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_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_256_double
38 * Electrostatics interaction: Ewald
39 * VdW interaction: LennardJones
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEwSw_VdwLJSw_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 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
85 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
87 __m256d rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
88 real rswitch_scalar,d_scalar;
89 __m256d dummy_mask,cutoff_mask;
90 __m128 tmpmask0,tmpmask1;
91 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
92 __m256d one = _mm256_set1_pd(1.0);
93 __m256d two = _mm256_set1_pd(2.0);
99 jindex = nlist->jindex;
101 shiftidx = nlist->shift;
103 shiftvec = fr->shift_vec[0];
104 fshift = fr->fshift[0];
105 facel = _mm256_set1_pd(fr->epsfac);
106 charge = mdatoms->chargeA;
107 nvdwtype = fr->ntype;
109 vdwtype = mdatoms->typeA;
111 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
112 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
113 beta2 = _mm256_mul_pd(beta,beta);
114 beta3 = _mm256_mul_pd(beta,beta2);
116 ewtab = fr->ic->tabq_coul_FDV0;
117 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
118 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
120 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
121 rcutoff_scalar = fr->rcoulomb;
122 rcutoff = _mm256_set1_pd(rcutoff_scalar);
123 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
125 rswitch_scalar = fr->rcoulomb_switch;
126 rswitch = _mm256_set1_pd(rswitch_scalar);
127 /* Setup switch parameters */
128 d_scalar = rcutoff_scalar-rswitch_scalar;
129 d = _mm256_set1_pd(d_scalar);
130 swV3 = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
131 swV4 = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
132 swV5 = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
133 swF2 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
134 swF3 = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
135 swF4 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
137 /* Avoid stupid compiler warnings */
138 jnrA = jnrB = jnrC = jnrD = 0;
147 for(iidx=0;iidx<4*DIM;iidx++)
152 /* Start outer loop over neighborlists */
153 for(iidx=0; iidx<nri; iidx++)
155 /* Load shift vector for this list */
156 i_shift_offset = DIM*shiftidx[iidx];
158 /* Load limits for loop over neighbors */
159 j_index_start = jindex[iidx];
160 j_index_end = jindex[iidx+1];
162 /* Get outer coordinate index */
164 i_coord_offset = DIM*inr;
166 /* Load i particle coords and add shift vector */
167 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
169 fix0 = _mm256_setzero_pd();
170 fiy0 = _mm256_setzero_pd();
171 fiz0 = _mm256_setzero_pd();
173 /* Load parameters for i particles */
174 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
175 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
177 /* Reset potential sums */
178 velecsum = _mm256_setzero_pd();
179 vvdwsum = _mm256_setzero_pd();
181 /* Start inner kernel loop */
182 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
185 /* Get j neighbor index, and coordinate index */
190 j_coord_offsetA = DIM*jnrA;
191 j_coord_offsetB = DIM*jnrB;
192 j_coord_offsetC = DIM*jnrC;
193 j_coord_offsetD = DIM*jnrD;
195 /* load j atom coordinates */
196 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
197 x+j_coord_offsetC,x+j_coord_offsetD,
200 /* Calculate displacement vector */
201 dx00 = _mm256_sub_pd(ix0,jx0);
202 dy00 = _mm256_sub_pd(iy0,jy0);
203 dz00 = _mm256_sub_pd(iz0,jz0);
205 /* Calculate squared distance and things based on it */
206 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
208 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
210 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
212 /* Load parameters for j particles */
213 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
214 charge+jnrC+0,charge+jnrD+0);
215 vdwjidx0A = 2*vdwtype[jnrA+0];
216 vdwjidx0B = 2*vdwtype[jnrB+0];
217 vdwjidx0C = 2*vdwtype[jnrC+0];
218 vdwjidx0D = 2*vdwtype[jnrD+0];
220 /**************************
221 * CALCULATE INTERACTIONS *
222 **************************/
224 if (gmx_mm256_any_lt(rsq00,rcutoff2))
227 r00 = _mm256_mul_pd(rsq00,rinv00);
229 /* Compute parameters for interactions between i and j atoms */
230 qq00 = _mm256_mul_pd(iq0,jq0);
231 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
232 vdwioffsetptr0+vdwjidx0B,
233 vdwioffsetptr0+vdwjidx0C,
234 vdwioffsetptr0+vdwjidx0D,
237 /* EWALD ELECTROSTATICS */
239 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
240 ewrt = _mm256_mul_pd(r00,ewtabscale);
241 ewitab = _mm256_cvttpd_epi32(ewrt);
242 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
243 ewitab = _mm_slli_epi32(ewitab,2);
244 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
245 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
246 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
247 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
248 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
249 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
250 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
251 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
252 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
