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_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
38 * Electrostatics interaction: ReactionField
39 * VdW interaction: LennardJones
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecRFCut_VdwLJSw_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);
82 __m128 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
83 real rswitch_scalar,d_scalar;
84 __m128 dummy_mask,cutoff_mask;
85 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
86 __m128 one = _mm_set1_ps(1.0);
87 __m128 two = _mm_set1_ps(2.0);
93 jindex = nlist->jindex;
95 shiftidx = nlist->shift;
97 shiftvec = fr->shift_vec[0];
98 fshift = fr->fshift[0];
99 facel = _mm_set1_ps(fr->epsfac);
100 charge = mdatoms->chargeA;
101 krf = _mm_set1_ps(fr->ic->k_rf);
102 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
103 crf = _mm_set1_ps(fr->ic->c_rf);
104 nvdwtype = fr->ntype;
106 vdwtype = mdatoms->typeA;
108 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
109 rcutoff_scalar = fr->rcoulomb;
110 rcutoff = _mm_set1_ps(rcutoff_scalar);
111 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
113 rswitch_scalar = fr->rvdw_switch;
114 rswitch = _mm_set1_ps(rswitch_scalar);
115 /* Setup switch parameters */
116 d_scalar = rcutoff_scalar-rswitch_scalar;
117 d = _mm_set1_ps(d_scalar);
118 swV3 = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
119 swV4 = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
120 swV5 = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
121 swF2 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
122 swF3 = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
123 swF4 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
125 /* Avoid stupid compiler warnings */
126 jnrA = jnrB = jnrC = jnrD = 0;
135 for(iidx=0;iidx<4*DIM;iidx++)
140 /* Start outer loop over neighborlists */
141 for(iidx=0; iidx<nri; iidx++)
143 /* Load shift vector for this list */
144 i_shift_offset = DIM*shiftidx[iidx];
146 /* Load limits for loop over neighbors */
147 j_index_start = jindex[iidx];
148 j_index_end = jindex[iidx+1];
150 /* Get outer coordinate index */
152 i_coord_offset = DIM*inr;
154 /* Load i particle coords and add shift vector */
155 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
157 fix0 = _mm_setzero_ps();
158 fiy0 = _mm_setzero_ps();
159 fiz0 = _mm_setzero_ps();
161 /* Load parameters for i particles */
162 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
163 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
165 /* Reset potential sums */
166 velecsum = _mm_setzero_ps();
167 vvdwsum = _mm_setzero_ps();
169 /* Start inner kernel loop */
170 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
173 /* Get j neighbor index, and coordinate index */
178 j_coord_offsetA = DIM*jnrA;
179 j_coord_offsetB = DIM*jnrB;
180 j_coord_offsetC = DIM*jnrC;
181 j_coord_offsetD = DIM*jnrD;
183 /* load j atom coordinates */
184 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
185 x+j_coord_offsetC,x+j_coord_offsetD,
188 /* Calculate displacement vector */
189 dx00 = _mm_sub_ps(ix0,jx0);
190 dy00 = _mm_sub_ps(iy0,jy0);
191 dz00 = _mm_sub_ps(iz0,jz0);
193 /* Calculate squared distance and things based on it */
194 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
196 rinv00 = gmx_mm_invsqrt_ps(rsq00);
198 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
200 /* Load parameters for j particles */
201 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
202 charge+jnrC+0,charge+jnrD+0);
203 vdwjidx0A = 2*vdwtype[jnrA+0];
204 vdwjidx0B = 2*vdwtype[jnrB+0];
205 vdwjidx0C = 2*vdwtype[jnrC+0];
206 vdwjidx0D = 2*vdwtype[jnrD+0];
208 /**************************
209 * CALCULATE INTERACTIONS *
210 **************************/
212 if (gmx_mm_any_lt(rsq00,rcutoff2))
215 r00 = _mm_mul_ps(rsq00,rinv00);
217 /* Compute parameters for interactions between i and j atoms */
218 qq00 = _mm_mul_ps(iq0,jq0);
219 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
220 vdwparam+vdwioffset0+vdwjidx0B,
221 vdwparam+vdwioffset0+vdwjidx0C,
222 vdwparam+vdwioffset0+vdwjidx0D,
225 /* REACTION-FIELD ELECTROSTATICS */
226 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
227 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
229 /* LENNARD-JONES DISPERSION/REPULSION */
231 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
232 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
