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_VdwNone_GeomP1P1_VF_avx_256_double
38 * Electrostatics interaction: Ewald
39 * VdW interaction: None
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEwSw_VdwNone_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 ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
79 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
81 __m256d rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
82 real rswitch_scalar,d_scalar;
83 __m256d dummy_mask,cutoff_mask;
84 __m128 tmpmask0,tmpmask1;
85 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
86 __m256d one = _mm256_set1_pd(1.0);
87 __m256d two = _mm256_set1_pd(2.0);
93 jindex = nlist->jindex;
95 shiftidx = nlist->shift;
97 shiftvec = fr->shift_vec[0];
98 fshift = fr->fshift[0];
99 facel = _mm256_set1_pd(fr->epsfac);
100 charge = mdatoms->chargeA;
102 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
103 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
104 beta2 = _mm256_mul_pd(beta,beta);
105 beta3 = _mm256_mul_pd(beta,beta2);
107 ewtab = fr->ic->tabq_coul_FDV0;
108 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
109 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
111 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
112 rcutoff_scalar = fr->rcoulomb;
113 rcutoff = _mm256_set1_pd(rcutoff_scalar);
114 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
116 rswitch_scalar = fr->rcoulomb_switch;
117 rswitch = _mm256_set1_pd(rswitch_scalar);
118 /* Setup switch parameters */
119 d_scalar = rcutoff_scalar-rswitch_scalar;
120 d = _mm256_set1_pd(d_scalar);
121 swV3 = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
122 swV4 = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
123 swV5 = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
124 swF2 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
125 swF3 = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
126 swF4 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
128 /* Avoid stupid compiler warnings */
129 jnrA = jnrB = jnrC = jnrD = 0;
138 for(iidx=0;iidx<4*DIM;iidx++)
143 /* Start outer loop over neighborlists */
144 for(iidx=0; iidx<nri; iidx++)
146 /* Load shift vector for this list */
147 i_shift_offset = DIM*shiftidx[iidx];
149 /* Load limits for loop over neighbors */
150 j_index_start = jindex[iidx];
151 j_index_end = jindex[iidx+1];
153 /* Get outer coordinate index */
155 i_coord_offset = DIM*inr;
157 /* Load i particle coords and add shift vector */
158 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
160 fix0 = _mm256_setzero_pd();
161 fiy0 = _mm256_setzero_pd();
162 fiz0 = _mm256_setzero_pd();
164 /* Load parameters for i particles */
165 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
167 /* Reset potential sums */
168 velecsum = _mm256_setzero_pd();
170 /* Start inner kernel loop */
171 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
174 /* Get j neighbor index, and coordinate index */
179 j_coord_offsetA = DIM*jnrA;
180 j_coord_offsetB = DIM*jnrB;
181 j_coord_offsetC = DIM*jnrC;
182 j_coord_offsetD = DIM*jnrD;
184 /* load j atom coordinates */
185 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
186 x+j_coord_offsetC,x+j_coord_offsetD,
189 /* Calculate displacement vector */
190 dx00 = _mm256_sub_pd(ix0,jx0);
191 dy00 = _mm256_sub_pd(iy0,jy0);
192 dz00 = _mm256_sub_pd(iz0,jz0);
194 /* Calculate squared distance and things based on it */
195 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
197 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
199 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
201 /* Load parameters for j particles */
202 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
203 charge+jnrC+0,charge+jnrD+0);
205 /**************************
206 * CALCULATE INTERACTIONS *
207 **************************/
209 if (gmx_mm256_any_lt(rsq00,rcutoff2))
212 r00 = _mm256_mul_pd(rsq00,rinv00);
214 /* Compute parameters for interactions between i and j atoms */
215 qq00 = _mm256_mul_pd(iq0,jq0);
217 /* EWALD ELECTROSTATICS */
219 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
220 ewrt = _mm256_mul_pd(r00,ewtabscale);
221 ewitab = _mm256_cvttpd_epi32(ewrt);
222 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
223 ewitab = _mm_slli_epi32(ewitab,2);
224 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
225 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
226 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
227 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
228 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
229 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
230 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
231 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
232 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
234 d = _mm256_sub_pd(r00,rswitch);
235 d = _mm256_max_pd(d,_mm256_setzero_pd());
236 d2 = _mm256_mul_pd(d,d);
237 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)))))));
239 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
241 /* Evaluate switch function */
242 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
243 felec = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
244 velec = _mm256_mul_pd(velec,sw);
245 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
247 /* Update potential sum for this i atom from the interaction with this j atom. */
248 velec = _mm256_and_pd(velec,cutoff_mask);
249 velecsum = _mm256_add_pd(velecsum,velec);
253 fscal = _mm256_and_pd(fscal,cutoff_mask);
255 /* Calculate temporary vectorial force */
256 tx = _mm256_mul_pd(fscal,dx00);
257 ty = _mm256_mul_pd(fscal,dy00);
258 tz = _mm256_mul_pd(fscal,dz00);
260 /* Update vectorial force */
261 fix0 = _mm256_add_pd(fix0,tx);
262 fiy0 = _mm256_add_pd(fiy0,ty);
263 fiz0 = _mm256_add_pd(fiz0,tz);
265 fjptrA = f+j_coord_offsetA;
266 fjptrB = f+j_coord_offsetB;
267 fjptrC = f+j_coord_offsetC;
268 fjptrD = f+j_coord_offsetD;
269 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
273 /* Inner loop uses 65 flops */
279 /* Get j neighbor index, and coordinate index */
280 jnrlistA = jjnr[jidx];
281 jnrlistB = jjnr[jidx+1];
282 jnrlistC = jjnr[jidx+2];
283 jnrlistD = jjnr[jidx+3];
284 /* Sign of each element will be negative for non-real atoms.
