2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 * Note: this file was generated by the GROMACS avx_256_double kernel generator.
42 #include "../nb_kernel.h"
43 #include "types/simple.h"
44 #include "gromacs/math/vec.h"
47 #include "gromacs/simd/math_x86_avx_256_double.h"
48 #include "kernelutil_x86_avx_256_double.h"
51 * Gromacs nonbonded kernel: nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_VF_avx_256_double
52 * Electrostatics interaction: CubicSplineTable
53 * VdW interaction: CubicSplineTable
54 * Geometry: Particle-Particle
55 * Calculate force/pot: PotentialAndForce
58 nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_VF_avx_256_double
59 (t_nblist * gmx_restrict nlist,
60 rvec * gmx_restrict xx,
61 rvec * gmx_restrict ff,
62 t_forcerec * gmx_restrict fr,
63 t_mdatoms * gmx_restrict mdatoms,
64 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
65 t_nrnb * gmx_restrict nrnb)
67 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
68 * just 0 for non-waters.
69 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
70 * jnr indices corresponding to data put in the four positions in the SIMD register.
72 int i_shift_offset,i_coord_offset,outeriter,inneriter;
73 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
74 int jnrA,jnrB,jnrC,jnrD;
75 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
76 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
77 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
78 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
80 real *shiftvec,*fshift,*x,*f;
81 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
83 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
84 real * vdwioffsetptr0;
85 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
86 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
87 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
88 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
89 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
92 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
95 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
96 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
98 __m128i ifour = _mm_set1_epi32(4);
99 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
101 __m256d dummy_mask,cutoff_mask;
102 __m128 tmpmask0,tmpmask1;
103 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
104 __m256d one = _mm256_set1_pd(1.0);
105 __m256d two = _mm256_set1_pd(2.0);
111 jindex = nlist->jindex;
113 shiftidx = nlist->shift;
115 shiftvec = fr->shift_vec[0];
116 fshift = fr->fshift[0];
117 facel = _mm256_set1_pd(fr->epsfac);
118 charge = mdatoms->chargeA;
119 nvdwtype = fr->ntype;
121 vdwtype = mdatoms->typeA;
123 vftab = kernel_data->table_elec_vdw->data;
124 vftabscale = _mm256_set1_pd(kernel_data->table_elec_vdw->scale);
126 /* Avoid stupid compiler warnings */
127 jnrA = jnrB = jnrC = jnrD = 0;
136 for(iidx=0;iidx<4*DIM;iidx++)
141 /* Start outer loop over neighborlists */
142 for(iidx=0; iidx<nri; iidx++)
144 /* Load shift vector for this list */
145 i_shift_offset = DIM*shiftidx[iidx];
147 /* Load limits for loop over neighbors */
148 j_index_start = jindex[iidx];
149 j_index_end = jindex[iidx+1];
151 /* Get outer coordinate index */
153 i_coord_offset = DIM*inr;
155 /* Load i particle coords and add shift vector */
156 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
158 fix0 = _mm256_setzero_pd();
159 fiy0 = _mm256_setzero_pd();
160 fiz0 = _mm256_setzero_pd();
162 /* Load parameters for i particles */
163 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
164 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
166 /* Reset potential sums */
167 velecsum = _mm256_setzero_pd();
168 vvdwsum = _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 /* Load parameters for j particles */
200 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
201 charge+jnrC+0,charge+jnrD+0);
202 vdwjidx0A = 2*vdwtype[jnrA+0];
203 vdwjidx0B = 2*vdwtype[jnrB+0];
204 vdwjidx0C = 2*vdwtype[jnrC+0];
205 vdwjidx0D = 2*vdwtype[jnrD+0];
207 /**************************
208 * CALCULATE INTERACTIONS *
209 **************************/
211 r00 = _mm256_mul_pd(rsq00,rinv00);
213 /* Compute parameters for interactions between i and j atoms */
214 qq00 = _mm256_mul_pd(iq0,jq0);
215 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
216 vdwioffsetptr0+vdwjidx0B,
217 vdwioffsetptr0+vdwjidx0C,
218 vdwioffsetptr0+vdwjidx0D,
221 /* Calculate table index by multiplying r with table scale and truncate to integer */
222 rt = _mm256_mul_pd(r00,vftabscale);
223 vfitab = _mm256_cvttpd_epi32(rt);
224 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
225 vfitab = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
227 /* CUBIC SPLINE TABLE ELECTROSTATICS */
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 velec = _mm256_mul_pd(qq00,VV);
