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_ElecEw_VdwLJ_GeomP1P1_VF_avx_128_fma_single
38 * Electrostatics interaction: Ewald
39 * VdW interaction: LennardJones
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEw_VdwLJ_GeomP1P1_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;
71 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
72 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
73 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
74 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
77 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
80 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
81 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
83 __m128 ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
84 __m128 beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
86 __m128 dummy_mask,cutoff_mask;
87 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
88 __m128 one = _mm_set1_ps(1.0);
89 __m128 two = _mm_set1_ps(2.0);
95 jindex = nlist->jindex;
97 shiftidx = nlist->shift;
99 shiftvec = fr->shift_vec[0];
100 fshift = fr->fshift[0];
101 facel = _mm_set1_ps(fr->epsfac);
102 charge = mdatoms->chargeA;
103 nvdwtype = fr->ntype;
105 vdwtype = mdatoms->typeA;
107 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
108 beta = _mm_set1_ps(fr->ic->ewaldcoeff);
109 beta2 = _mm_mul_ps(beta,beta);
110 beta3 = _mm_mul_ps(beta,beta2);
111 ewtab = fr->ic->tabq_coul_FDV0;
112 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
113 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
115 /* Avoid stupid compiler warnings */
116 jnrA = jnrB = jnrC = jnrD = 0;
125 for(iidx=0;iidx<4*DIM;iidx++)
130 /* Start outer loop over neighborlists */
131 for(iidx=0; iidx<nri; iidx++)
133 /* Load shift vector for this list */
134 i_shift_offset = DIM*shiftidx[iidx];
136 /* Load limits for loop over neighbors */
137 j_index_start = jindex[iidx];
138 j_index_end = jindex[iidx+1];
140 /* Get outer coordinate index */
142 i_coord_offset = DIM*inr;
144 /* Load i particle coords and add shift vector */
145 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
147 fix0 = _mm_setzero_ps();
148 fiy0 = _mm_setzero_ps();
149 fiz0 = _mm_setzero_ps();
151 /* Load parameters for i particles */
152 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
153 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
155 /* Reset potential sums */
156 velecsum = _mm_setzero_ps();
157 vvdwsum = _mm_setzero_ps();
159 /* Start inner kernel loop */
160 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
163 /* Get j neighbor index, and coordinate index */
168 j_coord_offsetA = DIM*jnrA;
169 j_coord_offsetB = DIM*jnrB;
170 j_coord_offsetC = DIM*jnrC;
171 j_coord_offsetD = DIM*jnrD;
173 /* load j atom coordinates */
174 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
175 x+j_coord_offsetC,x+j_coord_offsetD,
178 /* Calculate displacement vector */
179 dx00 = _mm_sub_ps(ix0,jx0);
180 dy00 = _mm_sub_ps(iy0,jy0);
181 dz00 = _mm_sub_ps(iz0,jz0);
183 /* Calculate squared distance and things based on it */
184 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
186 rinv00 = gmx_mm_invsqrt_ps(rsq00);
188 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
190 /* Load parameters for j particles */
191 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
192 charge+jnrC+0,charge+jnrD+0);
193 vdwjidx0A = 2*vdwtype[jnrA+0];
194 vdwjidx0B = 2*vdwtype[jnrB+0];
195 vdwjidx0C = 2*vdwtype[jnrC+0];
196 vdwjidx0D = 2*vdwtype[jnrD+0];
198 /**************************
199 * CALCULATE INTERACTIONS *
200 **************************/
202 r00 = _mm_mul_ps(rsq00,rinv00);
204 /* Compute parameters for interactions between i and j atoms */
205 qq00 = _mm_mul_ps(iq0,jq0);
206 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
207 vdwparam+vdwioffset0+vdwjidx0B,
208 vdwparam+vdwioffset0+vdwjidx0C,
209 vdwparam+vdwioffset0+vdwjidx0D,
212 /* EWALD ELECTROSTATICS */
214 /* Analytical PME correction */
215 zeta2 = _mm_mul_ps(beta2,rsq00);
216 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
217 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
218 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
219 felec = _mm_mul_ps(qq00,felec);
220 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
221 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
222 velec = _mm_mul_ps(qq00,velec);
224 /* LENNARD-JONES DISPERSION/REPULSION */
226 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
227 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
228 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
229 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
230 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
232 /* Update potential sum for this i atom from the interaction with this j atom. */
233 velecsum = _mm_add_ps(velecsum,velec);
234 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
236 fscal = _mm_add_ps(felec,fvdw);
238 /* Update vectorial force */
239 fix0 = _mm_macc_ps(dx00,fscal,fix0);
240 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
241 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
243 fjptrA = f+j_coord_offsetA;
244 fjptrB = f+j_coord_offsetB;
245 fjptrC = f+j_coord_offsetC;
246 fjptrD = f+j_coord_offsetD;
247 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
248 _mm_mul_ps(dx00,fscal),
249 _mm_mul_ps(dy00,fscal),
250 _mm_mul_ps(dz00,fscal));
252 /* Inner loop uses 41 flops */
258 /* Get j neighbor index, and coordinate index */
259 jnrlistA = jjnr[jidx];
260 jnrlistB = jjnr[jidx+1];
261 jnrlistC = jjnr[jidx+2];
262 jnrlistD = jjnr[jidx+3];
263 /* Sign of each element will be negative for non-real atoms.
