2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
44 #include "../nb_kernel.h"
45 #include "gromacs/legacyheaders/types/simple.h"
46 #include "gromacs/math/vec.h"
47 #include "gromacs/legacyheaders/nrnb.h"
49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
50 #include "kernelutil_x86_avx_128_fma_single.h"
53 * Gromacs nonbonded kernel: nb_kernel_ElecEwSw_VdwLJSw_GeomW3P1_VF_avx_128_fma_single
54 * Electrostatics interaction: Ewald
55 * VdW interaction: LennardJones
56 * Geometry: Water3-Particle
57 * Calculate force/pot: PotentialAndForce
60 nb_kernel_ElecEwSw_VdwLJSw_GeomW3P1_VF_avx_128_fma_single
61 (t_nblist * gmx_restrict nlist,
62 rvec * gmx_restrict xx,
63 rvec * gmx_restrict ff,
64 t_forcerec * gmx_restrict fr,
65 t_mdatoms * gmx_restrict mdatoms,
66 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
67 t_nrnb * gmx_restrict nrnb)
69 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
70 * just 0 for non-waters.
71 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
72 * jnr indices corresponding to data put in the four positions in the SIMD register.
74 int i_shift_offset,i_coord_offset,outeriter,inneriter;
75 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
76 int jnrA,jnrB,jnrC,jnrD;
77 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
78 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
79 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
81 real *shiftvec,*fshift,*x,*f;
82 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
84 __m128 fscal,rcutoff,rcutoff2,jidxall;
86 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
88 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
90 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
91 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
92 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
93 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
94 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
95 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
96 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
99 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
102 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
103 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
105 __m128 ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
106 __m128 beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
108 __m128 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
109 real rswitch_scalar,d_scalar;
110 __m128 dummy_mask,cutoff_mask;
111 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
112 __m128 one = _mm_set1_ps(1.0);
113 __m128 two = _mm_set1_ps(2.0);
119 jindex = nlist->jindex;
121 shiftidx = nlist->shift;
123 shiftvec = fr->shift_vec[0];
124 fshift = fr->fshift[0];
125 facel = _mm_set1_ps(fr->epsfac);
126 charge = mdatoms->chargeA;
127 nvdwtype = fr->ntype;
129 vdwtype = mdatoms->typeA;
131 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
132 beta = _mm_set1_ps(fr->ic->ewaldcoeff_q);
133 beta2 = _mm_mul_ps(beta,beta);
134 beta3 = _mm_mul_ps(beta,beta2);
135 ewtab = fr->ic->tabq_coul_FDV0;
136 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
137 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
139 /* Setup water-specific parameters */
140 inr = nlist->iinr[0];
141 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
142 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
143 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
144 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
146 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
147 rcutoff_scalar = fr->rcoulomb;
148 rcutoff = _mm_set1_ps(rcutoff_scalar);
149 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
151 rswitch_scalar = fr->rcoulomb_switch;
152 rswitch = _mm_set1_ps(rswitch_scalar);
153 /* Setup switch parameters */
154 d_scalar = rcutoff_scalar-rswitch_scalar;
155 d = _mm_set1_ps(d_scalar);
156 swV3 = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
157 swV4 = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
158 swV5 = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
159 swF2 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
160 swF3 = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
161 swF4 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
163 /* Avoid stupid compiler warnings */
164 jnrA = jnrB = jnrC = jnrD = 0;
173 for(iidx=0;iidx<4*DIM;iidx++)
178 /* Start outer loop over neighborlists */
179 for(iidx=0; iidx<nri; iidx++)
181 /* Load shift vector for this list */
182 i_shift_offset = DIM*shiftidx[iidx];
184 /* Load limits for loop over neighbors */
185 j_index_start = jindex[iidx];
186 j_index_end = jindex[iidx+1];
188 /* Get outer coordinate index */
190 i_coord_offset = DIM*inr;
192 /* Load i particle coords and add shift vector */
193 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
194 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
196 fix0 = _mm_setzero_ps();
197 fiy0 = _mm_setzero_ps();
198 fiz0 = _mm_setzero_ps();
199 fix1 = _mm_setzero_ps();
200 fiy1 = _mm_setzero_ps();
201 fiz1 = _mm_setzero_ps();
202 fix2 = _mm_setzero_ps();
203 fiy2 = _mm_setzero_ps();
204 fiz2 = _mm_setzero_ps();
206 /* Reset potential sums */
207 velecsum = _mm_setzero_ps();
208 vvdwsum = _mm_setzero_ps();
210 /* Start inner kernel loop */
211 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
214 /* Get j neighbor index, and coordinate index */
219 j_coord_offsetA = DIM*jnrA;
