2 * Note: this file was generated by the Gromacs avx_128_fma_double kernel generator.
4 * This source code is part of
8 * Copyright (c) 2001-2012, The GROMACS Development Team
10 * Gromacs is a library for molecular simulation and trajectory analysis,
11 * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
12 * a full list of developers and information, check out http://www.gromacs.org
14 * This program is free software; you can redistribute it and/or modify it under
15 * the terms of the GNU Lesser General Public License as published by the Free
16 * Software Foundation; either version 2 of the License, or (at your option) any
19 * To help fund GROMACS development, we humbly ask that you cite
20 * the papers people have written on it - you can find them on the website.
28 #include "../nb_kernel.h"
29 #include "types/simple.h"
33 #include "gmx_math_x86_avx_128_fma_double.h"
34 #include "kernelutil_x86_avx_128_fma_double.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwCSTab_GeomW3P1_VF_avx_128_fma_double
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_128_fma_double
45 (t_nblist * gmx_restrict nlist,
46 rvec * gmx_restrict xx,
47 rvec * gmx_restrict ff,
48 t_forcerec * gmx_restrict fr,
49 t_mdatoms * gmx_restrict mdatoms,
50 nb_kernel_data_t * gmx_restrict kernel_data,
51 t_nrnb * gmx_restrict nrnb)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
61 int j_coord_offsetA,j_coord_offsetB;
62 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
64 real *shiftvec,*fshift,*x,*f;
65 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
67 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
69 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
71 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
72 int vdwjidx0A,vdwjidx0B;
73 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
74 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
75 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
76 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
77 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
80 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
83 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
84 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
86 __m128i ifour = _mm_set1_epi32(4);
87 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
89 __m128d dummy_mask,cutoff_mask;
90 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
91 __m128d one = _mm_set1_pd(1.0);
92 __m128d two = _mm_set1_pd(2.0);
98 jindex = nlist->jindex;
100 shiftidx = nlist->shift;
102 shiftvec = fr->shift_vec[0];
103 fshift = fr->fshift[0];
104 facel = _mm_set1_pd(fr->epsfac);
105 charge = mdatoms->chargeA;
106 nvdwtype = fr->ntype;
108 vdwtype = mdatoms->typeA;
110 vftab = kernel_data->table_vdw->data;
111 vftabscale = _mm_set1_pd(kernel_data->table_vdw->scale);
113 /* Setup water-specific parameters */
114 inr = nlist->iinr[0];
115 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
116 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
117 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
118 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
120 /* Avoid stupid compiler warnings */
128 /* Start outer loop over neighborlists */
129 for(iidx=0; iidx<nri; iidx++)
131 /* Load shift vector for this list */
132 i_shift_offset = DIM*shiftidx[iidx];
134 /* Load limits for loop over neighbors */
135 j_index_start = jindex[iidx];
136 j_index_end = jindex[iidx+1];
138 /* Get outer coordinate index */
140 i_coord_offset = DIM*inr;
142 /* Load i particle coords and add shift vector */
143 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
144 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
146 fix0 = _mm_setzero_pd();
147 fiy0 = _mm_setzero_pd();
148 fiz0 = _mm_setzero_pd();
149 fix1 = _mm_setzero_pd();
150 fiy1 = _mm_setzero_pd();
151 fiz1 = _mm_setzero_pd();
152 fix2 = _mm_setzero_pd();
153 fiy2 = _mm_setzero_pd();
154 fiz2 = _mm_setzero_pd();
156 /* Reset potential sums */
157 velecsum = _mm_setzero_pd();
158 vvdwsum = _mm_setzero_pd();
160 /* Start inner kernel loop */
161 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
164 /* Get j neighbor index, and coordinate index */
167 j_coord_offsetA = DIM*jnrA;