264 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
265 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
267 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
268 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
269 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
270 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
271 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
272 j_coord_offsetA = DIM*jnrA;
273 j_coord_offsetB = DIM*jnrB;
274 j_coord_offsetC = DIM*jnrC;
275 j_coord_offsetD = DIM*jnrD;
277 /* load j atom coordinates */
278 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
279 x+j_coord_offsetC,x+j_coord_offsetD,
282 /* Calculate displacement vector */
283 dx00 = _mm_sub_ps(ix0,jx0);
284 dy00 = _mm_sub_ps(iy0,jy0);
285 dz00 = _mm_sub_ps(iz0,jz0);
287 /* Calculate squared distance and things based on it */
288 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
290 rinv00 = gmx_mm_invsqrt_ps(rsq00);
292 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
294 /* Load parameters for j particles */
295 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
296 charge+jnrC+0,charge+jnrD+0);
297 vdwjidx0A = 2*vdwtype[jnrA+0];
298 vdwjidx0B = 2*vdwtype[jnrB+0];
299 vdwjidx0C = 2*vdwtype[jnrC+0];
300 vdwjidx0D = 2*vdwtype[jnrD+0];
302 /**************************
303 * CALCULATE INTERACTIONS *
304 **************************/
306 r00 = _mm_mul_ps(rsq00,rinv00);
307 r00 = _mm_andnot_ps(dummy_mask,r00);
309 /* Compute parameters for interactions between i and j atoms */
310 qq00 = _mm_mul_ps(iq0,jq0);
311 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
312 vdwparam+vdwioffset0+vdwjidx0B,
313 vdwparam+vdwioffset0+vdwjidx0C,
314 vdwparam+vdwioffset0+vdwjidx0D,
317 /* EWALD ELECTROSTATICS */
319 /* Analytical PME correction */
320 zeta2 = _mm_mul_ps(beta2,rsq00);
321 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
322 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
323 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
324 felec = _mm_mul_ps(qq00,felec);
325 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
326 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
327 velec = _mm_mul_ps(qq00,velec);
329 /* LENNARD-JONES DISPERSION/REPULSION */
331 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
332 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
333 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
334 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
335 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
337 /* Update potential sum for this i atom from the interaction with this j atom. */
338 velec = _mm_andnot_ps(dummy_mask,velec);
339 velecsum = _mm_add_ps(velecsum,velec);
340 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
341 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
343 fscal = _mm_add_ps(felec,fvdw);
345 fscal = _mm_andnot_ps(dummy_mask,fscal);
347 /* Update vectorial force */
348 fix0 = _mm_macc_ps(dx00,fscal,fix0);
349 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
350 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
352 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
353 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
354 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
355 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
356 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
357 _mm_mul_ps(dx00,fscal),
358 _mm_mul_ps(dy00,fscal),
359 _mm_mul_ps(dz00,fscal));
361 /* Inner loop uses 42 flops */
364 /* End of innermost loop */
366 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
367 f+i_coord_offset,fshift+i_shift_offset);
370 /* Update potential energies */
371 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
372 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
374 /* Increment number of inner iterations */
375 inneriter += j_index_end - j_index_start;
377 /* Outer loop uses 9 flops */
380 /* Increment number of outer iterations */
383 /* Update outer/inner flops */
385 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*42);
388 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_128_fma_single
389 * Electrostatics interaction: Ewald
390 * VdW interaction: LennardJones
391 * Geometry: Particle-Particle
392 * Calculate force/pot: Force
395 nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_128_fma_single
396 (t_nblist * gmx_restrict nlist,
397 rvec * gmx_restrict xx,
398 rvec * gmx_restrict ff,
399 t_forcerec * gmx_restrict fr,
400 t_mdatoms * gmx_restrict mdatoms,
401 nb_kernel_data_t * gmx_restrict kernel_data,
402 t_nrnb * gmx_restrict nrnb)
404 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
405 * just 0 for non-waters.
