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 "gromacs/legacyheaders/types/simple.h"
44 #include "gromacs/math/vec.h"
45 #include "gromacs/legacyheaders/nrnb.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_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_256_double
52 * Electrostatics interaction: ReactionField
53 * VdW interaction: CubicSplineTable
54 * Geometry: Particle-Particle
55 * Calculate force/pot: PotentialAndForce
58 nb_kernel_ElecRFCut_VdwCSTab_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 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
95 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
96 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
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;
119 krf = _mm256_set1_pd(fr->ic->k_rf);
120 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
121 crf = _mm256_set1_pd(fr->ic->c_rf);
122 nvdwtype = fr->ntype;
124 vdwtype = mdatoms->typeA;
126 vftab = kernel_data->table_vdw->data;
127 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
129 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
130 rcutoff_scalar = fr->rcoulomb;
131 rcutoff = _mm256_set1_pd(rcutoff_scalar);
132 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
134 /* Avoid stupid compiler warnings */
135 jnrA = jnrB = jnrC = jnrD = 0;
144 for(iidx=0;iidx<4*DIM;iidx++)
149 /* Start outer loop over neighborlists */
150 for(iidx=0; iidx<nri; iidx++)
152 /* Load shift vector for this list */
153 i_shift_offset = DIM*shiftidx[iidx];
155 /* Load limits for loop over neighbors */
156 j_index_start = jindex[iidx];
157 j_index_end = jindex[iidx+1];
159 /* Get outer coordinate index */
161 i_coord_offset = DIM*inr;
163 /* Load i particle coords and add shift vector */
164 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
166 fix0 = _mm256_setzero_pd();
167 fiy0 = _mm256_setzero_pd();
168 fiz0 = _mm256_setzero_pd();
170 /* Load parameters for i particles */
171 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
172 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
174 /* Reset potential sums */
175 velecsum = _mm256_setzero_pd();
176 vvdwsum = _mm256_setzero_pd();
178 /* Start inner kernel loop */
179 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
182 /* Get j neighbor index, and coordinate index */
187 j_coord_offsetA = DIM*jnrA;
188 j_coord_offsetB = DIM*jnrB;
189 j_coord_offsetC = DIM*jnrC;
190 j_coord_offsetD = DIM*jnrD;
192 /* load j atom coordinates */
193 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
194 x+j_coord_offsetC,x+j_coord_offsetD,
197 /* Calculate displacement vector */
198 dx00 = _mm256_sub_pd(ix0,jx0);
199 dy00 = _mm256_sub_pd(iy0,jy0);
200 dz00 = _mm256_sub_pd(iz0,jz0);
202 /* Calculate squared distance and things based on it */
203 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
205 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
207 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
209 /* Load parameters for j particles */
210 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
211 charge+jnrC+0,charge+jnrD+0);
212 vdwjidx0A = 2*vdwtype[jnrA+0];
213 vdwjidx0B = 2*vdwtype[jnrB+0];
214 vdwjidx0C = 2*vdwtype[jnrC+0];
215 vdwjidx0D = 2*vdwtype[jnrD+0];
217 /**************************
218 * CALCULATE INTERACTIONS *
219 **************************/
221 if (gmx_mm256_any_lt(rsq00,rcutoff2))
224 r00 = _mm256_mul_pd(rsq00,rinv00);
226 /* Compute parameters for interactions between i and j atoms */
227 qq00 = _mm256_mul_pd(iq0,jq0);
228 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
229 vdwioffsetptr0+vdwjidx0B,
230 vdwioffsetptr0+vdwjidx0C,
231 vdwioffsetptr0+vdwjidx0D,
234 /* Calculate table index by multiplying r with table scale and truncate to integer */
235 rt = _mm256_mul_pd(r00,vftabscale);
236 vfitab = _mm256_cvttpd_epi32(rt);
237 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
238 vfitab = _mm_slli_epi32(vfitab,3);
240 /* REACTION-FIELD ELECTROSTATICS */
241 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
242 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
244 /* CUBIC SPLINE TABLE DISPERSION */
245 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
246 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
247 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
248 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
249 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
250 Heps = _mm256_mul_pd(vfeps,H);
251 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
252 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
253 vvdw6 = _mm256_mul_pd(c6_00,VV);
254 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
255 fvdw6 = _mm256_mul_pd(c6_00,FF);
257 /* CUBIC SPLINE TABLE REPULSION */
258 vfitab = _mm_add_epi32(vfitab,ifour);
259 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
260 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