220 j_coord_offsetB = DIM*jnrB;
221 j_coord_offsetC = DIM*jnrC;
222 j_coord_offsetD = DIM*jnrD;
224 /* load j atom coordinates */
225 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
226 x+j_coord_offsetC,x+j_coord_offsetD,
229 /* Calculate displacement vector */
230 dx00 = _mm_sub_ps(ix0,jx0);
231 dy00 = _mm_sub_ps(iy0,jy0);
232 dz00 = _mm_sub_ps(iz0,jz0);
233 dx10 = _mm_sub_ps(ix1,jx0);
234 dy10 = _mm_sub_ps(iy1,jy0);
235 dz10 = _mm_sub_ps(iz1,jz0);
236 dx20 = _mm_sub_ps(ix2,jx0);
237 dy20 = _mm_sub_ps(iy2,jy0);
238 dz20 = _mm_sub_ps(iz2,jz0);
240 /* Calculate squared distance and things based on it */
241 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
242 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
243 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
245 rinv00 = gmx_mm_invsqrt_ps(rsq00);
246 rinv10 = gmx_mm_invsqrt_ps(rsq10);
247 rinv20 = gmx_mm_invsqrt_ps(rsq20);
249 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
250 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
251 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
253 /* Load parameters for j particles */
254 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
255 charge+jnrC+0,charge+jnrD+0);
256 vdwjidx0A = 2*vdwtype[jnrA+0];
257 vdwjidx0B = 2*vdwtype[jnrB+0];
258 vdwjidx0C = 2*vdwtype[jnrC+0];
259 vdwjidx0D = 2*vdwtype[jnrD+0];
261 fjx0 = _mm_setzero_ps();
262 fjy0 = _mm_setzero_ps();
263 fjz0 = _mm_setzero_ps();
265 /**************************
266 * CALCULATE INTERACTIONS *
267 **************************/
269 if (gmx_mm_any_lt(rsq00,rcutoff2))
272 r00 = _mm_mul_ps(rsq00,rinv00);
274 /* Compute parameters for interactions between i and j atoms */
275 qq00 = _mm_mul_ps(iq0,jq0);
276 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
277 vdwparam+vdwioffset0+vdwjidx0B,
278 vdwparam+vdwioffset0+vdwjidx0C,
279 vdwparam+vdwioffset0+vdwjidx0D,
282 /* EWALD ELECTROSTATICS */
284 /* Analytical PME correction */
285 zeta2 = _mm_mul_ps(beta2,rsq00);
286 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
287 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
288 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
289 felec = _mm_mul_ps(qq00,felec);
290 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
291 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
292 velec = _mm_mul_ps(qq00,velec);
294 /* LENNARD-JONES DISPERSION/REPULSION */
296 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
297 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
298 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
299 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
300 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
302 d = _mm_sub_ps(r00,rswitch);
303 d = _mm_max_ps(d,_mm_setzero_ps());
304 d2 = _mm_mul_ps(d,d);
305 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
307 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
309 /* Evaluate switch function */
310 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
311 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
312 fvdw = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
313 velec = _mm_mul_ps(velec,sw);
314 vvdw = _mm_mul_ps(vvdw,sw);
315 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
317 /* Update potential sum for this i atom from the interaction with this j atom. */
318 velec = _mm_and_ps(velec,cutoff_mask);
319 velecsum = _mm_add_ps(velecsum,velec);
320 vvdw = _mm_and_ps(vvdw,cutoff_mask);
321 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
323 fscal = _mm_add_ps(felec,fvdw);
325 fscal = _mm_and_ps(fscal,cutoff_mask);
327 /* Update vectorial force */
328 fix0 = _mm_macc_ps(dx00,fscal,fix0);
329 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
330 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
332 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
333 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
334 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
338 /**************************
339 * CALCULATE INTERACTIONS *
340 **************************/
342 if (gmx_mm_any_lt(rsq10,rcutoff2))
345 r10 = _mm_mul_ps(rsq10,rinv10);
347 /* Compute parameters for interactions between i and j atoms */
348 qq10 = _mm_mul_ps(iq1,jq0);
350 /* EWALD ELECTROSTATICS */
352 /* Analytical PME correction */
353 zeta2 = _mm_mul_ps(beta2,rsq10);
354 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
355 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
356 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
357 felec = _mm_mul_ps(qq10,felec);
358 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
359 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
360 velec = _mm_mul_ps(qq10,velec);
362 d = _mm_sub_ps(r10,rswitch);
363 d = _mm_max_ps(d,_mm_setzero_ps());
364 d2 = _mm_mul_ps(d,d);
365 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
367 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
369 /* Evaluate switch function */
370 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
371 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
372 velec = _mm_mul_ps(velec,sw);
373 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
375 /* Update potential sum for this i atom from the interaction with this j atom. */
376 velec = _mm_and_ps(velec,cutoff_mask);
377 velecsum = _mm_add_ps(velecsum,velec);
381 fscal = _mm_and_ps(fscal,cutoff_mask);
383 /* Update vectorial force */
384 fix1 = _mm_macc_ps(dx10,fscal,fix1);