285 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
286 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
288 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
290 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
291 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
292 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
294 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
295 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
296 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
297 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
298 j_coord_offsetA = DIM*jnrA;
299 j_coord_offsetB = DIM*jnrB;
300 j_coord_offsetC = DIM*jnrC;
301 j_coord_offsetD = DIM*jnrD;
303 /* load j atom coordinates */
304 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
305 x+j_coord_offsetC,x+j_coord_offsetD,
308 /* Calculate displacement vector */
309 dx00 = _mm256_sub_pd(ix0,jx0);
310 dy00 = _mm256_sub_pd(iy0,jy0);
311 dz00 = _mm256_sub_pd(iz0,jz0);
313 /* Calculate squared distance and things based on it */
314 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
316 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
318 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
320 /* Load parameters for j particles */
321 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
322 charge+jnrC+0,charge+jnrD+0);
324 /**************************
325 * CALCULATE INTERACTIONS *
326 **************************/
328 if (gmx_mm256_any_lt(rsq00,rcutoff2))
331 r00 = _mm256_mul_pd(rsq00,rinv00);
332 r00 = _mm256_andnot_pd(dummy_mask,r00);
334 /* Compute parameters for interactions between i and j atoms */
335 qq00 = _mm256_mul_pd(iq0,jq0);
337 /* EWALD ELECTROSTATICS */
339 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
340 ewrt = _mm256_mul_pd(r00,ewtabscale);
341 ewitab = _mm256_cvttpd_epi32(ewrt);
342 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
343 ewitab = _mm_slli_epi32(ewitab,2);
344 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
345 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
346 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
347 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
348 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
349 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
350 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
351 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
352 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
354 d = _mm256_sub_pd(r00,rswitch);
355 d = _mm256_max_pd(d,_mm256_setzero_pd());
356 d2 = _mm256_mul_pd(d,d);
357 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)))))));
359 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
361 /* Evaluate switch function */
362 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
363 felec = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
364 velec = _mm256_mul_pd(velec,sw);
365 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
367 /* Update potential sum for this i atom from the interaction with this j atom. */
368 velec = _mm256_and_pd(velec,cutoff_mask);
369 velec = _mm256_andnot_pd(dummy_mask,velec);
370 velecsum = _mm256_add_pd(velecsum,velec);
374 fscal = _mm256_and_pd(fscal,cutoff_mask);
376 fscal = _mm256_andnot_pd(dummy_mask,fscal);
378 /* Calculate temporary vectorial force */
379 tx = _mm256_mul_pd(fscal,dx00);
380 ty = _mm256_mul_pd(fscal,dy00);
381 tz = _mm256_mul_pd(fscal,dz00);
383 /* Update vectorial force */
384 fix0 = _mm256_add_pd(fix0,tx);
385 fiy0 = _mm256_add_pd(fiy0,ty);
386 fiz0 = _mm256_add_pd(fiz0,tz);
388 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
389 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
390 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
391 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
392 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
396 /* Inner loop uses 66 flops */
399 /* End of innermost loop */
401 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
402 f+i_coord_offset,fshift+i_shift_offset);
405 /* Update potential energies */
406 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
408 /* Increment number of inner iterations */
409 inneriter += j_index_end - j_index_start;
411 /* Outer loop uses 8 flops */
414 /* Increment number of outer iterations */
417 /* Update outer/inner flops */
419 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*66);
422 * Gromacs nonbonded kernel: nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_double
423 * Electrostatics interaction: Ewald
424 * VdW interaction: None
425 * Geometry: Particle-Particle
426 * Calculate force/pot: Force
429 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_double
430 (t_nblist * gmx_restrict nlist,
431 rvec * gmx_restrict xx,
432 rvec * gmx_restrict ff,
433 t_forcerec * gmx_restrict fr,
434 t_mdatoms * gmx_restrict mdatoms,
435 nb_kernel_data_t * gmx_restrict kernel_data,
436 t_nrnb * gmx_restrict nrnb)
438 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
439 * just 0 for non-waters.
