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_ElecCSTab_VdwCSTab_GeomP1P1_VF_avx_256_double
38 * Electrostatics interaction: CubicSplineTable
39 * VdW interaction: CubicSplineTable
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecCSTab_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;
87 __m256d dummy_mask,cutoff_mask;
88 __m128 tmpmask0,tmpmask1;
89 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
90 __m256d one = _mm256_set1_pd(1.0);
91 __m256d two = _mm256_set1_pd(2.0);
97 jindex = nlist->jindex;
99 shiftidx = nlist->shift;
101 shiftvec = fr->shift_vec[0];
102 fshift = fr->fshift[0];
103 facel = _mm256_set1_pd(fr->epsfac);
104 charge = mdatoms->chargeA;
105 nvdwtype = fr->ntype;
107 vdwtype = mdatoms->typeA;
109 vftab = kernel_data->table_elec_vdw->data;
110 vftabscale = _mm256_set1_pd(kernel_data->table_elec_vdw->scale);
112 /* Avoid stupid compiler warnings */
113 jnrA = jnrB = jnrC = jnrD = 0;
122 for(iidx=0;iidx<4*DIM;iidx++)
127 /* Start outer loop over neighborlists */
128 for(iidx=0; iidx<nri; iidx++)
130 /* Load shift vector for this list */
131 i_shift_offset = DIM*shiftidx[iidx];
133 /* Load limits for loop over neighbors */
134 j_index_start = jindex[iidx];
135 j_index_end = jindex[iidx+1];
137 /* Get outer coordinate index */
139 i_coord_offset = DIM*inr;
141 /* Load i particle coords and add shift vector */
142 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
144 fix0 = _mm256_setzero_pd();
145 fiy0 = _mm256_setzero_pd();
146 fiz0 = _mm256_setzero_pd();
148 /* Load parameters for i particles */
149 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
150 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
152 /* Reset potential sums */
153 velecsum = _mm256_setzero_pd();
154 vvdwsum = _mm256_setzero_pd();
156 /* Start inner kernel loop */
157 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
160 /* Get j neighbor index, and coordinate index */
165 j_coord_offsetA = DIM*jnrA;
166 j_coord_offsetB = DIM*jnrB;
167 j_coord_offsetC = DIM*jnrC;
168 j_coord_offsetD = DIM*jnrD;
170 /* load j atom coordinates */
171 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
172 x+j_coord_offsetC,x+j_coord_offsetD,
175 /* Calculate displacement vector */
176 dx00 = _mm256_sub_pd(ix0,jx0);
177 dy00 = _mm256_sub_pd(iy0,jy0);
178 dz00 = _mm256_sub_pd(iz0,jz0);
180 /* Calculate squared distance and things based on it */
181 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
183 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
185 /* Load parameters for j particles */
186 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
187 charge+jnrC+0,charge+jnrD+0);
188 vdwjidx0A = 2*vdwtype[jnrA+0];
189 vdwjidx0B = 2*vdwtype[jnrB+0];
190 vdwjidx0C = 2*vdwtype[jnrC+0];
191 vdwjidx0D = 2*vdwtype[jnrD+0];
193 /**************************
194 * CALCULATE INTERACTIONS *
195 **************************/
197 r00 = _mm256_mul_pd(rsq00,rinv00);
199 /* Compute parameters for interactions between i and j atoms */
200 qq00 = _mm256_mul_pd(iq0,jq0);
201 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
202 vdwioffsetptr0+vdwjidx0B,
203 vdwioffsetptr0+vdwjidx0C,
204 vdwioffsetptr0+vdwjidx0D,
207 /* Calculate table index by multiplying r with table scale and truncate to integer */
208 rt = _mm256_mul_pd(r00,vftabscale);
209 vfitab = _mm256_cvttpd_epi32(rt);
210 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
211 vfitab = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
213 /* CUBIC SPLINE TABLE ELECTROSTATICS */
214 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
215 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
216 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
217 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
218 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
219 Heps = _mm256_mul_pd(vfeps,H);
220 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
221 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
222 velec = _mm256_mul_pd(qq00,VV);
223 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
224 felec = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
226 /* CUBIC SPLINE TABLE DISPERSION */
227 vfitab = _mm_add_epi32(vfitab,ifour);
228 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
229 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
230 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
