2 * Note: this file was generated by the Gromacs avx_128_fma_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_128_fma_single.h"
34 #include "kernelutil_x86_avx_128_fma_single.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecEwSw_VdwNone_GeomW3P1_VF_avx_128_fma_single
38 * Electrostatics interaction: Ewald
39 * VdW interaction: None
40 * Geometry: Water3-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEwSw_VdwNone_GeomW3P1_VF_avx_128_fma_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 refer to j loop unrolling done with AVX_128, e.g. for the four different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
62 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
63 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
65 real *shiftvec,*fshift,*x,*f;
66 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
68 __m128 fscal,rcutoff,rcutoff2,jidxall;
70 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
72 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
74 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
75 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
76 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
77 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
78 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
79 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
80 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
83 __m128 ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
84 __m128 beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
86 __m128 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
87 real rswitch_scalar,d_scalar;
88 __m128 dummy_mask,cutoff_mask;
89 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
90 __m128 one = _mm_set1_ps(1.0);
91 __m128 two = _mm_set1_ps(2.0);
97 jindex = nlist->jindex;
99 shiftidx = nlist->shift;
101 shiftvec = fr->shift_vec[0];
102 fshift = fr->fshift[0];
103 facel = _mm_set1_ps(fr->epsfac);
104 charge = mdatoms->chargeA;
106 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
107 beta = _mm_set1_ps(fr->ic->ewaldcoeff);
108 beta2 = _mm_mul_ps(beta,beta);
109 beta3 = _mm_mul_ps(beta,beta2);
110 ewtab = fr->ic->tabq_coul_FDV0;
111 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
112 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
114 /* Setup water-specific parameters */
115 inr = nlist->iinr[0];
116 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
117 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
118 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
120 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
121 rcutoff_scalar = fr->rcoulomb;
122 rcutoff = _mm_set1_ps(rcutoff_scalar);
123 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
125 rswitch_scalar = fr->rcoulomb_switch;
126 rswitch = _mm_set1_ps(rswitch_scalar);
127 /* Setup switch parameters */
128 d_scalar = rcutoff_scalar-rswitch_scalar;
129 d = _mm_set1_ps(d_scalar);
130 swV3 = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
131 swV4 = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
132 swV5 = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
133 swF2 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
134 swF3 = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
135 swF4 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
137 /* Avoid stupid compiler warnings */
138 jnrA = jnrB = jnrC = jnrD = 0;
147 for(iidx=0;iidx<4*DIM;iidx++)
152 /* Start outer loop over neighborlists */
153 for(iidx=0; iidx<nri; iidx++)
155 /* Load shift vector for this list */
156 i_shift_offset = DIM*shiftidx[iidx];
158 /* Load limits for loop over neighbors */
159 j_index_start = jindex[iidx];
160 j_index_end = jindex[iidx+1];
162 /* Get outer coordinate index */
164 i_coord_offset = DIM*inr;
166 /* Load i particle coords and add shift vector */
167 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
168 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
170 fix0 = _mm_setzero_ps();
171 fiy0 = _mm_setzero_ps();
172 fiz0 = _mm_setzero_ps();
173 fix1 = _mm_setzero_ps();
174 fiy1 = _mm_setzero_ps();
175 fiz1 = _mm_setzero_ps();
176 fix2 = _mm_setzero_ps();
177 fiy2 = _mm_setzero_ps();
178 fiz2 = _mm_setzero_ps();
180 /* Reset potential sums */
181 velecsum = _mm_setzero_ps();
183 /* Start inner kernel loop */
184 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
187 /* Get j neighbor index, and coordinate index */
192 j_coord_offsetA = DIM*jnrA;
193 j_coord_offsetB = DIM*jnrB;
194 j_coord_offsetC = DIM*jnrC;
195 j_coord_offsetD = DIM*jnrD;
197 /* load j atom coordinates */
198 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
199 x+j_coord_offsetC,x+j_coord_offsetD,
202 /* Calculate displacement vector */
203 dx00 = _mm_sub_ps(ix0,jx0);
204 dy00 = _mm_sub_ps(iy0,jy0);
205 dz00 = _mm_sub_ps(iz0,jz0);
206 dx10 = _mm_sub_ps(ix1,jx0);