168 j_coord_offsetB = DIM*jnrB;
170 /* load j atom coordinates */
171 gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
174 /* Calculate displacement vector */
175 dx00 = _mm_sub_pd(ix0,jx0);
176 dy00 = _mm_sub_pd(iy0,jy0);
177 dz00 = _mm_sub_pd(iz0,jz0);
178 dx10 = _mm_sub_pd(ix1,jx0);
179 dy10 = _mm_sub_pd(iy1,jy0);
180 dz10 = _mm_sub_pd(iz1,jz0);
181 dx20 = _mm_sub_pd(ix2,jx0);
182 dy20 = _mm_sub_pd(iy2,jy0);
183 dz20 = _mm_sub_pd(iz2,jz0);
185 /* Calculate squared distance and things based on it */
186 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
187 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
188 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
190 rinv00 = gmx_mm_invsqrt_pd(rsq00);
191 rinv10 = gmx_mm_invsqrt_pd(rsq10);
192 rinv20 = gmx_mm_invsqrt_pd(rsq20);
194 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
195 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
196 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
198 /* Load parameters for j particles */
199 jq0 = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
200 vdwjidx0A = 2*vdwtype[jnrA+0];
201 vdwjidx0B = 2*vdwtype[jnrB+0];
203 fjx0 = _mm_setzero_pd();
204 fjy0 = _mm_setzero_pd();
205 fjz0 = _mm_setzero_pd();
207 /**************************
208 * CALCULATE INTERACTIONS *
209 **************************/
211 r00 = _mm_mul_pd(rsq00,rinv00);
213 /* Compute parameters for interactions between i and j atoms */
214 qq00 = _mm_mul_pd(iq0,jq0);
215 gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
216 vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
218 /* Calculate table index by multiplying r with table scale and truncate to integer */
219 rt = _mm_mul_pd(r00,vftabscale);
220 vfitab = _mm_cvttpd_epi32(rt);
222 vfeps = _mm_frcz_pd(rt);
224 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
226 twovfeps = _mm_add_pd(vfeps,vfeps);
227 vfitab = _mm_slli_epi32(vfitab,3);
229 /* COULOMB ELECTROSTATICS */
230 velec = _mm_mul_pd(qq00,rinv00);
231 felec = _mm_mul_pd(velec,rinvsq00);
233 /* CUBIC SPLINE TABLE DISPERSION */
234 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
235 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
236 GMX_MM_TRANSPOSE2_PD(Y,F);
237 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
238 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
239 GMX_MM_TRANSPOSE2_PD(G,H);
240 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
241 VV = _mm_macc_pd(vfeps,Fp,Y);
242 vvdw6 = _mm_mul_pd(c6_00,VV);
243 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
244 fvdw6 = _mm_mul_pd(c6_00,FF);
246 /* CUBIC SPLINE TABLE REPULSION */
247 vfitab = _mm_add_epi32(vfitab,ifour);
248 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
249 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
250 GMX_MM_TRANSPOSE2_PD(Y,F);
251 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
252 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
253 GMX_MM_TRANSPOSE2_PD(G,H);
254 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
255 VV = _mm_macc_pd(vfeps,Fp,Y);
256 vvdw12 = _mm_mul_pd(c12_00,VV);
257 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
258 fvdw12 = _mm_mul_pd(c12_00,FF);
259 vvdw = _mm_add_pd(vvdw12,vvdw6);
260 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
262 /* Update potential sum for this i atom from the interaction with this j atom. */
263 velecsum = _mm_add_pd(velecsum,velec);
264 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
266 fscal = _mm_add_pd(felec,fvdw);
268 /* Update vectorial force */
269 fix0 = _mm_macc_pd(dx00,fscal,fix0);
270 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
271 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
273 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
274 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
275 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
277 /**************************
278 * CALCULATE INTERACTIONS *
279 **************************/
281 /* Compute parameters for interactions between i and j atoms */
282 qq10 = _mm_mul_pd(iq1,jq0);
284 /* COULOMB ELECTROSTATICS */
285 velec = _mm_mul_pd(qq10,rinv10);
286 felec = _mm_mul_pd(velec,rinvsq10);