231 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
232 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
233 Heps = _mm256_mul_pd(vfeps,H);
234 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
235 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
236 vvdw6 = _mm256_mul_pd(c6_00,VV);
237 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
238 fvdw6 = _mm256_mul_pd(c6_00,FF);
240 /* CUBIC SPLINE TABLE REPULSION */
241 vfitab = _mm_add_epi32(vfitab,ifour);
242 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
243 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
244 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
245 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
246 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
247 Heps = _mm256_mul_pd(vfeps,H);
248 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
249 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
250 vvdw12 = _mm256_mul_pd(c12_00,VV);
251 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
252 fvdw12 = _mm256_mul_pd(c12_00,FF);
253 vvdw = _mm256_add_pd(vvdw12,vvdw6);
254 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
256 /* Update potential sum for this i atom from the interaction with this j atom. */
257 velecsum = _mm256_add_pd(velecsum,velec);
258 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
260 fscal = _mm256_add_pd(felec,fvdw);
262 /* Calculate temporary vectorial force */
263 tx = _mm256_mul_pd(fscal,dx00);
264 ty = _mm256_mul_pd(fscal,dy00);
265 tz = _mm256_mul_pd(fscal,dz00);
267 /* Update vectorial force */
268 fix0 = _mm256_add_pd(fix0,tx);
269 fiy0 = _mm256_add_pd(fiy0,ty);
270 fiz0 = _mm256_add_pd(fiz0,tz);
272 fjptrA = f+j_coord_offsetA;
273 fjptrB = f+j_coord_offsetB;
274 fjptrC = f+j_coord_offsetC;
275 fjptrD = f+j_coord_offsetD;
276 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
278 /* Inner loop uses 73 flops */
284 /* Get j neighbor index, and coordinate index */
285 jnrlistA = jjnr[jidx];
286 jnrlistB = jjnr[jidx+1];
287 jnrlistC = jjnr[jidx+2];
288 jnrlistD = jjnr[jidx+3];
289 /* Sign of each element will be negative for non-real atoms.
290 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
291 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
293 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
295 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
296 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
297 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
299 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
300 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
301 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
302 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
303 j_coord_offsetA = DIM*jnrA;
304 j_coord_offsetB = DIM*jnrB;
305 j_coord_offsetC = DIM*jnrC;
306 j_coord_offsetD = DIM*jnrD;
308 /* load j atom coordinates */
309 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
310 x+j_coord_offsetC,x+j_coord_offsetD,
313 /* Calculate displacement vector */
314 dx00 = _mm256_sub_pd(ix0,jx0);
315 dy00 = _mm256_sub_pd(iy0,jy0);
316 dz00 = _mm256_sub_pd(iz0,jz0);
318 /* Calculate squared distance and things based on it */
319 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
321 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
323 /* Load parameters for j particles */
324 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
325 charge+jnrC+0,charge+jnrD+0);
326 vdwjidx0A = 2*vdwtype[jnrA+0];
327 vdwjidx0B = 2*vdwtype[jnrB+0];
328 vdwjidx0C = 2*vdwtype[jnrC+0];
329 vdwjidx0D = 2*vdwtype[jnrD+0];
331 /**************************
332 * CALCULATE INTERACTIONS *
333 **************************/
335 r00 = _mm256_mul_pd(rsq00,rinv00);
336 r00 = _mm256_andnot_pd(dummy_mask,r00);
338 /* Compute parameters for interactions between i and j atoms */
339 qq00 = _mm256_mul_pd(iq0,jq0);
340 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
341 vdwioffsetptr0+vdwjidx0B,
342 vdwioffsetptr0+vdwjidx0C,
343 vdwioffsetptr0+vdwjidx0D,
346 /* Calculate table index by multiplying r with table scale and truncate to integer */
347 rt = _mm256_mul_pd(r00,vftabscale);
348 vfitab = _mm256_cvttpd_epi32(rt);
349 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
350 vfitab = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
352 /* CUBIC SPLINE TABLE ELECTROSTATICS */
