2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 * Note: this file was generated by the GROMACS avx_256_double kernel generator.
42 #include "../nb_kernel.h"
43 #include "types/simple.h"
44 #include "gromacs/math/vec.h"
47 #include "gromacs/simd/math_x86_avx_256_double.h"
48 #include "kernelutil_x86_avx_256_double.h"
51 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwCSTab_GeomP1P1_VF_avx_256_double
52 * Electrostatics interaction: Ewald
53 * VdW interaction: CubicSplineTable
54 * Geometry: Particle-Particle
55 * Calculate force/pot: PotentialAndForce
58 nb_kernel_ElecEw_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;
102 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
103 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
105 __m256d dummy_mask,cutoff_mask;
106 __m128 tmpmask0,tmpmask1;
107 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
108 __m256d one = _mm256_set1_pd(1.0);
109 __m256d two = _mm256_set1_pd(2.0);
115 jindex = nlist->jindex;
117 shiftidx = nlist->shift;
119 shiftvec = fr->shift_vec[0];
120 fshift = fr->fshift[0];
121 facel = _mm256_set1_pd(fr->epsfac);
122 charge = mdatoms->chargeA;
123 nvdwtype = fr->ntype;
125 vdwtype = mdatoms->typeA;
127 vftab = kernel_data->table_vdw->data;
128 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
130 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
131 beta = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
132 beta2 = _mm256_mul_pd(beta,beta);
133 beta3 = _mm256_mul_pd(beta,beta2);
135 ewtab = fr->ic->tabq_coul_FDV0;
136 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
137 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
139 /* Avoid stupid compiler warnings */
140 jnrA = jnrB = jnrC = jnrD = 0;
149 for(iidx=0;iidx<4*DIM;iidx++)
154 /* Start outer loop over neighborlists */
155 for(iidx=0; iidx<nri; iidx++)
157 /* Load shift vector for this list */
158 i_shift_offset = DIM*shiftidx[iidx];
160 /* Load limits for loop over neighbors */
161 j_index_start = jindex[iidx];
162 j_index_end = jindex[iidx+1];
164 /* Get outer coordinate index */
166 i_coord_offset = DIM*inr;
168 /* Load i particle coords and add shift vector */
169 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
171 fix0 = _mm256_setzero_pd();
172 fiy0 = _mm256_setzero_pd();
173 fiz0 = _mm256_setzero_pd();
175 /* Load parameters for i particles */
176 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
177 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
179 /* Reset potential sums */
180 velecsum = _mm256_setzero_pd();
181 vvdwsum = _mm256_setzero_pd();
183 /* Start inner kernel loop */
184 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
187 /* Get j neighbor index, and coordinate index */
192 j_coord_offsetA = DIM*jnrA;
193 j_coord_offsetB = DIM*jnrB;
194 j_coord_offsetC = DIM*jnrC;
195 j_coord_offsetD = DIM*jnrD;
197 /* load j atom coordinates */
198 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
199 x+j_coord_offsetC,x+j_coord_offsetD,
202 /* Calculate displacement vector */
203 dx00 = _mm256_sub_pd(ix0,jx0);
204 dy00 = _mm256_sub_pd(iy0,jy0);
205 dz00 = _mm256_sub_pd(iz0,jz0);
207 /* Calculate squared distance and things based on it */
208 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
210 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
212 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
214 /* Load parameters for j particles */
215 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
216 charge+jnrC+0,charge+jnrD+0);
217 vdwjidx0A = 2*vdwtype[jnrA+0];
218 vdwjidx0B = 2*vdwtype[jnrB+0];
219 vdwjidx0C = 2*vdwtype[jnrC+0];
220 vdwjidx0D = 2*vdwtype[jnrD+0];
222 /**************************
223 * CALCULATE INTERACTIONS *
224 **************************/
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 /* EWALD ELECTROSTATICS */
244 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
245 ewrt = _mm256_mul_pd(r00,ewtabscale);
246 ewitab = _mm256_cvttpd_epi32(ewrt);
247 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
248 ewitab = _mm_slli_epi32(ewitab,2);
249 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
250 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
251 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
252 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
253 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
254 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
255 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
256 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
257 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
259 /* CUBIC SPLINE TABLE DISPERSION */
260 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
261 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
262 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
263 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
264 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
265 Heps = _mm256_mul_pd(vfeps,H);
266 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
267 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
268 vvdw6 = _mm256_mul_pd(c6_00,VV);
269 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
