2 * Note: this file was generated by the Gromacs avx_256_single 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_single.h"
34 #include "kernelutil_x86_avx_256_single.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwCSTab_GeomW3P1_VF_avx_256_single
38 * Electrostatics interaction: Coulomb
39 * VdW interaction: CubicSplineTable
40 * Geometry: Water3-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecCoul_VdwCSTab_GeomW3P1_VF_avx_256_single
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,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight 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 jnrE,jnrF,jnrG,jnrH;
62 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
63 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
64 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
65 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
66 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
68 real *shiftvec,*fshift,*x,*f;
69 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
71 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
72 real * vdwioffsetptr0;
73 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
74 real * vdwioffsetptr1;
75 __m256 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
76 real * vdwioffsetptr2;
77 __m256 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
78 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
79 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
80 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
81 __m256 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
82 __m256 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
83 __m256 velec,felec,velecsum,facel,crf,krf,krf2;
86 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
89 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
90 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
92 __m128i vfitab_lo,vfitab_hi;
93 __m128i ifour = _mm_set1_epi32(4);
94 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
96 __m256 dummy_mask,cutoff_mask;
97 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
98 __m256 one = _mm256_set1_ps(1.0);
99 __m256 two = _mm256_set1_ps(2.0);
105 jindex = nlist->jindex;
107 shiftidx = nlist->shift;
109 shiftvec = fr->shift_vec[0];
110 fshift = fr->fshift[0];
111 facel = _mm256_set1_ps(fr->epsfac);
112 charge = mdatoms->chargeA;
113 nvdwtype = fr->ntype;
115 vdwtype = mdatoms->typeA;
117 vftab = kernel_data->table_vdw->data;
118 vftabscale = _mm256_set1_ps(kernel_data->table_vdw->scale);
120 /* Setup water-specific parameters */
121 inr = nlist->iinr[0];
122 iq0 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
123 iq1 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+1]));
124 iq2 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+2]));
125 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
127 /* Avoid stupid compiler warnings */
128 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
141 for(iidx=0;iidx<4*DIM;iidx++)
146 /* Start outer loop over neighborlists */
147 for(iidx=0; iidx<nri; iidx++)
149 /* Load shift vector for this list */
150 i_shift_offset = DIM*shiftidx[iidx];
152 /* Load limits for loop over neighbors */
153 j_index_start = jindex[iidx];
154 j_index_end = jindex[iidx+1];
156 /* Get outer coordinate index */
158 i_coord_offset = DIM*inr;
160 /* Load i particle coords and add shift vector */
161 gmx_mm256_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
162 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
164 fix0 = _mm256_setzero_ps();
165 fiy0 = _mm256_setzero_ps();
166 fiz0 = _mm256_setzero_ps();
167 fix1 = _mm256_setzero_ps();
168 fiy1 = _mm256_setzero_ps();
169 fiz1 = _mm256_setzero_ps();
170 fix2 = _mm256_setzero_ps();
171 fiy2 = _mm256_setzero_ps();
172 fiz2 = _mm256_setzero_ps();
174 /* Reset potential sums */
175 velecsum = _mm256_setzero_ps();
176 vvdwsum = _mm256_setzero_ps();
178 /* Start inner kernel loop */
179 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
182 /* Get j neighbor index, and coordinate index */
191 j_coord_offsetA = DIM*jnrA;
192 j_coord_offsetB = DIM*jnrB;
193 j_coord_offsetC = DIM*jnrC;
194 j_coord_offsetD = DIM*jnrD;
195 j_coord_offsetE = DIM*jnrE;
196 j_coord_offsetF = DIM*jnrF;
197 j_coord_offsetG = DIM*jnrG;
198 j_coord_offsetH = DIM*jnrH;
200 /* load j atom coordinates */
201 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
202 x+j_coord_offsetC,x+j_coord_offsetD,
203 x+j_coord_offsetE,x+j_coord_offsetF,
204 x+j_coord_offsetG,x+j_coord_offsetH,
207 /* Calculate displacement vector */
208 dx00 = _mm256_sub_ps(ix0,jx0);
209 dy00 = _mm256_sub_ps(iy0,jy0);
210 dz00 = _mm256_sub_ps(iz0,jz0);
211 dx10 = _mm256_sub_ps(ix1,jx0);
212 dy10 = _mm256_sub_ps(iy1,jy0);
213 dz10 = _mm256_sub_ps(iz1,jz0);
214 dx20 = _mm256_sub_ps(ix2,jx0);
215 dy20 = _mm256_sub_ps(iy2,jy0);
216 dz20 = _mm256_sub_ps(iz2,jz0);
218 /* Calculate squared distance and things based on it */
219 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
220 rsq10 = gmx_mm256_calc_rsq_ps(dx10,dy10,dz10);
221 rsq20 = gmx_mm256_calc_rsq_ps(dx20,dy20,dz20);
223 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
224 rinv10 = gmx_mm256_invsqrt_ps(rsq10);
225 rinv20 = gmx_mm256_invsqrt_ps(rsq20);
227 rinvsq00 = _mm256_mul_ps(rinv00,rinv00);
228 rinvsq10 = _mm256_mul_ps(rinv10,rinv10);
229 rinvsq20 = _mm256_mul_ps(rinv20,rinv20);
231 /* Load parameters for j particles */
232 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
233 charge+jnrC+0,charge+jnrD+0,
234 charge+jnrE+0,charge+jnrF+0,