385 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
386 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
388 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
389 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
390 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
394 /**************************
395 * CALCULATE INTERACTIONS *
396 **************************/
398 if (gmx_mm_any_lt(rsq20,rcutoff2))
401 r20 = _mm_mul_ps(rsq20,rinv20);
403 /* Compute parameters for interactions between i and j atoms */
404 qq20 = _mm_mul_ps(iq2,jq0);
406 /* EWALD ELECTROSTATICS */
408 /* Analytical PME correction */
409 zeta2 = _mm_mul_ps(beta2,rsq20);
410 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
411 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
412 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
413 felec = _mm_mul_ps(qq20,felec);
414 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
415 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
416 velec = _mm_mul_ps(qq20,velec);
418 d = _mm_sub_ps(r20,rswitch);
419 d = _mm_max_ps(d,_mm_setzero_ps());
420 d2 = _mm_mul_ps(d,d);
421 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
423 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
425 /* Evaluate switch function */
426 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
427 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
428 velec = _mm_mul_ps(velec,sw);
429 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
431 /* Update potential sum for this i atom from the interaction with this j atom. */
432 velec = _mm_and_ps(velec,cutoff_mask);
433 velecsum = _mm_add_ps(velecsum,velec);
437 fscal = _mm_and_ps(fscal,cutoff_mask);
439 /* Update vectorial force */
440 fix2 = _mm_macc_ps(dx20,fscal,fix2);
441 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
442 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
444 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
445 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
446 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
450 fjptrA = f+j_coord_offsetA;
451 fjptrB = f+j_coord_offsetB;
452 fjptrC = f+j_coord_offsetC;
453 fjptrD = f+j_coord_offsetD;
455 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
457 /* Inner loop uses 177 flops */
463 /* Get j neighbor index, and coordinate index */
464 jnrlistA = jjnr[jidx];
465 jnrlistB = jjnr[jidx+1];
466 jnrlistC = jjnr[jidx+2];
467 jnrlistD = jjnr[jidx+3];
468 /* Sign of each element will be negative for non-real atoms.
469 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
470 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
472 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
473 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
474 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
475 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
476 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
477 j_coord_offsetA = DIM*jnrA;
478 j_coord_offsetB = DIM*jnrB;
479 j_coord_offsetC = DIM*jnrC;
480 j_coord_offsetD = DIM*jnrD;
482 /* load j atom coordinates */
483 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
484 x+j_coord_offsetC,x+j_coord_offsetD,
487 /* Calculate displacement vector */
488 dx00 = _mm_sub_ps(ix0,jx0);
489 dy00 = _mm_sub_ps(iy0,jy0);
490 dz00 = _mm_sub_ps(iz0,jz0);
491 dx10 = _mm_sub_ps(ix1,jx0);
492 dy10 = _mm_sub_ps(iy1,jy0);
493 dz10 = _mm_sub_ps(iz1,jz0);
494 dx20 = _mm_sub_ps(ix2,jx0);
495 dy20 = _mm_sub_ps(iy2,jy0);
496 dz20 = _mm_sub_ps(iz2,jz0);
498 /* Calculate squared distance and things based on it */
499 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
500 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
501 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
503 rinv00 = gmx_mm_invsqrt_ps(rsq00);
504 rinv10 = gmx_mm_invsqrt_ps(rsq10);
505 rinv20 = gmx_mm_invsqrt_ps(rsq20);
507 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
508 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
509 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
511 /* Load parameters for j particles */
512 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
513 charge+jnrC+0,charge+jnrD+0);
514 vdwjidx0A = 2*vdwtype[jnrA+0];
515 vdwjidx0B = 2*vdwtype[jnrB+0];
516 vdwjidx0C = 2*vdwtype[jnrC+0];
517 vdwjidx0D = 2*vdwtype[jnrD+0];
519 fjx0 = _mm_setzero_ps();
520 fjy0 = _mm_setzero_ps();
521 fjz0 = _mm_setzero_ps();
523 /**************************
524 * CALCULATE INTERACTIONS *
525 **************************/
527 if (gmx_mm_any_lt(rsq00,rcutoff2))
530 r00 = _mm_mul_ps(rsq00,rinv00);
531 r00 = _mm_andnot_ps(dummy_mask,r00);
533 /* Compute parameters for interactions between i and j atoms */
534 qq00 = _mm_mul_ps(iq0,jq0);
535 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
536 vdwparam+vdwioffset0+vdwjidx0B,
537 vdwparam+vdwioffset0+vdwjidx0C,
538 vdwparam+vdwioffset0+vdwjidx0D,
541 /* EWALD ELECTROSTATICS */
543 /* Analytical PME correction */
544 zeta2 = _mm_mul_ps(beta2,rsq00);
545 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
546 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
547 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
548 felec = _mm_mul_ps(qq00,felec);
549 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
550 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
551 velec = _mm_mul_ps(qq00,velec);
553 /* LENNARD-JONES DISPERSION/REPULSION */