353 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
354 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
355 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
356 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
357 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
358 Heps = _mm256_mul_pd(vfeps,H);
359 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
360 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
361 velec = _mm256_mul_pd(qq00,VV);
362 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
363 felec = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
365 /* CUBIC SPLINE TABLE DISPERSION */
366 vfitab = _mm_add_epi32(vfitab,ifour);
367 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
368 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
369 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
370 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
371 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
372 Heps = _mm256_mul_pd(vfeps,H);
373 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
374 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
375 vvdw6 = _mm256_mul_pd(c6_00,VV);
376 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
377 fvdw6 = _mm256_mul_pd(c6_00,FF);
379 /* CUBIC SPLINE TABLE REPULSION */
380 vfitab = _mm_add_epi32(vfitab,ifour);
381 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
382 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
383 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
384 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
385 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
386 Heps = _mm256_mul_pd(vfeps,H);
387 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
388 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
389 vvdw12 = _mm256_mul_pd(c12_00,VV);
390 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
391 fvdw12 = _mm256_mul_pd(c12_00,FF);
392 vvdw = _mm256_add_pd(vvdw12,vvdw6);
393 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
395 /* Update potential sum for this i atom from the interaction with this j atom. */
396 velec = _mm256_andnot_pd(dummy_mask,velec);
397 velecsum = _mm256_add_pd(velecsum,velec);
398 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
399 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
401 fscal = _mm256_add_pd(felec,fvdw);
403 fscal = _mm256_andnot_pd(dummy_mask,fscal);
405 /* Calculate temporary vectorial force */
406 tx = _mm256_mul_pd(fscal,dx00);
407 ty = _mm256_mul_pd(fscal,dy00);
408 tz = _mm256_mul_pd(fscal,dz00);
410 /* Update vectorial force */
411 fix0 = _mm256_add_pd(fix0,tx);
412 fiy0 = _mm256_add_pd(fiy0,ty);
413 fiz0 = _mm256_add_pd(fiz0,tz);
415 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
416 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
417 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
418 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
419 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
421 /* Inner loop uses 74 flops */
424 /* End of innermost loop */
426 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
427 f+i_coord_offset,fshift+i_shift_offset);
430 /* Update potential energies */
431 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
432 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
434 /* Increment number of inner iterations */
435 inneriter += j_index_end - j_index_start;
437 /* Outer loop uses 9 flops */
440 /* Increment number of outer iterations */
443 /* Update outer/inner flops */
445 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*74);
448 * Gromacs nonbonded kernel: nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_256_double
449 * Electrostatics interaction: CubicSplineTable
450 * VdW interaction: CubicSplineTable
451 * Geometry: Particle-Particle
452 * Calculate force/pot: Force
455 nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_256_double
456 (t_nblist * gmx_restrict nlist,
457 rvec * gmx_restrict xx,
458 rvec * gmx_restrict ff,
459 t_forcerec * gmx_restrict fr,
460 t_mdatoms * gmx_restrict mdatoms,
461 nb_kernel_data_t * gmx_restrict kernel_data,
462 t_nrnb * gmx_restrict nrnb)
464 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
465 * just 0 for non-waters.
466 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
467 * jnr indices corresponding to data put in the four positions in the SIMD register.