235 charge+jnrG+0,charge+jnrH+0);
236 vdwjidx0A = 2*vdwtype[jnrA+0];
237 vdwjidx0B = 2*vdwtype[jnrB+0];
238 vdwjidx0C = 2*vdwtype[jnrC+0];
239 vdwjidx0D = 2*vdwtype[jnrD+0];
240 vdwjidx0E = 2*vdwtype[jnrE+0];
241 vdwjidx0F = 2*vdwtype[jnrF+0];
242 vdwjidx0G = 2*vdwtype[jnrG+0];
243 vdwjidx0H = 2*vdwtype[jnrH+0];
245 fjx0 = _mm256_setzero_ps();
246 fjy0 = _mm256_setzero_ps();
247 fjz0 = _mm256_setzero_ps();
249 /**************************
250 * CALCULATE INTERACTIONS *
251 **************************/
253 r00 = _mm256_mul_ps(rsq00,rinv00);
255 /* Compute parameters for interactions between i and j atoms */
256 qq00 = _mm256_mul_ps(iq0,jq0);
257 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
258 vdwioffsetptr0+vdwjidx0B,
259 vdwioffsetptr0+vdwjidx0C,
260 vdwioffsetptr0+vdwjidx0D,
261 vdwioffsetptr0+vdwjidx0E,
262 vdwioffsetptr0+vdwjidx0F,
263 vdwioffsetptr0+vdwjidx0G,
264 vdwioffsetptr0+vdwjidx0H,
267 /* Calculate table index by multiplying r with table scale and truncate to integer */
268 rt = _mm256_mul_ps(r00,vftabscale);
269 vfitab = _mm256_cvttps_epi32(rt);
270 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
271 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
272 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
273 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
274 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
275 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
277 /* COULOMB ELECTROSTATICS */
278 velec = _mm256_mul_ps(qq00,rinv00);
279 felec = _mm256_mul_ps(velec,rinvsq00);
281 /* CUBIC SPLINE TABLE DISPERSION */
282 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
283 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
284 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
285 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
286 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
287 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
288 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
289 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
290 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
291 Heps = _mm256_mul_ps(vfeps,H);
292 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
293 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
294 vvdw6 = _mm256_mul_ps(c6_00,VV);
295 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
296 fvdw6 = _mm256_mul_ps(c6_00,FF);
298 /* CUBIC SPLINE TABLE REPULSION */
299 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
300 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
301 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
302 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
303 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
304 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
305 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
306 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
307 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
308 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
309 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
310 Heps = _mm256_mul_ps(vfeps,H);
311 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
312 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
313 vvdw12 = _mm256_mul_ps(c12_00,VV);
314 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
315 fvdw12 = _mm256_mul_ps(c12_00,FF);
316 vvdw = _mm256_add_ps(vvdw12,vvdw6);
317 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
319 /* Update potential sum for this i atom from the interaction with this j atom. */
320 velecsum = _mm256_add_ps(velecsum,velec);
321 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
323 fscal = _mm256_add_ps(felec,fvdw);
325 /* Calculate temporary vectorial force */
326 tx = _mm256_mul_ps(fscal,dx00);
327 ty = _mm256_mul_ps(fscal,dy00);
328 tz = _mm256_mul_ps(fscal,dz00);
330 /* Update vectorial force */
331 fix0 = _mm256_add_ps(fix0,tx);
332 fiy0 = _mm256_add_ps(fiy0,ty);
333 fiz0 = _mm256_add_ps(fiz0,tz);
335 fjx0 = _mm256_add_ps(fjx0,tx);
336 fjy0 = _mm256_add_ps(fjy0,ty);
337 fjz0 = _mm256_add_ps(fjz0,tz);
339 /**************************
340 * CALCULATE INTERACTIONS *
341 **************************/
343 /* Compute parameters for interactions between i and j atoms */
344 qq10 = _mm256_mul_ps(iq1,jq0);
346 /* COULOMB ELECTROSTATICS */
347 velec = _mm256_mul_ps(qq10,rinv10);
348 felec = _mm256_mul_ps(velec,rinvsq10);
350 /* Update potential sum for this i atom from the interaction with this j atom. */
351 velecsum = _mm256_add_ps(velecsum,velec);
355 /* Calculate temporary vectorial force */
356 tx = _mm256_mul_ps(fscal,dx10);
357 ty = _mm256_mul_ps(fscal,dy10);
358 tz = _mm256_mul_ps(fscal,dz10);
360 /* Update vectorial force */
361 fix1 = _mm256_add_ps(fix1,tx);
362 fiy1 = _mm256_add_ps(fiy1,ty);
363 fiz1 = _mm256_add_ps(fiz1,tz);
365 fjx0 = _mm256_add_ps(fjx0,tx);
366 fjy0 = _mm256_add_ps(fjy0,ty);
367 fjz0 = _mm256_add_ps(fjz0,tz);
369 /**************************
370 * CALCULATE INTERACTIONS *
371 **************************/
373 /* Compute parameters for interactions between i and j atoms */
374 qq20 = _mm256_mul_ps(iq2,jq0);
376 /* COULOMB ELECTROSTATICS */
377 velec = _mm256_mul_ps(qq20,rinv20);
378 felec = _mm256_mul_ps(velec,rinvsq20);
380 /* Update potential sum for this i atom from the interaction with this j atom. */
381 velecsum = _mm256_add_ps(velecsum,velec);
385 /* Calculate temporary vectorial force */
386 tx = _mm256_mul_ps(fscal,dx20);
387 ty = _mm256_mul_ps(fscal,dy20);
388 tz = _mm256_mul_ps(fscal,dz20);
