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_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_128_fma_single
38 * Electrostatics interaction: CubicSplineTable
39 * VdW interaction: LennardJones
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecCSTab_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 __m128i ifour = _mm_set1_epi32(4);
84 __m128 rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
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 vftab = kernel_data->table_elec->data;
108 vftabscale = _mm_set1_ps(kernel_data->table_elec->scale);
110 /* Avoid stupid compiler warnings */
111 jnrA = jnrB = jnrC = jnrD = 0;
120 for(iidx=0;iidx<4*DIM;iidx++)
125 /* Start outer loop over neighborlists */
126 for(iidx=0; iidx<nri; iidx++)
128 /* Load shift vector for this list */
129 i_shift_offset = DIM*shiftidx[iidx];
131 /* Load limits for loop over neighbors */
132 j_index_start = jindex[iidx];
133 j_index_end = jindex[iidx+1];
135 /* Get outer coordinate index */
137 i_coord_offset = DIM*inr;
139 /* Load i particle coords and add shift vector */
140 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
142 fix0 = _mm_setzero_ps();
143 fiy0 = _mm_setzero_ps();
144 fiz0 = _mm_setzero_ps();
146 /* Load parameters for i particles */
147 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
148 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
150 /* Reset potential sums */
151 velecsum = _mm_setzero_ps();
152 vvdwsum = _mm_setzero_ps();
154 /* Start inner kernel loop */
155 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
158 /* Get j neighbor index, and coordinate index */
163 j_coord_offsetA = DIM*jnrA;
164 j_coord_offsetB = DIM*jnrB;
165 j_coord_offsetC = DIM*jnrC;
166 j_coord_offsetD = DIM*jnrD;
168 /* load j atom coordinates */
169 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
170 x+j_coord_offsetC,x+j_coord_offsetD,
173 /* Calculate displacement vector */
174 dx00 = _mm_sub_ps(ix0,jx0);
175 dy00 = _mm_sub_ps(iy0,jy0);
176 dz00 = _mm_sub_ps(iz0,jz0);
178 /* Calculate squared distance and things based on it */
179 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
181 rinv00 = gmx_mm_invsqrt_ps(rsq00);
183 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
185 /* Load parameters for j particles */
186 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
187 charge+jnrC+0,charge+jnrD+0);
188 vdwjidx0A = 2*vdwtype[jnrA+0];
189 vdwjidx0B = 2*vdwtype[jnrB+0];
190 vdwjidx0C = 2*vdwtype[jnrC+0];
191 vdwjidx0D = 2*vdwtype[jnrD+0];
193 /**************************
194 * CALCULATE INTERACTIONS *
195 **************************/
197 r00 = _mm_mul_ps(rsq00,rinv00);
199 /* Compute parameters for interactions between i and j atoms */
200 qq00 = _mm_mul_ps(iq0,jq0);
201 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
202 vdwparam+vdwioffset0+vdwjidx0B,
203 vdwparam+vdwioffset0+vdwjidx0C,
204 vdwparam+vdwioffset0+vdwjidx0D,
207 /* Calculate table index by multiplying r with table scale and truncate to integer */
208 rt = _mm_mul_ps(r00,vftabscale);
209 vfitab = _mm_cvttps_epi32(rt);
211 vfeps = _mm_frcz_ps(rt);
213 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
215 twovfeps = _mm_add_ps(vfeps,vfeps);
216 vfitab = _mm_slli_epi32(vfitab,2);
218 /* CUBIC SPLINE TABLE ELECTROSTATICS */
219 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
220 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
221 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
222 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
223 _MM_TRANSPOSE4_PS(Y,F,G,H);
224 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
225 VV = _mm_macc_ps(vfeps,Fp,Y);
226 velec = _mm_mul_ps(qq00,VV);
227 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
228 felec = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
230 /* LENNARD-JONES DISPERSION/REPULSION */
232 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
233 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
234 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
235 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
236 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
238 /* Update potential sum for this i atom from the interaction with this j atom. */
239 velecsum = _mm_add_ps(velecsum,velec);
240 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
242 fscal = _mm_add_ps(felec,fvdw);
244 /* Update vectorial force */
245 fix0 = _mm_macc_ps(dx00,fscal,fix0);
246 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
247 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
249 fjptrA = f+j_coord_offsetA;
250 fjptrB = f+j_coord_offsetB;
251 fjptrC = f+j_coord_offsetC;
252 fjptrD = f+j_coord_offsetD;
253 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
254 _mm_mul_ps(dx00,fscal),
255 _mm_mul_ps(dy00,fscal),
256 _mm_mul_ps(dz00,fscal));
258 /* Inner loop uses 59 flops */
264 /* Get j neighbor index, and coordinate index */
265 jnrlistA = jjnr[jidx];
266 jnrlistB = jjnr[jidx+1];
267 jnrlistC = jjnr[jidx+2];
268 jnrlistD = jjnr[jidx+3];
269 /* Sign of each element will be negative for non-real atoms.
