnot_pd(dummy_mask,fscal);
380 /* Calculate temporary vectorial force */
381 tx = _mm256_mul_pd(fscal,dx00);
382 ty = _mm256_mul_pd(fscal,dy00);
383 tz = _mm256_mul_pd(fscal,dz00);
385 /* Update vectorial force */
386 fix0 = _mm256_add_pd(fix0,tx);
387 fiy0 = _mm256_add_pd(fiy0,ty);
388 fiz0 = _mm256_add_pd(fiz0,tz);
390 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
391 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
392 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
393 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
394 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
396 /* Inner loop uses 58 flops */
399 /* End of innermost loop */
401 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
402 f+i_coord_offset,fshift+i_shift_offset);
405 /* Update potential energies */
406 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
407 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
408 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
409 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
411 /* Increment number of inner iterations */
412 inneriter += j_index_end - j_index_start;
414 /* Outer loop uses 9 flops */
417 /* Increment number of outer iterations */
420 /* Update outer/inner flops */
422 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*58);
425 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
426 * Electrostatics interaction: GeneralizedBorn
427 * VdW interaction: None
428 * Geometry: Particle-Particle
429 * Calculate force/pot: Force
432 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_double
433 (t_nblist * gmx_restrict nlist,
434 rvec * gmx_restrict xx,
435 rvec * gmx_restrict ff,
436 t_forcerec * gmx_restrict fr,
437 t_mdatoms * gmx_restrict mdatoms,
438 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
439 t_nrnb * gmx_restrict nrnb)
441 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
442 * just 0 for non-waters.
443 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
444 * jnr indices corresponding to data put in the four positions in the SIMD register.
446 int i_shift_offset,i_coord_offset,outeriter,inneriter;
447 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
448 int jnrA,jnrB,jnrC,jnrD;
449 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
450 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
451 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
452 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
454 real *shiftvec,*fshift,*x,*f;
455 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
457 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
458 real * vdwioffsetptr0;
459 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
460 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
461 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
462 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
463 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
466 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
467 __m256d minushalf = _mm256_set1_pd(-0.5);
468 real *invsqrta,*dvda,*gbtab;
470 __m128i ifour = _mm_set1_epi32(4);
471 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
473 __m256d dummy_mask,cutoff_mask;
474 __m128 tmpmask0,tmpmask1;
475 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
476 __m256d one = _mm256_set1_pd(1.0);
477 __m256d two = _mm256_set1_pd(2.0);
483 jindex = nlist->jindex;
485 shiftidx = nlist->shift;
487 shiftvec = fr->shift_vec[0];
488 fshift = fr->fshift[0];
489 facel = _mm256_set1_pd(fr->epsfac);
490 charge = mdatoms->chargeA;
492 invsqrta = fr->invsqrta;
494 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
495 gbtab = fr->gbtab.data;
496 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
498 /* Avoid stupid compiler warnings */
499 jnrA = jnrB = jnrC = jnrD = 0;
508 for(iidx=0;iidx<4*DIM;iidx++)
513 /* Start outer loop over neighborlists */
514 for(iidx=0; iidx<nri; iidx++)
516 /* Load shift vector for this list */
517 i_shift_offset = DIM*shiftidx[iidx];
519 /* Load limits for loop over neighbors */
520 j_index_start = jindex[iidx];
521 j_index_end = jindex[iidx+1];
523 /* Get outer coordinate index */
525 i_coord_offset = DIM*inr;
527 /* Load i particle coords and add shift vector */
528 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
530 fix0 = _mm256_setzero_pd();
531 fiy0 = _mm256_setzero_pd();
532 fiz0 = _mm256_setzero_pd();
534 /* Load parameters for i particles */
535 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
536 isai0 = _mm256_set1_pd(invsqrta[inr+0]);
538 dvdasum = _mm256_setzero_pd();
540 /* Start inner kernel loop */
541 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
544 /* Get j neighbor index, and coordinate index */
549 j_coord_offsetA = DIM*jnrA;
550 j_coord_offsetB = DIM*jnrB;
551 j_coord_offsetC = DIM*jnrC;
552 j_coord_offsetD = DIM*jnrD;
554 /* load j atom coordinates */
555 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
556 x+j_coord_offsetC,x+j_coord_offsetD,
559 /* Calculate displacement vector */
560 dx00 = _mm256_sub_pd(ix0,jx0);
561 dy00 = _mm256_sub_pd(iy0,jy0);
562 dz00 = _mm256_sub_pd(iz0,jz0);
564 /* Calculate squared distance and things based on it */
565 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
567 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
569 /* Load parameters for j particles */
570 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
571 charge+jnrC+0,charge+jnrD+0);
572 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
573 invsqrta+jnrC+0,invsqrta+jnrD+0);
575 /**************************
576 * CALCULATE INTERACTIONS *
577 **************************/
579 r00 = _mm256_mul_pd(rsq00,rinv00);
581 /* Compute parameters for interactions between i and j atoms */
582 qq00 = _mm256_mul_pd(iq0,jq0);
584 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
585 isaprod = _mm256_mul_pd(isai0,isaj0);
586 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
587 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
589 /* Calculate generalized born table index - this is a separate table from the normal one,
590 * but we use the same procedure by multiplying r with scale and truncating to integer.