390 /* Update vectorial force */
391 fix2 = _mm256_add_ps(fix2,tx);
392 fiy2 = _mm256_add_ps(fiy2,ty);
393 fiz2 = _mm256_add_ps(fiz2,tz);
395 fjx0 = _mm256_add_ps(fjx0,tx);
396 fjy0 = _mm256_add_ps(fjy0,ty);
397 fjz0 = _mm256_add_ps(fjz0,tz);
399 fjptrA = f+j_coord_offsetA;
400 fjptrB = f+j_coord_offsetB;
401 fjptrC = f+j_coord_offsetC;
402 fjptrD = f+j_coord_offsetD;
403 fjptrE = f+j_coord_offsetE;
404 fjptrF = f+j_coord_offsetF;
405 fjptrG = f+j_coord_offsetG;
406 fjptrH = f+j_coord_offsetH;
408 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,fjx0,fjy0,fjz0);
410 /* Inner loop uses 119 flops */
416 /* Get j neighbor index, and coordinate index */
417 jnrlistA = jjnr[jidx];
418 jnrlistB = jjnr[jidx+1];
419 jnrlistC = jjnr[jidx+2];
420 jnrlistD = jjnr[jidx+3];
421 jnrlistE = jjnr[jidx+4];
422 jnrlistF = jjnr[jidx+5];
423 jnrlistG = jjnr[jidx+6];
424 jnrlistH = jjnr[jidx+7];
425 /* Sign of each element will be negative for non-real atoms.
426 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
427 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
429 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
430 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
432 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
433 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
434 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
435 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
436 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
437 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
438 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
439 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
440 j_coord_offsetA = DIM*jnrA;
441 j_coord_offsetB = DIM*jnrB;
442 j_coord_offsetC = DIM*jnrC;
443 j_coord_offsetD = DIM*jnrD;
444 j_coord_offsetE = DIM*jnrE;
445 j_coord_offsetF = DIM*jnrF;
446 j_coord_offsetG = DIM*jnrG;
447 j_coord_offsetH = DIM*jnrH;
449 /* load j atom coordinates */
450 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
451 x+j_coord_offsetC,x+j_coord_offsetD,
452 x+j_coord_offsetE,x+j_coord_offsetF,
453 x+j_coord_offsetG,x+j_coord_offsetH,
456 /* Calculate displacement vector */
457 dx00 = _mm256_sub_ps(ix0,jx0);
458 dy00 = _mm256_sub_ps(iy0,jy0);
459 dz00 = _mm256_sub_ps(iz0,jz0);
460 dx10 = _mm256_sub_ps(ix1,jx0);
461 dy10 = _mm256_sub_ps(iy1,jy0);
462 dz10 = _mm256_sub_ps(iz1,jz0);
463 dx20 = _mm256_sub_ps(ix2,jx0);
464 dy20 = _mm256_sub_ps(iy2,jy0);
465 dz20 = _mm256_sub_ps(iz2,jz0);
467 /* Calculate squared distance and things based on it */
468 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
469 rsq10 = gmx_mm256_calc_rsq_ps(dx10,dy10,dz10);
470 rsq20 = gmx_mm256_calc_rsq_ps(dx20,dy20,dz20);
472 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
473 rinv10 = gmx_mm256_invsqrt_ps(rsq10);
474 rinv20 = gmx_mm256_invsqrt_ps(rsq20);
476 rinvsq00 = _mm256_mul_ps(rinv00,rinv00);
477 rinvsq10 = _mm256_mul_ps(rinv10,rinv10);
478 rinvsq20 = _mm256_mul_ps(rinv20,rinv20);
480 /* Load parameters for j particles */
481 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
482 charge+jnrC+0,charge+jnrD+0,
483 charge+jnrE+0,charge+jnrF+0,
484 charge+jnrG+0,charge+jnrH+0);
485 vdwjidx0A = 2*vdwtype[jnrA+0];
486 vdwjidx0B = 2*vdwtype[jnrB+0];
487 vdwjidx0C = 2*vdwtype[jnrC+0];
488 vdwjidx0D = 2*vdwtype[jnrD+0];
489 vdwjidx0E = 2*vdwtype[jnrE+0];
490 vdwjidx0F = 2*vdwtype[jnrF+0];
491 vdwjidx0G = 2*vdwtype[jnrG+0];
492 vdwjidx0H = 2*vdwtype[jnrH+0];
494 fjx0 = _mm256_setzero_ps();
495 fjy0 = _mm256_setzero_ps();
496 fjz0 = _mm256_setzero_ps();
498 /**************************
499 * CALCULATE INTERACTIONS *
500 **************************/
502 r00 = _mm256_mul_ps(rsq00,rinv00);
503 r00 = _mm256_andnot_ps(dummy_mask,r00);
505 /* Compute parameters for interactions between i and j atoms */
506 qq00 = _mm256_mul_ps(iq0,jq0);
507 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
508 vdwioffsetptr0+vdwjidx0B,
509 vdwioffsetptr0+vdwjidx0C,
510 vdwioffsetptr0+vdwjidx0D,
511 vdwioffsetptr0+vdwjidx0E,
512 vdwioffsetptr0+vdwjidx0F,
513 vdwioffsetptr0+vdwjidx0G,
514 vdwioffsetptr0+vdwjidx0H,
517 /* Calculate table index by multiplying r with table scale and truncate to integer */
518 rt = _mm256_mul_ps(r00,vftabscale);
519 vfitab = _mm256_cvttps_epi32(rt);
520 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
521 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
522 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
523 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
524 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
525 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
527 /* COULOMB ELECTROSTATICS */
528 velec = _mm256_mul_ps(qq00,rinv00);
529 felec = _mm256_mul_ps(velec,rinvsq00);
531 /* CUBIC SPLINE TABLE DISPERSION */
532 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
533 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
534 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
535 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
536 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
537 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
538 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
539 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
540 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
541 Heps = _mm256_mul_ps(vfeps,H);
542 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
543 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
544 vvdw6 = _mm256_mul_ps(c6_00,VV);
545 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
