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.
44 #include "../nb_kernel.h"
45 #include "gromacs/legacyheaders/types/simple.h"
46 #include "gromacs/math/vec.h"
47 #include "gromacs/legacyheaders/nrnb.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_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_256_double
54 * Electrostatics interaction: ReactionField
55 * VdW interaction: CubicSplineTable
56 * Geometry: Particle-Particle
57 * Calculate force/pot: PotentialAndForce
60 nb_kernel_ElecRFCut_VdwCSTab_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 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
97 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
98 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
100 __m128i ifour = _mm_set1_epi32(4);
101 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
103 __m256d dummy_mask,cutoff_mask;
104 __m128 tmpmask0,tmpmask1;
105 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
106 __m256d one = _mm256_set1_pd(1.0);
107 __m256d two = _mm256_set1_pd(2.0);
113 jindex = nlist->jindex;
115 shiftidx = nlist->shift;
117 shiftvec = fr->shift_vec[0];
118 fshift = fr->fshift[0];
119 facel = _mm256_set1_pd(fr->epsfac);
120 charge = mdatoms->chargeA;
121 krf = _mm256_set1_pd(fr->ic->k_rf);
122 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
123 crf = _mm256_set1_pd(fr->ic->c_rf);
124 nvdwtype = fr->ntype;
126 vdwtype = mdatoms->typeA;
128 vftab = kernel_data->table_vdw->data;
129 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
131 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
132 rcutoff_scalar = fr->rcoulomb;
133 rcutoff = _mm256_set1_pd(rcutoff_scalar);
134 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
136 /* Avoid stupid compiler warnings */
137 jnrA = jnrB = jnrC = jnrD = 0;
146 for(iidx=0;iidx<4*DIM;iidx++)
151 /* Start outer loop over neighborlists */
152 for(iidx=0; iidx<nri; iidx++)
154 /* Load shift vector for this list */
155 i_shift_offset = DIM*shiftidx[iidx];
157 /* Load limits for loop over neighbors */
158 j_index_start = jindex[iidx];
159 j_index_end = jindex[iidx+1];
161 /* Get outer coordinate index */
163 i_coord_offset = DIM*inr;
165 /* Load i particle coords and add shift vector */
166 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
168 fix0 = _mm256_setzero_pd();
169 fiy0 = _mm256_setzero_pd();
170 fiz0 = _mm256_setzero_pd();
172 /* Load parameters for i particles */
173 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
174 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
176 /* Reset potential sums */
177 velecsum = _mm256_setzero_pd();
178 vvdwsum = _mm256_setzero_pd();
180 /* Start inner kernel loop */
181 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
184 /* Get j neighbor index, and coordinate index */
189 j_coord_offsetA = DIM*jnrA;
190 j_coord_offsetB = DIM*jnrB;
191 j_coord_offsetC = DIM*jnrC;
192 j_coord_offsetD = DIM*jnrD;
194 /* load j atom coordinates */
195 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
196 x+j_coord_offsetC,x+j_coord_offsetD,
199 /* Calculate displacement vector */
200 dx00 = _mm256_sub_pd(ix0,jx0);
201 dy00 = _mm256_sub_pd(iy0,jy0);
202 dz00 = _mm256_sub_pd(iz0,jz0);
204 /* Calculate squared distance and things based on it */
205 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
207 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
209 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
211 /* Load parameters for j particles */
212 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
213 charge+jnrC+0,charge+jnrD+0);
214 vdwjidx0A = 2*vdwtype[jnrA+0];
215 vdwjidx0B = 2*vdwtype[jnrB+0];
216 vdwjidx0C = 2*vdwtype[jnrC+0];
217 vdwjidx0D = 2*vdwtype[jnrD+0];
219 /**************************
220 * CALCULATE INTERACTIONS *
221 **************************/
223 if (gmx_mm256_any_lt(rsq00,rcutoff2))
226 r00 = _mm256_mul_pd(rsq00,rinv00);
228 /* Compute parameters for interactions between i and j atoms */
229 qq00 = _mm256_mul_pd(iq0,jq0);
230 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
231 vdwioffsetptr0+vdwjidx0B,
232 vdwioffsetptr0+vdwjidx0C,
233 vdwioffsetptr0+vdwjidx0D,
236 /* Calculate table index by multiplying r with table scale and truncate to integer */
237 rt = _mm256_mul_pd(r00,vftabscale);
238 vfitab = _mm256_cvttpd_epi32(rt);
239 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
240 vfitab = _mm_slli_epi32(vfitab,3);
242 /* REACTION-FIELD ELECTROSTATICS */
243 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