270 fvdw6 = _mm256_mul_pd(c6_00,FF);
272 /* CUBIC SPLINE TABLE REPULSION */
273 vfitab = _mm_add_epi32(vfitab,ifour);
274 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
275 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
276 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
277 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
278 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
279 Heps = _mm256_mul_pd(vfeps,H);
280 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
281 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
282 vvdw12 = _mm256_mul_pd(c12_00,VV);
283 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
284 fvdw12 = _mm256_mul_pd(c12_00,FF);
285 vvdw = _mm256_add_pd(vvdw12,vvdw6);
286 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
288 /* Update potential sum for this i atom from the interaction with this j atom. */
289 velecsum = _mm256_add_pd(velecsum,velec);
290 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
292 fscal = _mm256_add_pd(felec,fvdw);
294 /* Calculate temporary vectorial force */
295 tx = _mm256_mul_pd(fscal,dx00);
296 ty = _mm256_mul_pd(fscal,dy00);
297 tz = _mm256_mul_pd(fscal,dz00);
299 /* Update vectorial force */
300 fix0 = _mm256_add_pd(fix0,tx);
301 fiy0 = _mm256_add_pd(fiy0,ty);
302 fiz0 = _mm256_add_pd(fiz0,tz);
304 fjptrA = f+j_coord_offsetA;
305 fjptrB = f+j_coord_offsetB;
306 fjptrC = f+j_coord_offsetC;
307 fjptrD = f+j_coord_offsetD;
308 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
310 /* Inner loop uses 75 flops */
316 /* Get j neighbor index, and coordinate index */
317 jnrlistA = jjnr[jidx];
318 jnrlistB = jjnr[jidx+1];
319 jnrlistC = jjnr[jidx+2];
320 jnrlistD = jjnr[jidx+3];
321 /* Sign of each element will be negative for non-real atoms.
322 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
323 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
325 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
327 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
328 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
329 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
331 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
332 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
333 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
334 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
335 j_coord_offsetA = DIM*jnrA;
336 j_coord_offsetB = DIM*jnrB;
337 j_coord_offsetC = DIM*jnrC;
338 j_coord_offsetD = DIM*jnrD;
340 /* load j atom coordinates */
341 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
342 x+j_coord_offsetC,x+j_coord_offsetD,
345 /* Calculate displacement vector */
346 dx00 = _mm256_sub_pd(ix0,jx0);
347 dy00 = _mm256_sub_pd(iy0,jy0);
348 dz00 = _mm256_sub_pd(iz0,jz0);
350 /* Calculate squared distance and things based on it */
351 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
353 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
355 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
357 /* Load parameters for j particles */
358 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
359 charge+jnrC+0,charge+jnrD+0);
360 vdwjidx0A = 2*vdwtype[jnrA+0];
361 vdwjidx0B = 2*vdwtype[jnrB+0];
362 vdwjidx0C = 2*vdwtype[jnrC+0];
363 vdwjidx0D = 2*vdwtype[jnrD+0];
365 /**************************
366 * CALCULATE INTERACTIONS *
367 **************************/
369 r00 = _mm256_mul_pd(rsq00,rinv00);
370 r00 = _mm256_andnot_pd(dummy_mask,r00);
372 /* Compute parameters for interactions between i and j atoms */
373 qq00 = _mm256_mul_pd(iq0,jq0);
374 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
375 vdwioffsetptr0+vdwjidx0B,
376 vdwioffsetptr0+vdwjidx0C,
377 vdwioffsetptr0+vdwjidx0D,
380 /* Calculate table index by multiplying r with table scale and truncate to integer */
381 rt = _mm256_mul_pd(r00,vftabscale);
382 vfitab = _mm256_cvttpd_epi32(rt);
383 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
384 vfitab = _mm_slli_epi32(vfitab,3);
386 /* EWALD ELECTROSTATICS */
388 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
389 ewrt = _mm256_mul_pd(r00,ewtabscale);
390 ewitab = _mm256_cvttpd_epi32(ewrt);
391 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
392 ewitab = _mm_slli_epi32(ewitab,2);
393 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
394 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
395 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
396 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
397 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
398 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
399 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
400 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
401 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
403 /* CUBIC SPLINE TABLE DISPERSION */
404 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
405 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
406 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
407 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
408 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
409 Heps = _mm256_mul_pd(vfeps,H);
410 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