546 fvdw6 = _mm256_mul_ps(c6_00,FF);
548 /* CUBIC SPLINE TABLE REPULSION */
549 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
550 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
551 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
552 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
553 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
554 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
555 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
556 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
557 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
558 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
559 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
560 Heps = _mm256_mul_ps(vfeps,H);
561 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
562 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
563 vvdw12 = _mm256_mul_ps(c12_00,VV);
564 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
565 fvdw12 = _mm256_mul_ps(c12_00,FF);
566 vvdw = _mm256_add_ps(vvdw12,vvdw6);
567 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
569 /* Update potential sum for this i atom from the interaction with this j atom. */
570 velec = _mm256_andnot_ps(dummy_mask,velec);
571 velecsum = _mm256_add_ps(velecsum,velec);
572 vvdw = _mm256_andnot_ps(dummy_mask,vvdw);
573 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
575 fscal = _mm256_add_ps(felec,fvdw);
577 fscal = _mm256_andnot_ps(dummy_mask,fscal);
579 /* Calculate temporary vectorial force */
580 tx = _mm256_mul_ps(fscal,dx00);
581 ty = _mm256_mul_ps(fscal,dy00);
582 tz = _mm256_mul_ps(fscal,dz00);
584 /* Update vectorial force */
585 fix0 = _mm256_add_ps(fix0,tx);
586 fiy0 = _mm256_add_ps(fiy0,ty);
587 fiz0 = _mm256_add_ps(fiz0,tz);
589 fjx0 = _mm256_add_ps(fjx0,tx);
590 fjy0 = _mm256_add_ps(fjy0,ty);
591 fjz0 = _mm256_add_ps(fjz0,tz);
593 /**************************
594 * CALCULATE INTERACTIONS *
595 **************************/
597 /* Compute parameters for interactions between i and j atoms */
598 qq10 = _mm256_mul_ps(iq1,jq0);
600 /* COULOMB ELECTROSTATICS */
601 velec = _mm256_mul_ps(qq10,rinv10);
602 felec = _mm256_mul_ps(velec,rinvsq10);
604 /* Update potential sum for this i atom from the interaction with this j atom. */
605 velec = _mm256_andnot_ps(dummy_mask,velec);
606 velecsum = _mm256_add_ps(velecsum,velec);
610 fscal = _mm256_andnot_ps(dummy_mask,fscal);
612 /* Calculate temporary vectorial force */
613 tx = _mm256_mul_ps(fscal,dx10);
614 ty = _mm256_mul_ps(fscal,dy10);
615 tz = _mm256_mul_ps(fscal,dz10);
617 /* Update vectorial force */
618 fix1 = _mm256_add_ps(fix1,tx);
619 fiy1 = _mm256_add_ps(fiy1,ty);
620 fiz1 = _mm256_add_ps(fiz1,tz);
622 fjx0 = _mm256_add_ps(fjx0,tx);
623 fjy0 = _mm256_add_ps(fjy0,ty);
624 fjz0 = _mm256_add_ps(fjz0,tz);
626 /**************************
627 * CALCULATE INTERACTIONS *
628 **************************/
630 /* Compute parameters for interactions between i and j atoms */
631 qq20 = _mm256_mul_ps(iq2,jq0);
633 /* COULOMB ELECTROSTATICS */
634 velec = _mm256_mul_ps(qq20,rinv20);
635 felec = _mm256_mul_ps(velec,rinvsq20);
637 /* Update potential sum for this i atom from the interaction with this j atom. */
638 velec = _mm256_andnot_ps(dummy_mask,velec);
639 velecsum = _mm256_add_ps(velecsum,velec);
643 fscal = _mm256_andnot_ps(dummy_mask,fscal);
645 /* Calculate temporary vectorial force */
646 tx = _mm256_mul_ps(fscal,dx20);
647 ty = _mm256_mul_ps(fscal,dy20);
648 tz = _mm256_mul_ps(fscal,dz20);
650 /* Update vectorial force */
651 fix2 = _mm256_add_ps(fix2,tx);
652 fiy2 = _mm256_add_ps(fiy2,ty);
653 fiz2 = _mm256_add_ps(fiz2,tz);
655 fjx0 = _mm256_add_ps(fjx0,tx);
656 fjy0 = _mm256_add_ps(fjy0,ty);
657 fjz0 = _mm256_add_ps(fjz0,tz);
659 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
660 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
661 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
662 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
663 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
664 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
665 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
666 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
668 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,fjx0,fjy0,fjz0);
670 /* Inner loop uses 120 flops */
673 /* End of innermost loop */
675 gmx_mm256_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
676 f+i_coord_offset,fshift+i_shift_offset);
679 /* Update potential energies */
680 gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
681 gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
683 /* Increment number of inner iterations */
684 inneriter += j_index_end - j_index_start;
686 /* Outer loop uses 20 flops */
689 /* Increment number of outer iterations */
692 /* Update outer/inner flops */
694 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*120);
697 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwCSTab_GeomW3P1_F_avx_256_single
698 * Electrostatics interaction: Coulomb
699 * VdW interaction: CubicSplineTable
700 * Geometry: Water3-Particle
701 * Calculate force/pot: Force
704 nb_kernel_ElecCoul_VdwCSTab_GeomW3P1_F_avx_256_single
705 (t_nblist * gmx_restrict nlist,
706 rvec * gmx_restrict xx,
707 rvec * gmx_restrict ff,
708 t_forcerec * gmx_restrict fr,
709 t_mdatoms * gmx_restrict mdatoms,
710 nb_kernel_data_t * gmx_restrict kernel_data,
711 t_nrnb * gmx_restrict nrnb)
713 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
714 * just 0 for non-waters.