244 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
246 /* CUBIC SPLINE TABLE DISPERSION */
247 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
248 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
249 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
250 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
251 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
252 Heps = _mm256_mul_pd(vfeps,H);
253 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
254 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
255 vvdw6 = _mm256_mul_pd(c6_00,VV);
256 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
257 fvdw6 = _mm256_mul_pd(c6_00,FF);
259 /* CUBIC SPLINE TABLE REPULSION */
260 vfitab = _mm_add_epi32(vfitab,ifour);
261 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
262 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
263 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
264 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
265 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
266 Heps = _mm256_mul_pd(vfeps,H);
267 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
268 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
269 vvdw12 = _mm256_mul_pd(c12_00,VV);
270 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
271 fvdw12 = _mm256_mul_pd(c12_00,FF);
272 vvdw = _mm256_add_pd(vvdw12,vvdw6);
273 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
275 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
277 /* Update potential sum for this i atom from the interaction with this j atom. */
278 velec = _mm256_and_pd(velec,cutoff_mask);
279 velecsum = _mm256_add_pd(velecsum,velec);
280 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
281 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
283 fscal = _mm256_add_pd(felec,fvdw);
285 fscal = _mm256_and_pd(fscal,cutoff_mask);
287 /* Calculate temporary vectorial force */
288 tx = _mm256_mul_pd(fscal,dx00);
289 ty = _mm256_mul_pd(fscal,dy00);
290 tz = _mm256_mul_pd(fscal,dz00);
292 /* Update vectorial force */
293 fix0 = _mm256_add_pd(fix0,tx);
294 fiy0 = _mm256_add_pd(fiy0,ty);
295 fiz0 = _mm256_add_pd(fiz0,tz);
297 fjptrA = f+j_coord_offsetA;
298 fjptrB = f+j_coord_offsetB;
299 fjptrC = f+j_coord_offsetC;
300 fjptrD = f+j_coord_offsetD;
301 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
305 /* Inner loop uses 72 flops */
311 /* Get j neighbor index, and coordinate index */
312 jnrlistA = jjnr[jidx];
313 jnrlistB = jjnr[jidx+1];
314 jnrlistC = jjnr[jidx+2];
315 jnrlistD = jjnr[jidx+3];
316 /* Sign of each element will be negative for non-real atoms.
317 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
318 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
320 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
322 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
323 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
324 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
326 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
327 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
328 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
329 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
330 j_coord_offsetA = DIM*jnrA;
331 j_coord_offsetB = DIM*jnrB;
332 j_coord_offsetC = DIM*jnrC;
333 j_coord_offsetD = DIM*jnrD;
335 /* load j atom coordinates */
336 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
337 x+j_coord_offsetC,x+j_coord_offsetD,
340 /* Calculate displacement vector */
341 dx00 = _mm256_sub_pd(ix0,jx0);
342 dy00 = _mm256_sub_pd(iy0,jy0);
343 dz00 = _mm256_sub_pd(iz0,jz0);
345 /* Calculate squared distance and things based on it */
346 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
348 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
350 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
352 /* Load parameters for j particles */
353 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
354 charge+jnrC+0,charge+jnrD+0);
355 vdwjidx0A = 2*vdwtype[jnrA+0];
356 vdwjidx0B = 2*vdwtype[jnrB+0];
357 vdwjidx0C = 2*vdwtype[jnrC+0];
358 vdwjidx0D = 2*vdwtype[jnrD+0];
360 /**************************
361 * CALCULATE INTERACTIONS *
362 **************************/
364 if (gmx_mm256_any_lt(rsq00,rcutoff2))
367 r00 = _mm256_mul_pd(rsq00,rinv00);
368 r00 = _mm256_andnot_pd(dummy_mask,r00);
370 /* Compute parameters for interactions between i and j atoms */
371 qq00 = _mm256_mul_pd(iq0,jq0);
372 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
