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_ElecNone_VdwLJSw_GeomP1P1_VF_avx_256_single
38 * Electrostatics interaction: None
39 * VdW interaction: LennardJones
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_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 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
75 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
76 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
78 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
81 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
82 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
83 __m256 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
84 real rswitch_scalar,d_scalar;
85 __m256 dummy_mask,cutoff_mask;
86 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
87 __m256 one = _mm256_set1_ps(1.0);
88 __m256 two = _mm256_set1_ps(2.0);
94 jindex = nlist->jindex;
96 shiftidx = nlist->shift;
98 shiftvec = fr->shift_vec[0];
99 fshift = fr->fshift[0];
100 nvdwtype = fr->ntype;
102 vdwtype = mdatoms->typeA;
104 rcutoff_scalar = fr->rvdw;
105 rcutoff = _mm256_set1_ps(rcutoff_scalar);
106 rcutoff2 = _mm256_mul_ps(rcutoff,rcutoff);
108 rswitch_scalar = fr->rvdw_switch;
109 rswitch = _mm256_set1_ps(rswitch_scalar);
110 /* Setup switch parameters */
111 d_scalar = rcutoff_scalar-rswitch_scalar;
112 d = _mm256_set1_ps(d_scalar);
113 swV3 = _mm256_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
114 swV4 = _mm256_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
115 swV5 = _mm256_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
116 swF2 = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
117 swF3 = _mm256_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
118 swF4 = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
120 /* Avoid stupid compiler warnings */
121 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
134 for(iidx=0;iidx<4*DIM;iidx++)
139 /* Start outer loop over neighborlists */
140 for(iidx=0; iidx<nri; iidx++)
142 /* Load shift vector for this list */
143 i_shift_offset = DIM*shiftidx[iidx];
145 /* Load limits for loop over neighbors */
146 j_index_start = jindex[iidx];
147 j_index_end = jindex[iidx+1];
149 /* Get outer coordinate index */
151 i_coord_offset = DIM*inr;
153 /* Load i particle coords and add shift vector */
154 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
156 fix0 = _mm256_setzero_ps();
157 fiy0 = _mm256_setzero_ps();
158 fiz0 = _mm256_setzero_ps();
160 /* Load parameters for i particles */
161 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
163 /* Reset potential sums */
164 vvdwsum = _mm256_setzero_ps();
166 /* Start inner kernel loop */
167 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
170 /* 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;
183 j_coord_offsetE = DIM*jnrE;
184 j_coord_offsetF = DIM*jnrF;
185 j_coord_offsetG = DIM*jnrG;
186 j_coord_offsetH = DIM*jnrH;
188 /* load j atom coordinates */
189 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
190 x+j_coord_offsetC,x+j_coord_offsetD,
191 x+j_coord_offsetE,x+j_coord_offsetF,
192 x+j_coord_offsetG,x+j_coord_offsetH,
195 /* Calculate displacement vector */
196 dx00 = _mm256_sub_ps(ix0,jx0);
197 dy00 = _mm256_sub_ps(iy0,jy0);