715 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
716 * jnr indices corresponding to data put in the four positions in the SIMD register.
718 int i_shift_offset,i_coord_offset,outeriter,inneriter;
719 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
720 int jnrA,jnrB,jnrC,jnrD;
721 int jnrE,jnrF,jnrG,jnrH;
722 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
723 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
724 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
725 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
726 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
728 real *shiftvec,*fshift,*x,*f;
729 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
731 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
732 real * vdwioffsetptr0;
733 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
734 real * vdwioffsetptr1;
735 __m256 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
736 real * vdwioffsetptr2;
737 __m256 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
738 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
739 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
740 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
741 __m256 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
742 __m256 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
743 __m256 velec,felec,velecsum,facel,crf,krf,krf2;
746 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
749 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
750 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
752 __m128i vfitab_lo,vfitab_hi;
753 __m128i ifour = _mm_set1_epi32(4);
754 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
756 __m256 dummy_mask,cutoff_mask;
757 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
758 __m256 one = _mm256_set1_ps(1.0);
759 __m256 two = _mm256_set1_ps(2.0);
765 jindex = nlist->jindex;
767 shiftidx = nlist->shift;
769 shiftvec = fr->shift_vec[0];
770 fshift = fr->fshift[0];
771 facel = _mm256_set1_ps(fr->epsfac);
772 charge = mdatoms->chargeA;
773 nvdwtype = fr->ntype;
775 vdwtype = mdatoms->typeA;
777 vftab = kernel_data->table_vdw->data;
778 vftabscale = _mm256_set1_ps(kernel_data->table_vdw->scale);
780 /* Setup water-specific parameters */
781 inr = nlist->iinr[0];
782 iq0 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
783 iq1 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+1]));
784 iq2 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+2]));
785 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
787 /* Avoid stupid compiler warnings */
788 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
801 for(iidx=0;iidx<4*DIM;iidx++)
806 /* Start outer loop over neighborlists */
807 for(iidx=0; iidx<nri; iidx++)
809 /* Load shift vector for this list */
810 i_shift_offset = DIM*shiftidx[iidx];
812 /* Load limits for loop over neighbors */
813 j_index_start = jindex[iidx];
814 j_index_end = jindex[iidx+1];
816 /* Get outer coordinate index */
818 i_coord_offset = DIM*inr;
820 /* Load i particle coords and add shift vector */
821 gmx_mm256_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
822 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
824 fix0 = _mm256_setzero_ps();
825 fiy0 = _mm256_setzero_ps();
826 fiz0 = _mm256_setzero_ps();
827 fix1 = _mm256_setzero_ps();
828 fiy1 = _mm256_setzero_ps();
829 fiz1 = _mm256_setzero_ps();
830 fix2 = _mm256_setzero_ps();
831 fiy2 = _mm256_setzero_ps();
832 fiz2 = _mm256_setzero_ps();
834 /* Start inner kernel loop */
835 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
838 /* Get j neighbor index, and coordinate index */
847 j_coord_offsetA = DIM*jnrA;
848 j_coord_offsetB = DIM*jnrB;
849 j_coord_offsetC = DIM*jnrC;
850 j_coord_offsetD = DIM*jnrD;
851 j_coord_offsetE = DIM*jnrE;
852 j_coord_offsetF = DIM*jnrF;
853 j_coord_offsetG = DIM*jnrG;
854 j_coord_offsetH = DIM*jnrH;
856 /* load j atom coordinates */
857 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
858 x+j_coord_offsetC,x+j_coord_offsetD,
859 x+j_coord_offsetE,x+j_coord_offsetF,
860 x+j_coord_offsetG,x+j_coord_offsetH,
863 /* Calculate displacement vector */
864 dx00 = _mm256_sub_ps(ix0,jx0);
865 dy00 = _mm256_sub_ps(iy0,jy0);
866 dz00 = _mm256_sub_ps(iz0,jz0);
867 dx10 = _mm256_sub_ps(ix1,jx0);
868 dy10 = _mm256_sub_ps(iy1,jy0);
869 dz10 = _mm256_sub_ps(iz1,jz0);
870 dx20 = _mm256_sub_ps(ix2,jx0);
871 dy20 = _mm256_sub_ps(iy2,jy0);
872 dz20 = _mm256_sub_ps(iz2,jz0);
874 /* Calculate squared distance and things based on it */
875 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
876 rsq10 = gmx_mm256_calc_rsq_ps(dx10,dy10,dz10);
877 rsq20 = gmx_mm256_calc_rsq_ps(dx20,dy20,dz20);
879 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
880 rinv10 = gmx_mm256_invsqrt_ps(rsq10);
881 rinv20 = gmx_mm256_invsqrt_ps(rsq20);
883 rinvsq00 = _mm256_mul_ps(rinv00,rinv00);
884 rinvsq10 = _mm256_mul_ps(rinv10,rinv10);