237 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
238 felec = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
240 /* CUBIC SPLINE TABLE DISPERSION */
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 vvdw6 = _mm256_mul_pd(c6_00,VV);
251 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
252 fvdw6 = _mm256_mul_pd(c6_00,FF);
254 /* CUBIC SPLINE TABLE REPULSION */
255 vfitab = _mm_add_epi32(vfitab,ifour);
256 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
257 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
258 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
259 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
260 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
261 Heps = _mm256_mul_pd(vfeps,H);
262 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
263 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
264 vvdw12 = _mm256_mul_pd(c12_00,VV);
265 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
266 fvdw12 = _mm256_mul_pd(c12_00,FF);
267 vvdw = _mm256_add_pd(vvdw12,vvdw6);
268 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
270 /* Update potential sum for this i atom from the interaction with this j atom. */
271 velecsum = _mm256_add_pd(velecsum,velec);
272 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
274 fscal = _mm256_add_pd(felec,fvdw);
276 /* Calculate temporary vectorial force */
277 tx = _mm256_mul_pd(fscal,dx00);
278 ty = _mm256_mul_pd(fscal,dy00);
279 tz = _mm256_mul_pd(fscal,dz00);
281 /* Update vectorial force */
282 fix0 = _mm256_add_pd(fix0,tx);
283 fiy0 = _mm256_add_pd(fiy0,ty);
284 fiz0 = _mm256_add_pd(fiz0,tz);
286 fjptrA = f+j_coord_offsetA;
287 fjptrB = f+j_coord_offsetB;
288 fjptrC = f+j_coord_offsetC;
289 fjptrD = f+j_coord_offsetD;
290 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
292 /* Inner loop uses 73 flops */
298 /* Get j neighbor index, and coordinate index */
299 jnrlistA = jjnr[jidx];
300 jnrlistB = jjnr[jidx+1];
301 jnrlistC = jjnr[jidx+2];
302 jnrlistD = jjnr[jidx+3];
303 /* Sign of each element will be negative for non-real atoms.
304 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
305 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
307 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
309 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
310 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
311 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
313 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
314 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
315 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
316 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
317 j_coord_offsetA = DIM*jnrA;
318 j_coord_offsetB = DIM*jnrB;
319 j_coord_offsetC = DIM*jnrC;
320 j_coord_offsetD = DIM*jnrD;
322 /* load j atom coordinates */
323 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
324 x+j_coord_offsetC,x+j_coord_offsetD,
327 /* Calculate displacement vector */
328 dx00 = _mm256_sub_pd(ix0,jx0);
329 dy00 = _mm256_sub_pd(iy0,jy0);
330 dz00 = _mm256_sub_pd(iz0,jz0);
332 /* Calculate squared distance and things based on it */
333 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
335 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
337 /* Load parameters for j particles */
338 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
339 charge+jnrC+0,charge+jnrD+0);
340 vdwjidx0A = 2*vdwtype[jnrA+0];
341 vdwjidx0B = 2*vdwtype[jnrB+0];
342 vdwjidx0C = 2*vdwtype[jnrC+0];
343 vdwjidx0D = 2*vdwtype[jnrD+0];
345 /**************************
346 * CALCULATE INTERACTIONS *
347 **************************/
349 r00 = _mm256_mul_pd(rsq00,rinv00);
350 r00 = _mm256_andnot_pd(dummy_mask,r00);
352 /* Compute parameters for interactions between i and j atoms */
353 qq00 = _mm256_mul_pd(iq0,jq0);
354 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
355 vdwioffsetptr0+vdwjidx0B,
356 vdwioffsetptr0+vdwjidx0C,
357 vdwioffsetptr0+vdwjidx0D,
360 /* Calculate table index by multiplying r with table scale and truncate to integer */
361 rt = _mm256_mul_pd(r00,vftabscale);
362 vfitab = _mm256_cvttpd_epi32(rt);
363 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
364 vfitab = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
366 /* CUBIC SPLINE TABLE ELECTROSTATICS */
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 velec = _mm256_mul_pd(qq00,VV);
376 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
377 felec = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
379 /* CUBIC SPLINE TABLE DISPERSION */
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 vvdw6 = _mm256_mul_pd(c6_00,VV);
390 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
391 fvdw6 = _mm256_mul_pd(c6_00,FF);
393 /* CUBIC SPLINE TABLE REPULSION */
394 vfitab = _mm_add_epi32(vfitab,ifour);
395 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