198 dz00 = _mm256_sub_ps(iz0,jz0);
200 /* Calculate squared distance and things based on it */
201 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
203 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
205 rinvsq00 = _mm256_mul_ps(rinv00,rinv00);
207 /* Load parameters for j particles */
208 vdwjidx0A = 2*vdwtype[jnrA+0];
209 vdwjidx0B = 2*vdwtype[jnrB+0];
210 vdwjidx0C = 2*vdwtype[jnrC+0];
211 vdwjidx0D = 2*vdwtype[jnrD+0];
212 vdwjidx0E = 2*vdwtype[jnrE+0];
213 vdwjidx0F = 2*vdwtype[jnrF+0];
214 vdwjidx0G = 2*vdwtype[jnrG+0];
215 vdwjidx0H = 2*vdwtype[jnrH+0];
217 /**************************
218 * CALCULATE INTERACTIONS *
219 **************************/
221 if (gmx_mm256_any_lt(rsq00,rcutoff2))
224 r00 = _mm256_mul_ps(rsq00,rinv00);
226 /* Compute parameters for interactions between i and j atoms */
227 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
228 vdwioffsetptr0+vdwjidx0B,
229 vdwioffsetptr0+vdwjidx0C,
230 vdwioffsetptr0+vdwjidx0D,
231 vdwioffsetptr0+vdwjidx0E,
232 vdwioffsetptr0+vdwjidx0F,
233 vdwioffsetptr0+vdwjidx0G,
234 vdwioffsetptr0+vdwjidx0H,
237 /* LENNARD-JONES DISPERSION/REPULSION */
239 rinvsix = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
240 vvdw6 = _mm256_mul_ps(c6_00,rinvsix);
241 vvdw12 = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
242 vvdw = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
243 fvdw = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
245 d = _mm256_sub_ps(r00,rswitch);
246 d = _mm256_max_ps(d,_mm256_setzero_ps());
247 d2 = _mm256_mul_ps(d,d);
248 sw = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
250 dsw = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
252 /* Evaluate switch function */
253 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
254 fvdw = _mm256_sub_ps( _mm256_mul_ps(fvdw,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(vvdw,dsw)) );
255 vvdw = _mm256_mul_ps(vvdw,sw);
256 cutoff_mask = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
258 /* Update potential sum for this i atom from the interaction with this j atom. */
259 vvdw = _mm256_and_ps(vvdw,cutoff_mask);
260 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
264 fscal = _mm256_and_ps(fscal,cutoff_mask);
266 /* Calculate temporary vectorial force */
267 tx = _mm256_mul_ps(fscal,dx00);
268 ty = _mm256_mul_ps(fscal,dy00);
269 tz = _mm256_mul_ps(fscal,dz00);
271 /* Update vectorial force */
272 fix0 = _mm256_add_ps(fix0,tx);
273 fiy0 = _mm256_add_ps(fiy0,ty);
274 fiz0 = _mm256_add_ps(fiz0,tz);
276 fjptrA = f+j_coord_offsetA;
277 fjptrB = f+j_coord_offsetB;
278 fjptrC = f+j_coord_offsetC;
279 fjptrD = f+j_coord_offsetD;
280 fjptrE = f+j_coord_offsetE;
281 fjptrF = f+j_coord_offsetF;
282 fjptrG = f+j_coord_offsetG;
283 fjptrH = f+j_coord_offsetH;
284 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
288 /* Inner loop uses 59 flops */
294 /* Get j neighbor index, and coordinate index */
295 jnrlistA = jjnr[jidx];
296 jnrlistB = jjnr[jidx+1];
297 jnrlistC = jjnr[jidx+2];
298 jnrlistD = jjnr[jidx+3];
299 jnrlistE = jjnr[jidx+4];
300 jnrlistF = jjnr[jidx+5];
301 jnrlistG = jjnr[jidx+6];
302 jnrlistH = jjnr[jidx+7];
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_ps(mask,val) to clear dummy entries.