261 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
262 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
263 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
264 Heps = _mm256_mul_pd(vfeps,H);
265 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
266 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
267 vvdw12 = _mm256_mul_pd(c12_00,VV);
268 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
269 fvdw12 = _mm256_mul_pd(c12_00,FF);
270 vvdw = _mm256_add_pd(vvdw12,vvdw6);
271 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
273 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
275 /* Update potential sum for this i atom from the interaction with this j atom. */
276 velec = _mm256_and_pd(velec,cutoff_mask);
277 velecsum = _mm256_add_pd(velecsum,velec);
278 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
279 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
281 fscal = _mm256_add_pd(felec,fvdw);
283 fscal = _mm256_and_pd(fscal,cutoff_mask);
285 /* Calculate temporary vectorial force */
286 tx = _mm256_mul_pd(fscal,dx00);
287 ty = _mm256_mul_pd(fscal,dy00);
288 tz = _mm256_mul_pd(fscal,dz00);
290 /* Update vectorial force */
291 fix0 = _mm256_add_pd(fix0,tx);
292 fiy0 = _mm256_add_pd(fiy0,ty);
293 fiz0 = _mm256_add_pd(fiz0,tz);
295 fjptrA = f+j_coord_offsetA;
296 fjptrB = f+j_coord_offsetB;
297 fjptrC = f+j_coord_offsetC;
298 fjptrD = f+j_coord_offsetD;
299 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
303 /* Inner loop uses 72 flops */
309 /* Get j neighbor index, and coordinate index */
310 jnrlistA = jjnr[jidx];
311 jnrlistB = jjnr[jidx+1];
312 jnrlistC = jjnr[jidx+2];
313 jnrlistD = jjnr[jidx+3];
314 /* Sign of each element will be negative for non-real atoms.
315 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
316 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
318 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
320 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
321 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
322 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
324 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
325 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
326 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
327 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
328 j_coord_offsetA = DIM*jnrA;
329 j_coord_offsetB = DIM*jnrB;
330 j_coord_offsetC = DIM*jnrC;
331 j_coord_offsetD = DIM*jnrD;
333 /* load j atom coordinates */
334 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
335 x+j_coord_offsetC,x+j_coord_offsetD,
338 /* Calculate displacement vector */
339 dx00 = _mm256_sub_pd(ix0,jx0);
340 dy00 = _mm256_sub_pd(iy0,jy0);
341 dz00 = _mm256_sub_pd(iz0,jz0);
343 /* Calculate squared distance and things based on it */
344 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
346 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
348 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
350 /* Load parameters for j particles */
351 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
352 charge+jnrC+0,charge+jnrD+0);
353 vdwjidx0A = 2*vdwtype[jnrA+0];
354 vdwjidx0B = 2*vdwtype[jnrB+0];
355 vdwjidx0C = 2*vdwtype[jnrC+0];
356 vdwjidx0D = 2*vdwtype[jnrD+0];
358 /**************************
359 * CALCULATE INTERACTIONS *
360 **************************/
362 if (gmx_mm256_any_lt(rsq00,rcutoff2))
365 r00 = _mm256_mul_pd(rsq00,rinv00);
366 r00 = _mm256_andnot_pd(dummy_mask,r00);
368 /* Compute parameters for interactions between i and j atoms */
369 qq00 = _mm256_mul_pd(iq0,jq0);
370 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
371 vdwioffsetptr0+vdwjidx0B,
372 vdwioffsetptr0+vdwjidx0C,
373 vdwioffsetptr0+vdwjidx0D,
376 /* Calculate table index by multiplying r with table scale and truncate to integer */
377 rt = _mm256_mul_pd(r00,vftabscale);
378 vfitab = _mm256_cvttpd_epi32(rt);
379 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
380 vfitab = _mm_slli_epi32(vfitab,3);
382 /* REACTION-FIELD ELECTROSTATICS */
383 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
384 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
386 /* CUBIC SPLINE TABLE DISPERSION */
387 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
388 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
389 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
390 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
391 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
392 Heps = _mm256_mul_pd(vfeps,H);
393 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
394 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
395 vvdw6 = _mm256_mul_pd(c6_00,VV);
396 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
397 fvdw6 = _mm256_mul_pd(c6_00,FF);
399 /* CUBIC SPLINE TABLE REPULSION */
400 vfitab = _mm_add_epi32(vfitab,ifour);
401 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
402 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