254 /* LENNARD-JONES DISPERSION/REPULSION */
256 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
257 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
258 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
259 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
260 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
262 d = _mm256_sub_pd(r00,rswitch);
263 d = _mm256_max_pd(d,_mm256_setzero_pd());
264 d2 = _mm256_mul_pd(d,d);
265 sw = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
267 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
269 /* Evaluate switch function */
270 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
271 felec = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
272 fvdw = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
273 velec = _mm256_mul_pd(velec,sw);
274 vvdw = _mm256_mul_pd(vvdw,sw);
275 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
277 /* Update potential sum for this i atom from the interaction with this j atom. */
278 velec = _mm256_and_pd(velec,cutoff_mask);
279 velecsum = _mm256_add_pd(velecsum,velec);
280 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
281 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
283 fscal = _mm256_add_pd(felec,fvdw);
285 fscal = _mm256_and_pd(fscal,cutoff_mask);
287 /* Calculate temporary vectorial force */
288 tx = _mm256_mul_pd(fscal,dx00);
289 ty = _mm256_mul_pd(fscal,dy00);
290 tz = _mm256_mul_pd(fscal,dz00);
292 /* Update vectorial force */
293 fix0 = _mm256_add_pd(fix0,tx);
294 fiy0 = _mm256_add_pd(fiy0,ty);
295 fiz0 = _mm256_add_pd(fiz0,tz);
297 fjptrA = f+j_coord_offsetA;
298 fjptrB = f+j_coord_offsetB;
299 fjptrC = f+j_coord_offsetC;
300 fjptrD = f+j_coord_offsetD;
301 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
305 /* Inner loop uses 83 flops */
311 /* Get j neighbor index, and coordinate index */
312 jnrlistA = jjnr[jidx];
313 jnrlistB = jjnr[jidx+1];
314 jnrlistC = jjnr[jidx+2];
315 jnrlistD = jjnr[jidx+3];
316 /* Sign of each element will be negative for non-real atoms.
317 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
318 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
320 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
322 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
323 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
324 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
326 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
327 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
328 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
329 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
330 j_coord_offsetA = DIM*jnrA;
331 j_coord_offsetB = DIM*jnrB;
332 j_coord_offsetC = DIM*jnrC;
333 j_coord_offsetD = DIM*jnrD;
335 /* load j atom coordinates */
336 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
337 x+j_coord_offsetC,x+j_coord_offsetD,
340 /* Calculate displacement vector */
341 dx00 = _mm256_sub_pd(ix0,jx0);
342 dy00 = _mm256_sub_pd(iy0,jy0);
343 dz00 = _mm256_sub_pd(iz0,jz0);
345 /* Calculate squared distance and things based on it */
346 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
348 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
350 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
352 /* Load parameters for j particles */
353 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
354 charge+jnrC+0,charge+jnrD+0);
355 vdwjidx0A = 2*vdwtype[jnrA+0];
356 vdwjidx0B = 2*vdwtype[jnrB+0];
357 vdwjidx0C = 2*vdwtype[jnrC+0];
358 vdwjidx0D = 2*vdwtype[jnrD+0];
360 /**************************
361 * CALCULATE INTERACTIONS *
362 **************************/
364 if (gmx_mm256_any_lt(rsq00,rcutoff2))
367 r00 = _mm256_mul_pd(rsq00,rinv00);
368 r00 = _mm256_andnot_pd(dummy_mask,r00);
370 /* Compute parameters for interactions between i and j atoms */
371 qq00 = _mm256_mul_pd(iq0,jq0);
372 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
373 vdwioffsetptr0+vdwjidx0B,
374 vdwioffsetptr0+vdwjidx0C,
375 vdwioffsetptr0+vdwjidx0D,
378 /* EWALD ELECTROSTATICS */
380 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
381 ewrt = _mm256_mul_pd(r00,ewtabscale);
382 ewitab = _mm256_cvttpd_epi32(ewrt);
383 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
384 ewitab = _mm_slli_epi32(ewitab,2);
385 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
386 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
387 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
388 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
389 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
390 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
391 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
392 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
393 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
395 /* LENNARD-JONES DISPERSION/REPULSION */
397 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
398 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
399 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
400 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
401 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
403 d = _mm256_sub_pd(r00,rswitch);
404 d = _mm256_max_pd(d,_mm256_setzero_pd());
405 d2 = _mm256_mul_pd(d,d);