592 rt = _mm256_mul_pd(r00,gbscale);
593 gbitab = _mm256_cvttpd_epi32(rt);
594 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
595 gbitab = _mm_slli_epi32(gbitab,2);
596 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
597 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
598 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
599 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
600 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
601 Heps = _mm256_mul_pd(gbeps,H);
602 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
603 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
604 vgb = _mm256_mul_pd(gbqqfactor,VV);
606 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
607 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
608 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
609 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
614 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
615 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
616 velec = _mm256_mul_pd(qq00,rinv00);
617 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
621 /* Calculate temporary vectorial force */
622 tx = _mm256_mul_pd(fscal,dx00);
623 ty = _mm256_mul_pd(fscal,dy00);
624 tz = _mm256_mul_pd(fscal,dz00);
626 /* Update vectorial force */
627 fix0 = _mm256_add_pd(fix0,tx);
628 fiy0 = _mm256_add_pd(fiy0,ty);
629 fiz0 = _mm256_add_pd(fiz0,tz);
631 fjptrA = f+j_coord_offsetA;
632 fjptrB = f+j_coord_offsetB;
633 fjptrC = f+j_coord_offsetC;
634 fjptrD = f+j_coord_offsetD;
635 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
637 /* Inner loop uses 55 flops */
643 /* Get j neighbor index, and coordinate index */
644 jnrlistA = jjnr[jidx];
645 jnrlistB = jjnr[jidx+1];
646 jnrlistC = jjnr[jidx+2];
647 jnrlistD = jjnr[jidx+3];
648 /* Sign of each element will be negative for non-real atoms.
649 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
650 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
652 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
654 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
655 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
656 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
658 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
659 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
660 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
661 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
662 j_coord_offsetA = DIM*jnrA;
663 j_coord_offsetB = DIM*jnrB;
664 j_coord_offsetC = DIM*jnrC;
665 j_coord_offsetD = DIM*jnrD;
667 /* load j atom coordinates */
668 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
669 x+j_coord_offsetC,x+j_coord_offsetD,
672 /* Calculate displacement vector */
673 dx00 = _mm256_sub_pd(ix0,jx0);
674 dy00 = _mm256_sub_pd(iy0,jy0);
675 dz00 = _mm256_sub_pd(iz0,jz0);
677 /* Calculate squared distance and things based on it */
678 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
680 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
682 /* Load parameters for j particles */
683 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
684 charge+jnrC+0,charge+jnrD+0);
685 isaj0 = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0,
686 invsqrta+jnrC+0,invsqrta+jnrD+0);
688 /**************************
689 * CALCULATE INTERACTIONS *
690 **************************/
692 r00 = _mm256_mul_pd(rsq00,rinv00);
693 r00 = _mm256_andnot_pd(dummy_mask,r00);
695 /* Compute parameters for interactions between i and j atoms */
696 qq00 = _mm256_mul_pd(iq0,jq0);
698 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
699 isaprod = _mm256_mul_pd(isai0,isaj0);
700 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq00,_mm256_mul_pd(isaprod,gbinvepsdiff)));
701 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
703 /* Calculate generalized born table index - this is a separate table from the normal one,
704 * but we use the same procedure by multiplying r with scale and truncating to integer.
706 rt = _mm256_mul_pd(r00,gbscale);
707 gbitab = _mm256_cvttpd_epi32(rt);
708 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
709 gbitab = _mm_slli_epi32(gbitab,2);
710 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
711 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
712 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
713 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
714 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
715 Heps = _mm256_mul_pd(gbeps,H);
716 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
717 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
718 vgb = _mm256_mul_pd(gbqqfactor,VV);
720 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
721 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
722 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r00)));
723 dvdatmp = _mm256_andnot_pd(dummy_mask,dvdatmp);
724 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
725 /* 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. */
726 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
727 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
728 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
729 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
730 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
731 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj0,isaj0)));
732 velec = _mm256_mul_pd(qq00,rinv00);
733 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv00),fgb),rinv00);
737 fscal = _mm256_andnot_pd(dummy_mask,fscal);
739 /* Calculate temporary vectorial force */
740 tx = _mm256_mul_pd(fscal,dx00);
741 ty = _mm256_mul_pd(fscal,dy00);
742 tz = _mm256_mul_pd(fscal,dz00);
744 /* Update vectorial force */
745 fix0 = _mm256_add_pd(fix0,tx);
746 fiy0 = _mm256_add_pd(fiy0,ty);
747 fiz0 = _mm256_add_pd(fiz0,tz);
749 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
750 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
751 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
752 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
753 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
755 /* Inner loop uses 56 flops */
758 /* End of innermost loop */
760 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
761 f+i_coord_offset,fshift+i_shift_offset);
763 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai0,isai0));
764 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
766 /* Increment number of inner iterations */
767 inneriter += j_index_end - j_index_start;
769 /* Outer loop uses 7 flops */
772 /* Increment number of outer iterations */
775 /* Update outer/inner flops */
777 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*56);