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_ElecEw_VdwCSTab_GeomP1P1_VF_avx_256_double
54 * Electrostatics interaction: Ewald
55 * VdW interaction: CubicSplineTable
56 * Geometry: Particle-Particle
57 * Calculate force/pot: PotentialAndForce
60 nb_kernel_ElecEw_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;
104 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
105 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
107 __m256d dummy_mask,cutoff_mask;
108 __m128 tmpmask0,tmpmask1;
109 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
110 __m256d one = _mm256_set1_pd(1.0);
111 __m256d two = _mm256_set1_pd(2.0);
117 jindex = nlist->jindex;
119 shiftidx = nlist->shift;
121 shiftvec = fr->shift_vec[0];
122 fshift = fr->fshift[0];
123 facel = _mm256_set1_pd(fr->epsfac);
124 charge = mdatoms->chargeA;
125 nvdwtype = fr->ntype;
127 vdwtype = mdatoms->typeA;
129 vftab = kernel_data->table_vdw->data;
130 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
132 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
133 beta = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
134 beta2 = _mm256_mul_pd(beta,beta);
135 beta3 = _mm256_mul_pd(beta,beta2);
137 ewtab = fr->ic->tabq_coul_FDV0;
138 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
139 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
141 /* Avoid stupid compiler warnings */
142 jnrA = jnrB = jnrC = jnrD = 0;
151 for(iidx=0;iidx<4*DIM;iidx++)
156 /* Start outer loop over neighborlists */
157 for(iidx=0; iidx<nri; iidx++)
159 /* Load shift vector for this list */
160 i_shift_offset = DIM*shiftidx[iidx];
162 /* Load limits for loop over neighbors */
163 j_index_start = jindex[iidx];
164 j_index_end = jindex[iidx+1];
166 /* Get outer coordinate index */
168 i_coord_offset = DIM*inr;
170 /* Load i particle coords and add shift vector */
171 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
173 fix0 = _mm256_setzero_pd();
174 fiy0 = _mm256_setzero_pd();
175 fiz0 = _mm256_setzero_pd();
177 /* Load parameters for i particles */
178 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
179 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
181 /* Reset potential sums */
182 velecsum = _mm256_setzero_pd();
183 vvdwsum = _mm256_setzero_pd();
185 /* Start inner kernel loop */
186 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
189 /* Get j neighbor index, and coordinate index */
194 j_coord_offsetA = DIM*jnrA;
195 j_coord_offsetB = DIM*jnrB;
196 j_coord_offsetC = DIM*jnrC;
197 j_coord_offsetD = DIM*jnrD;
199 /* load j atom coordinates */
200 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
201 x+j_coord_offsetC,x+j_coord_offsetD,
204 /* Calculate displacement vector */
205 dx00 = _mm256_sub_pd(ix0,jx0);
206 dy00 = _mm256_sub_pd(iy0,jy0);
207 dz00 = _mm256_sub_pd(iz0,jz0);
209 /* Calculate squared distance and things based on it */
210 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
212 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
214 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
216 /* Load parameters for j particles */
217 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
218 charge+jnrC+0,charge+jnrD+0);
219 vdwjidx0A = 2*vdwtype[jnrA+0];
220 vdwjidx0B = 2*vdwtype[jnrB+0];
221 vdwjidx0C = 2*vdwtype[jnrC+0];
222 vdwjidx0D = 2*vdwtype[jnrD+0];
224 /**************************
225 * CALCULATE INTERACTIONS *
226 **************************/
228 r00 = _mm256_mul_pd(rsq00,rinv00);
230 /* Compute parameters for interactions between i and j atoms */
231 qq00 = _mm256_mul_pd(iq0,jq0);
232 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
233 vdwioffsetptr0+vdwjidx0B,
234 vdwioffsetptr0+vdwjidx0C,
235 vdwioffsetptr0+vdwjidx0D,
238 /* Calculate table index by multiplying r with table scale and truncate to integer */
239 rt = _mm256_mul_pd(r00,vftabscale);
240 vfitab = _mm256_cvttpd_epi32(rt);
241 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
242 vfitab = _mm_slli_epi32(vfitab,3);
244 /* EWALD ELECTROSTATICS */
246 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
247 ewrt = _mm256_mul_pd(r00,ewtabscale);
248 ewitab = _mm256_cvttpd_epi32(ewrt);
249 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
250 ewitab = _mm_slli_epi32(ewitab,2);
251 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
252 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
253 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
254 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
255 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
256 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