406 sw = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
408 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
410 /* Evaluate switch function */
411 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
412 felec = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
413 fvdw = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
414 velec = _mm256_mul_pd(velec,sw);
415 vvdw = _mm256_mul_pd(vvdw,sw);
416 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
418 /* Update potential sum for this i atom from the interaction with this j atom. */
419 velec = _mm256_and_pd(velec,cutoff_mask);
420 velec = _mm256_andnot_pd(dummy_mask,velec);
421 velecsum = _mm256_add_pd(velecsum,velec);
422 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
423 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
424 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
426 fscal = _mm256_add_pd(felec,fvdw);
428 fscal = _mm256_and_pd(fscal,cutoff_mask);
430 fscal = _mm256_andnot_pd(dummy_mask,fscal);
432 /* Calculate temporary vectorial force */
433 tx = _mm256_mul_pd(fscal,dx00);
434 ty = _mm256_mul_pd(fscal,dy00);
435 tz = _mm256_mul_pd(fscal,dz00);
437 /* Update vectorial force */
438 fix0 = _mm256_add_pd(fix0,tx);
439 fiy0 = _mm256_add_pd(fiy0,ty);
440 fiz0 = _mm256_add_pd(fiz0,tz);
442 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
443 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
444 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
445 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
446 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
450 /* Inner loop uses 84 flops */
453 /* End of innermost loop */
455 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
456 f+i_coord_offset,fshift+i_shift_offset);
459 /* Update potential energies */
460 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
461 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
463 /* Increment number of inner iterations */
464 inneriter += j_index_end - j_index_start;
466 /* Outer loop uses 9 flops */
469 /* Increment number of outer iterations */
472 /* Update outer/inner flops */
474 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*84);
477 * Gromacs nonbonded kernel: nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_256_double
478 * Electrostatics interaction: Ewald
479 * VdW interaction: LennardJones
480 * Geometry: Particle-Particle
481 * Calculate force/pot: Force
484 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_256_double
485 (t_nblist * gmx_restrict nlist,
486 rvec * gmx_restrict xx,
487 rvec * gmx_restrict ff,
488 t_forcerec * gmx_restrict fr,
489 t_mdatoms * gmx_restrict mdatoms,
490 nb_kernel_data_t * gmx_restrict kernel_data,
491 t_nrnb * gmx_restrict nrnb)
493 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
494 * just 0 for non-waters.
495 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
496 * jnr indices corresponding to data put in the four positions in the SIMD register.
498 int i_shift_offset,i_coord_offset,outeriter,inneriter;
499 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
500 int jnrA,jnrB,jnrC,jnrD;
501 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
502 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
503 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
504 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
506 real *shiftvec,*fshift,*x,*f;
507 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
509 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
510 real * vdwioffsetptr0;
511 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
512 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
513 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
514 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
515 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
518 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
521 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
522 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
524 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
525 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
527 __m256d rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
528 real rswitch_scalar,d_scalar;
529 __m256d dummy_mask,cutoff_mask;
530 __m128 tmpmask0,tmpmask1;
531 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
532 __m256d one = _mm256_set1_pd(1.0);
533 __m256d two = _mm256_set1_pd(2.0);
539 jindex = nlist->jindex;
541 shiftidx = nlist->shift;
543 shiftvec = fr->shift_vec[0];
544 fshift = fr->fshift[0];
545 facel = _mm256_set1_pd(fr->epsfac);
546 charge = mdatoms->chargeA;
547 nvdwtype = fr->ntype;
549 vdwtype = mdatoms->typeA;
551 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
552 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
553 beta2 = _mm256_mul_pd(beta,beta);
554 beta3 = _mm256_mul_pd(beta,beta2);
556 ewtab = fr->ic->tabq_coul_FDV0;
557 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
558 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
560 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
561 rcutoff_scalar = fr->rcoulomb;
562 rcutoff = _mm256_set1_pd(rcutoff_scalar);
563 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
565 rswitch_scalar = fr->rcoulomb_switch;
566 rswitch = _mm256_set1_pd(rswitch_scalar);