307 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
308 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
310 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
311 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
312 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
313 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
314 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
315 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
316 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
317 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
318 j_coord_offsetA = DIM*jnrA;
319 j_coord_offsetB = DIM*jnrB;
320 j_coord_offsetC = DIM*jnrC;
321 j_coord_offsetD = DIM*jnrD;
322 j_coord_offsetE = DIM*jnrE;
323 j_coord_offsetF = DIM*jnrF;
324 j_coord_offsetG = DIM*jnrG;
325 j_coord_offsetH = DIM*jnrH;
327 /* load j atom coordinates */
328 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
329 x+j_coord_offsetC,x+j_coord_offsetD,
330 x+j_coord_offsetE,x+j_coord_offsetF,
331 x+j_coord_offsetG,x+j_coord_offsetH,
334 /* Calculate displacement vector */
335 dx00 = _mm256_sub_ps(ix0,jx0);
336 dy00 = _mm256_sub_ps(iy0,jy0);
337 dz00 = _mm256_sub_ps(iz0,jz0);
339 /* Calculate squared distance and things based on it */
340 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
342 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
344 rinvsq00 = _mm256_mul_ps(rinv00,rinv00);
346 /* Load parameters for j particles */
347 vdwjidx0A = 2*vdwtype[jnrA+0];
348 vdwjidx0B = 2*vdwtype[jnrB+0];
349 vdwjidx0C = 2*vdwtype[jnrC+0];
350 vdwjidx0D = 2*vdwtype[jnrD+0];
351 vdwjidx0E = 2*vdwtype[jnrE+0];
352 vdwjidx0F = 2*vdwtype[jnrF+0];
353 vdwjidx0G = 2*vdwtype[jnrG+0];
354 vdwjidx0H = 2*vdwtype[jnrH+0];
356 /**************************
357 * CALCULATE INTERACTIONS *
358 **************************/
360 if (gmx_mm256_any_lt(rsq00,rcutoff2))
363 r00 = _mm256_mul_ps(rsq00,rinv00);
364 r00 = _mm256_andnot_ps(dummy_mask,r00);
366 /* Compute parameters for interactions between i and j atoms */
367 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
368 vdwioffsetptr0+vdwjidx0B,
369 vdwioffsetptr0+vdwjidx0C,
370 vdwioffsetptr0+vdwjidx0D,
371 vdwioffsetptr0+vdwjidx0E,
372 vdwioffsetptr0+vdwjidx0F,
373 vdwioffsetptr0+vdwjidx0G,
374 vdwioffsetptr0+vdwjidx0H,
377 /* LENNARD-JONES DISPERSION/REPULSION */
379 rinvsix = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
380 vvdw6 = _mm256_mul_ps(c6_00,rinvsix);
381 vvdw12 = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
382 vvdw = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
383 fvdw = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
385 d = _mm256_sub_ps(r00,rswitch);
386 d = _mm256_max_ps(d,_mm256_setzero_ps());
387 d2 = _mm256_mul_ps(d,d);
388 sw = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
390 dsw = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
392 /* Evaluate switch function */
393 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
394 fvdw = _mm256_sub_ps( _mm256_mul_ps(fvdw,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(vvdw,dsw)) );
395 vvdw = _mm256_mul_ps(vvdw,sw);
396 cutoff_mask = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
398 /* Update potential sum for this i atom from the interaction with this j atom. */
399 vvdw = _mm256_and_ps(vvdw,cutoff_mask);
400 vvdw = _mm256_andnot_ps(dummy_mask,vvdw);
401 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
405 fscal = _mm256_and_ps(fscal,cutoff_mask);
407 fscal = _mm256_andnot_ps(dummy_mask,fscal);
409 /* Calculate temporary vectorial force */
410 tx = _mm256_mul_ps(fscal,dx00);
411 ty = _mm256_mul_ps(fscal,dy00);
412 tz = _mm256_mul_ps(fscal,dz00);
414 /* Update vectorial force */
415 fix0 = _mm256_add_ps(fix0,tx);
416 fiy0 = _mm256_add_ps(fiy0,ty);
417 fiz0 = _mm256_add_ps(fiz0,tz);
419 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
420 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
421 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
422 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
423 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
424 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
425 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
426 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
427 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
431 /* Inner loop uses 60 flops */
434 /* End of innermost loop */
436 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
437 f+i_coord_offset,fshift+i_shift_offset);
440 /* Update potential energies */
441 gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
443 /* Increment number of inner iterations */
444 inneriter += j_index_end - j_index_start;
446 /* Outer loop uses 7 flops */
449 /* Increment number of outer iterations */
452 /* Update outer/inner flops */
454 inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*60);
457 * Gromacs nonbonded kernel: nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_256_single
458 * Electrostatics interaction: None
459 * VdW interaction: LennardJones
460 * Geometry: Particle-Particle
461 * Calculate force/pot: Force
464 nb_kernel_ElecNone_VdwLJSw_GeomP1P1_F_avx_256_single
465 (t_nblist * gmx_restrict nlist,
466 rvec * gmx_restrict xx,
467 rvec * gmx_restrict ff,
468 t_forcerec * gmx_restrict fr,
469 t_mdatoms * gmx_restrict mdatoms,
470 nb_kernel_data_t * gmx_restrict kernel_data,
471 t_nrnb * gmx_restrict nrnb)
473 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
474 * just 0 for non-waters.