288 /* Update potential sum for this i atom from the interaction with this j atom. */
289 velecsum = _mm_add_pd(velecsum,velec);
293 /* Update vectorial force */
294 fix1 = _mm_macc_pd(dx10,fscal,fix1);
295 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
296 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
298 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
299 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
300 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
302 /**************************
303 * CALCULATE INTERACTIONS *
304 **************************/
306 /* Compute parameters for interactions between i and j atoms */
307 qq20 = _mm_mul_pd(iq2,jq0);
309 /* COULOMB ELECTROSTATICS */
310 velec = _mm_mul_pd(qq20,rinv20);
311 felec = _mm_mul_pd(velec,rinvsq20);
313 /* Update potential sum for this i atom from the interaction with this j atom. */
314 velecsum = _mm_add_pd(velecsum,velec);
318 /* Update vectorial force */
319 fix2 = _mm_macc_pd(dx20,fscal,fix2);
320 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
321 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
323 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
324 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
325 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
327 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
329 /* Inner loop uses 131 flops */
336 j_coord_offsetA = DIM*jnrA;
338 /* load j atom coordinates */
339 gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
342 /* Calculate displacement vector */
343 dx00 = _mm_sub_pd(ix0,jx0);
344 dy00 = _mm_sub_pd(iy0,jy0);
345 dz00 = _mm_sub_pd(iz0,jz0);
346 dx10 = _mm_sub_pd(ix1,jx0);
347 dy10 = _mm_sub_pd(iy1,jy0);
348 dz10 = _mm_sub_pd(iz1,jz0);
349 dx20 = _mm_sub_pd(ix2,jx0);
350 dy20 = _mm_sub_pd(iy2,jy0);
351 dz20 = _mm_sub_pd(iz2,jz0);
353 /* Calculate squared distance and things based on it */
354 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
355 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
356 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
358 rinv00 = gmx_mm_invsqrt_pd(rsq00);
359 rinv10 = gmx_mm_invsqrt_pd(rsq10);
360 rinv20 = gmx_mm_invsqrt_pd(rsq20);
362 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
363 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
364 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
366 /* Load parameters for j particles */
367 jq0 = _mm_load_sd(charge+jnrA+0);
368 vdwjidx0A = 2*vdwtype[jnrA+0];
370 fjx0 = _mm_setzero_pd();
371 fjy0 = _mm_setzero_pd();
372 fjz0 = _mm_setzero_pd();
374 /**************************
375 * CALCULATE INTERACTIONS *
376 **************************/
378 r00 = _mm_mul_pd(rsq00,rinv00);
380 /* Compute parameters for interactions between i and j atoms */
381 qq00 = _mm_mul_pd(iq0,jq0);
382 gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
384 /* Calculate table index by multiplying r with table scale and truncate to integer */
385 rt = _mm_mul_pd(r00,vftabscale);
386 vfitab = _mm_cvttpd_epi32(rt);
388 vfeps = _mm_frcz_pd(rt);
390 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
392 twovfeps = _mm_add_pd(vfeps,vfeps);
393 vfitab = _mm_slli_epi32(vfitab,3);
395 /* COULOMB ELECTROSTATICS */
396 velec = _mm_mul_pd(qq00,rinv00);
397 felec = _mm_mul_pd(velec,rinvsq00);
399 /* CUBIC SPLINE TABLE DISPERSION */
400 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
401 F = _mm_setzero_pd();
402 GMX_MM_TRANSPOSE2_PD(Y,F);
403 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
404 H = _mm_setzero_pd();
405 GMX_MM_TRANSPOSE2_PD(G,H);
406 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
407 VV = _mm_macc_pd(vfeps,Fp,Y);
408 vvdw6 = _mm_mul_pd(c6_00,VV);
409 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
410 fvdw6 = _mm_mul_pd(c6_00,FF);
412 /* CUBIC SPLINE TABLE REPULSION */
413 vfitab = _mm_add_epi32(vfitab,ifour);
414 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
415 F = _mm_setzero_pd();
416 GMX_MM_TRANSPOSE2_PD(Y,F);
417 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
418 H = _mm_setzero_pd();
419 GMX_MM_TRANSPOSE2_PD(G,H);
420 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
421 VV = _mm_macc_pd(vfeps,Fp,Y);
422 vvdw12 = _mm_mul_pd(c12_00,VV);