403 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
404 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
405 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
406 Heps = _mm256_mul_pd(vfeps,H);
407 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
408 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
409 vvdw12 = _mm256_mul_pd(c12_00,VV);
410 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
411 fvdw12 = _mm256_mul_pd(c12_00,FF);
412 vvdw = _mm256_add_pd(vvdw12,vvdw6);
413 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
415 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
417 /* Update potential sum for this i atom from the interaction with this j atom. */
418 velec = _mm256_and_pd(velec,cutoff_mask);
419 velec = _mm256_andnot_pd(dummy_mask,velec);
420 velecsum = _mm256_add_pd(velecsum,velec);
421 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
422 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
423 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
425 fscal = _mm256_add_pd(felec,fvdw);
427 fscal = _mm256_and_pd(fscal,cutoff_mask);
429 fscal = _mm256_andnot_pd(dummy_mask,fscal);
431 /* Calculate temporary vectorial force */
432 tx = _mm256_mul_pd(fscal,dx00);
433 ty = _mm256_mul_pd(fscal,dy00);
434 tz = _mm256_mul_pd(fscal,dz00);
436 /* Update vectorial force */
437 fix0 = _mm256_add_pd(fix0,tx);
438 fiy0 = _mm256_add_pd(fiy0,ty);
439 fiz0 = _mm256_add_pd(fiz0,tz);
441 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
442 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
443 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
444 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
445 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
449 /* Inner loop uses 73 flops */
452 /* End of innermost loop */
454 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
455 f+i_coord_offset,fshift+i_shift_offset);
458 /* Update potential energies */
459 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
460 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
462 /* Increment number of inner iterations */
463 inneriter += j_index_end - j_index_start;
465 /* Outer loop uses 9 flops */
468 /* Increment number of outer iterations */
471 /* Update outer/inner flops */
473 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*73);
476 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_256_double
477 * Electrostatics interaction: ReactionField
478 * VdW interaction: CubicSplineTable
479 * Geometry: Particle-Particle
480 * Calculate force/pot: Force
483 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_256_double
484 (t_nblist * gmx_restrict nlist,
485 rvec * gmx_restrict xx,
486 rvec * gmx_restrict ff,
487 t_forcerec * gmx_restrict fr,
488 t_mdatoms * gmx_restrict mdatoms,
489 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
490 t_nrnb * gmx_restrict nrnb)
492 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
493 * just 0 for non-waters.
494 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
495 * jnr indices corresponding to data put in the four positions in the SIMD register.
497 int i_shift_offset,i_coord_offset,outeriter,inneriter;
498 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
499 int jnrA,jnrB,jnrC,jnrD;
500 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
501 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
502 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
503 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
505 real *shiftvec,*fshift,*x,*f;
506 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
508 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
509 real * vdwioffsetptr0;
510 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
511 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
512 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
513 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
514 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
517 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
520 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
521 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
523 __m128i ifour = _mm_set1_epi32(4);
524 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
526 __m256d dummy_mask,cutoff_mask;
527 __m128 tmpmask0,tmpmask1;
528 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
529 __m256d one = _mm256_set1_pd(1.0);
530 __m256d two = _mm256_set1_pd(2.0);
536 jindex = nlist->jindex;
538 shiftidx = nlist->shift;
540 shiftvec = fr->shift_vec[0];
541 fshift = fr->fshift[0];
542 facel = _mm256_set1_pd(fr->epsfac);
543 charge = mdatoms->chargeA;
544 krf = _mm256_set1_pd(fr->ic->k_rf);
545 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
546 crf = _mm256_set1_pd(fr->ic->c_rf);
547 nvdwtype = fr->ntype;
549 vdwtype = mdatoms->typeA;
551 vftab = kernel_data->table_vdw->data;
552 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