555 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
556 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
557 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
558 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
559 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
561 d = _mm_sub_ps(r00,rswitch);
562 d = _mm_max_ps(d,_mm_setzero_ps());
563 d2 = _mm_mul_ps(d,d);
564 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
566 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
568 /* Evaluate switch function */
569 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
570 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
571 fvdw = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
572 velec = _mm_mul_ps(velec,sw);
573 vvdw = _mm_mul_ps(vvdw,sw);
574 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
576 /* Update potential sum for this i atom from the interaction with this j atom. */
577 velec = _mm_and_ps(velec,cutoff_mask);
578 velec = _mm_andnot_ps(dummy_mask,velec);
579 velecsum = _mm_add_ps(velecsum,velec);
580 vvdw = _mm_and_ps(vvdw,cutoff_mask);
581 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
582 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
584 fscal = _mm_add_ps(felec,fvdw);
586 fscal = _mm_and_ps(fscal,cutoff_mask);
588 fscal = _mm_andnot_ps(dummy_mask,fscal);
590 /* Update vectorial force */
591 fix0 = _mm_macc_ps(dx00,fscal,fix0);
592 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
593 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
595 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
596 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
597 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
601 /**************************
602 * CALCULATE INTERACTIONS *
603 **************************/
605 if (gmx_mm_any_lt(rsq10,rcutoff2))
608 r10 = _mm_mul_ps(rsq10,rinv10);
609 r10 = _mm_andnot_ps(dummy_mask,r10);
611 /* Compute parameters for interactions between i and j atoms */
612 qq10 = _mm_mul_ps(iq1,jq0);
614 /* EWALD ELECTROSTATICS */
616 /* Analytical PME correction */
617 zeta2 = _mm_mul_ps(beta2,rsq10);
618 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
619 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
620 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
621 felec = _mm_mul_ps(qq10,felec);
622 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
623 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
624 velec = _mm_mul_ps(qq10,velec);
626 d = _mm_sub_ps(r10,rswitch);
627 d = _mm_max_ps(d,_mm_setzero_ps());
628 d2 = _mm_mul_ps(d,d);
629 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
631 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
633 /* Evaluate switch function */
634 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
635 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
636 velec = _mm_mul_ps(velec,sw);
637 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
639 /* Update potential sum for this i atom from the interaction with this j atom. */
640 velec = _mm_and_ps(velec,cutoff_mask);
641 velec = _mm_andnot_ps(dummy_mask,velec);
642 velecsum = _mm_add_ps(velecsum,velec);
646 fscal = _mm_and_ps(fscal,cutoff_mask);
648 fscal = _mm_andnot_ps(dummy_mask,fscal);
650 /* Update vectorial force */
651 fix1 = _mm_macc_ps(dx10,fscal,fix1);
652 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
653 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
655 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
656 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
657 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
661 /**************************
662 * CALCULATE INTERACTIONS *
663 **************************/
665 if (gmx_mm_any_lt(rsq20,rcutoff2))
668 r20 = _mm_mul_ps(rsq20,rinv20);
669 r20 = _mm_andnot_ps(dummy_mask,r20);
671 /* Compute parameters for interactions between i and j atoms */
672 qq20 = _mm_mul_ps(iq2,jq0);
674 /* EWALD ELECTROSTATICS */
676 /* Analytical PME correction */
677 zeta2 = _mm_mul_ps(beta2,rsq20);
678 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
679 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
680 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
681 felec = _mm_mul_ps(qq20,felec);
682 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
683 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
684 velec = _mm_mul_ps(qq20,velec);
686 d = _mm_sub_ps(r20,rswitch);
687 d = _mm_max_ps(d,_mm_setzero_ps());
688 d2 = _mm_mul_ps(d,d);
689 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
691 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
693 /* Evaluate switch function */
694 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
695 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
696 velec = _mm_mul_ps(velec,sw);
697 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
699 /* Update potential sum for this i atom from the interaction with this j atom. */
700 velec = _mm_and_ps(velec,cutoff_mask);
701 velec = _mm_andnot_ps(dummy_mask,velec);
702 velecsum = _mm_add_ps(velecsum,velec);
706 fscal = _mm_and_ps(fscal,cutoff_mask);
708 fscal = _mm_andnot_ps(dummy_mask,fscal);
710 /* Update vectorial force */
711 fix2 = _mm_macc_ps(dx20,fscal,fix2);
712 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
713 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
715 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
716 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
