2 * Note: this file was generated by the Gromacs avx_256_double kernel generator.
4 * This source code is part of
8 * Copyright (c) 2001-2012, The GROMACS Development Team
10 * Gromacs is a library for molecular simulation and trajectory analysis,
11 * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
12 * a full list of developers and information, check out http://www.gromacs.org
14 * This program is free software; you can redistribute it and/or modify it under
15 * the terms of the GNU Lesser General Public License as published by the Free
16 * Software Foundation; either version 2 of the License, or (at your option) any
19 * To help fund GROMACS development, we humbly ask that you cite
20 * the papers people have written on it - you can find them on the website.
28 #include "../nb_kernel.h"
29 #include "types/simple.h"
33 #include "gmx_math_x86_avx_256_double.h"
34 #include "kernelutil_x86_avx_256_double.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwCSTab_GeomP1P1_VF_avx_256_double
38 * Electrostatics interaction: Ewald
39 * VdW interaction: CubicSplineTable
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_VF_avx_256_double
45 (t_nblist * gmx_restrict nlist,
46 rvec * gmx_restrict xx,
47 rvec * gmx_restrict ff,
48 t_forcerec * gmx_restrict fr,
49 t_mdatoms * gmx_restrict mdatoms,
50 nb_kernel_data_t * gmx_restrict kernel_data,
51 t_nrnb * gmx_restrict nrnb)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
62 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
63 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
64 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
66 real *shiftvec,*fshift,*x,*f;
67 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
69 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
70 real * vdwioffsetptr0;
71 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
72 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
73 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
74 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
75 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
78 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
81 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
82 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
84 __m128i ifour = _mm_set1_epi32(4);
85 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
88 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
89 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
91 __m256d dummy_mask,cutoff_mask;
92 __m128 tmpmask0,tmpmask1;
93 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
94 __m256d one = _mm256_set1_pd(1.0);
95 __m256d two = _mm256_set1_pd(2.0);
101 jindex = nlist->jindex;
103 shiftidx = nlist->shift;
105 shiftvec = fr->shift_vec[0];
106 fshift = fr->fshift[0];
107 facel = _mm256_set1_pd(fr->epsfac);
108 charge = mdatoms->chargeA;
109 nvdwtype = fr->ntype;
111 vdwtype = mdatoms->typeA;
113 vftab = kernel_data->table_vdw->data;
114 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
116 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
117 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
118 beta2 = _mm256_mul_pd(beta,beta);
119 beta3 = _mm256_mul_pd(beta,beta2);
121 ewtab = fr->ic->tabq_coul_FDV0;
122 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
123 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
125 /* Avoid stupid compiler warnings */
126 jnrA = jnrB = jnrC = jnrD = 0;
135 for(iidx=0;iidx<4*DIM;iidx++)
140 /* Start outer loop over neighborlists */
141 for(iidx=0; iidx<nri; iidx++)
143 /* Load shift vector for this list */
144 i_shift_offset = DIM*shiftidx[iidx];
146 /* Load limits for loop over neighbors */
147 j_index_start = jindex[iidx];
148 j_index_end = jindex[iidx+1];
150 /* Get outer coordinate index */
152 i_coord_offset = DIM*inr;
154 /* Load i particle coords and add shift vector */
155 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
157 fix0 = _mm256_setzero_pd();
158 fiy0 = _mm256_setzero_pd();
159 fiz0 = _mm256_setzero_pd();
161 /* Load parameters for i particles */
162 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
163 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
165 /* Reset potential sums */
166 velecsum = _mm256_setzero_pd();
167 vvdwsum = _mm256_setzero_pd();
169 /* Start inner kernel loop */
170 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