207 dy10 = _mm_sub_ps(iy1,jy0);
208 dz10 = _mm_sub_ps(iz1,jz0);
209 dx20 = _mm_sub_ps(ix2,jx0);
210 dy20 = _mm_sub_ps(iy2,jy0);
211 dz20 = _mm_sub_ps(iz2,jz0);
213 /* Calculate squared distance and things based on it */
214 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
215 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
216 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
218 rinv00 = gmx_mm_invsqrt_ps(rsq00);
219 rinv10 = gmx_mm_invsqrt_ps(rsq10);
220 rinv20 = gmx_mm_invsqrt_ps(rsq20);
222 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
223 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
224 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
226 /* Load parameters for j particles */
227 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
228 charge+jnrC+0,charge+jnrD+0);
230 fjx0 = _mm_setzero_ps();
231 fjy0 = _mm_setzero_ps();
232 fjz0 = _mm_setzero_ps();
234 /**************************
235 * CALCULATE INTERACTIONS *
236 **************************/
238 if (gmx_mm_any_lt(rsq00,rcutoff2))
241 r00 = _mm_mul_ps(rsq00,rinv00);
243 /* Compute parameters for interactions between i and j atoms */
244 qq00 = _mm_mul_ps(iq0,jq0);
246 /* EWALD ELECTROSTATICS */
248 /* Analytical PME correction */
249 zeta2 = _mm_mul_ps(beta2,rsq00);
250 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
251 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
252 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
253 felec = _mm_mul_ps(qq00,felec);
254 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
255 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
256 velec = _mm_mul_ps(qq00,velec);
258 d = _mm_sub_ps(r00,rswitch);
259 d = _mm_max_ps(d,_mm_setzero_ps());
260 d2 = _mm_mul_ps(d,d);
261 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
263 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
265 /* Evaluate switch function */
266 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
267 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
268 velec = _mm_mul_ps(velec,sw);
269 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
271 /* Update potential sum for this i atom from the interaction with this j atom. */
272 velec = _mm_and_ps(velec,cutoff_mask);
273 velecsum = _mm_add_ps(velecsum,velec);
277 fscal = _mm_and_ps(fscal,cutoff_mask);
279 /* Update vectorial force */
280 fix0 = _mm_macc_ps(dx00,fscal,fix0);
281 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
282 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
284 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
285 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
286 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
290 /**************************
291 * CALCULATE INTERACTIONS *
292 **************************/
294 if (gmx_mm_any_lt(rsq10,rcutoff2))
297 r10 = _mm_mul_ps(rsq10,rinv10);
299 /* Compute parameters for interactions between i and j atoms */
300 qq10 = _mm_mul_ps(iq1,jq0);
302 /* EWALD ELECTROSTATICS */
304 /* Analytical PME correction */
305 zeta2 = _mm_mul_ps(beta2,rsq10);
306 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
307 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
308 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
309 felec = _mm_mul_ps(qq10,felec);
310 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
311 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
312 velec = _mm_mul_ps(qq10,velec);
314 d = _mm_sub_ps(r10,rswitch);
315 d = _mm_max_ps(d,_mm_setzero_ps());
316 d2 = _mm_mul_ps(d,d);
317 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
319 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
321 /* Evaluate switch function */
322 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
323 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
324 velec = _mm_mul_ps(velec,sw);
325 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
327 /* Update potential sum for this i atom from the interaction with this j atom. */
328 velec = _mm_and_ps(velec,cutoff_mask);
329 velecsum = _mm_add_ps(velecsum,velec);
333 fscal = _mm_and_ps(fscal,cutoff_mask);
335 /* Update vectorial force */
336 fix1 = _mm_macc_ps(dx10,fscal,fix1);
337 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
338 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
340 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
341 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
342 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
346 /**************************
347 * CALCULATE INTERACTIONS *
348 **************************/
350 if (gmx_mm_any_lt(rsq20,rcutoff2))
353 r20 = _mm_mul_ps(rsq20,rinv20);
355 /* Compute parameters for interactions between i and j atoms */
356 qq20 = _mm_mul_ps(iq2,jq0);
358 /* EWALD ELECTROSTATICS */
360 /* Analytical PME correction */
361 zeta2 = _mm_mul_ps(beta2,rsq20);