373 vdwioffsetptr0+vdwjidx0B,
374 vdwioffsetptr0+vdwjidx0C,
375 vdwioffsetptr0+vdwjidx0D,
378 /* Calculate table index by multiplying r with table scale and truncate to integer */
379 rt = _mm256_mul_pd(r00,vftabscale);
380 vfitab = _mm256_cvttpd_epi32(rt);
381 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
382 vfitab = _mm_slli_epi32(vfitab,3);
384 /* REACTION-FIELD ELECTROSTATICS */
385 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_add_pd(rinv00,_mm256_mul_pd(krf,rsq00)),crf));
386 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
388 /* CUBIC SPLINE TABLE DISPERSION */
389 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
390 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
391 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
392 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
393 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
394 Heps = _mm256_mul_pd(vfeps,H);
395 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
396 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
397 vvdw6 = _mm256_mul_pd(c6_00,VV);
398 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
399 fvdw6 = _mm256_mul_pd(c6_00,FF);
401 /* CUBIC SPLINE TABLE REPULSION */
402 vfitab = _mm_add_epi32(vfitab,ifour);
403 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
404 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
405 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
406 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
407 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
408 Heps = _mm256_mul_pd(vfeps,H);
409 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
410 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
411 vvdw12 = _mm256_mul_pd(c12_00,VV);
412 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
413 fvdw12 = _mm256_mul_pd(c12_00,FF);
414 vvdw = _mm256_add_pd(vvdw12,vvdw6);
415 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
417 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
419 /* Update potential sum for this i atom from the interaction with this j atom. */
420 velec = _mm256_and_pd(velec,cutoff_mask);
421 velec = _mm256_andnot_pd(dummy_mask,velec);
422 velecsum = _mm256_add_pd(velecsum,velec);
423 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
424 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
425 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
427 fscal = _mm256_add_pd(felec,fvdw);
429 fscal = _mm256_and_pd(fscal,cutoff_mask);
431 fscal = _mm256_andnot_pd(dummy_mask,fscal);
433 /* Calculate temporary vectorial force */
434 tx = _mm256_mul_pd(fscal,dx00);
435 ty = _mm256_mul_pd(fscal,dy00);
436 tz = _mm256_mul_pd(fscal,dz00);
438 /* Update vectorial force */
439 fix0 = _mm256_add_pd(fix0,tx);
440 fiy0 = _mm256_add_pd(fiy0,ty);
441 fiz0 = _mm256_add_pd(fiz0,tz);
443 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
444 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
445 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
446 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
447 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
451 /* Inner loop uses 73 flops */
454 /* End of innermost loop */
456 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
457 f+i_coord_offset,fshift+i_shift_offset);
460 /* Update potential energies */
461 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
462 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
464 /* Increment number of inner iterations */
465 inneriter += j_index_end - j_index_start;
467 /* Outer loop uses 9 flops */
470 /* Increment number of outer iterations */
473 /* Update outer/inner flops */
475 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*73);
478 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_256_double
479 * Electrostatics interaction: ReactionField
480 * VdW interaction: CubicSplineTable
481 * Geometry: Particle-Particle
482 * Calculate force/pot: Force
485 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_256_double
486 (t_nblist * gmx_restrict nlist,
487 rvec * gmx_restrict xx,
488 rvec * gmx_restrict ff,
489 t_forcerec * gmx_restrict fr,
490 t_mdatoms * gmx_restrict mdatoms,
491 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
492 t_nrnb * gmx_restrict nrnb)
494 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
495 * just 0 for non-waters.
496 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
497 * jnr indices corresponding to data put in the four positions in the SIMD register.