173 /* Get j neighbor index, and coordinate index */
178 j_coord_offsetA = DIM*jnrA;
179 j_coord_offsetB = DIM*jnrB;
180 j_coord_offsetC = DIM*jnrC;
181 j_coord_offsetD = DIM*jnrD;
183 /* load j atom coordinates */
184 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
185 x+j_coord_offsetC,x+j_coord_offsetD,
188 /* Calculate displacement vector */
189 dx00 = _mm256_sub_pd(ix0,jx0);
190 dy00 = _mm256_sub_pd(iy0,jy0);
191 dz00 = _mm256_sub_pd(iz0,jz0);
193 /* Calculate squared distance and things based on it */
194 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
196 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
198 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
200 /* Load parameters for j particles */
201 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
202 charge+jnrC+0,charge+jnrD+0);
203 vdwjidx0A = 2*vdwtype[jnrA+0];
204 vdwjidx0B = 2*vdwtype[jnrB+0];
205 vdwjidx0C = 2*vdwtype[jnrC+0];
206 vdwjidx0D = 2*vdwtype[jnrD+0];
208 /**************************
209 * CALCULATE INTERACTIONS *
210 **************************/
212 r00 = _mm256_mul_pd(rsq00,rinv00);
214 /* Compute parameters for interactions between i and j atoms */
215 qq00 = _mm256_mul_pd(iq0,jq0);
216 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
217 vdwioffsetptr0+vdwjidx0B,
218 vdwioffsetptr0+vdwjidx0C,
219 vdwioffsetptr0+vdwjidx0D,
222 /* Calculate table index by multiplying r with table scale and truncate to integer */
223 rt = _mm256_mul_pd(r00,vftabscale);
224 vfitab = _mm256_cvttpd_epi32(rt);
225 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
226 vfitab = _mm_slli_epi32(vfitab,3);
228 /* EWALD ELECTROSTATICS */
230 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
231 ewrt = _mm256_mul_pd(r00,ewtabscale);
232 ewitab = _mm256_cvttpd_epi32(ewrt);
233 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
234 ewitab = _mm_slli_epi32(ewitab,2);
235 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
236 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
237 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
238 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
239 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
240 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
241 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
242 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
243 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
245 /* CUBIC SPLINE TABLE DISPERSION */
246 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
247 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
248 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
249 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
250 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
251 Heps = _mm256_mul_pd(vfeps,H);
252 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
253 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
254 vvdw6 = _mm256_mul_pd(c6_00,VV);
255 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
256 fvdw6 = _mm256_mul_pd(c6_00,FF);
258 /* CUBIC SPLINE TABLE REPULSION */
259 vfitab = _mm_add_epi32(vfitab,ifour);
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 vvdw12 = _mm256_mul_pd(c12_00,VV);
269 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
270 fvdw12 = _mm256_mul_pd(c12_00,FF);
271 vvdw = _mm256_add_pd(vvdw12,vvdw6);
272 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
274 /* Update potential sum for this i atom from the interaction with this j atom. */
275 velecsum = _mm256_add_pd(velecsum,velec);
276 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
278 fscal = _mm256_add_pd(felec,fvdw);
280 /* Calculate temporary vectorial force */
281 tx = _mm256_mul_pd(fscal,dx00);
282 ty = _mm256_mul_pd(fscal,dy00);
283 tz = _mm256_mul_pd(fscal,dz00);
285 /* Update vectorial force */
286 fix0 = _mm256_add_pd(fix0,tx);
287 fiy0 = _mm256_add_pd(fiy0,ty);
288 fiz0 = _mm256_add_pd(fiz0,tz);
290 fjptrA = f+j_coord_offsetA;
291 fjptrB = f+j_coord_offsetB;
292 fjptrC = f+j_coord_offsetC;
293 fjptrD = f+j_coord_offsetD;
294 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
296 /* Inner loop uses 75 flops */
302 /* Get j neighbor index, and coordinate index */
303 jnrlistA = jjnr[jidx];
304 jnrlistB = jjnr[jidx+1];
305 jnrlistC = jjnr[jidx+2];
306 jnrlistD = jjnr[jidx+3];
307 /* Sign of each element will be negative for non-real atoms.