257 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
258 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
259 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
261 /* CUBIC SPLINE TABLE DISPERSION */
262 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
263 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
264 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
265 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
266 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
267 Heps = _mm256_mul_pd(vfeps,H);
268 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
269 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
270 vvdw6 = _mm256_mul_pd(c6_00,VV);
271 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
272 fvdw6 = _mm256_mul_pd(c6_00,FF);
274 /* CUBIC SPLINE TABLE REPULSION */
275 vfitab = _mm_add_epi32(vfitab,ifour);
276 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
277 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
278 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
279 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
280 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
281 Heps = _mm256_mul_pd(vfeps,H);
282 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
283 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
284 vvdw12 = _mm256_mul_pd(c12_00,VV);
285 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
286 fvdw12 = _mm256_mul_pd(c12_00,FF);
287 vvdw = _mm256_add_pd(vvdw12,vvdw6);
288 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
290 /* Update potential sum for this i atom from the interaction with this j atom. */
291 velecsum = _mm256_add_pd(velecsum,velec);
292 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
294 fscal = _mm256_add_pd(felec,fvdw);
296 /* Calculate temporary vectorial force */
297 tx = _mm256_mul_pd(fscal,dx00);
298 ty = _mm256_mul_pd(fscal,dy00);
299 tz = _mm256_mul_pd(fscal,dz00);
301 /* Update vectorial force */
302 fix0 = _mm256_add_pd(fix0,tx);
303 fiy0 = _mm256_add_pd(fiy0,ty);
304 fiz0 = _mm256_add_pd(fiz0,tz);
306 fjptrA = f+j_coord_offsetA;
307 fjptrB = f+j_coord_offsetB;
308 fjptrC = f+j_coord_offsetC;
309 fjptrD = f+j_coord_offsetD;
310 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
312 /* Inner loop uses 75 flops */
318 /* Get j neighbor index, and coordinate index */
319 jnrlistA = jjnr[jidx];
320 jnrlistB = jjnr[jidx+1];
321 jnrlistC = jjnr[jidx+2];
322 jnrlistD = jjnr[jidx+3];
323 /* Sign of each element will be negative for non-real atoms.
324 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
325 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
327 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
329 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
330 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
331 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
333 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
334 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
335 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
336 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
337 j_coord_offsetA = DIM*jnrA;
338 j_coord_offsetB = DIM*jnrB;
339 j_coord_offsetC = DIM*jnrC;
340 j_coord_offsetD = DIM*jnrD;
342 /* load j atom coordinates */
343 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
344 x+j_coord_offsetC,x+j_coord_offsetD,
347 /* Calculate displacement vector */
348 dx00 = _mm256_sub_pd(ix0,jx0);
349 dy00 = _mm256_sub_pd(iy0,jy0);
350 dz00 = _mm256_sub_pd(iz0,jz0);
352 /* Calculate squared distance and things based on it */
353 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
355 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
357 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
359 /* Load parameters for j particles */
360 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
361 charge+jnrC+0,charge+jnrD+0);
362 vdwjidx0A = 2*vdwtype[jnrA+0];
363 vdwjidx0B = 2*vdwtype[jnrB+0];
364 vdwjidx0C = 2*vdwtype[jnrC+0];
365 vdwjidx0D = 2*vdwtype[jnrD+0];
367 /**************************
368 * CALCULATE INTERACTIONS *
369 **************************/
371 r00 = _mm256_mul_pd(rsq00,rinv00);
372 r00 = _mm256_andnot_pd(dummy_mask,r00);
374 /* Compute parameters for interactions between i and j atoms */
375 qq00 = _mm256_mul_pd(iq0,jq0);
376 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
377 vdwioffsetptr0+vdwjidx0B,
378 vdwioffsetptr0+vdwjidx0C,
379 vdwioffsetptr0+vdwjidx0D,
382 /* Calculate table index by multiplying r with table scale and truncate to integer */
383 rt = _mm256_mul_pd(r00,vftabscale);
384 vfitab = _mm256_cvttpd_epi32(rt);