233 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
234 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
235 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
237 d = _mm_sub_ps(r00,rswitch);
238 d = _mm_max_ps(d,_mm_setzero_ps());
239 d2 = _mm_mul_ps(d,d);
240 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
242 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
244 /* Evaluate switch function */
245 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
246 fvdw = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
247 vvdw = _mm_mul_ps(vvdw,sw);
248 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
250 /* Update potential sum for this i atom from the interaction with this j atom. */
251 velec = _mm_and_ps(velec,cutoff_mask);
252 velecsum = _mm_add_ps(velecsum,velec);
253 vvdw = _mm_and_ps(vvdw,cutoff_mask);
254 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
256 fscal = _mm_add_ps(felec,fvdw);
258 fscal = _mm_and_ps(fscal,cutoff_mask);
260 /* Update vectorial force */
261 fix0 = _mm_macc_ps(dx00,fscal,fix0);
262 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
263 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
265 fjptrA = f+j_coord_offsetA;
266 fjptrB = f+j_coord_offsetB;
267 fjptrC = f+j_coord_offsetC;
268 fjptrD = f+j_coord_offsetD;
269 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
270 _mm_mul_ps(dx00,fscal),
271 _mm_mul_ps(dy00,fscal),
272 _mm_mul_ps(dz00,fscal));
276 /* Inner loop uses 73 flops */
282 /* Get j neighbor index, and coordinate index */
283 jnrlistA = jjnr[jidx];
284 jnrlistB = jjnr[jidx+1];
285 jnrlistC = jjnr[jidx+2];
286 jnrlistD = jjnr[jidx+3];
287 /* Sign of each element will be negative for non-real atoms.
288 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
289 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
291 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
292 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
293 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
294 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
295 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
296 j_coord_offsetA = DIM*jnrA;
297 j_coord_offsetB = DIM*jnrB;
298 j_coord_offsetC = DIM*jnrC;
299 j_coord_offsetD = DIM*jnrD;
301 /* load j atom coordinates */
302 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
303 x+j_coord_offsetC,x+j_coord_offsetD,
306 /* Calculate displacement vector */
307 dx00 = _mm_sub_ps(ix0,jx0);
308 dy00 = _mm_sub_ps(iy0,jy0);
309 dz00 = _mm_sub_ps(iz0,jz0);
311 /* Calculate squared distance and things based on it */
312 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
314 rinv00 = gmx_mm_invsqrt_ps(rsq00);
316 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
318 /* Load parameters for j particles */
319 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
320 charge+jnrC+0,charge+jnrD+0);
321 vdwjidx0A = 2*vdwtype[jnrA+0];
322 vdwjidx0B = 2*vdwtype[jnrB+0];
323 vdwjidx0C = 2*vdwtype[jnrC+0];
324 vdwjidx0D = 2*vdwtype[jnrD+0];
326 /**************************
327 * CALCULATE INTERACTIONS *
328 **************************/
330 if (gmx_mm_any_lt(rsq00,rcutoff2))
333 r00 = _mm_mul_ps(rsq00,rinv00);
334 r00 = _mm_andnot_ps(dummy_mask,r00);
336 /* Compute parameters for interactions between i and j atoms */
337 qq00 = _mm_mul_ps(iq0,jq0);
338 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
339 vdwparam+vdwioffset0+vdwjidx0B,
340 vdwparam+vdwioffset0+vdwjidx0C,
341 vdwparam+vdwioffset0+vdwjidx0D,
344 /* REACTION-FIELD ELECTROSTATICS */
345 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
346 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
348 /* LENNARD-JONES DISPERSION/REPULSION */
350 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
351 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
352 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
353 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
354 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
356 d = _mm_sub_ps(r00,rswitch);
357 d = _mm_max_ps(d,_mm_setzero_ps());
358 d2 = _mm_mul_ps(d,d);
359 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
361 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
363 /* Evaluate switch function */
364 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
365 fvdw = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
366 vvdw = _mm_mul_ps(vvdw,sw);