362 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
363 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
364 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
365 felec = _mm_mul_ps(qq20,felec);
366 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
367 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
368 velec = _mm_mul_ps(qq20,velec);
370 d = _mm_sub_ps(r20,rswitch);
371 d = _mm_max_ps(d,_mm_setzero_ps());
372 d2 = _mm_mul_ps(d,d);
373 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
375 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
377 /* Evaluate switch function */
378 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
379 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
380 velec = _mm_mul_ps(velec,sw);
381 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
383 /* Update potential sum for this i atom from the interaction with this j atom. */
384 velec = _mm_and_ps(velec,cutoff_mask);
385 velecsum = _mm_add_ps(velecsum,velec);
389 fscal = _mm_and_ps(fscal,cutoff_mask);
391 /* Update vectorial force */
392 fix2 = _mm_macc_ps(dx20,fscal,fix2);
393 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
394 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
396 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
397 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
398 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
402 fjptrA = f+j_coord_offsetA;
403 fjptrB = f+j_coord_offsetB;
404 fjptrC = f+j_coord_offsetC;
405 fjptrD = f+j_coord_offsetD;
407 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
409 /* Inner loop uses 159 flops */
415 /* Get j neighbor index, and coordinate index */
416 jnrlistA = jjnr[jidx];
417 jnrlistB = jjnr[jidx+1];
418 jnrlistC = jjnr[jidx+2];
419 jnrlistD = jjnr[jidx+3];
420 /* Sign of each element will be negative for non-real atoms.
421 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
422 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
424 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
425 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
426 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
427 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
428 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
429 j_coord_offsetA = DIM*jnrA;
430 j_coord_offsetB = DIM*jnrB;
431 j_coord_offsetC = DIM*jnrC;
432 j_coord_offsetD = DIM*jnrD;
434 /* load j atom coordinates */
435 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
436 x+j_coord_offsetC,x+j_coord_offsetD,
439 /* Calculate displacement vector */
440 dx00 = _mm_sub_ps(ix0,jx0);
441 dy00 = _mm_sub_ps(iy0,jy0);
442 dz00 = _mm_sub_ps(iz0,jz0);
443 dx10 = _mm_sub_ps(ix1,jx0);
444 dy10 = _mm_sub_ps(iy1,jy0);
445 dz10 = _mm_sub_ps(iz1,jz0);
446 dx20 = _mm_sub_ps(ix2,jx0);
447 dy20 = _mm_sub_ps(iy2,jy0);
448 dz20 = _mm_sub_ps(iz2,jz0);
450 /* Calculate squared distance and things based on it */
451 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
452 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
453 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
455 rinv00 = gmx_mm_invsqrt_ps(rsq00);
456 rinv10 = gmx_mm_invsqrt_ps(rsq10);
457 rinv20 = gmx_mm_invsqrt_ps(rsq20);
459 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
460 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
461 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
463 /* Load parameters for j particles */
464 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
465 charge+jnrC+0,charge+jnrD+0);
467 fjx0 = _mm_setzero_ps();
468 fjy0 = _mm_setzero_ps();
469 fjz0 = _mm_setzero_ps();
471 /**************************
472 * CALCULATE INTERACTIONS *
473 **************************/
475 if (gmx_mm_any_lt(rsq00,rcutoff2))
478 r00 = _mm_mul_ps(rsq00,rinv00);
479 r00 = _mm_andnot_ps(dummy_mask,r00);
481 /* Compute parameters for interactions between i and j atoms */
482 qq00 = _mm_mul_ps(iq0,jq0);
484 /* EWALD ELECTROSTATICS */
486 /* Analytical PME correction */
487 zeta2 = _mm_mul_ps(beta2,rsq00);
488 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
489 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
490 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
491 felec = _mm_mul_ps(qq00,felec);
492 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
493 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
494 velec = _mm_mul_ps(qq00,velec);
496 d = _mm_sub_ps(r00,rswitch);
497 d = _mm_max_ps(d,_mm_setzero_ps());
498 d2 = _mm_mul_ps(d,d);
499 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
501 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
503 /* Evaluate switch function */
504 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
505 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
506 velec = _mm_mul_ps(velec,sw);
507 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