308 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
309 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
311 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
313 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
314 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
315 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
317 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
318 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
319 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
320 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
321 j_coord_offsetA = DIM*jnrA;
322 j_coord_offsetB = DIM*jnrB;
323 j_coord_offsetC = DIM*jnrC;
324 j_coord_offsetD = DIM*jnrD;
326 /* load j atom coordinates */
327 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
328 x+j_coord_offsetC,x+j_coord_offsetD,
331 /* Calculate displacement vector */
332 dx00 = _mm256_sub_pd(ix0,jx0);
333 dy00 = _mm256_sub_pd(iy0,jy0);
334 dz00 = _mm256_sub_pd(iz0,jz0);
336 /* Calculate squared distance and things based on it */
337 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
339 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
341 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
343 /* Load parameters for j particles */
344 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
345 charge+jnrC+0,charge+jnrD+0);
346 vdwjidx0A = 2*vdwtype[jnrA+0];
347 vdwjidx0B = 2*vdwtype[jnrB+0];
348 vdwjidx0C = 2*vdwtype[jnrC+0];
349 vdwjidx0D = 2*vdwtype[jnrD+0];
351 /**************************
352 * CALCULATE INTERACTIONS *
353 **************************/
355 r00 = _mm256_mul_pd(rsq00,rinv00);
356 r00 = _mm256_andnot_pd(dummy_mask,r00);
358 /* Compute parameters for interactions between i and j atoms */
359 qq00 = _mm256_mul_pd(iq0,jq0);
360 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
361 vdwioffsetptr0+vdwjidx0B,
362 vdwioffsetptr0+vdwjidx0C,
363 vdwioffsetptr0+vdwjidx0D,
366 /* Calculate table index by multiplying r with table scale and truncate to integer */
367 rt = _mm256_mul_pd(r00,vftabscale);
368 vfitab = _mm256_cvttpd_epi32(rt);
369 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
370 vfitab = _mm_slli_epi32(vfitab,3);
372 /* EWALD ELECTROSTATICS */
374 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
375 ewrt = _mm256_mul_pd(r00,ewtabscale);
376 ewitab = _mm256_cvttpd_epi32(ewrt);
377 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
378 ewitab = _mm_slli_epi32(ewitab,2);
379 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
380 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
381 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
382 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
383 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
384 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
385 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
386 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
387 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
389 /* CUBIC SPLINE TABLE DISPERSION */
390 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
391 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
392 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
393 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
394 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
395 Heps = _mm256_mul_pd(vfeps,H);
396 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
397 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
398 vvdw6 = _mm256_mul_pd(c6_00,VV);
399 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
400 fvdw6 = _mm256_mul_pd(c6_00,FF);
402 /* CUBIC SPLINE TABLE REPULSION */
403 vfitab = _mm_add_epi32(vfitab,ifour);
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 vvdw12 = _mm256_mul_pd(c12_00,VV);
413 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
414 fvdw12 = _mm256_mul_pd(c12_00,FF);
415 vvdw = _mm256_add_pd(vvdw12,vvdw6);
416 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
418 /* Update potential sum for this i atom from the interaction with this j atom. */
419 velec = _mm256_andnot_pd(dummy_mask,velec);
420 velecsum = _mm256_add_pd(velecsum,velec);
421 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
422 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
424 fscal = _mm256_add_pd(felec,fvdw);
426 fscal = _mm256_andnot_pd(dummy_mask,fscal);
428 /* Calculate temporary vectorial force */
429 tx = _mm256_mul_pd(fscal,dx00);
430 ty = _mm256_mul_pd(fscal,dy00);
431 tz = _mm256_mul_pd(fscal,dz00);
433 /* Update vectorial force */
434 fix0 = _mm256_add_pd(fix0,tx);
435 fiy0 = _mm256_add_pd(fiy0,ty);
436 fiz0 = _mm256_add_pd(fiz0,tz);
438 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
439 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
440 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
441 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
442 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
444 /* Inner loop uses 76 flops */
447 /* End of innermost loop */
449 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
450 f+i_coord_offset,fshift+i_shift_offset);
453 /* Update potential energies */
454 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
455 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
457 /* Increment number of inner iterations */
458 inneriter += j_index_end - j_index_start;
460 /* Outer loop uses 9 flops */
463 /* Increment number of outer iterations */
466 /* Update outer/inner flops */
468 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*76);
471 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_256_double
472 * Electrostatics interaction: Ewald
473 * VdW interaction: CubicSplineTable
474 * Geometry: Particle-Particle
475 * Calculate force/pot: Force
478 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_avx_256_double
479 (t_nblist * gmx_restrict nlist,
480 rvec * gmx_restrict xx,
481 rvec * gmx_restrict ff,
482 t_forcerec * gmx_restrict fr,
483 t_mdatoms * gmx_restrict mdatoms,
484 nb_kernel_data_t * gmx_restrict kernel_data,
485 t_nrnb * gmx_restrict nrnb)
487 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
488 * just 0 for non-waters.
489 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
490 * jnr indices corresponding to data put in the four positions in the SIMD register.