411 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
412 vvdw6 = _mm256_mul_pd(c6_00,VV);
413 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
414 fvdw6 = _mm256_mul_pd(c6_00,FF);
416 /* CUBIC SPLINE TABLE REPULSION */
417 vfitab = _mm_add_epi32(vfitab,ifour);
418 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
419 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
420 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
421 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
422 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
423 Heps = _mm256_mul_pd(vfeps,H);
424 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
425 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
426 vvdw12 = _mm256_mul_pd(c12_00,VV);
427 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
428 fvdw12 = _mm256_mul_pd(c12_00,FF);
429 vvdw = _mm256_add_pd(vvdw12,vvdw6);
430 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
432 /* Update potential sum for this i atom from the interaction with this j atom. */
433 velec = _mm256_andnot_pd(dummy_mask,velec);
434 velecsum = _mm256_add_pd(velecsum,velec);
435 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
436 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
438 fscal = _mm256_add_pd(felec,fvdw);
440 fscal = _mm256_andnot_pd(dummy_mask,fscal);
442 /* Calculate temporary vectorial force */
443 tx = _mm256_mul_pd(fscal,dx00);
444 ty = _mm256_mul_pd(fscal,dy00);
445 tz = _mm256_mul_pd(fscal,dz00);
447 /* Update vectorial force */
448 fix0 = _mm256_add_pd(fix0,tx);
449 fiy0 = _mm256_add_pd(fiy0,ty);
450 fiz0 = _mm256_add_pd(fiz0,tz);
452 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
453 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
454 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
455 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
456 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
458 /* Inner loop uses 76 flops */
461 /* End of innermost loop */
463 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
464 f+i_coord_offset,fshift+i_shift_offset);
467 /* Update potential energies */
468 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
469 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
471 /* Increment number of inner iterations */
472 inneriter += j_index_end - j_index_start;
474 /* Outer loop uses 9 flops */
477 /* Increment number of outer iterations */
480 /* Update outer/inner flops */
482 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*76);
485 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_256_double
486 * Electrostatics interaction: Ewald
487 * VdW interaction: CubicSplineTable
488 * Geometry: Particle-Particle
489 * Calculate force/pot: Force
492 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_256_double
493 (t_nblist * gmx_restrict nlist,
494 rvec * gmx_restrict xx,
495 rvec * gmx_restrict ff,
496 t_forcerec * gmx_restrict fr,
497 t_mdatoms * gmx_restrict mdatoms,
498 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
499 t_nrnb * gmx_restrict nrnb)
501 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
502 * just 0 for non-waters.
503 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
504 * jnr indices corresponding to data put in the four positions in the SIMD register.
506 int i_shift_offset,i_coord_offset,outeriter,inneriter;
507 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
508 int jnrA,jnrB,jnrC,jnrD;
509 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
510 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
511 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
512 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
514 real *shiftvec,*fshift,*x,*f;
515 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
517 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
518 real * vdwioffsetptr0;
519 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
520 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
521 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
522 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
523 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
526 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
529 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
530 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
532 __m128i ifour = _mm_set1_epi32(4);
533 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
536 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
537 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
539 __m256d dummy_mask,cutoff_mask;
540 __m128 tmpmask0,tmpmask1;
541 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
542 __m256d one = _mm256_set1_pd(1.0);
543 __m256d two = _mm256_set1_pd(2.0);
549 jindex = nlist->jindex;
551 shiftidx = nlist->shift;
553 shiftvec = fr->shift_vec[0];
554 fshift = fr->fshift[0];
555 facel = _mm256_set1_pd(fr->epsfac);
556 charge = mdatoms->chargeA;
557 nvdwtype = fr->ntype;
559 vdwtype = mdatoms->typeA;
561 vftab = kernel_data->table_vdw->data;
562 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
564 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
565 beta = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