423 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
424 fvdw12 = _mm_mul_pd(c12_00,FF);
425 vvdw = _mm_add_pd(vvdw12,vvdw6);
426 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
428 /* Update potential sum for this i atom from the interaction with this j atom. */
429 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
430 velecsum = _mm_add_pd(velecsum,velec);
431 vvdw = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
432 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
434 fscal = _mm_add_pd(felec,fvdw);
436 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
438 /* Update vectorial force */
439 fix0 = _mm_macc_pd(dx00,fscal,fix0);
440 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
441 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
443 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
444 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
445 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
447 /**************************
448 * CALCULATE INTERACTIONS *
449 **************************/
451 /* Compute parameters for interactions between i and j atoms */
452 qq10 = _mm_mul_pd(iq1,jq0);
454 /* COULOMB ELECTROSTATICS */
455 velec = _mm_mul_pd(qq10,rinv10);
456 felec = _mm_mul_pd(velec,rinvsq10);
458 /* Update potential sum for this i atom from the interaction with this j atom. */
459 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
460 velecsum = _mm_add_pd(velecsum,velec);
464 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
466 /* Update vectorial force */
467 fix1 = _mm_macc_pd(dx10,fscal,fix1);
468 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
469 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
471 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
472 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
473 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
475 /**************************
476 * CALCULATE INTERACTIONS *
477 **************************/
479 /* Compute parameters for interactions between i and j atoms */
480 qq20 = _mm_mul_pd(iq2,jq0);
482 /* COULOMB ELECTROSTATICS */
483 velec = _mm_mul_pd(qq20,rinv20);
484 felec = _mm_mul_pd(velec,rinvsq20);
486 /* Update potential sum for this i atom from the interaction with this j atom. */
487 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
488 velecsum = _mm_add_pd(velecsum,velec);
492 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
494 /* Update vectorial force */
495 fix2 = _mm_macc_pd(dx20,fscal,fix2);
496 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
497 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
499 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
500 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
501 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
503 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
505 /* Inner loop uses 131 flops */
508 /* End of innermost loop */
510 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
511 f+i_coord_offset,fshift+i_shift_offset);
514 /* Update potential energies */
515 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
516 gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
518 /* Increment number of inner iterations */
519 inneriter += j_index_end - j_index_start;
521 /* Outer loop uses 20 flops */
524 /* Increment number of outer iterations */
527 /* Update outer/inner flops */
529 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*131);
532 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwCSTab_GeomW3P1_F_avx_128_fma_double
533 * Electrostatics interaction: Coulomb
534 * VdW interaction: CubicSplineTable
535 * Geometry: Water3-Particle
536 * Calculate force/pot: Force
539 nb_kernel_ElecCoul_VdwCSTab_GeomW3P1_F_avx_128_fma_double
540 (t_nblist * gmx_restrict nlist,
541 rvec * gmx_restrict xx,
542 rvec * gmx_restrict ff,
543 t_forcerec * gmx_restrict fr,
544 t_mdatoms * gmx_restrict mdatoms,
545 nb_kernel_data_t * gmx_restrict kernel_data,
546 t_nrnb * gmx_restrict nrnb)
548 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
549 * just 0 for non-waters.
550 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
551 * jnr indices corresponding to data put in the four positions in the SIMD register.