270 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
271 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
273 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
274 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
275 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
276 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
277 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
278 j_coord_offsetA = DIM*jnrA;
279 j_coord_offsetB = DIM*jnrB;
280 j_coord_offsetC = DIM*jnrC;
281 j_coord_offsetD = DIM*jnrD;
283 /* load j atom coordinates */
284 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
285 x+j_coord_offsetC,x+j_coord_offsetD,
288 /* Calculate displacement vector */
289 dx00 = _mm_sub_ps(ix0,jx0);
290 dy00 = _mm_sub_ps(iy0,jy0);
291 dz00 = _mm_sub_ps(iz0,jz0);
293 /* Calculate squared distance and things based on it */
294 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
296 rinv00 = gmx_mm_invsqrt_ps(rsq00);
298 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
300 /* Load parameters for j particles */
301 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
302 charge+jnrC+0,charge+jnrD+0);
303 vdwjidx0A = 2*vdwtype[jnrA+0];
304 vdwjidx0B = 2*vdwtype[jnrB+0];
305 vdwjidx0C = 2*vdwtype[jnrC+0];
306 vdwjidx0D = 2*vdwtype[jnrD+0];
308 /**************************
309 * CALCULATE INTERACTIONS *
310 **************************/
312 r00 = _mm_mul_ps(rsq00,rinv00);
313 r00 = _mm_andnot_ps(dummy_mask,r00);
315 /* Compute parameters for interactions between i and j atoms */
316 qq00 = _mm_mul_ps(iq0,jq0);
317 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
318 vdwparam+vdwioffset0+vdwjidx0B,
319 vdwparam+vdwioffset0+vdwjidx0C,
320 vdwparam+vdwioffset0+vdwjidx0D,
323 /* Calculate table index by multiplying r with table scale and truncate to integer */
324 rt = _mm_mul_ps(r00,vftabscale);
325 vfitab = _mm_cvttps_epi32(rt);
327 vfeps = _mm_frcz_ps(rt);
329 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
331 twovfeps = _mm_add_ps(vfeps,vfeps);
332 vfitab = _mm_slli_epi32(vfitab,2);
334 /* CUBIC SPLINE TABLE ELECTROSTATICS */
335 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
336 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
337 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
338 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
339 _MM_TRANSPOSE4_PS(Y,F,G,H);
340 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
341 VV = _mm_macc_ps(vfeps,Fp,Y);
342 velec = _mm_mul_ps(qq00,VV);
343 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
344 felec = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
346 /* LENNARD-JONES DISPERSION/REPULSION */
348 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
349 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
350 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
351 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
352 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
354 /* Update potential sum for this i atom from the interaction with this j atom. */
355 velec = _mm_andnot_ps(dummy_mask,velec);
356 velecsum = _mm_add_ps(velecsum,velec);
357 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
358 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
360 fscal = _mm_add_ps(felec,fvdw);
362 fscal = _mm_andnot_ps(dummy_mask,fscal);
364 /* Update vectorial force */
365 fix0 = _mm_macc_ps(dx00,fscal,fix0);
366 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
367 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
369 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
370 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
371 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
372 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
373 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
374 _mm_mul_ps(dx00,fscal),
375 _mm_mul_ps(dy00,fscal),
376 _mm_mul_ps(dz00,fscal));
378 /* Inner loop uses 60 flops */
381 /* End of innermost loop */
383 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
384 f+i_coord_offset,fshift+i_shift_offset);
387 /* Update potential energies */
388 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
389 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
391 /* Increment number of inner iterations */
392 inneriter += j_index_end - j_index_start;
394 /* Outer loop uses 9 flops */
397 /* Increment number of outer iterations */
400 /* Update outer/inner flops */
402 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*60);
405 * Gromacs nonbonded kernel: nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_128_fma_single
406 * Electrostatics interaction: CubicSplineTable
407 * VdW interaction: LennardJones
408 * Geometry: Particle-Particle
409 * Calculate force/pot: Force
412 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_128_fma_single
413 (t_nblist * gmx_restrict nlist,
414 rvec * gmx_restrict xx,
415 rvec * gmx_restrict ff,
416 t_forcerec * gmx_restrict fr,
417 t_mdatoms * gmx_restrict mdatoms,
418 nb_kernel_data_t * gmx_restrict kernel_data,
419 t_nrnb * gmx_restrict nrnb)
421 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
422 * just 0 for non-waters.