509 /* Update potential sum for this i atom from the interaction with this j atom. */
510 velec = _mm_and_ps(velec,cutoff_mask);
511 velec = _mm_andnot_ps(dummy_mask,velec);
512 velecsum = _mm_add_ps(velecsum,velec);
516 fscal = _mm_and_ps(fscal,cutoff_mask);
518 fscal = _mm_andnot_ps(dummy_mask,fscal);
520 /* Update vectorial force */
521 fix0 = _mm_macc_ps(dx00,fscal,fix0);
522 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
523 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
525 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
526 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
527 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
531 /**************************
532 * CALCULATE INTERACTIONS *
533 **************************/
535 if (gmx_mm_any_lt(rsq10,rcutoff2))
538 r10 = _mm_mul_ps(rsq10,rinv10);
539 r10 = _mm_andnot_ps(dummy_mask,r10);
541 /* Compute parameters for interactions between i and j atoms */
542 qq10 = _mm_mul_ps(iq1,jq0);
544 /* EWALD ELECTROSTATICS */
546 /* Analytical PME correction */
547 zeta2 = _mm_mul_ps(beta2,rsq10);
548 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
549 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
550 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
551 felec = _mm_mul_ps(qq10,felec);
552 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
553 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
554 velec = _mm_mul_ps(qq10,velec);
556 d = _mm_sub_ps(r10,rswitch);
557 d = _mm_max_ps(d,_mm_setzero_ps());
558 d2 = _mm_mul_ps(d,d);
559 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
561 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
563 /* Evaluate switch function */
564 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
565 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
566 velec = _mm_mul_ps(velec,sw);
567 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
569 /* Update potential sum for this i atom from the interaction with this j atom. */
570 velec = _mm_and_ps(velec,cutoff_mask);
571 velec = _mm_andnot_ps(dummy_mask,velec);
572 velecsum = _mm_add_ps(velecsum,velec);
576 fscal = _mm_and_ps(fscal,cutoff_mask);
578 fscal = _mm_andnot_ps(dummy_mask,fscal);
580 /* Update vectorial force */
581 fix1 = _mm_macc_ps(dx10,fscal,fix1);
582 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
583 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
585 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
586 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
587 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
591 /**************************
592 * CALCULATE INTERACTIONS *
593 **************************/
595 if (gmx_mm_any_lt(rsq20,rcutoff2))
598 r20 = _mm_mul_ps(rsq20,rinv20);
599 r20 = _mm_andnot_ps(dummy_mask,r20);
601 /* Compute parameters for interactions between i and j atoms */
602 qq20 = _mm_mul_ps(iq2,jq0);
604 /* EWALD ELECTROSTATICS */
606 /* Analytical PME correction */
607 zeta2 = _mm_mul_ps(beta2,rsq20);
608 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
609 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
610 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
611 felec = _mm_mul_ps(qq20,felec);
612 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
613 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
614 velec = _mm_mul_ps(qq20,velec);
616 d = _mm_sub_ps(r20,rswitch);
617 d = _mm_max_ps(d,_mm_setzero_ps());
618 d2 = _mm_mul_ps(d,d);
619 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
621 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
623 /* Evaluate switch function */
624 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
625 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
626 velec = _mm_mul_ps(velec,sw);
627 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
629 /* Update potential sum for this i atom from the interaction with this j atom. */
630 velec = _mm_and_ps(velec,cutoff_mask);
631 velec = _mm_andnot_ps(dummy_mask,velec);
632 velecsum = _mm_add_ps(velecsum,velec);
636 fscal = _mm_and_ps(fscal,cutoff_mask);
638 fscal = _mm_andnot_ps(dummy_mask,fscal);
640 /* Update vectorial force */
641 fix2 = _mm_macc_ps(dx20,fscal,fix2);
642 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
643 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
645 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
646 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
647 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
651 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
652 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
653 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
654 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
656 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
658 /* Inner loop uses 162 flops */
661 /* End of innermost loop */