499 int i_shift_offset,i_coord_offset,outeriter,inneriter;
500 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
501 int jnrA,jnrB,jnrC,jnrD;
502 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
503 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
504 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
505 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
507 real *shiftvec,*fshift,*x,*f;
508 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
510 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
511 real * vdwioffsetptr0;
512 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
513 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
514 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
515 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
516 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
519 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
522 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
523 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
525 __m128i ifour = _mm_set1_epi32(4);
526 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
528 __m256d dummy_mask,cutoff_mask;
529 __m128 tmpmask0,tmpmask1;
530 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
531 __m256d one = _mm256_set1_pd(1.0);
532 __m256d two = _mm256_set1_pd(2.0);
538 jindex = nlist->jindex;
540 shiftidx = nlist->shift;
542 shiftvec = fr->shift_vec[0];
543 fshift = fr->fshift[0];
544 facel = _mm256_set1_pd(fr->epsfac);
545 charge = mdatoms->chargeA;
546 krf = _mm256_set1_pd(fr->ic->k_rf);
547 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
548 crf = _mm256_set1_pd(fr->ic->c_rf);
549 nvdwtype = fr->ntype;
551 vdwtype = mdatoms->typeA;
553 vftab = kernel_data->table_vdw->data;
554 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
556 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
557 rcutoff_scalar = fr->rcoulomb;
558 rcutoff = _mm256_set1_pd(rcutoff_scalar);
559 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
561 /* Avoid stupid compiler warnings */
562 jnrA = jnrB = jnrC = jnrD = 0;
571 for(iidx=0;iidx<4*DIM;iidx++)
576 /* Start outer loop over neighborlists */
577 for(iidx=0; iidx<nri; iidx++)
579 /* Load shift vector for this list */
580 i_shift_offset = DIM*shiftidx[iidx];
582 /* Load limits for loop over neighbors */
583 j_index_start = jindex[iidx];
584 j_index_end = jindex[iidx+1];
586 /* Get outer coordinate index */
588 i_coord_offset = DIM*inr;
590 /* Load i particle coords and add shift vector */
591 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
593 fix0 = _mm256_setzero_pd();
594 fiy0 = _mm256_setzero_pd();
595 fiz0 = _mm256_setzero_pd();
597 /* Load parameters for i particles */
598 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
599 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
601 /* Start inner kernel loop */
602 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
605 /* Get j neighbor index, and coordinate index */
610 j_coord_offsetA = DIM*jnrA;
611 j_coord_offsetB = DIM*jnrB;
612 j_coord_offsetC = DIM*jnrC;
613 j_coord_offsetD = DIM*jnrD;
615 /* load j atom coordinates */
616 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
617 x+j_coord_offsetC,x+j_coord_offsetD,
620 /* Calculate displacement vector */
621 dx00 = _mm256_sub_pd(ix0,jx0);
622 dy00 = _mm256_sub_pd(iy0,jy0);
623 dz00 = _mm256_sub_pd(iz0,jz0);
625 /* Calculate squared distance and things based on it */
626 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
628 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
630 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
632 /* Load parameters for j particles */
633 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
634 charge+jnrC+0,charge+jnrD+0);
635 vdwjidx0A = 2*vdwtype[jnrA+0];
636 vdwjidx0B = 2*vdwtype[jnrB+0];
637 vdwjidx0C = 2*vdwtype[jnrC+0];
638 vdwjidx0D = 2*vdwtype[jnrD+0];
640 /**************************
641 * CALCULATE INTERACTIONS *
642 **************************/
644 if (gmx_mm256_any_lt(rsq00,rcutoff2))
647 r00 = _mm256_mul_pd(rsq00,rinv00);
649 /* Compute parameters for interactions between i and j atoms */
650 qq00 = _mm256_mul_pd(iq0,jq0);
651 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
652 vdwioffsetptr0+vdwjidx0B,
653 vdwioffsetptr0+vdwjidx0C,
654 vdwioffsetptr0+vdwjidx0D,
657 /* Calculate table index by multiplying r with table scale and truncate to integer */
658 rt = _mm256_mul_pd(r00,vftabscale);
659 vfitab = _mm256_cvttpd_epi32(rt);
660 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
661 vfitab = _mm_slli_epi32(vfitab,3);
663 /* REACTION-FIELD ELECTROSTATICS */
664 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
666 /* CUBIC SPLINE TABLE DISPERSION */
667 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
668 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
669 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
670 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
671 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
672 Heps = _mm256_mul_pd(vfeps,H);
673 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
674 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
675 fvdw6 = _mm256_mul_pd(c6_00,FF);
677 /* CUBIC SPLINE TABLE REPULSION */
678 vfitab = _mm_add_epi32(vfitab,ifour);
679 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
680 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
681 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
682 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
683 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
684 Heps = _mm256_mul_pd(vfeps,H);
685 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
686 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
687 fvdw12 = _mm256_mul_pd(c12_00,FF);
688 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
690 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
692 fscal = _mm256_add_pd(felec,fvdw);
694 fscal = _mm256_and_pd(fscal,cutoff_mask);
696 /* Calculate temporary vectorial force */
697 tx = _mm256_mul_pd(fscal,dx00);
698 ty = _mm256_mul_pd(fscal,dy00);
699 tz = _mm256_mul_pd(fscal,dz00);
701 /* Update vectorial force */
702 fix0 = _mm256_add_pd(fix0,tx);
703 fiy0 = _mm256_add_pd(fiy0,ty);
704 fiz0 = _mm256_add_pd(fiz0,tz);
706 fjptrA = f+j_coord_offsetA;
707 fjptrB = f+j_coord_offsetB;
708 fjptrC = f+j_coord_offsetC;
709 fjptrD = f+j_coord_offsetD;
710 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
714 /* Inner loop uses 57 flops */
720 /* Get j neighbor index, and coordinate index */
721 jnrlistA = jjnr[jidx];
722 jnrlistB = jjnr[jidx+1];
723 jnrlistC = jjnr[jidx+2];
724 jnrlistD = jjnr[jidx+3];
725 /* Sign of each element will be negative for non-real atoms.