567 /* Setup switch parameters */
568 d_scalar = rcutoff_scalar-rswitch_scalar;
569 d = _mm256_set1_pd(d_scalar);
570 swV3 = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
571 swV4 = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
572 swV5 = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
573 swF2 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
574 swF3 = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
575 swF4 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
577 /* Avoid stupid compiler warnings */
578 jnrA = jnrB = jnrC = jnrD = 0;
587 for(iidx=0;iidx<4*DIM;iidx++)
592 /* Start outer loop over neighborlists */
593 for(iidx=0; iidx<nri; iidx++)
595 /* Load shift vector for this list */
596 i_shift_offset = DIM*shiftidx[iidx];
598 /* Load limits for loop over neighbors */
599 j_index_start = jindex[iidx];
600 j_index_end = jindex[iidx+1];
602 /* Get outer coordinate index */
604 i_coord_offset = DIM*inr;
606 /* Load i particle coords and add shift vector */
607 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
609 fix0 = _mm256_setzero_pd();
610 fiy0 = _mm256_setzero_pd();
611 fiz0 = _mm256_setzero_pd();
613 /* Load parameters for i particles */
614 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
615 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
617 /* Start inner kernel loop */
618 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
621 /* Get j neighbor index, and coordinate index */
626 j_coord_offsetA = DIM*jnrA;
627 j_coord_offsetB = DIM*jnrB;
628 j_coord_offsetC = DIM*jnrC;
629 j_coord_offsetD = DIM*jnrD;
631 /* load j atom coordinates */
632 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
633 x+j_coord_offsetC,x+j_coord_offsetD,
636 /* Calculate displacement vector */
637 dx00 = _mm256_sub_pd(ix0,jx0);
638 dy00 = _mm256_sub_pd(iy0,jy0);
639 dz00 = _mm256_sub_pd(iz0,jz0);
641 /* Calculate squared distance and things based on it */
642 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
644 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
646 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
648 /* Load parameters for j particles */
649 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
650 charge+jnrC+0,charge+jnrD+0);
651 vdwjidx0A = 2*vdwtype[jnrA+0];
652 vdwjidx0B = 2*vdwtype[jnrB+0];
653 vdwjidx0C = 2*vdwtype[jnrC+0];
654 vdwjidx0D = 2*vdwtype[jnrD+0];
656 /**************************
657 * CALCULATE INTERACTIONS *
658 **************************/
660 if (gmx_mm256_any_lt(rsq00,rcutoff2))
663 r00 = _mm256_mul_pd(rsq00,rinv00);
665 /* Compute parameters for interactions between i and j atoms */
666 qq00 = _mm256_mul_pd(iq0,jq0);
667 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
668 vdwioffsetptr0+vdwjidx0B,
669 vdwioffsetptr0+vdwjidx0C,
670 vdwioffsetptr0+vdwjidx0D,
673 /* EWALD ELECTROSTATICS */
675 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
676 ewrt = _mm256_mul_pd(r00,ewtabscale);
677 ewitab = _mm256_cvttpd_epi32(ewrt);
678 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
679 ewitab = _mm_slli_epi32(ewitab,2);
680 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
681 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
682 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
683 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
684 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
685 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
686 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
687 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
688 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
690 /* LENNARD-JONES DISPERSION/REPULSION */
692 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
693 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
694 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
695 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
696 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
698 d = _mm256_sub_pd(r00,rswitch);
699 d = _mm256_max_pd(d,_mm256_setzero_pd());
700 d2 = _mm256_mul_pd(d,d);
701 sw = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
703 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
705 /* Evaluate switch function */
706 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
707 felec = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
708 fvdw = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
709 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
711 fscal = _mm256_add_pd(felec,fvdw);
713 fscal = _mm256_and_pd(fscal,cutoff_mask);
715 /* Calculate temporary vectorial force */
716 tx = _mm256_mul_pd(fscal,dx00);
717 ty = _mm256_mul_pd(fscal,dy00);
718 tz = _mm256_mul_pd(fscal,dz00);
720 /* Update vectorial force */
721 fix0 = _mm256_add_pd(fix0,tx);
722 fiy0 = _mm256_add_pd(fiy0,ty);
723 fiz0 = _mm256_add_pd(fiz0,tz);
725 fjptrA = f+j_coord_offsetA;
726 fjptrB = f+j_coord_offsetB;
727 fjptrC = f+j_coord_offsetC;
728 fjptrD = f+j_coord_offsetD;
729 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
733 /* Inner loop uses 77 flops */
739 /* Get j neighbor index, and coordinate index */
740 jnrlistA = jjnr[jidx];
741 jnrlistB = jjnr[jidx+1];
742 jnrlistC = jjnr[jidx+2];
743 jnrlistD = jjnr[jidx+3];
744 /* Sign of each element will be negative for non-real atoms.