492 int i_shift_offset,i_coord_offset,outeriter,inneriter;
493 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
494 int jnrA,jnrB,jnrC,jnrD;
495 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
496 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
497 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
498 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
500 real *shiftvec,*fshift,*x,*f;
501 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
503 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
504 real * vdwioffsetptr0;
505 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
506 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
507 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
508 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
509 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
512 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
515 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
516 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
518 __m128i ifour = _mm_set1_epi32(4);
519 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
522 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
523 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
525 __m256d dummy_mask,cutoff_mask;
526 __m128 tmpmask0,tmpmask1;
527 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
528 __m256d one = _mm256_set1_pd(1.0);
529 __m256d two = _mm256_set1_pd(2.0);
535 jindex = nlist->jindex;
537 shiftidx = nlist->shift;
539 shiftvec = fr->shift_vec[0];
540 fshift = fr->fshift[0];
541 facel = _mm256_set1_pd(fr->epsfac);
542 charge = mdatoms->chargeA;
543 nvdwtype = fr->ntype;
545 vdwtype = mdatoms->typeA;
547 vftab = kernel_data->table_vdw->data;
548 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
550 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
551 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
552 beta2 = _mm256_mul_pd(beta,beta);
553 beta3 = _mm256_mul_pd(beta,beta2);
555 ewtab = fr->ic->tabq_coul_F;
556 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
557 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
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 r00 = _mm256_mul_pd(rsq00,rinv00);
644 /* Compute parameters for interactions between i and j atoms */
645 qq00 = _mm256_mul_pd(iq0,jq0);
646 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
647 vdwioffsetptr0+vdwjidx0B,
648 vdwioffsetptr0+vdwjidx0C,
649 vdwioffsetptr0+vdwjidx0D,
652 /* Calculate table index by multiplying r with table scale and truncate to integer */
653 rt = _mm256_mul_pd(r00,vftabscale);
654 vfitab = _mm256_cvttpd_epi32(rt);
655 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
656 vfitab = _mm_slli_epi32(vfitab,3);
658 /* EWALD ELECTROSTATICS */
660 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
661 ewrt = _mm256_mul_pd(r00,ewtabscale);
662 ewitab = _mm256_cvttpd_epi32(ewrt);
663 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
664 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
665 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
667 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
668 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
670 /* CUBIC SPLINE TABLE DISPERSION */
671 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
672 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
673 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
674 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
675 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
676 Heps = _mm256_mul_pd(vfeps,H);
677 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
678 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
679 fvdw6 = _mm256_mul_pd(c6_00,FF);
681 /* CUBIC SPLINE TABLE REPULSION */
682 vfitab = _mm_add_epi32(vfitab,ifour);
683 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
684 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
685 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
686 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
687 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
688 Heps = _mm256_mul_pd(vfeps,H);
689 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
690 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
691 fvdw12 = _mm256_mul_pd(c12_00,FF);
692 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
694 fscal = _mm256_add_pd(felec,fvdw);
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);
712 /* Inner loop uses 62 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_pd(_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 r00 = _mm256_mul_pd(rsq00,rinv00);
772 r00 = _mm256_andnot_pd(dummy_mask,r00);
774 /* Compute parameters for interactions between i and j atoms */
775 qq00 = _mm256_mul_pd(iq0,jq0);
776 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
777 vdwioffsetptr0+vdwjidx0B,
778 vdwioffsetptr0+vdwjidx0C,
779 vdwioffsetptr0+vdwjidx0D,
782 /* Calculate table index by multiplying r with table scale and truncate to integer */
783 rt = _mm256_mul_pd(r00,vftabscale);
784 vfitab = _mm256_cvttpd_epi32(rt);
785 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
786 vfitab = _mm_slli_epi32(vfitab,3);
788 /* EWALD ELECTROSTATICS */
790 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
791 ewrt = _mm256_mul_pd(r00,ewtabscale);
792 ewitab = _mm256_cvttpd_epi32(ewrt);
793 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
794 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
795 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
797 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
798 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
800 /* CUBIC SPLINE TABLE DISPERSION */
801 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
802 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
803 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
804 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
805 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
806 Heps = _mm256_mul_pd(vfeps,H);
807 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
808 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
809 fvdw6 = _mm256_mul_pd(c6_00,FF);
811 /* CUBIC SPLINE TABLE REPULSION */
812 vfitab = _mm_add_epi32(vfitab,ifour);
813 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
814 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
815 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
816 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
817 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
818 Heps = _mm256_mul_pd(vfeps,H);
819 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
820 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
821 fvdw12 = _mm256_mul_pd(c12_00,FF);
822 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
824 fscal = _mm256_add_pd(felec,fvdw);
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);
844 /* Inner loop uses 63 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*63);