406 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
407 * jnr indices corresponding to data put in the four positions in the SIMD register.
409 int i_shift_offset,i_coord_offset,outeriter,inneriter;
410 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
411 int jnrA,jnrB,jnrC,jnrD;
412 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
413 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
414 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
416 real *shiftvec,*fshift,*x,*f;
417 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
419 __m128 fscal,rcutoff,rcutoff2,jidxall;
421 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
422 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
423 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
424 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
425 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
428 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
431 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
432 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
434 __m128 ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
435 __m128 beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
437 __m128 dummy_mask,cutoff_mask;
438 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
439 __m128 one = _mm_set1_ps(1.0);
440 __m128 two = _mm_set1_ps(2.0);
446 jindex = nlist->jindex;
448 shiftidx = nlist->shift;
450 shiftvec = fr->shift_vec[0];
451 fshift = fr->fshift[0];
452 facel = _mm_set1_ps(fr->epsfac);
453 charge = mdatoms->chargeA;
454 nvdwtype = fr->ntype;
456 vdwtype = mdatoms->typeA;
458 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
459 beta = _mm_set1_ps(fr->ic->ewaldcoeff);
460 beta2 = _mm_mul_ps(beta,beta);
461 beta3 = _mm_mul_ps(beta,beta2);
462 ewtab = fr->ic->tabq_coul_F;
463 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
464 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
466 /* Avoid stupid compiler warnings */
467 jnrA = jnrB = jnrC = jnrD = 0;
476 for(iidx=0;iidx<4*DIM;iidx++)
481 /* Start outer loop over neighborlists */
482 for(iidx=0; iidx<nri; iidx++)
484 /* Load shift vector for this list */
485 i_shift_offset = DIM*shiftidx[iidx];
487 /* Load limits for loop over neighbors */
488 j_index_start = jindex[iidx];
489 j_index_end = jindex[iidx+1];
491 /* Get outer coordinate index */
493 i_coord_offset = DIM*inr;
495 /* Load i particle coords and add shift vector */
496 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
498 fix0 = _mm_setzero_ps();
499 fiy0 = _mm_setzero_ps();
500 fiz0 = _mm_setzero_ps();
502 /* Load parameters for i particles */
503 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
504 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
506 /* Start inner kernel loop */
507 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
510 /* Get j neighbor index, and coordinate index */
515 j_coord_offsetA = DIM*jnrA;
516 j_coord_offsetB = DIM*jnrB;
517 j_coord_offsetC = DIM*jnrC;
518 j_coord_offsetD = DIM*jnrD;
520 /* load j atom coordinates */
521 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
522 x+j_coord_offsetC,x+j_coord_offsetD,
525 /* Calculate displacement vector */
526 dx00 = _mm_sub_ps(ix0,jx0);
527 dy00 = _mm_sub_ps(iy0,jy0);
528 dz00 = _mm_sub_ps(iz0,jz0);
530 /* Calculate squared distance and things based on it */
531 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
533 rinv00 = gmx_mm_invsqrt_ps(rsq00);
535 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
537 /* Load parameters for j particles */
538 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
539 charge+jnrC+0,charge+jnrD+0);
540 vdwjidx0A = 2*vdwtype[jnrA+0];
541 vdwjidx0B = 2*vdwtype[jnrB+0];
542 vdwjidx0C = 2*vdwtype[jnrC+0];
543 vdwjidx0D = 2*vdwtype[jnrD+0];
545 /**************************
546 * CALCULATE INTERACTIONS *
547 **************************/
549 r00 = _mm_mul_ps(rsq00,rinv00);
551 /* Compute parameters for interactions between i and j atoms */
552 qq00 = _mm_mul_ps(iq0,jq0);
553 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
554 vdwparam+vdwioffset0+vdwjidx0B,
555 vdwparam+vdwioffset0+vdwjidx0C,
556 vdwparam+vdwioffset0+vdwjidx0D,
559 /* EWALD ELECTROSTATICS */
561 /* Analytical PME correction */
562 zeta2 = _mm_mul_ps(beta2,rsq00);
563 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