367 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
369 /* Update potential sum for this i atom from the interaction with this j atom. */
370 velec = _mm_and_ps(velec,cutoff_mask);
371 velec = _mm_andnot_ps(dummy_mask,velec);
372 velecsum = _mm_add_ps(velecsum,velec);
373 vvdw = _mm_and_ps(vvdw,cutoff_mask);
374 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
375 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
377 fscal = _mm_add_ps(felec,fvdw);
379 fscal = _mm_and_ps(fscal,cutoff_mask);
381 fscal = _mm_andnot_ps(dummy_mask,fscal);
383 /* Update vectorial force */
384 fix0 = _mm_macc_ps(dx00,fscal,fix0);
385 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
386 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
388 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
389 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
390 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
391 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
392 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
393 _mm_mul_ps(dx00,fscal),
394 _mm_mul_ps(dy00,fscal),
395 _mm_mul_ps(dz00,fscal));
399 /* Inner loop uses 74 flops */
402 /* End of innermost loop */
404 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
405 f+i_coord_offset,fshift+i_shift_offset);
408 /* Update potential energies */
409 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
410 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
412 /* Increment number of inner iterations */
413 inneriter += j_index_end - j_index_start;
415 /* Outer loop uses 9 flops */
418 /* Increment number of outer iterations */
421 /* Update outer/inner flops */
423 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*74);
426 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_128_fma_single
427 * Electrostatics interaction: ReactionField
428 * VdW interaction: LennardJones
429 * Geometry: Particle-Particle
430 * Calculate force/pot: Force
433 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_128_fma_single
434 (t_nblist * gmx_restrict nlist,
435 rvec * gmx_restrict xx,
436 rvec * gmx_restrict ff,
437 t_forcerec * gmx_restrict fr,
438 t_mdatoms * gmx_restrict mdatoms,
439 nb_kernel_data_t * gmx_restrict kernel_data,
440 t_nrnb * gmx_restrict nrnb)
442 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
443 * just 0 for non-waters.
444 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
445 * jnr indices corresponding to data put in the four positions in the SIMD register.
447 int i_shift_offset,i_coord_offset,outeriter,inneriter;
448 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
449 int jnrA,jnrB,jnrC,jnrD;
450 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
451 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
452 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
454 real *shiftvec,*fshift,*x,*f;
455 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
457 __m128 fscal,rcutoff,rcutoff2,jidxall;
459 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
460 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
461 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
462 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
463 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
466 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
469 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
470 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
471 __m128 rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
472 real rswitch_scalar,d_scalar;
473 __m128 dummy_mask,cutoff_mask;
474 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
475 __m128 one = _mm_set1_ps(1.0);
476 __m128 two = _mm_set1_ps(2.0);
482 jindex = nlist->jindex;
484 shiftidx = nlist->shift;
486 shiftvec = fr->shift_vec[0];
487 fshift = fr->fshift[0];
488 facel = _mm_set1_ps(fr->epsfac);
489 charge = mdatoms->chargeA;
490 krf = _mm_set1_ps(fr->ic->k_rf);
491 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
492 crf = _mm_set1_ps(fr->ic->c_rf);
493 nvdwtype = fr->ntype;
495 vdwtype = mdatoms->typeA;
497 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
498 rcutoff_scalar = fr->rcoulomb;
499 rcutoff = _mm_set1_ps(rcutoff_scalar);
500 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
502 rswitch_scalar = fr->rvdw_switch;
503 rswitch = _mm_set1_ps(rswitch_scalar);
504 /* Setup switch parameters */