396 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
397 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
398 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
399 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
400 Heps = _mm256_mul_pd(vfeps,H);
401 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
402 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
403 vvdw12 = _mm256_mul_pd(c12_00,VV);
404 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
405 fvdw12 = _mm256_mul_pd(c12_00,FF);
406 vvdw = _mm256_add_pd(vvdw12,vvdw6);
407 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
409 /* Update potential sum for this i atom from the interaction with this j atom. */
410 velec = _mm256_andnot_pd(dummy_mask,velec);
411 velecsum = _mm256_add_pd(velecsum,velec);
412 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
413 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
415 fscal = _mm256_add_pd(felec,fvdw);
417 fscal = _mm256_andnot_pd(dummy_mask,fscal);
419 /* Calculate temporary vectorial force */
420 tx = _mm256_mul_pd(fscal,dx00);
421 ty = _mm256_mul_pd(fscal,dy00);
422 tz = _mm256_mul_pd(fscal,dz00);
424 /* Update vectorial force */
425 fix0 = _mm256_add_pd(fix0,tx);
426 fiy0 = _mm256_add_pd(fiy0,ty);
427 fiz0 = _mm256_add_pd(fiz0,tz);
429 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
430 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
431 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
432 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
433 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
435 /* Inner loop uses 74 flops */
438 /* End of innermost loop */
440 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
441 f+i_coord_offset,fshift+i_shift_offset);
444 /* Update potential energies */
445 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
446 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
448 /* Increment number of inner iterations */
449 inneriter += j_index_end - j_index_start;
451 /* Outer loop uses 9 flops */
454 /* Increment number of outer iterations */
457 /* Update outer/inner flops */
459 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*74);
462 * Gromacs nonbonded kernel: nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_256_double
463 * Electrostatics interaction: CubicSplineTable
464 * VdW interaction: CubicSplineTable
465 * Geometry: Particle-Particle
466 * Calculate force/pot: Force
469 nb_kernel_ElecCSTab_VdwCSTab_GeomP1P1_F_avx_256_double
470 (t_nblist * gmx_restrict nlist,
471 rvec * gmx_restrict xx,
472 rvec * gmx_restrict ff,
473 t_forcerec * gmx_restrict fr,
474 t_mdatoms * gmx_restrict mdatoms,
475 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
476 t_nrnb * gmx_restrict nrnb)
478 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
479 * just 0 for non-waters.
480 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
481 * jnr indices corresponding to data put in the four positions in the SIMD register.
483 int i_shift_offset,i_coord_offset,outeriter,inneriter;
484 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
485 int jnrA,jnrB,jnrC,jnrD;
486 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
487 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
488 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
489 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
491 real *shiftvec,*fshift,*x,*f;
492 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
494 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
495 real * vdwioffsetptr0;
496 __m256d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
497 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
498 __m256d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
499 __m256d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
500 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
503 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
506 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
507 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
509 __m128i ifour = _mm_set1_epi32(4);
510 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
512 __m256d dummy_mask,cutoff_mask;
513 __m128 tmpmask0,tmpmask1;
514 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
515 __m256d one = _mm256_set1_pd(1.0);
516 __m256d two = _mm256_set1_pd(2.0);
522 jindex = nlist->jindex;
524 shiftidx = nlist->shift;
526 shiftvec = fr->shift_vec[0];
527 fshift = fr->fshift[0];
528 facel = _mm256_set1_pd(fr->epsfac);
529 charge = mdatoms->chargeA;
530 nvdwtype = fr->ntype;
532 vdwtype = mdatoms->typeA;
534 vftab = kernel_data->table_elec_vdw->data;
535 vftabscale = _mm256_set1_pd(kernel_data->table_elec_vdw->scale);
537 /* Avoid stupid compiler warnings */
538 jnrA = jnrB = jnrC = jnrD = 0;
547 for(iidx=0;iidx<4*DIM;iidx++)
552 /* Start outer loop over neighborlists */