745 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
746 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
748 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
750 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
751 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
752 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
754 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
755 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
756 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
757 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
758 j_coord_offsetA = DIM*jnrA;
759 j_coord_offsetB = DIM*jnrB;
760 j_coord_offsetC = DIM*jnrC;
761 j_coord_offsetD = DIM*jnrD;
763 /* load j atom coordinates */
764 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
765 x+j_coord_offsetC,x+j_coord_offsetD,
768 /* Calculate displacement vector */
769 dx00 = _mm256_sub_pd(ix0,jx0);
770 dy00 = _mm256_sub_pd(iy0,jy0);
771 dz00 = _mm256_sub_pd(iz0,jz0);
773 /* Calculate squared distance and things based on it */
774 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
776 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
778 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
780 /* Load parameters for j particles */
781 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
782 charge+jnrC+0,charge+jnrD+0);
783 vdwjidx0A = 2*vdwtype[jnrA+0];
784 vdwjidx0B = 2*vdwtype[jnrB+0];
785 vdwjidx0C = 2*vdwtype[jnrC+0];
786 vdwjidx0D = 2*vdwtype[jnrD+0];
788 /**************************
789 * CALCULATE INTERACTIONS *
790 **************************/
792 if (gmx_mm256_any_lt(rsq00,rcutoff2))
795 r00 = _mm256_mul_pd(rsq00,rinv00);
796 r00 = _mm256_andnot_pd(dummy_mask,r00);
798 /* Compute parameters for interactions between i and j atoms */
799 qq00 = _mm256_mul_pd(iq0,jq0);
800 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
801 vdwioffsetptr0+vdwjidx0B,
802 vdwioffsetptr0+vdwjidx0C,
803 vdwioffsetptr0+vdwjidx0D,
806 /* EWALD ELECTROSTATICS */
808 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
809 ewrt = _mm256_mul_pd(r00,ewtabscale);
810 ewitab = _mm256_cvttpd_epi32(ewrt);
811 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
812 ewitab = _mm_slli_epi32(ewitab,2);
813 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
814 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
815 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
816 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
817 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
818 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
819 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
820 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
821 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
823 /* LENNARD-JONES DISPERSION/REPULSION */
825 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
826 vvdw6 = _mm256_mul_pd(c6_00,rinvsix);
827 vvdw12 = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
828 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
829 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
831 d = _mm256_sub_pd(r00,rswitch);
832 d = _mm256_max_pd(d,_mm256_setzero_pd());
833 d2 = _mm256_mul_pd(d,d);
834 sw = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
836 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
838 /* Evaluate switch function */
839 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
840 felec = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
841 fvdw = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(vvdw,dsw)) );
842 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
844 fscal = _mm256_add_pd(felec,fvdw);
846 fscal = _mm256_and_pd(fscal,cutoff_mask);
848 fscal = _mm256_andnot_pd(dummy_mask,fscal);
850 /* Calculate temporary vectorial force */
851 tx = _mm256_mul_pd(fscal,dx00);
852 ty = _mm256_mul_pd(fscal,dy00);
853 tz = _mm256_mul_pd(fscal,dz00);
855 /* Update vectorial force */
856 fix0 = _mm256_add_pd(fix0,tx);
857 fiy0 = _mm256_add_pd(fiy0,ty);
858 fiz0 = _mm256_add_pd(fiz0,tz);
860 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
861 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
862 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
863 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
864 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
868 /* Inner loop uses 78 flops */
871 /* End of innermost loop */
873 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
874 f+i_coord_offset,fshift+i_shift_offset);
876 /* Increment number of inner iterations */
877 inneriter += j_index_end - j_index_start;
879 /* Outer loop uses 7 flops */
882 /* Increment number of outer iterations */
885 /* Update outer/inner flops */
887 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*78);