885 rinvsq20 = _mm256_mul_ps(rinv20,rinv20);
887 /* Load parameters for j particles */
888 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
889 charge+jnrC+0,charge+jnrD+0,
890 charge+jnrE+0,charge+jnrF+0,
891 charge+jnrG+0,charge+jnrH+0);
892 vdwjidx0A = 2*vdwtype[jnrA+0];
893 vdwjidx0B = 2*vdwtype[jnrB+0];
894 vdwjidx0C = 2*vdwtype[jnrC+0];
895 vdwjidx0D = 2*vdwtype[jnrD+0];
896 vdwjidx0E = 2*vdwtype[jnrE+0];
897 vdwjidx0F = 2*vdwtype[jnrF+0];
898 vdwjidx0G = 2*vdwtype[jnrG+0];
899 vdwjidx0H = 2*vdwtype[jnrH+0];
901 fjx0 = _mm256_setzero_ps();
902 fjy0 = _mm256_setzero_ps();
903 fjz0 = _mm256_setzero_ps();
905 /**************************
906 * CALCULATE INTERACTIONS *
907 **************************/
909 r00 = _mm256_mul_ps(rsq00,rinv00);
911 /* Compute parameters for interactions between i and j atoms */
912 qq00 = _mm256_mul_ps(iq0,jq0);
913 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
914 vdwioffsetptr0+vdwjidx0B,
915 vdwioffsetptr0+vdwjidx0C,
916 vdwioffsetptr0+vdwjidx0D,
917 vdwioffsetptr0+vdwjidx0E,
918 vdwioffsetptr0+vdwjidx0F,
919 vdwioffsetptr0+vdwjidx0G,
920 vdwioffsetptr0+vdwjidx0H,
923 /* Calculate table index by multiplying r with table scale and truncate to integer */
924 rt = _mm256_mul_ps(r00,vftabscale);
925 vfitab = _mm256_cvttps_epi32(rt);
926 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
927 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
928 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
929 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
930 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
931 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
933 /* COULOMB ELECTROSTATICS */
934 velec = _mm256_mul_ps(qq00,rinv00);
935 felec = _mm256_mul_ps(velec,rinvsq00);
937 /* CUBIC SPLINE TABLE DISPERSION */
938 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
939 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
940 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
941 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
942 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
943 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
944 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
945 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
946 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
947 Heps = _mm256_mul_ps(vfeps,H);
948 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
949 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
950 fvdw6 = _mm256_mul_ps(c6_00,FF);
952 /* CUBIC SPLINE TABLE REPULSION */
953 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
954 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
955 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
956 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
957 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
958 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
959 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
960 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
961 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
962 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
963 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
964 Heps = _mm256_mul_ps(vfeps,H);
965 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
966 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
967 fvdw12 = _mm256_mul_ps(c12_00,FF);
968 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
970 fscal = _mm256_add_ps(felec,fvdw);
972 /* Calculate temporary vectorial force */
973 tx = _mm256_mul_ps(fscal,dx00);
974 ty = _mm256_mul_ps(fscal,dy00);
975 tz = _mm256_mul_ps(fscal,dz00);
977 /* Update vectorial force */
978 fix0 = _mm256_add_ps(fix0,tx);
979 fiy0 = _mm256_add_ps(fiy0,ty);
980 fiz0 = _mm256_add_ps(fiz0,tz);
982 fjx0 = _mm256_add_ps(fjx0,tx);
983 fjy0 = _mm256_add_ps(fjy0,ty);
984 fjz0 = _mm256_add_ps(fjz0,tz);
986 /**************************
987 * CALCULATE INTERACTIONS *
988 **************************/
990 /* Compute parameters for interactions between i and j atoms */
991 qq10 = _mm256_mul_ps(iq1,jq0);
993 /* COULOMB ELECTROSTATICS */
994 velec = _mm256_mul_ps(qq10,rinv10);
995 felec = _mm256_mul_ps(velec,rinvsq10);
999 /* Calculate temporary vectorial force */
1000 tx = _mm256_mul_ps(fscal,dx10);
1001 ty = _mm256_mul_ps(fscal,dy10);
1002 tz = _mm256_mul_ps(fscal,dz10);
1004 /* Update vectorial force */
1005 fix1 = _mm256_add_ps(fix1,tx);
1006 fiy1 = _mm256_add_ps(fiy1,ty);
1007 fiz1 = _mm256_add_ps(fiz1,tz);
1009 fjx0 = _mm256_add_ps(fjx0,tx);
1010 fjy0 = _mm256_add_ps(fjy0,ty);
1011 fjz0 = _mm256_add_ps(fjz0,tz);
1013 /**************************
1014 * CALCULATE INTERACTIONS *
1015 **************************/
1017 /* Compute parameters for interactions between i and j atoms */
1018 qq20 = _mm256_mul_ps(iq2,jq0);
1020 /* COULOMB ELECTROSTATICS */
1021 velec = _mm256_mul_ps(qq20,rinv20);
1022 felec = _mm256_mul_ps(velec,rinvsq20);