440 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
441 * jnr indices corresponding to data put in the four positions in the SIMD register.
443 int i_shift_offset,i_coord_offset,outeriter,inneriter;
444 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
445 int jnrA,jnrB,jnrC,jnrD;
446 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
447 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
448 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
449 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
451 real *shiftvec,*fshift,*x,*f;
452 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
454 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
455 real * vdwioffsetptr0;
456 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
457 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
458 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
459 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
460 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
463 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
464 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
466 __m256d rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
467 real rswitch_scalar,d_scalar;
468 __m256d dummy_mask,cutoff_mask;
469 __m128 tmpmask0,tmpmask1;
470 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
471 __m256d one = _mm256_set1_pd(1.0);
472 __m256d two = _mm256_set1_pd(2.0);
478 jindex = nlist->jindex;
480 shiftidx = nlist->shift;
482 shiftvec = fr->shift_vec[0];
483 fshift = fr->fshift[0];
484 facel = _mm256_set1_pd(fr->epsfac);
485 charge = mdatoms->chargeA;
487 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
488 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
489 beta2 = _mm256_mul_pd(beta,beta);
490 beta3 = _mm256_mul_pd(beta,beta2);
492 ewtab = fr->ic->tabq_coul_FDV0;
493 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
494 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
496 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
497 rcutoff_scalar = fr->rcoulomb;
498 rcutoff = _mm256_set1_pd(rcutoff_scalar);
499 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
501 rswitch_scalar = fr->rcoulomb_switch;
502 rswitch = _mm256_set1_pd(rswitch_scalar);
503 /* Setup switch parameters */
504 d_scalar = rcutoff_scalar-rswitch_scalar;
505 d = _mm256_set1_pd(d_scalar);
506 swV3 = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
507 swV4 = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
508 swV5 = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
509 swF2 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
510 swF3 = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
511 swF4 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
513 /* Avoid stupid compiler warnings */
514 jnrA = jnrB = jnrC = jnrD = 0;
523 for(iidx=0;iidx<4*DIM;iidx++)
528 /* Start outer loop over neighborlists */
529 for(iidx=0; iidx<nri; iidx++)
531 /* Load shift vector for this list */
532 i_shift_offset = DIM*shiftidx[iidx];
534 /* Load limits for loop over neighbors */
535 j_index_start = jindex[iidx];
536 j_index_end = jindex[iidx+1];
538 /* Get outer coordinate index */
540 i_coord_offset = DIM*inr;
542 /* Load i particle coords and add shift vector */
543 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
545 fix0 = _mm256_setzero_pd();
546 fiy0 = _mm256_setzero_pd();
547 fiz0 = _mm256_setzero_pd();
549 /* Load parameters for i particles */
550 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
552 /* Start inner kernel loop */
553 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
556 /* Get j neighbor index, and coordinate index */
561 j_coord_offsetA = DIM*jnrA;
562 j_coord_offsetB = DIM*jnrB;
563 j_coord_offsetC = DIM*jnrC;
564 j_coord_offsetD = DIM*jnrD;
566 /* load j atom coordinates */
567 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
568 x+j_coord_offsetC,x+j_coord_offsetD,
571 /* Calculate displacement vector */
572 dx00 = _mm256_sub_pd(ix0,jx0);
573 dy00 = _mm256_sub_pd(iy0,jy0);
574 dz00 = _mm256_sub_pd(iz0,jz0);
576 /* Calculate squared distance and things based on it */
577 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
579 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
581 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
583 /* Load parameters for j particles */
584 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
585 charge+jnrC+0,charge+jnrD+0);
587 /**************************
588 * CALCULATE INTERACTIONS *
589 **************************/
591 if (gmx_mm256_any_lt(rsq00,rcutoff2))
594 r00 = _mm256_mul_pd(rsq00,rinv00);
596 /* Compute parameters for interactions between i and j atoms */
597 qq00 = _mm256_mul_pd(iq0,jq0);
599 /* EWALD ELECTROSTATICS */
601 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
602 ewrt = _mm256_mul_pd(r00,ewtabscale);
603 ewitab = _mm256_cvttpd_epi32(ewrt);
604 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
605 ewitab = _mm_slli_epi32(ewitab,2);
606 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
607 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
608 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
609 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
610 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
611 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
612 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