553 int i_shift_offset,i_coord_offset,outeriter,inneriter;
554 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
556 int j_coord_offsetA,j_coord_offsetB;
557 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
559 real *shiftvec,*fshift,*x,*f;
560 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
562 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
564 __m128d ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
566 __m128d ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
567 int vdwjidx0A,vdwjidx0B;
568 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
569 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
570 __m128d dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
571 __m128d dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
572 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
575 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
578 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
579 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
581 __m128i ifour = _mm_set1_epi32(4);
582 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
584 __m128d dummy_mask,cutoff_mask;
585 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
586 __m128d one = _mm_set1_pd(1.0);
587 __m128d two = _mm_set1_pd(2.0);
593 jindex = nlist->jindex;
595 shiftidx = nlist->shift;
597 shiftvec = fr->shift_vec[0];
598 fshift = fr->fshift[0];
599 facel = _mm_set1_pd(fr->epsfac);
600 charge = mdatoms->chargeA;
601 nvdwtype = fr->ntype;
603 vdwtype = mdatoms->typeA;
605 vftab = kernel_data->table_vdw->data;
606 vftabscale = _mm_set1_pd(kernel_data->table_vdw->scale);
608 /* Setup water-specific parameters */
609 inr = nlist->iinr[0];
610 iq0 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
611 iq1 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
612 iq2 = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
613 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
615 /* Avoid stupid compiler warnings */
623 /* Start outer loop over neighborlists */
624 for(iidx=0; iidx<nri; iidx++)
626 /* Load shift vector for this list */
627 i_shift_offset = DIM*shiftidx[iidx];
629 /* Load limits for loop over neighbors */
630 j_index_start = jindex[iidx];
631 j_index_end = jindex[iidx+1];
633 /* Get outer coordinate index */
635 i_coord_offset = DIM*inr;
637 /* Load i particle coords and add shift vector */
638 gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
639 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
641 fix0 = _mm_setzero_pd();
642 fiy0 = _mm_setzero_pd();
643 fiz0 = _mm_setzero_pd();
644 fix1 = _mm_setzero_pd();
645 fiy1 = _mm_setzero_pd();
646 fiz1 = _mm_setzero_pd();
647 fix2 = _mm_setzero_pd();
648 fiy2 = _mm_setzero_pd();
649 fiz2 = _mm_setzero_pd();
651 /* Start inner kernel loop */
652 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
655 /* Get j neighbor index, and coordinate index */
658 j_coord_offsetA = DIM*jnrA;
659 j_coord_offsetB = DIM*jnrB;
661 /* load j atom coordinates */
662 gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
665 /* Calculate displacement vector */
666 dx00 = _mm_sub_pd(ix0,jx0);
667 dy00 = _mm_sub_pd(iy0,jy0);
668 dz00 = _mm_sub_pd(iz0,jz0);
669 dx10 = _mm_sub_pd(ix1,jx0);
670 dy10 = _mm_sub_pd(iy1,jy0);
671 dz10 = _mm_sub_pd(iz1,jz0);
672 dx20 = _mm_sub_pd(ix2,jx0);
673 dy20 = _mm_sub_pd(iy2,jy0);
674 dz20 = _mm_sub_pd(iz2,jz0);
676 /* Calculate squared distance and things based on it */
677 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
678 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
679 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
681 rinv00 = gmx_mm_invsqrt_pd(rsq00);
682 rinv10 = gmx_mm_invsqrt_pd(rsq10);
683 rinv20 = gmx_mm_invsqrt_pd(rsq20);
685 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
686 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
687 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
689 /* Load parameters for j particles */
690 jq0 = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
691 vdwjidx0A = 2*vdwtype[jnrA+0];
692 vdwjidx0B = 2*vdwtype[jnrB+0];
694 fjx0 = _mm_setzero_pd();
695 fjy0 = _mm_setzero_pd();
696 fjz0 = _mm_setzero_pd();
698 /**************************
699 * CALCULATE INTERACTIONS *
700 **************************/
702 r00 = _mm_mul_pd(rsq00,rinv00);