385 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
386 vfitab = _mm_slli_epi32(vfitab,3);
388 /* EWALD ELECTROSTATICS */
390 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
391 ewrt = _mm256_mul_pd(r00,ewtabscale);
392 ewitab = _mm256_cvttpd_epi32(ewrt);
393 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
394 ewitab = _mm_slli_epi32(ewitab,2);
395 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
396 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
397 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
398 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
399 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
400 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
401 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
402 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
403 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
405 /* CUBIC SPLINE TABLE DISPERSION */
406 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
407 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
408 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
409 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
410 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
411 Heps = _mm256_mul_pd(vfeps,H);
412 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
413 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
414 vvdw6 = _mm256_mul_pd(c6_00,VV);
415 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
416 fvdw6 = _mm256_mul_pd(c6_00,FF);
418 /* CUBIC SPLINE TABLE REPULSION */
419 vfitab = _mm_add_epi32(vfitab,ifour);
420 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
421 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
422 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
423 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
424 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
425 Heps = _mm256_mul_pd(vfeps,H);
426 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
427 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
428 vvdw12 = _mm256_mul_pd(c12_00,VV);
429 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
430 fvdw12 = _mm256_mul_pd(c12_00,FF);
431 vvdw = _mm256_add_pd(vvdw12,vvdw6);
432 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
434 /* Update potential sum for this i atom from the interaction with this j atom. */
435 velec = _mm256_andnot_pd(dummy_mask,velec);
436 velecsum = _mm256_add_pd(velecsum,velec);
437 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
438 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
440 fscal = _mm256_add_pd(felec,fvdw);
442 fscal = _mm256_andnot_pd(dummy_mask,fscal);
444 /* Calculate temporary vectorial force */
445 tx = _mm256_mul_pd(fscal,dx00);
446 ty = _mm256_mul_pd(fscal,dy00);
447 tz = _mm256_mul_pd(fscal,dz00);
449 /* Update vectorial force */
450 fix0 = _mm256_add_pd(fix0,tx);
451 fiy0 = _mm256_add_pd(fiy0,ty);
452 fiz0 = _mm256_add_pd(fiz0,tz);
454 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
455 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
456 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
457 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
458 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
460 /* Inner loop uses 76 flops */
463 /* End of innermost loop */
465 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
466 f+i_coord_offset,fshift+i_shift_offset);
469 /* Update potential energies */
470 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
471 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
473 /* Increment number of inner iterations */
474 inneriter += j_index_end - j_index_start;
476 /* Outer loop uses 9 flops */
479 /* Increment number of outer iterations */
482 /* Update outer/inner flops */
484 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*76);
487 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_256_double
488 * Electrostatics interaction: Ewald
489 * VdW interaction: CubicSplineTable
490 * Geometry: Particle-Particle
491 * Calculate force/pot: Force
494 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_256_double
495 (t_nblist * gmx_restrict nlist,
496 rvec * gmx_restrict xx,
497 rvec * gmx_restrict ff,
498 t_forcerec * gmx_restrict fr,
499 t_mdatoms * gmx_restrict mdatoms,
500 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
501 t_nrnb * gmx_restrict nrnb)
503 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
504 * just 0 for non-waters.
505 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
506 * jnr indices corresponding to data put in the four positions in the SIMD register.