423 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
424 * jnr indices corresponding to data put in the four positions in the SIMD register.
426 int i_shift_offset,i_coord_offset,outeriter,inneriter;
427 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
428 int jnrA,jnrB,jnrC,jnrD;
429 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
430 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
431 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
433 real *shiftvec,*fshift,*x,*f;
434 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
436 __m128 fscal,rcutoff,rcutoff2,jidxall;
438 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
439 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
440 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
441 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
442 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
445 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
448 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
449 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
451 __m128i ifour = _mm_set1_epi32(4);
452 __m128 rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
454 __m128 dummy_mask,cutoff_mask;
455 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
456 __m128 one = _mm_set1_ps(1.0);
457 __m128 two = _mm_set1_ps(2.0);
463 jindex = nlist->jindex;
465 shiftidx = nlist->shift;
467 shiftvec = fr->shift_vec[0];
468 fshift = fr->fshift[0];
469 facel = _mm_set1_ps(fr->epsfac);
470 charge = mdatoms->chargeA;
471 nvdwtype = fr->ntype;
473 vdwtype = mdatoms->typeA;
475 vftab = kernel_data->table_elec->data;
476 vftabscale = _mm_set1_ps(kernel_data->table_elec->scale);
478 /* Avoid stupid compiler warnings */
479 jnrA = jnrB = jnrC = jnrD = 0;
488 for(iidx=0;iidx<4*DIM;iidx++)
493 /* Start outer loop over neighborlists */
494 for(iidx=0; iidx<nri; iidx++)
496 /* Load shift vector for this list */
497 i_shift_offset = DIM*shiftidx[iidx];
499 /* Load limits for loop over neighbors */
500 j_index_start = jindex[iidx];
501 j_index_end = jindex[iidx+1];
503 /* Get outer coordinate index */
505 i_coord_offset = DIM*inr;
507 /* Load i particle coords and add shift vector */
508 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
510 fix0 = _mm_setzero_ps();
511 fiy0 = _mm_setzero_ps();
512 fiz0 = _mm_setzero_ps();
514 /* Load parameters for i particles */
515 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
516 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
518 /* Start inner kernel loop */
519 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
522 /* Get j neighbor index, and coordinate index */
527 j_coord_offsetA = DIM*jnrA;
528 j_coord_offsetB = DIM*jnrB;
529 j_coord_offsetC = DIM*jnrC;
530 j_coord_offsetD = DIM*jnrD;
532 /* load j atom coordinates */
533 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
534 x+j_coord_offsetC,x+j_coord_offsetD,
537 /* Calculate displacement vector */
538 dx00 = _mm_sub_ps(ix0,jx0);
539 dy00 = _mm_sub_ps(iy0,jy0);
540 dz00 = _mm_sub_ps(iz0,jz0);
542 /* Calculate squared distance and things based on it */
543 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
545 rinv00 = gmx_mm_invsqrt_ps(rsq00);
547 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
549 /* Load parameters for j particles */
550 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
551 charge+jnrC+0,charge+jnrD+0);
552 vdwjidx0A = 2*vdwtype[jnrA+0];
553 vdwjidx0B = 2*vdwtype[jnrB+0];
554 vdwjidx0C = 2*vdwtype[jnrC+0];
555 vdwjidx0D = 2*vdwtype[jnrD+0];
557 /**************************
558 * CALCULATE INTERACTIONS *
559 **************************/
561 r00 = _mm_mul_ps(rsq00,rinv00);
563 /* Compute parameters for interactions between i and j atoms */
564 qq00 = _mm_mul_ps(iq0,jq0);
565 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
566 vdwparam+vdwioffset0+vdwjidx0B,
567 vdwparam+vdwioffset0+vdwjidx0C,
568 vdwparam+vdwioffset0+vdwjidx0D,
571 /* Calculate table index by multiplying r with table scale and truncate to integer */
572 rt = _mm_mul_ps(r00,vftabscale);
573 vfitab = _mm_cvttps_epi32(rt);
575 vfeps = _mm_frcz_ps(rt);
577 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
579 twovfeps = _mm_add_ps(vfeps,vfeps);
580 vfitab = _mm_slli_epi32(vfitab,2);
582 /* CUBIC SPLINE TABLE ELECTROSTATICS */
583 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
584 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
585 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
586 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
587 _MM_TRANSPOSE4_PS(Y,F,G,H);
588 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
589 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
590 felec = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
592 /* LENNARD-JONES DISPERSION/REPULSION */
594 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
595 fvdw = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
597 fscal = _mm_add_ps(felec,fvdw);
599 /* Update vectorial force */
600 fix0 = _mm_macc_ps(dx00,fscal,fix0);
601 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
602 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
604 fjptrA = f+j_coord_offsetA;
605 fjptrB = f+j_coord_offsetB;
606 fjptrC = f+j_coord_offsetC;
607 fjptrD = f+j_coord_offsetD;
608 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
609 _mm_mul_ps(dx00,fscal),
610 _mm_mul_ps(dy00,fscal),
611 _mm_mul_ps(dz00,fscal));
613 /* Inner loop uses 50 flops */
619 /* Get j neighbor index, and coordinate index */
620 jnrlistA = jjnr[jidx];
621 jnrlistB = jjnr[jidx+1];
622 jnrlistC = jjnr[jidx+2];
623 jnrlistD = jjnr[jidx+3];
624 /* Sign of each element will be negative for non-real atoms.