717 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
721 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
722 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
723 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
724 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
726 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
728 /* Inner loop uses 180 flops */
731 /* End of innermost loop */
733 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
734 f+i_coord_offset,fshift+i_shift_offset);
737 /* Update potential energies */
738 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
739 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
741 /* Increment number of inner iterations */
742 inneriter += j_index_end - j_index_start;
744 /* Outer loop uses 20 flops */
747 /* Increment number of outer iterations */
750 /* Update outer/inner flops */
752 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*180);
755 * Gromacs nonbonded kernel: nb_kernel_ElecEwSw_VdwLJSw_GeomW3P1_F_avx_128_fma_single
756 * Electrostatics interaction: Ewald
757 * VdW interaction: LennardJones
758 * Geometry: Water3-Particle
759 * Calculate force/pot: Force
762 nb_kernel_ElecEwSw_VdwLJSw_GeomW3P1_F_avx_128_fma_single
763 (t_nblist * gmx_restrict nlist,
764 rvec * gmx_restrict xx,
765 rvec * gmx_restrict ff,
766 t_forcerec * gmx_restrict fr,
767 t_mdatoms * gmx_restrict mdatoms,
768 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
769 t_nrnb * gmx_restrict nrnb)
771 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
772 * just 0 for non-waters.
773 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
774 * jnr indices corresponding to data put in the four positions in the SIMD register.
776 int i_shift_offset,i_coord_offset,outeriter,inneriter;
777 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
778 int jnrA,jnrB,jnrC,jnrD;
779 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
780 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
781 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
783 real *shiftvec,*fshift,*x,*f;
784 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
786 __m128 fscal,rcutoff,rcutoff2,jidxall;
788 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
790 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
792 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
793 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
794 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
795 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
796 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
797 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
798 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
801 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
804 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
805 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
807 __m128 ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
808 __m128 beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
810 __m128 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
811 real rswitch_scalar,d_scalar;
812 __m128 dummy_mask,cutoff_mask;
813 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
814 __m128 one = _mm_set1_ps(1.0);
815 __m128 two = _mm_set1_ps(2.0);
821 jindex = nlist->jindex;
823 shiftidx = nlist->shift;
825 shiftvec = fr->shift_vec[0];
826 fshift = fr->fshift[0];
827 facel = _mm_set1_ps(fr->epsfac);
828 charge = mdatoms->chargeA;
829 nvdwtype = fr->ntype;
831 vdwtype = mdatoms->typeA;
833 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
834 beta = _mm_set1_ps(fr->ic->ewaldcoeff_q);
835 beta2 = _mm_mul_ps(beta,beta);
836 beta3 = _mm_mul_ps(beta,beta2);
837 ewtab = fr->ic->tabq_coul_FDV0;
838 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
839 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
841 /* Setup water-specific parameters */
842 inr = nlist->iinr[0];
843 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
844 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
845 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
846 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
848 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
849 rcutoff_scalar = fr->rcoulomb;
850 rcutoff = _mm_set1_ps(rcutoff_scalar);
851 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
853 rswitch_scalar = fr->rcoulomb_switch;
854 rswitch = _mm_set1_ps(rswitch_scalar);
855 /* Setup switch parameters */
856 d_scalar = rcutoff_scalar-rswitch_scalar;
857 d = _mm_set1_ps(d_scalar);
858 swV3 = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
859 swV4 = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
860 swV5 = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
861 swF2 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
862 swF3 = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
863 swF4 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
865 /* Avoid stupid compiler warnings */
866 jnrA = jnrB = jnrC = jnrD = 0;
875 for(iidx=0;iidx<4*DIM;iidx++)
880 /* Start outer loop over neighborlists */
881 for(iidx=0; iidx<nri; iidx++)
883 /* Load shift vector for this list */
884 i_shift_offset = DIM*shiftidx[iidx];
886 /* Load limits for loop over neighbors */
887 j_index_start = jindex[iidx];
888 j_index_end = jindex[iidx+1];
890 /* Get outer coordinate index */
892 i_coord_offset = DIM*inr;
894 /* Load i particle coords and add shift vector */