663 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
664 f+i_coord_offset,fshift+i_shift_offset);
667 /* Update potential energies */
668 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
670 /* Increment number of inner iterations */
671 inneriter += j_index_end - j_index_start;
673 /* Outer loop uses 19 flops */
676 /* Increment number of outer iterations */
679 /* Update outer/inner flops */
681 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_VF,outeriter*19 + inneriter*162);
684 * Gromacs nonbonded kernel: nb_kernel_ElecEwSw_VdwNone_GeomW3P1_F_avx_128_fma_single
685 * Electrostatics interaction: Ewald
686 * VdW interaction: None
687 * Geometry: Water3-Particle
688 * Calculate force/pot: Force
691 nb_kernel_ElecEwSw_VdwNone_GeomW3P1_F_avx_128_fma_single
692 (t_nblist * gmx_restrict nlist,
693 rvec * gmx_restrict xx,
694 rvec * gmx_restrict ff,
695 t_forcerec * gmx_restrict fr,
696 t_mdatoms * gmx_restrict mdatoms,
697 nb_kernel_data_t * gmx_restrict kernel_data,
698 t_nrnb * gmx_restrict nrnb)
700 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
701 * just 0 for non-waters.
702 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
703 * jnr indices corresponding to data put in the four positions in the SIMD register.
705 int i_shift_offset,i_coord_offset,outeriter,inneriter;
706 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
707 int jnrA,jnrB,jnrC,jnrD;
708 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
709 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
710 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
712 real *shiftvec,*fshift,*x,*f;
713 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
715 __m128 fscal,rcutoff,rcutoff2,jidxall;
717 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
719 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
721 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
722 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
723 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
724 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
725 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
726 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
727 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
730 __m128 ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
731 __m128 beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
733 __m128 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
734 real rswitch_scalar,d_scalar;
735 __m128 dummy_mask,cutoff_mask;
736 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
737 __m128 one = _mm_set1_ps(1.0);
738 __m128 two = _mm_set1_ps(2.0);
744 jindex = nlist->jindex;
746 shiftidx = nlist->shift;
748 shiftvec = fr->shift_vec[0];
749 fshift = fr->fshift[0];
750 facel = _mm_set1_ps(fr->epsfac);
751 charge = mdatoms->chargeA;
753 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
754 beta = _mm_set1_ps(fr->ic->ewaldcoeff);
755 beta2 = _mm_mul_ps(beta,beta);
756 beta3 = _mm_mul_ps(beta,beta2);
757 ewtab = fr->ic->tabq_coul_FDV0;
758 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
759 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
761 /* Setup water-specific parameters */
762 inr = nlist->iinr[0];
763 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
764 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
765 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
767 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
768 rcutoff_scalar = fr->rcoulomb;
769 rcutoff = _mm_set1_ps(rcutoff_scalar);
770 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
772 rswitch_scalar = fr->rcoulomb_switch;
773 rswitch = _mm_set1_ps(rswitch_scalar);
774 /* Setup switch parameters */
775 d_scalar = rcutoff_scalar-rswitch_scalar;
776 d = _mm_set1_ps(d_scalar);
777 swV3 = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
778 swV4 = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
779 swV5 = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
780 swF2 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
781 swF3 = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
782 swF4 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
784 /* Avoid stupid compiler warnings */
785 jnrA = jnrB = jnrC = jnrD = 0;
794 for(iidx=0;iidx<4*DIM;iidx++)
799 /* Start outer loop over neighborlists */
800 for(iidx=0; iidx<nri; iidx++)
802 /* Load shift vector for this list */
803 i_shift_offset = DIM*shiftidx[iidx];
805 /* Load limits for loop over neighbors */
806 j_index_start = jindex[iidx];
807 j_index_end = jindex[iidx+1];
809 /* Get outer coordinate index */
811 i_coord_offset = DIM*inr;
813 /* Load i particle coords and add shift vector */
814 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
815 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
817 fix0 = _mm_setzero_ps();
818 fiy0 = _mm_setzero_ps();