475 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
476 * jnr indices corresponding to data put in the four positions in the SIMD register.
478 int i_shift_offset,i_coord_offset,outeriter,inneriter;
479 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
480 int jnrA,jnrB,jnrC,jnrD;
481 int jnrE,jnrF,jnrG,jnrH;
482 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
483 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
484 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
485 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
486 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
488 real *shiftvec,*fshift,*x,*f;
489 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
491 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
492 real * vdwioffsetptr0;
493 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
494 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
495 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
496 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
498 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
501 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
502 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
503 __m256 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
504 real rswitch_scalar,d_scalar;
505 __m256 dummy_mask,cutoff_mask;
506 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
507 __m256 one = _mm256_set1_ps(1.0);
508 __m256 two = _mm256_set1_ps(2.0);
514 jindex = nlist->jindex;
516 shiftidx = nlist->shift;
518 shiftvec = fr->shift_vec[0];
519 fshift = fr->fshift[0];
520 nvdwtype = fr->ntype;
522 vdwtype = mdatoms->typeA;
524 rcutoff_scalar = fr->rvdw;
525 rcutoff = _mm256_set1_ps(rcutoff_scalar);
526 rcutoff2 = _mm256_mul_ps(rcutoff,rcutoff);
528 rswitch_scalar = fr->rvdw_switch;
529 rswitch = _mm256_set1_ps(rswitch_scalar);
530 /* Setup switch parameters */
531 d_scalar = rcutoff_scalar-rswitch_scalar;
532 d = _mm256_set1_ps(d_scalar);
533 swV3 = _mm256_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
534 swV4 = _mm256_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
535 swV5 = _mm256_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
536 swF2 = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
537 swF3 = _mm256_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
538 swF4 = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
540 /* Avoid stupid compiler warnings */
541 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
554 for(iidx=0;iidx<4*DIM;iidx++)
559 /* Start outer loop over neighborlists */
560 for(iidx=0; iidx<nri; iidx++)
562 /* Load shift vector for this list */
563 i_shift_offset = DIM*shiftidx[iidx];
565 /* Load limits for loop over neighbors */
566 j_index_start = jindex[iidx];
567 j_index_end = jindex[iidx+1];
569 /* Get outer coordinate index */
571 i_coord_offset = DIM*inr;
573 /* Load i particle coords and add shift vector */
574 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
576 fix0 = _mm256_setzero_ps();
577 fiy0 = _mm256_setzero_ps();
578 fiz0 = _mm256_setzero_ps();
580 /* Load parameters for i particles */
581 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
583 /* Start inner kernel loop */
584 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
587 /* Get j neighbor index, and coordinate index */
596 j_coord_offsetA = DIM*jnrA;
597 j_coord_offsetB = DIM*jnrB;
598 j_coord_offsetC = DIM*jnrC;
599 j_coord_offsetD = DIM*jnrD;