505 d_scalar = rcutoff_scalar-rswitch_scalar;
506 d = _mm_set1_ps(d_scalar);
507 swV3 = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
508 swV4 = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
509 swV5 = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
510 swF2 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
511 swF3 = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
512 swF4 = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
514 /* Avoid stupid compiler warnings */
515 jnrA = jnrB = jnrC = jnrD = 0;
524 for(iidx=0;iidx<4*DIM;iidx++)
529 /* Start outer loop over neighborlists */
530 for(iidx=0; iidx<nri; iidx++)
532 /* Load shift vector for this list */
533 i_shift_offset = DIM*shiftidx[iidx];
535 /* Load limits for loop over neighbors */
536 j_index_start = jindex[iidx];
537 j_index_end = jindex[iidx+1];
539 /* Get outer coordinate index */
541 i_coord_offset = DIM*inr;
543 /* Load i particle coords and add shift vector */
544 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
546 fix0 = _mm_setzero_ps();
547 fiy0 = _mm_setzero_ps();
548 fiz0 = _mm_setzero_ps();
550 /* Load parameters for i particles */
551 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
552 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
554 /* Start inner kernel loop */
555 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
558 /* Get j neighbor index, and coordinate index */
563 j_coord_offsetA = DIM*jnrA;
564 j_coord_offsetB = DIM*jnrB;
565 j_coord_offsetC = DIM*jnrC;
566 j_coord_offsetD = DIM*jnrD;
568 /* load j atom coordinates */
569 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
570 x+j_coord_offsetC,x+j_coord_offsetD,
573 /* Calculate displacement vector */
574 dx00 = _mm_sub_ps(ix0,jx0);
575 dy00 = _mm_sub_ps(iy0,jy0);
576 dz00 = _mm_sub_ps(iz0,jz0);
578 /* Calculate squared distance and things based on it */
579 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
581 rinv00 = gmx_mm_invsqrt_ps(rsq00);
583 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
585 /* Load parameters for j particles */
586 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
587 charge+jnrC+0,charge+jnrD+0);
588 vdwjidx0A = 2*vdwtype[jnrA+0];
589 vdwjidx0B = 2*vdwtype[jnrB+0];
590 vdwjidx0C = 2*vdwtype[jnrC+0];
591 vdwjidx0D = 2*vdwtype[jnrD+0];
593 /**************************
594 * CALCULATE INTERACTIONS *
595 **************************/
597 if (gmx_mm_any_lt(rsq00,rcutoff2))
600 r00 = _mm_mul_ps(rsq00,rinv00);
602 /* Compute parameters for interactions between i and j atoms */
603 qq00 = _mm_mul_ps(iq0,jq0);
604 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
605 vdwparam+vdwioffset0+vdwjidx0B,
606 vdwparam+vdwioffset0+vdwjidx0C,
607 vdwparam+vdwioffset0+vdwjidx0D,
610 /* REACTION-FIELD ELECTROSTATICS */
611 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
613 /* LENNARD-JONES DISPERSION/REPULSION */
615 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
616 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
617 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
618 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
619 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
621 d = _mm_sub_ps(r00,rswitch);
622 d = _mm_max_ps(d,_mm_setzero_ps());
623 d2 = _mm_mul_ps(d,d);
624 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
626 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
628 /* Evaluate switch function */
629 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
630 fvdw = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
631 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
633 fscal = _mm_add_ps(felec,fvdw);
635 fscal = _mm_and_ps(fscal,cutoff_mask);
637 /* Update vectorial force */
638 fix0 = _mm_macc_ps(dx00,fscal,fix0);
639 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
640 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
642 fjptrA = f+j_coord_offsetA;
643 fjptrB = f+j_coord_offsetB;
644 fjptrC = f+j_coord_offsetC;
645 fjptrD = f+j_coord_offsetD;
646 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
647 _mm_mul_ps(dx00,fscal),
648 _mm_mul_ps(dy00,fscal),
649 _mm_mul_ps(dz00,fscal));
653 /* Inner loop uses 64 flops */
659 /* Get j neighbor index, and coordinate index */
660 jnrlistA = jjnr[jidx];
661 jnrlistB = jjnr[jidx+1];
662 jnrlistC = jjnr[jidx+2];
663 jnrlistD = jjnr[jidx+3];
664 /* Sign of each element will be negative for non-real atoms.