819 fiz0 = _mm_setzero_ps();
820 fix1 = _mm_setzero_ps();
821 fiy1 = _mm_setzero_ps();
822 fiz1 = _mm_setzero_ps();
823 fix2 = _mm_setzero_ps();
824 fiy2 = _mm_setzero_ps();
825 fiz2 = _mm_setzero_ps();
827 /* Start inner kernel loop */
828 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
831 /* Get j neighbor index, and coordinate index */
836 j_coord_offsetA = DIM*jnrA;
837 j_coord_offsetB = DIM*jnrB;
838 j_coord_offsetC = DIM*jnrC;
839 j_coord_offsetD = DIM*jnrD;
841 /* load j atom coordinates */
842 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
843 x+j_coord_offsetC,x+j_coord_offsetD,
846 /* Calculate displacement vector */
847 dx00 = _mm_sub_ps(ix0,jx0);
848 dy00 = _mm_sub_ps(iy0,jy0);
849 dz00 = _mm_sub_ps(iz0,jz0);
850 dx10 = _mm_sub_ps(ix1,jx0);
851 dy10 = _mm_sub_ps(iy1,jy0);
852 dz10 = _mm_sub_ps(iz1,jz0);
853 dx20 = _mm_sub_ps(ix2,jx0);
854 dy20 = _mm_sub_ps(iy2,jy0);
855 dz20 = _mm_sub_ps(iz2,jz0);
857 /* Calculate squared distance and things based on it */
858 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
859 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
860 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
862 rinv00 = gmx_mm_invsqrt_ps(rsq00);
863 rinv10 = gmx_mm_invsqrt_ps(rsq10);
864 rinv20 = gmx_mm_invsqrt_ps(rsq20);
866 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
867 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
868 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
870 /* Load parameters for j particles */
871 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
872 charge+jnrC+0,charge+jnrD+0);
874 fjx0 = _mm_setzero_ps();
875 fjy0 = _mm_setzero_ps();
876 fjz0 = _mm_setzero_ps();
878 /**************************
879 * CALCULATE INTERACTIONS *
880 **************************/
882 if (gmx_mm_any_lt(rsq00,rcutoff2))
885 r00 = _mm_mul_ps(rsq00,rinv00);
887 /* Compute parameters for interactions between i and j atoms */
888 qq00 = _mm_mul_ps(iq0,jq0);
890 /* EWALD ELECTROSTATICS */
892 /* Analytical PME correction */
893 zeta2 = _mm_mul_ps(beta2,rsq00);
894 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
895 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
896 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
897 felec = _mm_mul_ps(qq00,felec);
898 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
899 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
900 velec = _mm_mul_ps(qq00,velec);
902 d = _mm_sub_ps(r00,rswitch);
903 d = _mm_max_ps(d,_mm_setzero_ps());
904 d2 = _mm_mul_ps(d,d);
905 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
907 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
909 /* Evaluate switch function */
910 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
911 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
912 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
916 fscal = _mm_and_ps(fscal,cutoff_mask);
918 /* Update vectorial force */
919 fix0 = _mm_macc_ps(dx00,fscal,fix0);
920 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
921 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
923 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
924 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
925 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
929 /**************************
930 * CALCULATE INTERACTIONS *
931 **************************/
933 if (gmx_mm_any_lt(rsq10,rcutoff2))
936 r10 = _mm_mul_ps(rsq10,rinv10);
938 /* Compute parameters for interactions between i and j atoms */
939 qq10 = _mm_mul_ps(iq1,jq0);
941 /* EWALD ELECTROSTATICS */
943 /* Analytical PME correction */
944 zeta2 = _mm_mul_ps(beta2,rsq10);
945 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
946 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
947 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
948 felec = _mm_mul_ps(qq10,felec);
949 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
950 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
951 velec = _mm_mul_ps(qq10,velec);
953 d = _mm_sub_ps(r10,rswitch);
954 d = _mm_max_ps(d,_mm_setzero_ps());
955 d2 = _mm_mul_ps(d,d);
956 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
958 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
960 /* Evaluate switch function */
961 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
962 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
963 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
967 fscal = _mm_and_ps(fscal,cutoff_mask);
969 /* Update vectorial force */
970 fix1 = _mm_macc_ps(dx10,fscal,fix1);
971 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
972 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
974 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
975 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
976 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