895 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
896 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
898 fix0 = _mm_setzero_ps();
899 fiy0 = _mm_setzero_ps();
900 fiz0 = _mm_setzero_ps();
901 fix1 = _mm_setzero_ps();
902 fiy1 = _mm_setzero_ps();
903 fiz1 = _mm_setzero_ps();
904 fix2 = _mm_setzero_ps();
905 fiy2 = _mm_setzero_ps();
906 fiz2 = _mm_setzero_ps();
908 /* Start inner kernel loop */
909 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
912 /* Get j neighbor index, and coordinate index */
917 j_coord_offsetA = DIM*jnrA;
918 j_coord_offsetB = DIM*jnrB;
919 j_coord_offsetC = DIM*jnrC;
920 j_coord_offsetD = DIM*jnrD;
922 /* load j atom coordinates */
923 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
924 x+j_coord_offsetC,x+j_coord_offsetD,
927 /* Calculate displacement vector */
928 dx00 = _mm_sub_ps(ix0,jx0);
929 dy00 = _mm_sub_ps(iy0,jy0);
930 dz00 = _mm_sub_ps(iz0,jz0);
931 dx10 = _mm_sub_ps(ix1,jx0);
932 dy10 = _mm_sub_ps(iy1,jy0);
933 dz10 = _mm_sub_ps(iz1,jz0);
934 dx20 = _mm_sub_ps(ix2,jx0);
935 dy20 = _mm_sub_ps(iy2,jy0);
936 dz20 = _mm_sub_ps(iz2,jz0);
938 /* Calculate squared distance and things based on it */
939 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
940 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
941 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
943 rinv00 = gmx_mm_invsqrt_ps(rsq00);
944 rinv10 = gmx_mm_invsqrt_ps(rsq10);
945 rinv20 = gmx_mm_invsqrt_ps(rsq20);
947 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
948 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
949 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
951 /* Load parameters for j particles */
952 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
953 charge+jnrC+0,charge+jnrD+0);
954 vdwjidx0A = 2*vdwtype[jnrA+0];
955 vdwjidx0B = 2*vdwtype[jnrB+0];
956 vdwjidx0C = 2*vdwtype[jnrC+0];
957 vdwjidx0D = 2*vdwtype[jnrD+0];
959 fjx0 = _mm_setzero_ps();
960 fjy0 = _mm_setzero_ps();
961 fjz0 = _mm_setzero_ps();
963 /**************************
964 * CALCULATE INTERACTIONS *
965 **************************/
967 if (gmx_mm_any_lt(rsq00,rcutoff2))
970 r00 = _mm_mul_ps(rsq00,rinv00);
972 /* Compute parameters for interactions between i and j atoms */
973 qq00 = _mm_mul_ps(iq0,jq0);
974 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
975 vdwparam+vdwioffset0+vdwjidx0B,
976 vdwparam+vdwioffset0+vdwjidx0C,
977 vdwparam+vdwioffset0+vdwjidx0D,
980 /* EWALD ELECTROSTATICS */
982 /* Analytical PME correction */
983 zeta2 = _mm_mul_ps(beta2,rsq00);
984 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
985 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
986 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
987 felec = _mm_mul_ps(qq00,felec);
988 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
989 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
990 velec = _mm_mul_ps(qq00,velec);
992 /* LENNARD-JONES DISPERSION/REPULSION */
994 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
995 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
996 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
997 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
998 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1000 d = _mm_sub_ps(r00,rswitch);
1001 d = _mm_max_ps(d,_mm_setzero_ps());
1002 d2 = _mm_mul_ps(d,d);
1003 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1005 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1007 /* Evaluate switch function */
1008 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1009 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
1010 fvdw = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1011 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
1013 fscal = _mm_add_ps(felec,fvdw);
1015 fscal = _mm_and_ps(fscal,cutoff_mask);
1017 /* Update vectorial force */
1018 fix0 = _mm_macc_ps(dx00,fscal,fix0);
1019 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
1020 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
1022 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
1023 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
1024 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
1028 /**************************
1029 * CALCULATE INTERACTIONS *
1030 **************************/
1032 if (gmx_mm_any_lt(rsq10,rcutoff2))
1035 r10 = _mm_mul_ps(rsq10,rinv10);
1037 /* Compute parameters for interactions between i and j atoms */
1038 qq10 = _mm_mul_ps(iq1,jq0);
1040 /* EWALD ELECTROSTATICS */
1042 /* Analytical PME correction */
1043 zeta2 = _mm_mul_ps(beta2,rsq10);
1044 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
1045 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1046 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1047 felec = _mm_mul_ps(qq10,felec);
1048 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
1049 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
1050 velec = _mm_mul_ps(qq10,velec);
1052 d = _mm_sub_ps(r10,rswitch);
1053 d = _mm_max_ps(d,_mm_setzero_ps());
1054 d2 = _mm_mul_ps(d,d);
1055 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1057 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1059 /* Evaluate switch function */
1060 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1061 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
1062 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
1066 fscal = _mm_and_ps(fscal,cutoff_mask);