508 int i_shift_offset,i_coord_offset,outeriter,inneriter;
509 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
510 int jnrA,jnrB,jnrC,jnrD;
511 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
512 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
513 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
514 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
516 real *shiftvec,*fshift,*x,*f;
517 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
519 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
520 real * vdwioffsetptr0;
521 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
522 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
523 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
524 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
525 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
528 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
531 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
532 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
534 __m128i ifour = _mm_set1_epi32(4);
535 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
538 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
539 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
541 __m256d dummy_mask,cutoff_mask;
542 __m128 tmpmask0,tmpmask1;
543 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
544 __m256d one = _mm256_set1_pd(1.0);
545 __m256d two = _mm256_set1_pd(2.0);
551 jindex = nlist->jindex;
553 shiftidx = nlist->shift;
555 shiftvec = fr->shift_vec[0];
556 fshift = fr->fshift[0];
557 facel = _mm256_set1_pd(fr->epsfac);
558 charge = mdatoms->chargeA;
559 nvdwtype = fr->ntype;
561 vdwtype = mdatoms->typeA;
563 vftab = kernel_data->table_vdw->data;
564 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
566 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
567 beta = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
568 beta2 = _mm256_mul_pd(beta,beta);
569 beta3 = _mm256_mul_pd(beta,beta2);
571 ewtab = fr->ic->tabq_coul_F;
572 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
573 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
575 /* Avoid stupid compiler warnings */
576 jnrA = jnrB = jnrC = jnrD = 0;
585 for(iidx=0;iidx<4*DIM;iidx++)
590 /* Start outer loop over neighborlists */
591 for(iidx=0; iidx<nri; iidx++)
593 /* Load shift vector for this list */
594 i_shift_offset = DIM*shiftidx[iidx];
596 /* Load limits for loop over neighbors */
597 j_index_start = jindex[iidx];
598 j_index_end = jindex[iidx+1];
600 /* Get outer coordinate index */
602 i_coord_offset = DIM*inr;
604 /* Load i particle coords and add shift vector */
605 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
607 fix0 = _mm256_setzero_pd();
608 fiy0 = _mm256_setzero_pd();
609 fiz0 = _mm256_setzero_pd();
611 /* Load parameters for i particles */
612 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
613 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
615 /* Start inner kernel loop */
616 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
619 /* Get j neighbor index, and coordinate index */
624 j_coord_offsetA = DIM*jnrA;
625 j_coord_offsetB = DIM*jnrB;
626 j_coord_offsetC = DIM*jnrC;
627 j_coord_offsetD = DIM*jnrD;
629 /* load j atom coordinates */
630 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
631 x+j_coord_offsetC,x+j_coord_offsetD,
634 /* Calculate displacement vector */
635 dx00 = _mm256_sub_pd(ix0,jx0);
636 dy00 = _mm256_sub_pd(iy0,jy0);
637 dz00 = _mm256_sub_pd(iz0,jz0);
639 /* Calculate squared distance and things based on it */
640 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
642 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
644 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
646 /* Load parameters for j particles */
647 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
648 charge+jnrC+0,charge+jnrD+0);
649 vdwjidx0A = 2*vdwtype[jnrA+0];
650 vdwjidx0B = 2*vdwtype[jnrB+0];
651 vdwjidx0C = 2*vdwtype[jnrC+0];
652 vdwjidx0D = 2*vdwtype[jnrD+0];
654 /**************************
655 * CALCULATE INTERACTIONS *
656 **************************/
658 r00 = _mm256_mul_pd(rsq00,rinv00);
660 /* Compute parameters for interactions between i and j atoms */
661 qq00 = _mm256_mul_pd(iq0,jq0);
662 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
663 vdwioffsetptr0+vdwjidx0B,
664 vdwioffsetptr0+vdwjidx0C,
665 vdwioffsetptr0+vdwjidx0D,
668 /* Calculate table index by multiplying r with table scale and truncate to integer */
669 rt = _mm256_mul_pd(r00,vftabscale);
670 vfitab = _mm256_cvttpd_epi32(rt);
671 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
672 vfitab = _mm_slli_epi32(vfitab,3);
674 /* EWALD ELECTROSTATICS */
676 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
677 ewrt = _mm256_mul_pd(r00,ewtabscale);
678 ewitab = _mm256_cvttpd_epi32(ewrt);
679 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
680 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
681 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
683 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
684 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
686 /* CUBIC SPLINE TABLE DISPERSION */
687 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
688 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
689 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
690 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
691 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
692 Heps = _mm256_mul_pd(vfeps,H);
693 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
694 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
695 fvdw6 = _mm256_mul_pd(c6_00,FF);
697 /* CUBIC SPLINE TABLE REPULSION */
698 vfitab = _mm_add_epi32(vfitab,ifour);
699 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
700 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
701 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
702 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
703 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
704 Heps = _mm256_mul_pd(vfeps,H);
705 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
706 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
707 fvdw12 = _mm256_mul_pd(c12_00,FF);
708 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
710 fscal = _mm256_add_pd(felec,fvdw);
712 /* Calculate temporary vectorial force */
713 tx = _mm256_mul_pd(fscal,dx00);
714 ty = _mm256_mul_pd(fscal,dy00);
715 tz = _mm256_mul_pd(fscal,dz00);
717 /* Update vectorial force */
718 fix0 = _mm256_add_pd(fix0,tx);
719 fiy0 = _mm256_add_pd(fiy0,ty);
720 fiz0 = _mm256_add_pd(fiz0,tz);
722 fjptrA = f+j_coord_offsetA;
723 fjptrB = f+j_coord_offsetB;
724 fjptrC = f+j_coord_offsetC;
725 fjptrD = f+j_coord_offsetD;
726 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
728 /* Inner loop uses 62 flops */
734 /* Get j neighbor index, and coordinate index */
735 jnrlistA = jjnr[jidx];
736 jnrlistB = jjnr[jidx+1];
737 jnrlistC = jjnr[jidx+2];
738 jnrlistD = jjnr[jidx+3];
739 /* Sign of each element will be negative for non-real atoms.