566 beta2 = _mm256_mul_pd(beta,beta);
567 beta3 = _mm256_mul_pd(beta,beta2);
569 ewtab = fr->ic->tabq_coul_F;
570 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
571 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
573 /* Avoid stupid compiler warnings */
574 jnrA = jnrB = jnrC = jnrD = 0;
583 for(iidx=0;iidx<4*DIM;iidx++)
588 /* Start outer loop over neighborlists */
589 for(iidx=0; iidx<nri; iidx++)
591 /* Load shift vector for this list */
592 i_shift_offset = DIM*shiftidx[iidx];
594 /* Load limits for loop over neighbors */
595 j_index_start = jindex[iidx];
596 j_index_end = jindex[iidx+1];
598 /* Get outer coordinate index */
600 i_coord_offset = DIM*inr;
602 /* Load i particle coords and add shift vector */
603 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
605 fix0 = _mm256_setzero_pd();
606 fiy0 = _mm256_setzero_pd();
607 fiz0 = _mm256_setzero_pd();
609 /* Load parameters for i particles */
610 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
611 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
613 /* Start inner kernel loop */
614 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
617 /* Get j neighbor index, and coordinate index */
622 j_coord_offsetA = DIM*jnrA;
623 j_coord_offsetB = DIM*jnrB;
624 j_coord_offsetC = DIM*jnrC;
625 j_coord_offsetD = DIM*jnrD;
627 /* load j atom coordinates */
628 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
629 x+j_coord_offsetC,x+j_coord_offsetD,
632 /* Calculate displacement vector */
633 dx00 = _mm256_sub_pd(ix0,jx0);
634 dy00 = _mm256_sub_pd(iy0,jy0);
635 dz00 = _mm256_sub_pd(iz0,jz0);
637 /* Calculate squared distance and things based on it */
638 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
640 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
642 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
644 /* Load parameters for j particles */
645 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
646 charge+jnrC+0,charge+jnrD+0);
647 vdwjidx0A = 2*vdwtype[jnrA+0];
648 vdwjidx0B = 2*vdwtype[jnrB+0];
649 vdwjidx0C = 2*vdwtype[jnrC+0];
650 vdwjidx0D = 2*vdwtype[jnrD+0];
652 /**************************
653 * CALCULATE INTERACTIONS *
654 **************************/
656 r00 = _mm256_mul_pd(rsq00,rinv00);
658 /* Compute parameters for interactions between i and j atoms */
659 qq00 = _mm256_mul_pd(iq0,jq0);
660 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
661 vdwioffsetptr0+vdwjidx0B,
662 vdwioffsetptr0+vdwjidx0C,
663 vdwioffsetptr0+vdwjidx0D,
666 /* Calculate table index by multiplying r with table scale and truncate to integer */
667 rt = _mm256_mul_pd(r00,vftabscale);
668 vfitab = _mm256_cvttpd_epi32(rt);
669 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
670 vfitab = _mm_slli_epi32(vfitab,3);
672 /* EWALD ELECTROSTATICS */
674 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
675 ewrt = _mm256_mul_pd(r00,ewtabscale);
676 ewitab = _mm256_cvttpd_epi32(ewrt);
677 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
678 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
679 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
681 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
682 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
684 /* CUBIC SPLINE TABLE DISPERSION */
685 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
686 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
687 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
688 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
689 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
690 Heps = _mm256_mul_pd(vfeps,H);
691 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
692 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
693 fvdw6 = _mm256_mul_pd(c6_00,FF);
695 /* CUBIC SPLINE TABLE REPULSION */
696 vfitab = _mm_add_epi32(vfitab,ifour);
697 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
698 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
699 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
700 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
701 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
702 Heps = _mm256_mul_pd(vfeps,H);
703 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
704 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
705 fvdw12 = _mm256_mul_pd(c12_00,FF);
706 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
708 fscal = _mm256_add_pd(felec,fvdw);
710 /* Calculate temporary vectorial force */
711 tx = _mm256_mul_pd(fscal,dx00);
712 ty = _mm256_mul_pd(fscal,dy00);
713 tz = _mm256_mul_pd(fscal,dz00);
715 /* Update vectorial force */
716 fix0 = _mm256_add_pd(fix0,tx);
717 fiy0 = _mm256_add_pd(fiy0,ty);
718 fiz0 = _mm256_add_pd(fiz0,tz);
720 fjptrA = f+j_coord_offsetA;
721 fjptrB = f+j_coord_offsetB;
722 fjptrC = f+j_coord_offsetC;
723 fjptrD = f+j_coord_offsetD;
724 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
726 /* Inner loop uses 62 flops */
732 /* Get j neighbor index, and coordinate index */
733 jnrlistA = jjnr[jidx];
734 jnrlistB = jjnr[jidx+1];
735 jnrlistC = jjnr[jidx+2];
736 jnrlistD = jjnr[jidx+3];
737 /* Sign of each element will be negative for non-real atoms.