600 j_coord_offsetE = DIM*jnrE;
601 j_coord_offsetF = DIM*jnrF;
602 j_coord_offsetG = DIM*jnrG;
603 j_coord_offsetH = DIM*jnrH;
605 /* load j atom coordinates */
606 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
607 x+j_coord_offsetC,x+j_coord_offsetD,
608 x+j_coord_offsetE,x+j_coord_offsetF,
609 x+j_coord_offsetG,x+j_coord_offsetH,
612 /* Calculate displacement vector */
613 dx00 = _mm256_sub_ps(ix0,jx0);
614 dy00 = _mm256_sub_ps(iy0,jy0);
615 dz00 = _mm256_sub_ps(iz0,jz0);
617 /* Calculate squared distance and things based on it */
618 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
620 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
622 rinvsq00 = _mm256_mul_ps(rinv00,rinv00);
624 /* Load parameters for j particles */
625 vdwjidx0A = 2*vdwtype[jnrA+0];
626 vdwjidx0B = 2*vdwtype[jnrB+0];
627 vdwjidx0C = 2*vdwtype[jnrC+0];
628 vdwjidx0D = 2*vdwtype[jnrD+0];
629 vdwjidx0E = 2*vdwtype[jnrE+0];
630 vdwjidx0F = 2*vdwtype[jnrF+0];
631 vdwjidx0G = 2*vdwtype[jnrG+0];
632 vdwjidx0H = 2*vdwtype[jnrH+0];
634 /**************************
635 * CALCULATE INTERACTIONS *
636 **************************/
638 if (gmx_mm256_any_lt(rsq00,rcutoff2))
641 r00 = _mm256_mul_ps(rsq00,rinv00);
643 /* Compute parameters for interactions between i and j atoms */
644 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
645 vdwioffsetptr0+vdwjidx0B,
646 vdwioffsetptr0+vdwjidx0C,
647 vdwioffsetptr0+vdwjidx0D,
648 vdwioffsetptr0+vdwjidx0E,
649 vdwioffsetptr0+vdwjidx0F,
650 vdwioffsetptr0+vdwjidx0G,
651 vdwioffsetptr0+vdwjidx0H,
654 /* LENNARD-JONES DISPERSION/REPULSION */
656 rinvsix = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
657 vvdw6 = _mm256_mul_ps(c6_00,rinvsix);
658 vvdw12 = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
659 vvdw = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
660 fvdw = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
662 d = _mm256_sub_ps(r00,rswitch);
663 d = _mm256_max_ps(d,_mm256_setzero_ps());
664 d2 = _mm256_mul_ps(d,d);
665 sw = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
667 dsw = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
669 /* Evaluate switch function */
670 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
671 fvdw = _mm256_sub_ps( _mm256_mul_ps(fvdw,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(vvdw,dsw)) );
672 cutoff_mask = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
676 fscal = _mm256_and_ps(fscal,cutoff_mask);
678 /* Calculate temporary vectorial force */
679 tx = _mm256_mul_ps(fscal,dx00);
680 ty = _mm256_mul_ps(fscal,dy00);
681 tz = _mm256_mul_ps(fscal,dz00);
683 /* Update vectorial force */
684 fix0 = _mm256_add_ps(fix0,tx);
685 fiy0 = _mm256_add_ps(fiy0,ty);
686 fiz0 = _mm256_add_ps(fiz0,tz);
688 fjptrA = f+j_coord_offsetA;
689 fjptrB = f+j_coord_offsetB;
690 fjptrC = f+j_coord_offsetC;
691 fjptrD = f+j_coord_offsetD;
692 fjptrE = f+j_coord_offsetE;
693 fjptrF = f+j_coord_offsetF;
694 fjptrG = f+j_coord_offsetG;
695 fjptrH = f+j_coord_offsetH;
696 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
700 /* Inner loop uses 56 flops */
706 /* Get j neighbor index, and coordinate index */
707 jnrlistA = jjnr[jidx];
708 jnrlistB = jjnr[jidx+1];
709 jnrlistC = jjnr[jidx+2];
710 jnrlistD = jjnr[jidx+3];
711 jnrlistE = jjnr[jidx+4];
712 jnrlistF = jjnr[jidx+5];
713 jnrlistG = jjnr[jidx+6];
714 jnrlistH = jjnr[jidx+7];
715 /* Sign of each element will be negative for non-real atoms.