554 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
555 rcutoff_scalar = fr->rcoulomb;
556 rcutoff = _mm256_set1_pd(rcutoff_scalar);
557 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
559 /* Avoid stupid compiler warnings */
560 jnrA = jnrB = jnrC = jnrD = 0;
569 for(iidx=0;iidx<4*DIM;iidx++)
574 /* Start outer loop over neighborlists */
575 for(iidx=0; iidx<nri; iidx++)
577 /* Load shift vector for this list */
578 i_shift_offset = DIM*shiftidx[iidx];
580 /* Load limits for loop over neighbors */
581 j_index_start = jindex[iidx];
582 j_index_end = jindex[iidx+1];
584 /* Get outer coordinate index */
586 i_coord_offset = DIM*inr;
588 /* Load i particle coords and add shift vector */
589 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
591 fix0 = _mm256_setzero_pd();
592 fiy0 = _mm256_setzero_pd();
593 fiz0 = _mm256_setzero_pd();
595 /* Load parameters for i particles */
596 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
597 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
599 /* Start inner kernel loop */
600 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
603 /* Get j neighbor index, and coordinate index */
608 j_coord_offsetA = DIM*jnrA;
609 j_coord_offsetB = DIM*jnrB;
610 j_coord_offsetC = DIM*jnrC;
611 j_coord_offsetD = DIM*jnrD;
613 /* load j atom coordinates */
614 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
615 x+j_coord_offsetC,x+j_coord_offsetD,
618 /* Calculate displacement vector */
619 dx00 = _mm256_sub_pd(ix0,jx0);
620 dy00 = _mm256_sub_pd(iy0,jy0);
621 dz00 = _mm256_sub_pd(iz0,jz0);
623 /* Calculate squared distance and things based on it */
624 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
626 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
628 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
630 /* Load parameters for j particles */
631 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
632 charge+jnrC+0,charge+jnrD+0);
633 vdwjidx0A = 2*vdwtype[jnrA+0];
634 vdwjidx0B = 2*vdwtype[jnrB+0];
635 vdwjidx0C = 2*vdwtype[jnrC+0];
636 vdwjidx0D = 2*vdwtype[jnrD+0];
638 /**************************
639 * CALCULATE INTERACTIONS *
640 **************************/
642 if (gmx_mm256_any_lt(rsq00,rcutoff2))
645 r00 = _mm256_mul_pd(rsq00,rinv00);
647 /* Compute parameters for interactions between i and j atoms */
648 qq00 = _mm256_mul_pd(iq0,jq0);
649 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
650 vdwioffsetptr0+vdwjidx0B,
651 vdwioffsetptr0+vdwjidx0C,
652 vdwioffsetptr0+vdwjidx0D,
655 /* Calculate table index by multiplying r with table scale and truncate to integer */
656 rt = _mm256_mul_pd(r00,vftabscale);
657 vfitab = _mm256_cvttpd_epi32(rt);
658 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
659 vfitab = _mm_slli_epi32(vfitab,3);
661 /* REACTION-FIELD ELECTROSTATICS */
662 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
664 /* CUBIC SPLINE TABLE DISPERSION */
665 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
666 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
667 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
668 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
669 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
670 Heps = _mm256_mul_pd(vfeps,H);
671 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
672 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
673 fvdw6 = _mm256_mul_pd(c6_00,FF);
675 /* CUBIC SPLINE TABLE REPULSION */
676 vfitab = _mm_add_epi32(vfitab,ifour);
677 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
678 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
679 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
680 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
681 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
682 Heps = _mm256_mul_pd(vfeps,H);
683 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
684 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
685 fvdw12 = _mm256_mul_pd(c12_00,FF);
686 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
688 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
690 fscal = _mm256_add_pd(felec,fvdw);
692 fscal = _mm256_and_pd(fscal,cutoff_mask);
694 /* Calculate temporary vectorial force */
695 tx = _mm256_mul_pd(fscal,dx00);
696 ty = _mm256_mul_pd(fscal,dy00);
697 tz = _mm256_mul_pd(fscal,dz00);
699 /* Update vectorial force */
700 fix0 = _mm256_add_pd(fix0,tx);
701 fiy0 = _mm256_add_pd(fiy0,ty);
702 fiz0 = _mm256_add_pd(fiz0,tz);
704 fjptrA = f+j_coord_offsetA;
705 fjptrB = f+j_coord_offsetB;
706 fjptrC = f+j_coord_offsetC;
707 fjptrD = f+j_coord_offsetD;
708 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
712 /* Inner loop uses 57 flops */
718 /* Get j neighbor index, and coordinate index */
719 jnrlistA = jjnr[jidx];
720 jnrlistB = jjnr[jidx+1];
721 jnrlistC = jjnr[jidx+2];
722 jnrlistD = jjnr[jidx+3];
723 /* Sign of each element will be negative for non-real atoms.