613 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
614 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
616 d = _mm256_sub_pd(r00,rswitch);
617 d = _mm256_max_pd(d,_mm256_setzero_pd());
618 d2 = _mm256_mul_pd(d,d);
619 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)))))));
621 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
623 /* Evaluate switch function */
624 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
625 felec = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
626 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
630 fscal = _mm256_and_pd(fscal,cutoff_mask);
632 /* Calculate temporary vectorial force */
633 tx = _mm256_mul_pd(fscal,dx00);
634 ty = _mm256_mul_pd(fscal,dy00);
635 tz = _mm256_mul_pd(fscal,dz00);
637 /* Update vectorial force */
638 fix0 = _mm256_add_pd(fix0,tx);
639 fiy0 = _mm256_add_pd(fiy0,ty);
640 fiz0 = _mm256_add_pd(fiz0,tz);
642 fjptrA = f+j_coord_offsetA;
643 fjptrB = f+j_coord_offsetB;
644 fjptrC = f+j_coord_offsetC;
645 fjptrD = f+j_coord_offsetD;
646 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
650 /* Inner loop uses 62 flops */
656 /* Get j neighbor index, and coordinate index */
657 jnrlistA = jjnr[jidx];
658 jnrlistB = jjnr[jidx+1];
659 jnrlistC = jjnr[jidx+2];
660 jnrlistD = jjnr[jidx+3];
661 /* Sign of each element will be negative for non-real atoms.
662 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
663 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
665 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
667 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
668 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
669 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
671 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
672 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
673 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
674 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
675 j_coord_offsetA = DIM*jnrA;
676 j_coord_offsetB = DIM*jnrB;
677 j_coord_offsetC = DIM*jnrC;
678 j_coord_offsetD = DIM*jnrD;
680 /* load j atom coordinates */
681 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
682 x+j_coord_offsetC,x+j_coord_offsetD,
685 /* Calculate displacement vector */
686 dx00 = _mm256_sub_pd(ix0,jx0);
687 dy00 = _mm256_sub_pd(iy0,jy0);
688 dz00 = _mm256_sub_pd(iz0,jz0);
690 /* Calculate squared distance and things based on it */
691 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
693 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
695 rinvsq00 = _mm256_mul_pd(rinv00,rinv00);
697 /* Load parameters for j particles */
698 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
699 charge+jnrC+0,charge+jnrD+0);
701 /**************************
702 * CALCULATE INTERACTIONS *
703 **************************/
705 if (gmx_mm256_any_lt(rsq00,rcutoff2))
708 r00 = _mm256_mul_pd(rsq00,rinv00);
709 r00 = _mm256_andnot_pd(dummy_mask,r00);
711 /* Compute parameters for interactions between i and j atoms */
712 qq00 = _mm256_mul_pd(iq0,jq0);
714 /* EWALD ELECTROSTATICS */
716 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
717 ewrt = _mm256_mul_pd(r00,ewtabscale);
718 ewitab = _mm256_cvttpd_epi32(ewrt);
719 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
720 ewitab = _mm_slli_epi32(ewitab,2);
721 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
722 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
723 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
724 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
725 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
726 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
727 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
728 velec = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
729 felec = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
731 d = _mm256_sub_pd(r00,rswitch);
732 d = _mm256_max_pd(d,_mm256_setzero_pd());
733 d2 = _mm256_mul_pd(d,d);
734 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)))))));
736 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
738 /* Evaluate switch function */
739 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
740 felec = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
741 cutoff_mask = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
745 fscal = _mm256_and_pd(fscal,cutoff_mask);
747 fscal = _mm256_andnot_pd(dummy_mask,fscal);
749 /* Calculate temporary vectorial force */
750 tx = _mm256_mul_pd(fscal,dx00);
751 ty = _mm256_mul_pd(fscal,dy00);
752 tz = _mm256_mul_pd(fscal,dz00);
754 /* Update vectorial force */
755 fix0 = _mm256_add_pd(fix0,tx);
756 fiy0 = _mm256_add_pd(fiy0,ty);
757 fiz0 = _mm256_add_pd(fiz0,tz);
759 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
760 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
761 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
762 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
763 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
767 /* Inner loop uses 63 flops */
770 /* End of innermost loop */
772 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
773 f+i_coord_offset,fshift+i_shift_offset);
775 /* Increment number of inner iterations */
776 inneriter += j_index_end - j_index_start;
778 /* Outer loop uses 7 flops */
781 /* Increment number of outer iterations */
784 /* Update outer/inner flops */
786 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*63);