740 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
741 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
743 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
745 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
746 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
747 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
749 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
750 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
751 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
752 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
753 j_coord_offsetA = DIM*jnrA;
754 j_coord_offsetB = DIM*jnrB;
755 j_coord_offsetC = DIM*jnrC;
756 j_coord_offsetD = DIM*jnrD;
758 /* load j atom coordinates */
759 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
760 x+j_coord_offsetC,x+j_coord_offsetD,
763 /* Calculate displacement vector */
764 dx00 = _mm256_sub_pd(ix0,jx0);
765 dy00 = _mm256_sub_pd(iy0,jy0);
766 dz00 = _mm256_sub_pd(iz0,jz0);
768 /* Calculate squared distance and things based on it */
769 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
771 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
773 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
775 /* Load parameters for j particles */
776 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
777 charge+jnrC+0,charge+jnrD+0);
778 vdwjidx0A = 2*vdwtype[jnrA+0];
779 vdwjidx0B = 2*vdwtype[jnrB+0];
780 vdwjidx0C = 2*vdwtype[jnrC+0];
781 vdwjidx0D = 2*vdwtype[jnrD+0];
783 /**************************
784 * CALCULATE INTERACTIONS *
785 **************************/
787 r00 = _mm256_mul_pd(rsq00,rinv00);
788 r00 = _mm256_andnot_pd(dummy_mask,r00);
790 /* Compute parameters for interactions between i and j atoms */
791 qq00 = _mm256_mul_pd(iq0,jq0);
792 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
793 vdwioffsetptr0+vdwjidx0B,
794 vdwioffsetptr0+vdwjidx0C,
795 vdwioffsetptr0+vdwjidx0D,
798 /* Calculate table index by multiplying r with table scale and truncate to integer */
799 rt = _mm256_mul_pd(r00,vftabscale);
800 vfitab = _mm256_cvttpd_epi32(rt);
801 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
802 vfitab = _mm_slli_epi32(vfitab,3);
804 /* EWALD ELECTROSTATICS */
806 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
807 ewrt = _mm256_mul_pd(r00,ewtabscale);
808 ewitab = _mm256_cvttpd_epi32(ewrt);
809 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
810 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
811 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
813 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
814 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
816 /* CUBIC SPLINE TABLE DISPERSION */
817 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
818 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
819 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
820 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
821 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
822 Heps = _mm256_mul_pd(vfeps,H);
823 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
824 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
825 fvdw6 = _mm256_mul_pd(c6_00,FF);
827 /* CUBIC SPLINE TABLE REPULSION */
828 vfitab = _mm_add_epi32(vfitab,ifour);
829 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
830 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
831 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
832 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
833 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
834 Heps = _mm256_mul_pd(vfeps,H);
835 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
836 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
837 fvdw12 = _mm256_mul_pd(c12_00,FF);
838 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
840 fscal = _mm256_add_pd(felec,fvdw);
842 fscal = _mm256_andnot_pd(dummy_mask,fscal);
844 /* Calculate temporary vectorial force */
845 tx = _mm256_mul_pd(fscal,dx00);
846 ty = _mm256_mul_pd(fscal,dy00);
847 tz = _mm256_mul_pd(fscal,dz00);
849 /* Update vectorial force */
850 fix0 = _mm256_add_pd(fix0,tx);
851 fiy0 = _mm256_add_pd(fiy0,ty);
852 fiz0 = _mm256_add_pd(fiz0,tz);
854 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
855 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
856 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
857 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
858 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
860 /* Inner loop uses 63 flops */
863 /* End of innermost loop */
865 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
866 f+i_coord_offset,fshift+i_shift_offset);
868 /* Increment number of inner iterations */
869 inneriter += j_index_end - j_index_start;
871 /* Outer loop uses 7 flops */
874 /* Increment number of outer iterations */
877 /* Update outer/inner flops */
879 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*63);