724 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
725 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
727 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
729 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
730 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
731 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
733 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
734 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
735 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
736 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
737 j_coord_offsetA = DIM*jnrA;
738 j_coord_offsetB = DIM*jnrB;
739 j_coord_offsetC = DIM*jnrC;
740 j_coord_offsetD = DIM*jnrD;
742 /* load j atom coordinates */
743 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
744 x+j_coord_offsetC,x+j_coord_offsetD,
747 /* Calculate displacement vector */
748 dx00 = _mm256_sub_pd(ix0,jx0);
749 dy00 = _mm256_sub_pd(iy0,jy0);
750 dz00 = _mm256_sub_pd(iz0,jz0);
752 /* Calculate squared distance and things based on it */
753 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
755 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
757 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
759 /* Load parameters for j particles */
760 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
761 charge+jnrC+0,charge+jnrD+0);
762 vdwjidx0A = 2*vdwtype[jnrA+0];
763 vdwjidx0B = 2*vdwtype[jnrB+0];
764 vdwjidx0C = 2*vdwtype[jnrC+0];
765 vdwjidx0D = 2*vdwtype[jnrD+0];
767 /**************************
768 * CALCULATE INTERACTIONS *
769 **************************/
771 if (gmx_mm256_any_lt(rsq00,rcutoff2))
774 r00 = _mm256_mul_pd(rsq00,rinv00);
775 r00 = _mm256_andnot_pd(dummy_mask,r00);
777 /* Compute parameters for interactions between i and j atoms */
778 qq00 = _mm256_mul_pd(iq0,jq0);
779 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
780 vdwioffsetptr0+vdwjidx0B,
781 vdwioffsetptr0+vdwjidx0C,
782 vdwioffsetptr0+vdwjidx0D,
785 /* Calculate table index by multiplying r with table scale and truncate to integer */
786 rt = _mm256_mul_pd(r00,vftabscale);
787 vfitab = _mm256_cvttpd_epi32(rt);
788 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
789 vfitab = _mm_slli_epi32(vfitab,3);
791 /* REACTION-FIELD ELECTROSTATICS */
792 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
794 /* CUBIC SPLINE TABLE DISPERSION */
795 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
796 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
797 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
798 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
799 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
800 Heps = _mm256_mul_pd(vfeps,H);
801 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
802 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
803 fvdw6 = _mm256_mul_pd(c6_00,FF);
805 /* CUBIC SPLINE TABLE REPULSION */
806 vfitab = _mm_add_epi32(vfitab,ifour);
807 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
808 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
809 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
810 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
811 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
812 Heps = _mm256_mul_pd(vfeps,H);
813 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
814 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
815 fvdw12 = _mm256_mul_pd(c12_00,FF);
816 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
818 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
820 fscal = _mm256_add_pd(felec,fvdw);
822 fscal = _mm256_and_pd(fscal,cutoff_mask);
824 fscal = _mm256_andnot_pd(dummy_mask,fscal);
826 /* Calculate temporary vectorial force */
827 tx = _mm256_mul_pd(fscal,dx00);
828 ty = _mm256_mul_pd(fscal,dy00);
829 tz = _mm256_mul_pd(fscal,dz00);
831 /* Update vectorial force */
832 fix0 = _mm256_add_pd(fix0,tx);
833 fiy0 = _mm256_add_pd(fiy0,ty);
834 fiz0 = _mm256_add_pd(fiz0,tz);
836 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
837 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
838 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
839 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
840 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
844 /* Inner loop uses 58 flops */
847 /* End of innermost loop */
849 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
850 f+i_coord_offset,fshift+i_shift_offset);
852 /* Increment number of inner iterations */
853 inneriter += j_index_end - j_index_start;
855 /* Outer loop uses 7 flops */
858 /* Increment number of outer iterations */
861 /* Update outer/inner flops */
863 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*58);