980 /**************************
981 * CALCULATE INTERACTIONS *
982 **************************/
984 if (gmx_mm_any_lt(rsq20,rcutoff2))
987 r20 = _mm_mul_ps(rsq20,rinv20);
989 /* Compute parameters for interactions between i and j atoms */
990 qq20 = _mm_mul_ps(iq2,jq0);
992 /* EWALD ELECTROSTATICS */
994 /* Analytical PME correction */
995 zeta2 = _mm_mul_ps(beta2,rsq20);
996 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
997 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
998 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
999 felec = _mm_mul_ps(qq20,felec);
1000 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
1001 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
1002 velec = _mm_mul_ps(qq20,velec);
1004 d = _mm_sub_ps(r20,rswitch);
1005 d = _mm_max_ps(d,_mm_setzero_ps());
1006 d2 = _mm_mul_ps(d,d);
1007 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1009 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1011 /* Evaluate switch function */
1012 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1013 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
1014 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
1018 fscal = _mm_and_ps(fscal,cutoff_mask);
1020 /* Update vectorial force */
1021 fix2 = _mm_macc_ps(dx20,fscal,fix2);
1022 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
1023 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
1025 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
1026 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
1027 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
1031 fjptrA = f+j_coord_offsetA;
1032 fjptrB = f+j_coord_offsetB;
1033 fjptrC = f+j_coord_offsetC;
1034 fjptrD = f+j_coord_offsetD;
1036 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1038 /* Inner loop uses 150 flops */
1041 if(jidx<j_index_end)
1044 /* Get j neighbor index, and coordinate index */
1045 jnrlistA = jjnr[jidx];
1046 jnrlistB = jjnr[jidx+1];
1047 jnrlistC = jjnr[jidx+2];
1048 jnrlistD = jjnr[jidx+3];
1049 /* Sign of each element will be negative for non-real atoms.
1050 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1051 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1053 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1054 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1055 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1056 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1057 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1058 j_coord_offsetA = DIM*jnrA;
1059 j_coord_offsetB = DIM*jnrB;
1060 j_coord_offsetC = DIM*jnrC;
1061 j_coord_offsetD = DIM*jnrD;
1063 /* load j atom coordinates */
1064 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1065 x+j_coord_offsetC,x+j_coord_offsetD,
1068 /* Calculate displacement vector */
1069 dx00 = _mm_sub_ps(ix0,jx0);
1070 dy00 = _mm_sub_ps(iy0,jy0);
1071 dz00 = _mm_sub_ps(iz0,jz0);
1072 dx10 = _mm_sub_ps(ix1,jx0);
1073 dy10 = _mm_sub_ps(iy1,jy0);
1074 dz10 = _mm_sub_ps(iz1,jz0);
1075 dx20 = _mm_sub_ps(ix2,jx0);
1076 dy20 = _mm_sub_ps(iy2,jy0);
1077 dz20 = _mm_sub_ps(iz2,jz0);
1079 /* Calculate squared distance and things based on it */
1080 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1081 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1082 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1084 rinv00 = gmx_mm_invsqrt_ps(rsq00);
1085 rinv10 = gmx_mm_invsqrt_ps(rsq10);
1086 rinv20 = gmx_mm_invsqrt_ps(rsq20);
1088 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
1089 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
1090 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
1092 /* Load parameters for j particles */
1093 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1094 charge+jnrC+0,charge+jnrD+0);
1096 fjx0 = _mm_setzero_ps();
1097 fjy0 = _mm_setzero_ps();
1098 fjz0 = _mm_setzero_ps();
1100 /**************************
1101 * CALCULATE INTERACTIONS *
1102 **************************/
1104 if (gmx_mm_any_lt(rsq00,rcutoff2))
1107 r00 = _mm_mul_ps(rsq00,rinv00);
1108 r00 = _mm_andnot_ps(dummy_mask,r00);
1110 /* Compute parameters for interactions between i and j atoms */
1111 qq00 = _mm_mul_ps(iq0,jq0);
1113 /* EWALD ELECTROSTATICS */
1115 /* Analytical PME correction */
1116 zeta2 = _mm_mul_ps(beta2,rsq00);
1117 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
1118 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1119 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1120 felec = _mm_mul_ps(qq00,felec);
1121 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
1122 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
1123 velec = _mm_mul_ps(qq00,velec);
1125 d = _mm_sub_ps(r00,rswitch);
1126 d = _mm_max_ps(d,_mm_setzero_ps());
1127 d2 = _mm_mul_ps(d,d);
1128 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1130 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1132 /* Evaluate switch function */