564 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
565 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
566 felec = _mm_mul_ps(qq00,felec);
568 /* LENNARD-JONES DISPERSION/REPULSION */
570 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
571 fvdw = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
573 fscal = _mm_add_ps(felec,fvdw);
575 /* Update vectorial force */
576 fix0 = _mm_macc_ps(dx00,fscal,fix0);
577 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
578 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
580 fjptrA = f+j_coord_offsetA;
581 fjptrB = f+j_coord_offsetB;
582 fjptrC = f+j_coord_offsetC;
583 fjptrD = f+j_coord_offsetD;
584 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
585 _mm_mul_ps(dx00,fscal),
586 _mm_mul_ps(dy00,fscal),
587 _mm_mul_ps(dz00,fscal));
589 /* Inner loop uses 35 flops */
595 /* Get j neighbor index, and coordinate index */
596 jnrlistA = jjnr[jidx];
597 jnrlistB = jjnr[jidx+1];
598 jnrlistC = jjnr[jidx+2];
599 jnrlistD = jjnr[jidx+3];
600 /* Sign of each element will be negative for non-real atoms.
601 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
602 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
604 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
605 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
606 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
607 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
608 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
609 j_coord_offsetA = DIM*jnrA;
610 j_coord_offsetB = DIM*jnrB;
611 j_coord_offsetC = DIM*jnrC;
612 j_coord_offsetD = DIM*jnrD;
614 /* load j atom coordinates */
615 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
616 x+j_coord_offsetC,x+j_coord_offsetD,
619 /* Calculate displacement vector */
620 dx00 = _mm_sub_ps(ix0,jx0);
621 dy00 = _mm_sub_ps(iy0,jy0);
622 dz00 = _mm_sub_ps(iz0,jz0);
624 /* Calculate squared distance and things based on it */
625 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
627 rinv00 = gmx_mm_invsqrt_ps(rsq00);
629 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
631 /* Load parameters for j particles */
632 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
633 charge+jnrC+0,charge+jnrD+0);
634 vdwjidx0A = 2*vdwtype[jnrA+0];
635 vdwjidx0B = 2*vdwtype[jnrB+0];
636 vdwjidx0C = 2*vdwtype[jnrC+0];
637 vdwjidx0D = 2*vdwtype[jnrD+0];
639 /**************************
640 * CALCULATE INTERACTIONS *
641 **************************/
643 r00 = _mm_mul_ps(rsq00,rinv00);
644 r00 = _mm_andnot_ps(dummy_mask,r00);
646 /* Compute parameters for interactions between i and j atoms */
647 qq00 = _mm_mul_ps(iq0,jq0);
648 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
649 vdwparam+vdwioffset0+vdwjidx0B,
650 vdwparam+vdwioffset0+vdwjidx0C,
651 vdwparam+vdwioffset0+vdwjidx0D,
654 /* EWALD ELECTROSTATICS */
656 /* Analytical PME correction */
657 zeta2 = _mm_mul_ps(beta2,rsq00);
658 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
659 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
660 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
661 felec = _mm_mul_ps(qq00,felec);
663 /* LENNARD-JONES DISPERSION/REPULSION */
665 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
666 fvdw = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
668 fscal = _mm_add_ps(felec,fvdw);
670 fscal = _mm_andnot_ps(dummy_mask,fscal);
672 /* Update vectorial force */
673 fix0 = _mm_macc_ps(dx00,fscal,fix0);
674 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
675 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
677 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
678 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
679 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
680 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
681 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
682 _mm_mul_ps(dx00,fscal),
683 _mm_mul_ps(dy00,fscal),
684 _mm_mul_ps(dz00,fscal));
686 /* Inner loop uses 36 flops */
689 /* End of innermost loop */
691 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
692 f+i_coord_offset,fshift+i_shift_offset);
694 /* Increment number of inner iterations */
695 inneriter += j_index_end - j_index_start;
697 /* Outer loop uses 7 flops */
700 /* Increment number of outer iterations */
703 /* Update outer/inner flops */
705 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*36);