1026 /* Calculate temporary vectorial force */
1027 tx = _mm256_mul_ps(fscal,dx20);
1028 ty = _mm256_mul_ps(fscal,dy20);
1029 tz = _mm256_mul_ps(fscal,dz20);
1031 /* Update vectorial force */
1032 fix2 = _mm256_add_ps(fix2,tx);
1033 fiy2 = _mm256_add_ps(fiy2,ty);
1034 fiz2 = _mm256_add_ps(fiz2,tz);
1036 fjx0 = _mm256_add_ps(fjx0,tx);
1037 fjy0 = _mm256_add_ps(fjy0,ty);
1038 fjz0 = _mm256_add_ps(fjz0,tz);
1040 fjptrA = f+j_coord_offsetA;
1041 fjptrB = f+j_coord_offsetB;
1042 fjptrC = f+j_coord_offsetC;
1043 fjptrD = f+j_coord_offsetD;
1044 fjptrE = f+j_coord_offsetE;
1045 fjptrF = f+j_coord_offsetF;
1046 fjptrG = f+j_coord_offsetG;
1047 fjptrH = f+j_coord_offsetH;
1049 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,fjx0,fjy0,fjz0);
1051 /* Inner loop uses 108 flops */
1054 if(jidx<j_index_end)
1057 /* Get j neighbor index, and coordinate index */
1058 jnrlistA = jjnr[jidx];
1059 jnrlistB = jjnr[jidx+1];
1060 jnrlistC = jjnr[jidx+2];
1061 jnrlistD = jjnr[jidx+3];
1062 jnrlistE = jjnr[jidx+4];
1063 jnrlistF = jjnr[jidx+5];
1064 jnrlistG = jjnr[jidx+6];
1065 jnrlistH = jjnr[jidx+7];
1066 /* Sign of each element will be negative for non-real atoms.
1067 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1068 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1070 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
1071 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
1073 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1074 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1075 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1076 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1077 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
1078 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
1079 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
1080 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
1081 j_coord_offsetA = DIM*jnrA;
1082 j_coord_offsetB = DIM*jnrB;
1083 j_coord_offsetC = DIM*jnrC;
1084 j_coord_offsetD = DIM*jnrD;
1085 j_coord_offsetE = DIM*jnrE;
1086 j_coord_offsetF = DIM*jnrF;
1087 j_coord_offsetG = DIM*jnrG;
1088 j_coord_offsetH = DIM*jnrH;
1090 /* load j atom coordinates */
1091 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1092 x+j_coord_offsetC,x+j_coord_offsetD,
1093 x+j_coord_offsetE,x+j_coord_offsetF,
1094 x+j_coord_offsetG,x+j_coord_offsetH,
1097 /* Calculate displacement vector */
1098 dx00 = _mm256_sub_ps(ix0,jx0);
1099 dy00 = _mm256_sub_ps(iy0,jy0);
1100 dz00 = _mm256_sub_ps(iz0,jz0);
1101 dx10 = _mm256_sub_ps(ix1,jx0);
1102 dy10 = _mm256_sub_ps(iy1,jy0);
1103 dz10 = _mm256_sub_ps(iz1,jz0);
1104 dx20 = _mm256_sub_ps(ix2,jx0);
1105 dy20 = _mm256_sub_ps(iy2,jy0);
1106 dz20 = _mm256_sub_ps(iz2,jz0);
1108 /* Calculate squared distance and things based on it */
1109 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
1110 rsq10 = gmx_mm256_calc_rsq_ps(dx10,dy10,dz10);
1111 rsq20 = gmx_mm256_calc_rsq_ps(dx20,dy20,dz20);
1113 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
1114 rinv10 = gmx_mm256_invsqrt_ps(rsq10);
1115 rinv20 = gmx_mm256_invsqrt_ps(rsq20);
1117 rinvsq00 = _mm256_mul_ps(rinv00,rinv00);
1118 rinvsq10 = _mm256_mul_ps(rinv10,rinv10);
1119 rinvsq20 = _mm256_mul_ps(rinv20,rinv20);
1121 /* Load parameters for j particles */
1122 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1123 charge+jnrC+0,charge+jnrD+0,
1124 charge+jnrE+0,charge+jnrF+0,
1125 charge+jnrG+0,charge+jnrH+0);
1126 vdwjidx0A = 2*vdwtype[jnrA+0];
1127 vdwjidx0B = 2*vdwtype[jnrB+0];
1128 vdwjidx0C = 2*vdwtype[jnrC+0];
1129 vdwjidx0D = 2*vdwtype[jnrD+0];
1130 vdwjidx0E = 2*vdwtype[jnrE+0];
1131 vdwjidx0F = 2*vdwtype[jnrF+0];
1132 vdwjidx0G = 2*vdwtype[jnrG+0];
1133 vdwjidx0H = 2*vdwtype[jnrH+0];
1135 fjx0 = _mm256_setzero_ps();
1136 fjy0 = _mm256_setzero_ps();
1137 fjz0 = _mm256_setzero_ps();
1139 /**************************
1140 * CALCULATE INTERACTIONS *
1141 **************************/
1143 r00 = _mm256_mul_ps(rsq00,rinv00);
1144 r00 = _mm256_andnot_ps(dummy_mask,r00);
1146 /* Compute parameters for interactions between i and j atoms */
1147 qq00 = _mm256_mul_ps(iq0,jq0);
1148 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
1149 vdwioffsetptr0+vdwjidx0B,
1150 vdwioffsetptr0+vdwjidx0C,
1151 vdwioffsetptr0+vdwjidx0D,
1152 vdwioffsetptr0+vdwjidx0E,
1153 vdwioffsetptr0+vdwjidx0F,
1154 vdwioffsetptr0+vdwjidx0G,
1155 vdwioffsetptr0+vdwjidx0H,
1158 /* Calculate table index by multiplying r with table scale and truncate to integer */
1159 rt = _mm256_mul_ps(r00,vftabscale);
1160 vfitab = _mm256_cvttps_epi32(rt);
1161 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
1162 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
1163 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
1164 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
1165 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
1166 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