738 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
739 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
741 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
743 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
744 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
745 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
747 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
748 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
749 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
750 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
751 j_coord_offsetA = DIM*jnrA;
752 j_coord_offsetB = DIM*jnrB;
753 j_coord_offsetC = DIM*jnrC;
754 j_coord_offsetD = DIM*jnrD;
756 /* load j atom coordinates */
757 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
758 x+j_coord_offsetC,x+j_coord_offsetD,
761 /* Calculate displacement vector */
762 dx00 = _mm256_sub_pd(ix0,jx0);
763 dy00 = _mm256_sub_pd(iy0,jy0);
764 dz00 = _mm256_sub_pd(iz0,jz0);
766 /* Calculate squared distance and things based on it */
767 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
769 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
771 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
773 /* Load parameters for j particles */
774 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
775 charge+jnrC+0,charge+jnrD+0);
776 vdwjidx0A = 2*vdwtype[jnrA+0];
777 vdwjidx0B = 2*vdwtype[jnrB+0];
778 vdwjidx0C = 2*vdwtype[jnrC+0];
779 vdwjidx0D = 2*vdwtype[jnrD+0];
781 /**************************
782 * CALCULATE INTERACTIONS *
783 **************************/
785 r00 = _mm256_mul_pd(rsq00,rinv00);
786 r00 = _mm256_andnot_pd(dummy_mask,r00);
788 /* Compute parameters for interactions between i and j atoms */
789 qq00 = _mm256_mul_pd(iq0,jq0);
790 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
791 vdwioffsetptr0+vdwjidx0B,
792 vdwioffsetptr0+vdwjidx0C,
793 vdwioffsetptr0+vdwjidx0D,
796 /* Calculate table index by multiplying r with table scale and truncate to integer */
797 rt = _mm256_mul_pd(r00,vftabscale);
798 vfitab = _mm256_cvttpd_epi32(rt);
799 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
800 vfitab = _mm_slli_epi32(vfitab,3);
802 /* EWALD ELECTROSTATICS */
804 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
805 ewrt = _mm256_mul_pd(r00,ewtabscale);
806 ewitab = _mm256_cvttpd_epi32(ewrt);
807 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
808 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
809 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
811 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
812 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
814 /* CUBIC SPLINE TABLE DISPERSION */
815 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
816 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
817 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
818 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
819 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
820 Heps = _mm256_mul_pd(vfeps,H);
821 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
822 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
823 fvdw6 = _mm256_mul_pd(c6_00,FF);
825 /* CUBIC SPLINE TABLE REPULSION */
826 vfitab = _mm_add_epi32(vfitab,ifour);
827 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
828 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
829 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
830 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
831 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
832 Heps = _mm256_mul_pd(vfeps,H);
833 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
834 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
835 fvdw12 = _mm256_mul_pd(c12_00,FF);
836 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
838 fscal = _mm256_add_pd(felec,fvdw);
840 fscal = _mm256_andnot_pd(dummy_mask,fscal);
842 /* Calculate temporary vectorial force */
843 tx = _mm256_mul_pd(fscal,dx00);
844 ty = _mm256_mul_pd(fscal,dy00);
845 tz = _mm256_mul_pd(fscal,dz00);
847 /* Update vectorial force */
848 fix0 = _mm256_add_pd(fix0,tx);
849 fiy0 = _mm256_add_pd(fiy0,ty);
850 fiz0 = _mm256_add_pd(fiz0,tz);
852 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
853 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
854 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
855 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
856 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
858 /* Inner loop uses 63 flops */
861 /* End of innermost loop */
863 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
864 f+i_coord_offset,fshift+i_shift_offset);
866 /* Increment number of inner iterations */
867 inneriter += j_index_end - j_index_start;
869 /* Outer loop uses 7 flops */
872 /* Increment number of outer iterations */
875 /* Update outer/inner flops */
877 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*63);