716 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
717 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
719 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
720 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
722 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
723 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
724 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
725 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
726 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
727 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
728 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
729 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
730 j_coord_offsetA = DIM*jnrA;
731 j_coord_offsetB = DIM*jnrB;
732 j_coord_offsetC = DIM*jnrC;
733 j_coord_offsetD = DIM*jnrD;
734 j_coord_offsetE = DIM*jnrE;
735 j_coord_offsetF = DIM*jnrF;
736 j_coord_offsetG = DIM*jnrG;
737 j_coord_offsetH = DIM*jnrH;
739 /* load j atom coordinates */
740 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
741 x+j_coord_offsetC,x+j_coord_offsetD,
742 x+j_coord_offsetE,x+j_coord_offsetF,
743 x+j_coord_offsetG,x+j_coord_offsetH,
746 /* Calculate displacement vector */
747 dx00 = _mm256_sub_ps(ix0,jx0);
748 dy00 = _mm256_sub_ps(iy0,jy0);
749 dz00 = _mm256_sub_ps(iz0,jz0);
751 /* Calculate squared distance and things based on it */
752 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
754 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
756 rinvsq00 = _mm256_mul_ps(rinv00,rinv00);
758 /* Load parameters for j particles */
759 vdwjidx0A = 2*vdwtype[jnrA+0];
760 vdwjidx0B = 2*vdwtype[jnrB+0];
761 vdwjidx0C = 2*vdwtype[jnrC+0];
762 vdwjidx0D = 2*vdwtype[jnrD+0];
763 vdwjidx0E = 2*vdwtype[jnrE+0];
764 vdwjidx0F = 2*vdwtype[jnrF+0];
765 vdwjidx0G = 2*vdwtype[jnrG+0];
766 vdwjidx0H = 2*vdwtype[jnrH+0];
768 /**************************
769 * CALCULATE INTERACTIONS *
770 **************************/
772 if (gmx_mm256_any_lt(rsq00,rcutoff2))
775 r00 = _mm256_mul_ps(rsq00,rinv00);
776 r00 = _mm256_andnot_ps(dummy_mask,r00);
778 /* Compute parameters for interactions between i and j atoms */
779 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
780 vdwioffsetptr0+vdwjidx0B,
781 vdwioffsetptr0+vdwjidx0C,
782 vdwioffsetptr0+vdwjidx0D,
783 vdwioffsetptr0+vdwjidx0E,
784 vdwioffsetptr0+vdwjidx0F,
785 vdwioffsetptr0+vdwjidx0G,
786 vdwioffsetptr0+vdwjidx0H,
789 /* LENNARD-JONES DISPERSION/REPULSION */
791 rinvsix = _mm256_mul_ps(_mm256_mul_ps(rinvsq00,rinvsq00),rinvsq00);
792 vvdw6 = _mm256_mul_ps(c6_00,rinvsix);
793 vvdw12 = _mm256_mul_ps(c12_00,_mm256_mul_ps(rinvsix,rinvsix));
794 vvdw = _mm256_sub_ps( _mm256_mul_ps(vvdw12,one_twelfth) , _mm256_mul_ps(vvdw6,one_sixth) );
795 fvdw = _mm256_mul_ps(_mm256_sub_ps(vvdw12,vvdw6),rinvsq00);
797 d = _mm256_sub_ps(r00,rswitch);
798 d = _mm256_max_ps(d,_mm256_setzero_ps());
799 d2 = _mm256_mul_ps(d,d);
800 sw = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
802 dsw = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
804 /* Evaluate switch function */
805 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
806 fvdw = _mm256_sub_ps( _mm256_mul_ps(fvdw,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(vvdw,dsw)) );
807 cutoff_mask = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
811 fscal = _mm256_and_ps(fscal,cutoff_mask);
813 fscal = _mm256_andnot_ps(dummy_mask,fscal);
815 /* Calculate temporary vectorial force */
816 tx = _mm256_mul_ps(fscal,dx00);
817 ty = _mm256_mul_ps(fscal,dy00);
818 tz = _mm256_mul_ps(fscal,dz00);
820 /* Update vectorial force */
821 fix0 = _mm256_add_ps(fix0,tx);
822 fiy0 = _mm256_add_ps(fiy0,ty);
823 fiz0 = _mm256_add_ps(fiz0,tz);
825 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
826 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
827 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
828 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
829 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
830 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
831 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
832 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
833 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
837 /* Inner loop uses 57 flops */
840 /* End of innermost loop */
842 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
843 f+i_coord_offset,fshift+i_shift_offset);
845 /* Increment number of inner iterations */
846 inneriter += j_index_end - j_index_start;
848 /* Outer loop uses 6 flops */
851 /* Increment number of outer iterations */
854 /* Update outer/inner flops */
856 inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*57);