726 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
727 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
729 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
731 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
732 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
733 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
735 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
736 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
737 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
738 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
739 j_coord_offsetA = DIM*jnrA;
740 j_coord_offsetB = DIM*jnrB;
741 j_coord_offsetC = DIM*jnrC;
742 j_coord_offsetD = DIM*jnrD;
744 /* load j atom coordinates */
745 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
746 x+j_coord_offsetC,x+j_coord_offsetD,
749 /* Calculate displacement vector */
750 dx00 = _mm256_sub_pd(ix0,jx0);
751 dy00 = _mm256_sub_pd(iy0,jy0);
752 dz00 = _mm256_sub_pd(iz0,jz0);
754 /* Calculate squared distance and things based on it */
755 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
757 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
759 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
761 /* Load parameters for j particles */
762 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
763 charge+jnrC+0,charge+jnrD+0);
764 vdwjidx0A = 2*vdwtype[jnrA+0];
765 vdwjidx0B = 2*vdwtype[jnrB+0];
766 vdwjidx0C = 2*vdwtype[jnrC+0];
767 vdwjidx0D = 2*vdwtype[jnrD+0];
769 /**************************
770 * CALCULATE INTERACTIONS *
771 **************************/
773 if (gmx_mm256_any_lt(rsq00,rcutoff2))
776 r00 = _mm256_mul_pd(rsq00,rinv00);
777 r00 = _mm256_andnot_pd(dummy_mask,r00);
779 /* Compute parameters for interactions between i and j atoms */
780 qq00 = _mm256_mul_pd(iq0,jq0);
781 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
782 vdwioffsetptr0+vdwjidx0B,
783 vdwioffsetptr0+vdwjidx0C,
784 vdwioffsetptr0+vdwjidx0D,
787 /* Calculate table index by multiplying r with table scale and truncate to integer */
788 rt = _mm256_mul_pd(r00,vftabscale);
789 vfitab = _mm256_cvttpd_epi32(rt);
790 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
791 vfitab = _mm_slli_epi32(vfitab,3);
793 /* REACTION-FIELD ELECTROSTATICS */
794 felec = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_mul_pd(rinv00,rinvsq00),krf2));
796 /* CUBIC SPLINE TABLE DISPERSION */
797 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
798 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
799 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
800 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
801 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
802 Heps = _mm256_mul_pd(vfeps,H);
803 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
804 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
805 fvdw6 = _mm256_mul_pd(c6_00,FF);
807 /* CUBIC SPLINE TABLE REPULSION */
808 vfitab = _mm_add_epi32(vfitab,ifour);
809 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
810 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
811 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
812 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
813 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
814 Heps = _mm256_mul_pd(vfeps,H);
815 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
816 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
817 fvdw12 = _mm256_mul_pd(c12_00,FF);
818 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
820 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
822 fscal = _mm256_add_pd(felec,fvdw);
824 fscal = _mm256_and_pd(fscal,cutoff_mask);
826 fscal = _mm256_andnot_pd(dummy_mask,fscal);
828 /* Calculate temporary vectorial force */
829 tx = _mm256_mul_pd(fscal,dx00);
830 ty = _mm256_mul_pd(fscal,dy00);
831 tz = _mm256_mul_pd(fscal,dz00);
833 /* Update vectorial force */
834 fix0 = _mm256_add_pd(fix0,tx);
835 fiy0 = _mm256_add_pd(fiy0,ty);
836 fiz0 = _mm256_add_pd(fiz0,tz);
838 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
839 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
840 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
841 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
842 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
846 /* Inner loop uses 58 flops */
849 /* End of innermost loop */
851 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
852 f+i_coord_offset,fshift+i_shift_offset);
854 /* Increment number of inner iterations */
855 inneriter += j_index_end - j_index_start;
857 /* Outer loop uses 7 flops */
860 /* Increment number of outer iterations */
863 /* Update outer/inner flops */
865 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*58);