704 /* Compute parameters for interactions between i and j atoms */
705 qq00 = _mm_mul_pd(iq0,jq0);
706 gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
707 vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
709 /* Calculate table index by multiplying r with table scale and truncate to integer */
710 rt = _mm_mul_pd(r00,vftabscale);
711 vfitab = _mm_cvttpd_epi32(rt);
713 vfeps = _mm_frcz_pd(rt);
715 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
717 twovfeps = _mm_add_pd(vfeps,vfeps);
718 vfitab = _mm_slli_epi32(vfitab,3);
720 /* COULOMB ELECTROSTATICS */
721 velec = _mm_mul_pd(qq00,rinv00);
722 felec = _mm_mul_pd(velec,rinvsq00);
724 /* CUBIC SPLINE TABLE DISPERSION */
725 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
726 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
727 GMX_MM_TRANSPOSE2_PD(Y,F);
728 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
729 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
730 GMX_MM_TRANSPOSE2_PD(G,H);
731 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
732 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
733 fvdw6 = _mm_mul_pd(c6_00,FF);
735 /* CUBIC SPLINE TABLE REPULSION */
736 vfitab = _mm_add_epi32(vfitab,ifour);
737 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
738 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
739 GMX_MM_TRANSPOSE2_PD(Y,F);
740 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
741 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
742 GMX_MM_TRANSPOSE2_PD(G,H);
743 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
744 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
745 fvdw12 = _mm_mul_pd(c12_00,FF);
746 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
748 fscal = _mm_add_pd(felec,fvdw);
750 /* Update vectorial force */
751 fix0 = _mm_macc_pd(dx00,fscal,fix0);
752 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
753 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
755 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
756 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
757 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
759 /**************************
760 * CALCULATE INTERACTIONS *
761 **************************/
763 /* Compute parameters for interactions between i and j atoms */
764 qq10 = _mm_mul_pd(iq1,jq0);
766 /* COULOMB ELECTROSTATICS */
767 velec = _mm_mul_pd(qq10,rinv10);
768 felec = _mm_mul_pd(velec,rinvsq10);
772 /* Update vectorial force */
773 fix1 = _mm_macc_pd(dx10,fscal,fix1);
774 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
775 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
777 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
778 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
779 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
781 /**************************
782 * CALCULATE INTERACTIONS *
783 **************************/
785 /* Compute parameters for interactions between i and j atoms */
786 qq20 = _mm_mul_pd(iq2,jq0);
788 /* COULOMB ELECTROSTATICS */
789 velec = _mm_mul_pd(qq20,rinv20);
790 felec = _mm_mul_pd(velec,rinvsq20);
794 /* Update vectorial force */
795 fix2 = _mm_macc_pd(dx20,fscal,fix2);
796 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
797 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
799 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
800 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
801 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
803 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0);
805 /* Inner loop uses 120 flops */
812 j_coord_offsetA = DIM*jnrA;
814 /* load j atom coordinates */
815 gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
818 /* Calculate displacement vector */
819 dx00 = _mm_sub_pd(ix0,jx0);
820 dy00 = _mm_sub_pd(iy0,jy0);
821 dz00 = _mm_sub_pd(iz0,jz0);
822 dx10 = _mm_sub_pd(ix1,jx0);
823 dy10 = _mm_sub_pd(iy1,jy0);
824 dz10 = _mm_sub_pd(iz1,jz0);
825 dx20 = _mm_sub_pd(ix2,jx0);
826 dy20 = _mm_sub_pd(iy2,jy0);
827 dz20 = _mm_sub_pd(iz2,jz0);
829 /* Calculate squared distance and things based on it */
830 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
831 rsq10 = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
832 rsq20 = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
834 rinv00 = gmx_mm_invsqrt_pd(rsq00);