1168 /* COULOMB ELECTROSTATICS */
1169 velec = _mm256_mul_ps(qq00,rinv00);
1170 felec = _mm256_mul_ps(velec,rinvsq00);
1172 /* CUBIC SPLINE TABLE DISPERSION */
1173 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1174 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1175 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1176 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1177 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1178 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1179 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1180 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1181 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1182 Heps = _mm256_mul_ps(vfeps,H);
1183 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1184 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1185 fvdw6 = _mm256_mul_ps(c6_00,FF);
1187 /* CUBIC SPLINE TABLE REPULSION */
1188 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
1189 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
1190 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
1191 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
1192 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
1193 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
1194 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
1195 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
1196 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
1197 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
1198 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
1199 Heps = _mm256_mul_ps(vfeps,H);
1200 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
1201 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
1202 fvdw12 = _mm256_mul_ps(c12_00,FF);
1203 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
1205 fscal = _mm256_add_ps(felec,fvdw);
1207 fscal = _mm256_andnot_ps(dummy_mask,fscal);
1209 /* Calculate temporary vectorial force */
1210 tx = _mm256_mul_ps(fscal,dx00);
1211 ty = _mm256_mul_ps(fscal,dy00);
1212 tz = _mm256_mul_ps(fscal,dz00);
1214 /* Update vectorial force */
1215 fix0 = _mm256_add_ps(fix0,tx);
1216 fiy0 = _mm256_add_ps(fiy0,ty);
1217 fiz0 = _mm256_add_ps(fiz0,tz);
1219 fjx0 = _mm256_add_ps(fjx0,tx);
1220 fjy0 = _mm256_add_ps(fjy0,ty);
1221 fjz0 = _mm256_add_ps(fjz0,tz);
1223 /**************************
1224 * CALCULATE INTERACTIONS *
1225 **************************/
1227 /* Compute parameters for interactions between i and j atoms */
1228 qq10 = _mm256_mul_ps(iq1,jq0);
1230 /* COULOMB ELECTROSTATICS */
1231 velec = _mm256_mul_ps(qq10,rinv10);
1232 felec = _mm256_mul_ps(velec,rinvsq10);
1236 fscal = _mm256_andnot_ps(dummy_mask,fscal);
1238 /* Calculate temporary vectorial force */
1239 tx = _mm256_mul_ps(fscal,dx10);
1240 ty = _mm256_mul_ps(fscal,dy10);
1241 tz = _mm256_mul_ps(fscal,dz10);
1243 /* Update vectorial force */
1244 fix1 = _mm256_add_ps(fix1,tx);
1245 fiy1 = _mm256_add_ps(fiy1,ty);
1246 fiz1 = _mm256_add_ps(fiz1,tz);
1248 fjx0 = _mm256_add_ps(fjx0,tx);
1249 fjy0 = _mm256_add_ps(fjy0,ty);
1250 fjz0 = _mm256_add_ps(fjz0,tz);
1252 /**************************
1253 * CALCULATE INTERACTIONS *
1254 **************************/
1256 /* Compute parameters for interactions between i and j atoms */
1257 qq20 = _mm256_mul_ps(iq2,jq0);
1259 /* COULOMB ELECTROSTATICS */
1260 velec = _mm256_mul_ps(qq20,rinv20);
1261 felec = _mm256_mul_ps(velec,rinvsq20);
1265 fscal = _mm256_andnot_ps(dummy_mask,fscal);
1267 /* Calculate temporary vectorial force */
1268 tx = _mm256_mul_ps(fscal,dx20);
1269 ty = _mm256_mul_ps(fscal,dy20);
1270 tz = _mm256_mul_ps(fscal,dz20);
1272 /* Update vectorial force */
1273 fix2 = _mm256_add_ps(fix2,tx);
1274 fiy2 = _mm256_add_ps(fiy2,ty);
1275 fiz2 = _mm256_add_ps(fiz2,tz);
1277 fjx0 = _mm256_add_ps(fjx0,tx);
1278 fjy0 = _mm256_add_ps(fjy0,ty);
1279 fjz0 = _mm256_add_ps(fjz0,tz);
1281 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1282 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1283 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1284 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1285 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
1286 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
1287 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
1288 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
1290 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,fjx0,fjy0,fjz0);
1292 /* Inner loop uses 109 flops */
1295 /* End of innermost loop */
1297 gmx_mm256_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1298 f+i_coord_offset,fshift+i_shift_offset);
1300 /* Increment number of inner iterations */
1301 inneriter += j_index_end - j_index_start;
1303 /* Outer loop uses 18 flops */
1306 /* Increment number of outer iterations */
1309 /* Update outer/inner flops */
1311 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*109);