469 int i_shift_offset,i_coord_offset,outeriter,inneriter;
470 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
471 int jnrA,jnrB,jnrC,jnrD;
472 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
473 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
474 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
475 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
477 real *shiftvec,*fshift,*x,*f;
478 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
480 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
481 real * vdwioffsetptr0;
482 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
483 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
484 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
485 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
486 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
489 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
492 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
493 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
495 __m128i ifour = _mm_set1_epi32(4);
496 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
498 __m256d dummy_mask,cutoff_mask;
499 __m128 tmpmask0,tmpmask1;
500 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
501 __m256d one = _mm256_set1_pd(1.0);
502 __m256d two = _mm256_set1_pd(2.0);
508 jindex = nlist->jindex;
510 shiftidx = nlist->shift;
512 shiftvec = fr->shift_vec[0];
513 fshift = fr->fshift[0];
514 facel = _mm256_set1_pd(fr->epsfac);
515 charge = mdatoms->chargeA;
516 nvdwtype = fr->ntype;
518 vdwtype = mdatoms->typeA;
520 vftab = kernel_data->table_elec_vdw->data;
521 vftabscale = _mm256_set1_pd(kernel_data->table_elec_vdw->scale);
523 /* Avoid stupid compiler warnings */
524 jnrA = jnrB = jnrC = jnrD = 0;
533 for(iidx=0;iidx<4*DIM;iidx++)
538 /* Start outer loop over neighborlists */
539 for(iidx=0; iidx<nri; iidx++)
541 /* Load shift vector for this list */
542 i_shift_offset = DIM*shiftidx[iidx];
544 /* Load limits for loop over neighbors */
545 j_index_start = jindex[iidx];
546 j_index_end = jindex[iidx+1];
548 /* Get outer coordinate index */
550 i_coord_offset = DIM*inr;
552 /* Load i particle coords and add shift vector */
553 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
555 fix0 = _mm256_setzero_pd();
556 fiy0 = _mm256_setzero_pd();
557 fiz0 = _mm256_setzero_pd();
559 /* Load parameters for i particles */
560 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
561 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
563 /* Start inner kernel loop */
564 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
567 /* Get j neighbor index, and coordinate index */
572 j_coord_offsetA = DIM*jnrA;
573 j_coord_offsetB = DIM*jnrB;
574 j_coord_offsetC = DIM*jnrC;
575 j_coord_offsetD = DIM*jnrD;
577 /* load j atom coordinates */
578 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
579 x+j_coord_offsetC,x+j_coord_offsetD,
582 /* Calculate displacement vector */
583 dx00 = _mm256_sub_pd(ix0,jx0);
584 dy00 = _mm256_sub_pd(iy0,jy0);
585 dz00 = _mm256_sub_pd(iz0,jz0);
587 /* Calculate squared distance and things based on it */
588 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
590 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
592 /* Load parameters for j particles */
593 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
594 charge+jnrC+0,charge+jnrD+0);
595 vdwjidx0A = 2*vdwtype[jnrA+0];
596 vdwjidx0B = 2*vdwtype[jnrB+0];
597 vdwjidx0C = 2*vdwtype[jnrC+0];
598 vdwjidx0D = 2*vdwtype[jnrD+0];
600 /**************************
601 * CALCULATE INTERACTIONS *
602 **************************/
604 r00 = _mm256_mul_pd(rsq00,rinv00);
606 /* Compute parameters for interactions between i and j atoms */
607 qq00 = _mm256_mul_pd(iq0,jq0);
608 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
609 vdwioffsetptr0+vdwjidx0B,
610 vdwioffsetptr0+vdwjidx0C,
611 vdwioffsetptr0+vdwjidx0D,
614 /* Calculate table index by multiplying r with table scale and truncate to integer */
615 rt = _mm256_mul_pd(r00,vftabscale);
616 vfitab = _mm256_cvttpd_epi32(rt);
617 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
618 vfitab = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
620 /* CUBIC SPLINE TABLE ELECTROSTATICS */
621 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
622 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
623 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
624 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
625 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
626 Heps = _mm256_mul_pd(vfeps,H);
627 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
628 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
629 felec = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
631 /* CUBIC SPLINE TABLE DISPERSION */
632 vfitab = _mm_add_epi32(vfitab,ifour);
633 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
634 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
635 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
636 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
637 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
638 Heps = _mm256_mul_pd(vfeps,H);
639 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
640 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
641 fvdw6 = _mm256_mul_pd(c6_00,FF);
643 /* CUBIC SPLINE TABLE REPULSION */
644 vfitab = _mm_add_epi32(vfitab,ifour);
645 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
646 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
647 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
648 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
649 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
650 Heps = _mm256_mul_pd(vfeps,H);
651 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
652 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
653 fvdw12 = _mm256_mul_pd(c12_00,FF);
654 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
656 fscal = _mm256_add_pd(felec,fvdw);
658 /* Calculate temporary vectorial force */
659 tx = _mm256_mul_pd(fscal,dx00);
660 ty = _mm256_mul_pd(fscal,dy00);
661 tz = _mm256_mul_pd(fscal,dz00);
663 /* Update vectorial force */
664 fix0 = _mm256_add_pd(fix0,tx);
665 fiy0 = _mm256_add_pd(fiy0,ty);
666 fiz0 = _mm256_add_pd(fiz0,tz);
668 fjptrA = f+j_coord_offsetA;
669 fjptrB = f+j_coord_offsetB;
670 fjptrC = f+j_coord_offsetC;
671 fjptrD = f+j_coord_offsetD;
672 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
674 /* Inner loop uses 61 flops */
680 /* Get j neighbor index, and coordinate index */
681 jnrlistA = jjnr[jidx];
682 jnrlistB = jjnr[jidx+1];
683 jnrlistC = jjnr[jidx+2];
684 jnrlistD = jjnr[jidx+3];
685 /* Sign of each element will be negative for non-real atoms.