835 rinv10 = gmx_mm_invsqrt_pd(rsq10);
836 rinv20 = gmx_mm_invsqrt_pd(rsq20);
838 rinvsq00 = _mm_mul_pd(rinv00,rinv00);
839 rinvsq10 = _mm_mul_pd(rinv10,rinv10);
840 rinvsq20 = _mm_mul_pd(rinv20,rinv20);
842 /* Load parameters for j particles */
843 jq0 = _mm_load_sd(charge+jnrA+0);
844 vdwjidx0A = 2*vdwtype[jnrA+0];
846 fjx0 = _mm_setzero_pd();
847 fjy0 = _mm_setzero_pd();
848 fjz0 = _mm_setzero_pd();
850 /**************************
851 * CALCULATE INTERACTIONS *
852 **************************/
854 r00 = _mm_mul_pd(rsq00,rinv00);
856 /* Compute parameters for interactions between i and j atoms */
857 qq00 = _mm_mul_pd(iq0,jq0);
858 gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
860 /* Calculate table index by multiplying r with table scale and truncate to integer */
861 rt = _mm_mul_pd(r00,vftabscale);
862 vfitab = _mm_cvttpd_epi32(rt);
864 vfeps = _mm_frcz_pd(rt);
866 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
868 twovfeps = _mm_add_pd(vfeps,vfeps);
869 vfitab = _mm_slli_epi32(vfitab,3);
871 /* COULOMB ELECTROSTATICS */
872 velec = _mm_mul_pd(qq00,rinv00);
873 felec = _mm_mul_pd(velec,rinvsq00);
875 /* CUBIC SPLINE TABLE DISPERSION */
876 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
877 F = _mm_setzero_pd();
878 GMX_MM_TRANSPOSE2_PD(Y,F);
879 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
880 H = _mm_setzero_pd();
881 GMX_MM_TRANSPOSE2_PD(G,H);
882 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
883 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
884 fvdw6 = _mm_mul_pd(c6_00,FF);
886 /* CUBIC SPLINE TABLE REPULSION */
887 vfitab = _mm_add_epi32(vfitab,ifour);
888 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
889 F = _mm_setzero_pd();
890 GMX_MM_TRANSPOSE2_PD(Y,F);
891 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
892 H = _mm_setzero_pd();
893 GMX_MM_TRANSPOSE2_PD(G,H);
894 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
895 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
896 fvdw12 = _mm_mul_pd(c12_00,FF);
897 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
899 fscal = _mm_add_pd(felec,fvdw);
901 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
903 /* Update vectorial force */
904 fix0 = _mm_macc_pd(dx00,fscal,fix0);
905 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
906 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
908 fjx0 = _mm_macc_pd(dx00,fscal,fjx0);
909 fjy0 = _mm_macc_pd(dy00,fscal,fjy0);
910 fjz0 = _mm_macc_pd(dz00,fscal,fjz0);
912 /**************************
913 * CALCULATE INTERACTIONS *
914 **************************/
916 /* Compute parameters for interactions between i and j atoms */
917 qq10 = _mm_mul_pd(iq1,jq0);
919 /* COULOMB ELECTROSTATICS */
920 velec = _mm_mul_pd(qq10,rinv10);
921 felec = _mm_mul_pd(velec,rinvsq10);
925 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
927 /* Update vectorial force */
928 fix1 = _mm_macc_pd(dx10,fscal,fix1);
929 fiy1 = _mm_macc_pd(dy10,fscal,fiy1);
930 fiz1 = _mm_macc_pd(dz10,fscal,fiz1);
932 fjx0 = _mm_macc_pd(dx10,fscal,fjx0);
933 fjy0 = _mm_macc_pd(dy10,fscal,fjy0);
934 fjz0 = _mm_macc_pd(dz10,fscal,fjz0);
936 /**************************
937 * CALCULATE INTERACTIONS *
938 **************************/
940 /* Compute parameters for interactions between i and j atoms */
941 qq20 = _mm_mul_pd(iq2,jq0);
943 /* COULOMB ELECTROSTATICS */
944 velec = _mm_mul_pd(qq20,rinv20);
945 felec = _mm_mul_pd(velec,rinvsq20);
949 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
951 /* Update vectorial force */
952 fix2 = _mm_macc_pd(dx20,fscal,fix2);
953 fiy2 = _mm_macc_pd(dy20,fscal,fiy2);
954 fiz2 = _mm_macc_pd(dz20,fscal,fiz2);
956 fjx0 = _mm_macc_pd(dx20,fscal,fjx0);
957 fjy0 = _mm_macc_pd(dy20,fscal,fjy0);
958 fjz0 = _mm_macc_pd(dz20,fscal,fjz0);
960 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0);
962 /* Inner loop uses 120 flops */
965 /* End of innermost loop */
967 gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
968 f+i_coord_offset,fshift+i_shift_offset);
970 /* Increment number of inner iterations */
971 inneriter += j_index_end - j_index_start;
973 /* Outer loop uses 18 flops */
976 /* Increment number of outer iterations */
979 /* Update outer/inner flops */
981 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*120);