1068 /* Update vectorial force */
1069 fix1 = _mm_macc_ps(dx10,fscal,fix1);
1070 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
1071 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
1073 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
1074 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
1075 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
1079 /**************************
1080 * CALCULATE INTERACTIONS *
1081 **************************/
1083 if (gmx_mm_any_lt(rsq20,rcutoff2))
1086 r20 = _mm_mul_ps(rsq20,rinv20);
1088 /* Compute parameters for interactions between i and j atoms */
1089 qq20 = _mm_mul_ps(iq2,jq0);
1091 /* EWALD ELECTROSTATICS */
1093 /* Analytical PME correction */
1094 zeta2 = _mm_mul_ps(beta2,rsq20);
1095 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
1096 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1097 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1098 felec = _mm_mul_ps(qq20,felec);
1099 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
1100 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
1101 velec = _mm_mul_ps(qq20,velec);
1103 d = _mm_sub_ps(r20,rswitch);
1104 d = _mm_max_ps(d,_mm_setzero_ps());
1105 d2 = _mm_mul_ps(d,d);
1106 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1108 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1110 /* Evaluate switch function */
1111 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1112 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
1113 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
1117 fscal = _mm_and_ps(fscal,cutoff_mask);
1119 /* Update vectorial force */
1120 fix2 = _mm_macc_ps(dx20,fscal,fix2);
1121 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
1122 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
1124 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
1125 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
1126 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
1130 fjptrA = f+j_coord_offsetA;
1131 fjptrB = f+j_coord_offsetB;
1132 fjptrC = f+j_coord_offsetC;
1133 fjptrD = f+j_coord_offsetD;
1135 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1137 /* Inner loop uses 165 flops */
1140 if(jidx<j_index_end)
1143 /* Get j neighbor index, and coordinate index */
1144 jnrlistA = jjnr[jidx];
1145 jnrlistB = jjnr[jidx+1];
1146 jnrlistC = jjnr[jidx+2];
1147 jnrlistD = jjnr[jidx+3];
1148 /* Sign of each element will be negative for non-real atoms.
1149 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1150 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1152 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1153 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1154 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1155 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1156 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1157 j_coord_offsetA = DIM*jnrA;
1158 j_coord_offsetB = DIM*jnrB;
1159 j_coord_offsetC = DIM*jnrC;
1160 j_coord_offsetD = DIM*jnrD;
1162 /* load j atom coordinates */
1163 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1164 x+j_coord_offsetC,x+j_coord_offsetD,
1167 /* Calculate displacement vector */
1168 dx00 = _mm_sub_ps(ix0,jx0);
1169 dy00 = _mm_sub_ps(iy0,jy0);
1170 dz00 = _mm_sub_ps(iz0,jz0);
1171 dx10 = _mm_sub_ps(ix1,jx0);
1172 dy10 = _mm_sub_ps(iy1,jy0);
1173 dz10 = _mm_sub_ps(iz1,jz0);
1174 dx20 = _mm_sub_ps(ix2,jx0);
1175 dy20 = _mm_sub_ps(iy2,jy0);
1176 dz20 = _mm_sub_ps(iz2,jz0);
1178 /* Calculate squared distance and things based on it */
1179 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1180 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1181 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1183 rinv00 = gmx_mm_invsqrt_ps(rsq00);
1184 rinv10 = gmx_mm_invsqrt_ps(rsq10);
1185 rinv20 = gmx_mm_invsqrt_ps(rsq20);
1187 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
1188 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
1189 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
1191 /* Load parameters for j particles */
1192 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1193 charge+jnrC+0,charge+jnrD+0);
1194 vdwjidx0A = 2*vdwtype[jnrA+0];
1195 vdwjidx0B = 2*vdwtype[jnrB+0];
1196 vdwjidx0C = 2*vdwtype[jnrC+0];
1197 vdwjidx0D = 2*vdwtype[jnrD+0];
1199 fjx0 = _mm_setzero_ps();
1200 fjy0 = _mm_setzero_ps();
1201 fjz0 = _mm_setzero_ps();
1203 /**************************
1204 * CALCULATE INTERACTIONS *
1205 **************************/
1207 if (gmx_mm_any_lt(rsq00,rcutoff2))
1210 r00 = _mm_mul_ps(rsq00,rinv00);
1211 r00 = _mm_andnot_ps(dummy_mask,r00);
1213 /* Compute parameters for interactions between i and j atoms */
1214 qq00 = _mm_mul_ps(iq0,jq0);
1215 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1216 vdwparam+vdwioffset0+vdwjidx0B,
1217 vdwparam+vdwioffset0+vdwjidx0C,
1218 vdwparam+vdwioffset0+vdwjidx0D,
1221 /* EWALD ELECTROSTATICS */
1223 /* Analytical PME correction */
1224 zeta2 = _mm_mul_ps(beta2,rsq00);
1225 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
1226 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1227 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1228 felec = _mm_mul_ps(qq00,felec);
1229 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
1230 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
1231 velec = _mm_mul_ps(qq00,velec);
1233 /* LENNARD-JONES DISPERSION/REPULSION */