553 for(iidx=0; iidx<nri; iidx++)
555 /* Load shift vector for this list */
556 i_shift_offset = DIM*shiftidx[iidx];
558 /* Load limits for loop over neighbors */
559 j_index_start = jindex[iidx];
560 j_index_end = jindex[iidx+1];
562 /* Get outer coordinate index */
564 i_coord_offset = DIM*inr;
566 /* Load i particle coords and add shift vector */
567 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
569 fix0 = _mm256_setzero_pd();
570 fiy0 = _mm256_setzero_pd();
571 fiz0 = _mm256_setzero_pd();
573 /* Load parameters for i particles */
574 iq0 = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
575 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
577 /* Start inner kernel loop */
578 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
581 /* Get j neighbor index, and coordinate index */
586 j_coord_offsetA = DIM*jnrA;
587 j_coord_offsetB = DIM*jnrB;
588 j_coord_offsetC = DIM*jnrC;
589 j_coord_offsetD = DIM*jnrD;
591 /* load j atom coordinates */
592 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
593 x+j_coord_offsetC,x+j_coord_offsetD,
596 /* Calculate displacement vector */
597 dx00 = _mm256_sub_pd(ix0,jx0);
598 dy00 = _mm256_sub_pd(iy0,jy0);
599 dz00 = _mm256_sub_pd(iz0,jz0);
601 /* Calculate squared distance and things based on it */
602 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
604 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
606 /* Load parameters for j particles */
607 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
608 charge+jnrC+0,charge+jnrD+0);
609 vdwjidx0A = 2*vdwtype[jnrA+0];
610 vdwjidx0B = 2*vdwtype[jnrB+0];
611 vdwjidx0C = 2*vdwtype[jnrC+0];
612 vdwjidx0D = 2*vdwtype[jnrD+0];
614 /**************************
615 * CALCULATE INTERACTIONS *
616 **************************/
618 r00 = _mm256_mul_pd(rsq00,rinv00);
620 /* Compute parameters for interactions between i and j atoms */
621 qq00 = _mm256_mul_pd(iq0,jq0);
622 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
623 vdwioffsetptr0+vdwjidx0B,
624 vdwioffsetptr0+vdwjidx0C,
625 vdwioffsetptr0+vdwjidx0D,
628 /* Calculate table index by multiplying r with table scale and truncate to integer */
629 rt = _mm256_mul_pd(r00,vftabscale);
630 vfitab = _mm256_cvttpd_epi32(rt);
631 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
632 vfitab = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
634 /* CUBIC SPLINE TABLE ELECTROSTATICS */
635 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
636 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
637 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
638 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
639 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
640 Heps = _mm256_mul_pd(vfeps,H);
641 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
642 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
643 felec = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
645 /* CUBIC SPLINE TABLE DISPERSION */
646 vfitab = _mm_add_epi32(vfitab,ifour);
647 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
648 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
649 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
650 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
651 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
652 Heps = _mm256_mul_pd(vfeps,H);
653 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
654 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
655 fvdw6 = _mm256_mul_pd(c6_00,FF);
657 /* CUBIC SPLINE TABLE REPULSION */
658 vfitab = _mm_add_epi32(vfitab,ifour);
659 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
660 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
661 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
662 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
663 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
664 Heps = _mm256_mul_pd(vfeps,H);
665 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
666 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
667 fvdw12 = _mm256_mul_pd(c12_00,FF);
668 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
670 fscal = _mm256_add_pd(felec,fvdw);
672 /* Calculate temporary vectorial force */
673 tx = _mm256_mul_pd(fscal,dx00);
674 ty = _mm256_mul_pd(fscal,dy00);
675 tz = _mm256_mul_pd(fscal,dz00);
677 /* Update vectorial force */
678 fix0 = _mm256_add_pd(fix0,tx);
679 fiy0 = _mm256_add_pd(fiy0,ty);
680 fiz0 = _mm256_add_pd(fiz0,tz);
682 fjptrA = f+j_coord_offsetA;
683 fjptrB = f+j_coord_offsetB;
684 fjptrC = f+j_coord_offsetC;
685 fjptrD = f+j_coord_offsetD;
686 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
688 /* Inner loop uses 61 flops */
694 /* Get j neighbor index, and coordinate index */
695 jnrlistA = jjnr[jidx];
696 jnrlistB = jjnr[jidx+1];
697 jnrlistC = jjnr[jidx+2];
698 jnrlistD = jjnr[jidx+3];
699 /* Sign of each element will be negative for non-real atoms.