665 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
666 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
668 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
669 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
670 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
671 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
672 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
673 j_coord_offsetA = DIM*jnrA;
674 j_coord_offsetB = DIM*jnrB;
675 j_coord_offsetC = DIM*jnrC;
676 j_coord_offsetD = DIM*jnrD;
678 /* load j atom coordinates */
679 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
680 x+j_coord_offsetC,x+j_coord_offsetD,
683 /* Calculate displacement vector */
684 dx00 = _mm_sub_ps(ix0,jx0);
685 dy00 = _mm_sub_ps(iy0,jy0);
686 dz00 = _mm_sub_ps(iz0,jz0);
688 /* Calculate squared distance and things based on it */
689 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
691 rinv00 = gmx_mm_invsqrt_ps(rsq00);
693 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
695 /* Load parameters for j particles */
696 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
697 charge+jnrC+0,charge+jnrD+0);
698 vdwjidx0A = 2*vdwtype[jnrA+0];
699 vdwjidx0B = 2*vdwtype[jnrB+0];
700 vdwjidx0C = 2*vdwtype[jnrC+0];
701 vdwjidx0D = 2*vdwtype[jnrD+0];
703 /**************************
704 * CALCULATE INTERACTIONS *
705 **************************/
707 if (gmx_mm_any_lt(rsq00,rcutoff2))
710 r00 = _mm_mul_ps(rsq00,rinv00);
711 r00 = _mm_andnot_ps(dummy_mask,r00);
713 /* Compute parameters for interactions between i and j atoms */
714 qq00 = _mm_mul_ps(iq0,jq0);
715 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
716 vdwparam+vdwioffset0+vdwjidx0B,
717 vdwparam+vdwioffset0+vdwjidx0C,
718 vdwparam+vdwioffset0+vdwjidx0D,
721 /* REACTION-FIELD ELECTROSTATICS */
722 felec = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
724 /* LENNARD-JONES DISPERSION/REPULSION */
726 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
727 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
728 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
729 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
730 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
732 d = _mm_sub_ps(r00,rswitch);
733 d = _mm_max_ps(d,_mm_setzero_ps());
734 d2 = _mm_mul_ps(d,d);
735 sw = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
737 dsw = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
739 /* Evaluate switch function */
740 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
741 fvdw = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
742 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
744 fscal = _mm_add_ps(felec,fvdw);
746 fscal = _mm_and_ps(fscal,cutoff_mask);
748 fscal = _mm_andnot_ps(dummy_mask,fscal);
750 /* Update vectorial force */
751 fix0 = _mm_macc_ps(dx00,fscal,fix0);
752 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
753 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
755 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
756 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
757 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
758 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
759 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
760 _mm_mul_ps(dx00,fscal),
761 _mm_mul_ps(dy00,fscal),
762 _mm_mul_ps(dz00,fscal));
766 /* Inner loop uses 65 flops */
769 /* End of innermost loop */
771 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
772 f+i_coord_offset,fshift+i_shift_offset);
774 /* Increment number of inner iterations */
775 inneriter += j_index_end - j_index_start;
777 /* Outer loop uses 7 flops */
780 /* Increment number of outer iterations */
783 /* Update outer/inner flops */
785 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*65);