686 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
687 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
689 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
691 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
692 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
693 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
695 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
696 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
697 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
698 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
699 j_coord_offsetA = DIM*jnrA;
700 j_coord_offsetB = DIM*jnrB;
701 j_coord_offsetC = DIM*jnrC;
702 j_coord_offsetD = DIM*jnrD;
704 /* load j atom coordinates */
705 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
706 x+j_coord_offsetC,x+j_coord_offsetD,
709 /* Calculate displacement vector */
710 dx00 = _mm256_sub_pd(ix0,jx0);
711 dy00 = _mm256_sub_pd(iy0,jy0);
712 dz00 = _mm256_sub_pd(iz0,jz0);
714 /* Calculate squared distance and things based on it */
715 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
717 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
719 /* Load parameters for j particles */
720 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
721 charge+jnrC+0,charge+jnrD+0);
722 vdwjidx0A = 2*vdwtype[jnrA+0];
723 vdwjidx0B = 2*vdwtype[jnrB+0];
724 vdwjidx0C = 2*vdwtype[jnrC+0];
725 vdwjidx0D = 2*vdwtype[jnrD+0];
727 /**************************
728 * CALCULATE INTERACTIONS *
729 **************************/
731 r00 = _mm256_mul_pd(rsq00,rinv00);
732 r00 = _mm256_andnot_pd(dummy_mask,r00);
734 /* Compute parameters for interactions between i and j atoms */
735 qq00 = _mm256_mul_pd(iq0,jq0);
736 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
737 vdwioffsetptr0+vdwjidx0B,
738 vdwioffsetptr0+vdwjidx0C,
739 vdwioffsetptr0+vdwjidx0D,
742 /* Calculate table index by multiplying r with table scale and truncate to integer */
743 rt = _mm256_mul_pd(r00,vftabscale);
744 vfitab = _mm256_cvttpd_epi32(rt);
745 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
746 vfitab = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
748 /* CUBIC SPLINE TABLE ELECTROSTATICS */
749 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
750 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
751 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
752 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
753 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
754 Heps = _mm256_mul_pd(vfeps,H);
755 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
756 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
757 felec = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
759 /* CUBIC SPLINE TABLE DISPERSION */
760 vfitab = _mm_add_epi32(vfitab,ifour);
761 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
762 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
763 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
764 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
765 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
766 Heps = _mm256_mul_pd(vfeps,H);
767 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
768 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
769 fvdw6 = _mm256_mul_pd(c6_00,FF);
771 /* CUBIC SPLINE TABLE REPULSION */
772 vfitab = _mm_add_epi32(vfitab,ifour);
773 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
774 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
775 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
776 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
777 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
778 Heps = _mm256_mul_pd(vfeps,H);
779 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
780 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
781 fvdw12 = _mm256_mul_pd(c12_00,FF);
782 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
784 fscal = _mm256_add_pd(felec,fvdw);
786 fscal = _mm256_andnot_pd(dummy_mask,fscal);
788 /* Calculate temporary vectorial force */
789 tx = _mm256_mul_pd(fscal,dx00);
790 ty = _mm256_mul_pd(fscal,dy00);
791 tz = _mm256_mul_pd(fscal,dz00);
793 /* Update vectorial force */
794 fix0 = _mm256_add_pd(fix0,tx);
795 fiy0 = _mm256_add_pd(fiy0,ty);
796 fiz0 = _mm256_add_pd(fiz0,tz);
798 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
799 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
800 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
801 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
802 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
804 /* Inner loop uses 62 flops */
807 /* End of innermost loop */
809 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
810 f+i_coord_offset,fshift+i_shift_offset);
812 /* Increment number of inner iterations */
813 inneriter += j_index_end - j_index_start;
815 /* Outer loop uses 7 flops */
818 /* Increment number of outer iterations */
821 /* Update outer/inner flops */
823 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*62);