700 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
701 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
703 tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
705 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
706 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
707 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
709 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
710 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
711 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
712 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
713 j_coord_offsetA = DIM*jnrA;
714 j_coord_offsetB = DIM*jnrB;
715 j_coord_offsetC = DIM*jnrC;
716 j_coord_offsetD = DIM*jnrD;
718 /* load j atom coordinates */
719 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
720 x+j_coord_offsetC,x+j_coord_offsetD,
723 /* Calculate displacement vector */
724 dx00 = _mm256_sub_pd(ix0,jx0);
725 dy00 = _mm256_sub_pd(iy0,jy0);
726 dz00 = _mm256_sub_pd(iz0,jz0);
728 /* Calculate squared distance and things based on it */
729 rsq00 = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
731 rinv00 = gmx_mm256_invsqrt_pd(rsq00);
733 /* Load parameters for j particles */
734 jq0 = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
735 charge+jnrC+0,charge+jnrD+0);
736 vdwjidx0A = 2*vdwtype[jnrA+0];
737 vdwjidx0B = 2*vdwtype[jnrB+0];
738 vdwjidx0C = 2*vdwtype[jnrC+0];
739 vdwjidx0D = 2*vdwtype[jnrD+0];
741 /**************************
742 * CALCULATE INTERACTIONS *
743 **************************/
745 r00 = _mm256_mul_pd(rsq00,rinv00);
746 r00 = _mm256_andnot_pd(dummy_mask,r00);
748 /* Compute parameters for interactions between i and j atoms */
749 qq00 = _mm256_mul_pd(iq0,jq0);
750 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
751 vdwioffsetptr0+vdwjidx0B,
752 vdwioffsetptr0+vdwjidx0C,
753 vdwioffsetptr0+vdwjidx0D,
756 /* Calculate table index by multiplying r with table scale and truncate to integer */
757 rt = _mm256_mul_pd(r00,vftabscale);
758 vfitab = _mm256_cvttpd_epi32(rt);
759 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
760 vfitab = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
762 /* CUBIC SPLINE TABLE ELECTROSTATICS */
763 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
764 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
765 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
766 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
767 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
768 Heps = _mm256_mul_pd(vfeps,H);
769 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
770 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
771 felec = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq00,FF),_mm256_mul_pd(vftabscale,rinv00)));
773 /* CUBIC SPLINE TABLE DISPERSION */
774 vfitab = _mm_add_epi32(vfitab,ifour);
775 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
776 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
777 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
778 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
779 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
780 Heps = _mm256_mul_pd(vfeps,H);
781 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
782 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
783 fvdw6 = _mm256_mul_pd(c6_00,FF);
785 /* CUBIC SPLINE TABLE REPULSION */
786 vfitab = _mm_add_epi32(vfitab,ifour);
787 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
788 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
789 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
790 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
791 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
792 Heps = _mm256_mul_pd(vfeps,H);
793 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
794 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
795 fvdw12 = _mm256_mul_pd(c12_00,FF);
796 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv00)));
798 fscal = _mm256_add_pd(felec,fvdw);
800 fscal = _mm256_andnot_pd(dummy_mask,fscal);
802 /* Calculate temporary vectorial force */
803 tx = _mm256_mul_pd(fscal,dx00);
804 ty = _mm256_mul_pd(fscal,dy00);
805 tz = _mm256_mul_pd(fscal,dz00);
807 /* Update vectorial force */
808 fix0 = _mm256_add_pd(fix0,tx);
809 fiy0 = _mm256_add_pd(fiy0,ty);
810 fiz0 = _mm256_add_pd(fiz0,tz);
812 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
813 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
814 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
815 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
816 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
818 /* Inner loop uses 62 flops */
821 /* End of innermost loop */
823 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
824 f+i_coord_offset,fshift+i_shift_offset);
826 /* Increment number of inner iterations */
827 inneriter += j_index_end - j_index_start;
829 /* Outer loop uses 7 flops */
832 /* Increment number of outer iterations */
835 /* Update outer/inner flops */
837 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*62);