1133 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1134 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
1135 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
1139 fscal = _mm_and_ps(fscal,cutoff_mask);
1141 fscal = _mm_andnot_ps(dummy_mask,fscal);
1143 /* Update vectorial force */
1144 fix0 = _mm_macc_ps(dx00,fscal,fix0);
1145 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
1146 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
1148 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
1149 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
1150 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
1154 /**************************
1155 * CALCULATE INTERACTIONS *
1156 **************************/
1158 if (gmx_mm_any_lt(rsq10,rcutoff2))
1161 r10 = _mm_mul_ps(rsq10,rinv10);
1162 r10 = _mm_andnot_ps(dummy_mask,r10);
1164 /* Compute parameters for interactions between i and j atoms */
1165 qq10 = _mm_mul_ps(iq1,jq0);
1167 /* EWALD ELECTROSTATICS */
1169 /* Analytical PME correction */
1170 zeta2 = _mm_mul_ps(beta2,rsq10);
1171 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
1172 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1173 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1174 felec = _mm_mul_ps(qq10,felec);
1175 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
1176 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
1177 velec = _mm_mul_ps(qq10,velec);
1179 d = _mm_sub_ps(r10,rswitch);
1180 d = _mm_max_ps(d,_mm_setzero_ps());
1181 d2 = _mm_mul_ps(d,d);
1182 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1184 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1186 /* Evaluate switch function */
1187 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1188 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
1189 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
1193 fscal = _mm_and_ps(fscal,cutoff_mask);
1195 fscal = _mm_andnot_ps(dummy_mask,fscal);
1197 /* Update vectorial force */
1198 fix1 = _mm_macc_ps(dx10,fscal,fix1);
1199 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
1200 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
1202 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
1203 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
1204 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
1208 /**************************
1209 * CALCULATE INTERACTIONS *
1210 **************************/
1212 if (gmx_mm_any_lt(rsq20,rcutoff2))
1215 r20 = _mm_mul_ps(rsq20,rinv20);
1216 r20 = _mm_andnot_ps(dummy_mask,r20);
1218 /* Compute parameters for interactions between i and j atoms */
1219 qq20 = _mm_mul_ps(iq2,jq0);
1221 /* EWALD ELECTROSTATICS */
1223 /* Analytical PME correction */
1224 zeta2 = _mm_mul_ps(beta2,rsq20);
1225 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
1226 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1227 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1228 felec = _mm_mul_ps(qq20,felec);
1229 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
1230 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
1231 velec = _mm_mul_ps(qq20,velec);
1233 d = _mm_sub_ps(r20,rswitch);
1234 d = _mm_max_ps(d,_mm_setzero_ps());
1235 d2 = _mm_mul_ps(d,d);
1236 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1238 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1240 /* Evaluate switch function */
1241 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1242 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
1243 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
1247 fscal = _mm_and_ps(fscal,cutoff_mask);
1249 fscal = _mm_andnot_ps(dummy_mask,fscal);
1251 /* Update vectorial force */
1252 fix2 = _mm_macc_ps(dx20,fscal,fix2);
1253 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
1254 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
1256 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
1257 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
1258 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
1262 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1263 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1264 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1265 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1267 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1269 /* Inner loop uses 153 flops */
1272 /* End of innermost loop */
1274 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1275 f+i_coord_offset,fshift+i_shift_offset);
1277 /* Increment number of inner iterations */
1278 inneriter += j_index_end - j_index_start;
1280 /* Outer loop uses 18 flops */
1283 /* Increment number of outer iterations */
1286 /* Update outer/inner flops */
1288 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3_F,outeriter*18 + inneriter*153);