1235 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1236 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
1237 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1238 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
1239 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1241 d = _mm_sub_ps(r00,rswitch);
1242 d = _mm_max_ps(d,_mm_setzero_ps());
1243 d2 = _mm_mul_ps(d,d);
1244 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1246 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1248 /* Evaluate switch function */
1249 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1250 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
1251 fvdw = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1252 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
1254 fscal = _mm_add_ps(felec,fvdw);
1256 fscal = _mm_and_ps(fscal,cutoff_mask);
1258 fscal = _mm_andnot_ps(dummy_mask,fscal);
1260 /* Update vectorial force */
1261 fix0 = _mm_macc_ps(dx00,fscal,fix0);
1262 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
1263 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
1265 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
1266 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
1267 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
1271 /**************************
1272 * CALCULATE INTERACTIONS *
1273 **************************/
1275 if (gmx_mm_any_lt(rsq10,rcutoff2))
1278 r10 = _mm_mul_ps(rsq10,rinv10);
1279 r10 = _mm_andnot_ps(dummy_mask,r10);
1281 /* Compute parameters for interactions between i and j atoms */
1282 qq10 = _mm_mul_ps(iq1,jq0);
1284 /* EWALD ELECTROSTATICS */
1286 /* Analytical PME correction */
1287 zeta2 = _mm_mul_ps(beta2,rsq10);
1288 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
1289 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1290 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1291 felec = _mm_mul_ps(qq10,felec);
1292 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
1293 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
1294 velec = _mm_mul_ps(qq10,velec);
1296 d = _mm_sub_ps(r10,rswitch);
1297 d = _mm_max_ps(d,_mm_setzero_ps());
1298 d2 = _mm_mul_ps(d,d);
1299 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1301 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1303 /* Evaluate switch function */
1304 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1305 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv10,_mm_mul_ps(velec,dsw)) );
1306 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
1310 fscal = _mm_and_ps(fscal,cutoff_mask);
1312 fscal = _mm_andnot_ps(dummy_mask,fscal);
1314 /* Update vectorial force */
1315 fix1 = _mm_macc_ps(dx10,fscal,fix1);
1316 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
1317 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
1319 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
1320 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
1321 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
1325 /**************************
1326 * CALCULATE INTERACTIONS *
1327 **************************/
1329 if (gmx_mm_any_lt(rsq20,rcutoff2))
1332 r20 = _mm_mul_ps(rsq20,rinv20);
1333 r20 = _mm_andnot_ps(dummy_mask,r20);
1335 /* Compute parameters for interactions between i and j atoms */
1336 qq20 = _mm_mul_ps(iq2,jq0);
1338 /* EWALD ELECTROSTATICS */
1340 /* Analytical PME correction */
1341 zeta2 = _mm_mul_ps(beta2,rsq20);
1342 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
1343 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1344 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1345 felec = _mm_mul_ps(qq20,felec);
1346 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
1347 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
1348 velec = _mm_mul_ps(qq20,velec);
1350 d = _mm_sub_ps(r20,rswitch);
1351 d = _mm_max_ps(d,_mm_setzero_ps());
1352 d2 = _mm_mul_ps(d,d);
1353 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1355 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1357 /* Evaluate switch function */
1358 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1359 felec = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv20,_mm_mul_ps(velec,dsw)) );
1360 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
1364 fscal = _mm_and_ps(fscal,cutoff_mask);
1366 fscal = _mm_andnot_ps(dummy_mask,fscal);
1368 /* Update vectorial force */
1369 fix2 = _mm_macc_ps(dx20,fscal,fix2);
1370 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
1371 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
1373 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
1374 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
1375 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
1379 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1380 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1381 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1382 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1384 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1386 /* Inner loop uses 168 flops */
1389 /* End of innermost loop */
1391 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1392 f+i_coord_offset,fshift+i_shift_offset);
1394 /* Increment number of inner iterations */
1395 inneriter += j_index_end - j_index_start;
1397 /* Outer loop uses 18 flops */
1400 /* Increment number of outer iterations */
1403 /* Update outer/inner flops */
1405 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*168);