625 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
626 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
628 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
629 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
630 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
631 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
632 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
633 j_coord_offsetA = DIM*jnrA;
634 j_coord_offsetB = DIM*jnrB;
635 j_coord_offsetC = DIM*jnrC;
636 j_coord_offsetD = DIM*jnrD;
638 /* load j atom coordinates */
639 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
640 x+j_coord_offsetC,x+j_coord_offsetD,
643 /* Calculate displacement vector */
644 dx00 = _mm_sub_ps(ix0,jx0);
645 dy00 = _mm_sub_ps(iy0,jy0);
646 dz00 = _mm_sub_ps(iz0,jz0);
648 /* Calculate squared distance and things based on it */
649 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
651 rinv00 = gmx_mm_invsqrt_ps(rsq00);
653 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
655 /* Load parameters for j particles */
656 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
657 charge+jnrC+0,charge+jnrD+0);
658 vdwjidx0A = 2*vdwtype[jnrA+0];
659 vdwjidx0B = 2*vdwtype[jnrB+0];
660 vdwjidx0C = 2*vdwtype[jnrC+0];
661 vdwjidx0D = 2*vdwtype[jnrD+0];
663 /**************************
664 * CALCULATE INTERACTIONS *
665 **************************/
667 r00 = _mm_mul_ps(rsq00,rinv00);
668 r00 = _mm_andnot_ps(dummy_mask,r00);
670 /* Compute parameters for interactions between i and j atoms */
671 qq00 = _mm_mul_ps(iq0,jq0);
672 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
673 vdwparam+vdwioffset0+vdwjidx0B,
674 vdwparam+vdwioffset0+vdwjidx0C,
675 vdwparam+vdwioffset0+vdwjidx0D,
678 /* Calculate table index by multiplying r with table scale and truncate to integer */
679 rt = _mm_mul_ps(r00,vftabscale);
680 vfitab = _mm_cvttps_epi32(rt);
682 vfeps = _mm_frcz_ps(rt);
684 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
686 twovfeps = _mm_add_ps(vfeps,vfeps);
687 vfitab = _mm_slli_epi32(vfitab,2);
689 /* CUBIC SPLINE TABLE ELECTROSTATICS */
690 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
691 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
692 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
693 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
694 _MM_TRANSPOSE4_PS(Y,F,G,H);
695 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
696 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
697 felec = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
699 /* LENNARD-JONES DISPERSION/REPULSION */
701 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
702 fvdw = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
704 fscal = _mm_add_ps(felec,fvdw);
706 fscal = _mm_andnot_ps(dummy_mask,fscal);
708 /* Update vectorial force */
709 fix0 = _mm_macc_ps(dx00,fscal,fix0);
710 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
711 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
713 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
714 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
715 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
716 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
717 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
718 _mm_mul_ps(dx00,fscal),
719 _mm_mul_ps(dy00,fscal),
720 _mm_mul_ps(dz00,fscal));
722 /* Inner loop uses 51 flops */
725 /* End of innermost loop */
727 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
728 f+i_coord_offset,fshift+i_shift_offset);
730 /* Increment number of inner iterations */
731 inneriter += j_index_end - j_index_start;
733 /* Outer loop uses 7 flops */
736 /* Increment number of outer iterations */
739 /* Update outer/inner flops */
741 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*51);