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_VdwCSTab_GeomW3P1_VF_avx_128_fma_single
38 * Electrostatics interaction: Ewald
39 * VdW interaction: CubicSplineTable
40 * Geometry: Water3-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEw_VdwCSTab_GeomW3P1_VF_avx_128_fma_single
45 (t_nblist * gmx_restrict nlist,
46 rvec * gmx_restrict xx,
47 rvec * gmx_restrict ff,
48 t_forcerec * gmx_restrict fr,
49 t_mdatoms * gmx_restrict mdatoms,
50 nb_kernel_data_t * gmx_restrict kernel_data,
51 t_nrnb * gmx_restrict nrnb)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
62 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
63 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
65 real *shiftvec,*fshift,*x,*f;
66 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
68 __m128 fscal,rcutoff,rcutoff2,jidxall;
70 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
72 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
74 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
75 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
76 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
77 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
78 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
79 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
80 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
83 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
86 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
87 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
89 __m128i ifour = _mm_set1_epi32(4);
90 __m128 rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
93 __m128 ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
94 __m128 beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
96 __m128 dummy_mask,cutoff_mask;
97 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
98 __m128 one = _mm_set1_ps(1.0);
99 __m128 two = _mm_set1_ps(2.0);
105 jindex = nlist->jindex;
107 shiftidx = nlist->shift;
109 shiftvec = fr->shift_vec[0];
110 fshift = fr->fshift[0];
111 facel = _mm_set1_ps(fr->epsfac);
112 charge = mdatoms->chargeA;
113 nvdwtype = fr->ntype;
115 vdwtype = mdatoms->typeA;
117 vftab = kernel_data->table_vdw->data;
118 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
120 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
121 beta = _mm_set1_ps(fr->ic->ewaldcoeff);
122 beta2 = _mm_mul_ps(beta,beta);
123 beta3 = _mm_mul_ps(beta,beta2);
124 ewtab = fr->ic->tabq_coul_FDV0;
125 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
126 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
128 /* Setup water-specific parameters */
129 inr = nlist->iinr[0];
130 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
131 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
132 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
133 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
135 /* Avoid stupid compiler warnings */
136 jnrA = jnrB = jnrC = jnrD = 0;
145 for(iidx=0;iidx<4*DIM;iidx++)
150 /* Start outer loop over neighborlists */
151 for(iidx=0; iidx<nri; iidx++)
153 /* Load shift vector for this list */
154 i_shift_offset = DIM*shiftidx[iidx];
156 /* Load limits for loop over neighbors */
157 j_index_start = jindex[iidx];
158 j_index_end = jindex[iidx+1];
160 /* Get outer coordinate index */
162 i_coord_offset = DIM*inr;
164 /* Load i particle coords and add shift vector */
165 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
166 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
168 fix0 = _mm_setzero_ps();
169 fiy0 = _mm_setzero_ps();
170 fiz0 = _mm_setzero_ps();
171 fix1 = _mm_setzero_ps();
172 fiy1 = _mm_setzero_ps();
173 fiz1 = _mm_setzero_ps();
174 fix2 = _mm_setzero_ps();
175 fiy2 = _mm_setzero_ps();
176 fiz2 = _mm_setzero_ps();
178 /* Reset potential sums */
179 velecsum = _mm_setzero_ps();
180 vvdwsum = _mm_setzero_ps();
182 /* Start inner kernel loop */
183 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
186 /* Get j neighbor index, and coordinate index */
191 j_coord_offsetA = DIM*jnrA;
192 j_coord_offsetB = DIM*jnrB;
193 j_coord_offsetC = DIM*jnrC;
194 j_coord_offsetD = DIM*jnrD;
196 /* load j atom coordinates */
197 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
198 x+j_coord_offsetC,x+j_coord_offsetD,
201 /* Calculate displacement vector */
202 dx00 = _mm_sub_ps(ix0,jx0);
203 dy00 = _mm_sub_ps(iy0,jy0);
204 dz00 = _mm_sub_ps(iz0,jz0);
205 dx10 = _mm_sub_ps(ix1,jx0);
206 dy10 = _mm_sub_ps(iy1,jy0);
207 dz10 = _mm_sub_ps(iz1,jz0);
208 dx20 = _mm_sub_ps(ix2,jx0);
209 dy20 = _mm_sub_ps(iy2,jy0);
210 dz20 = _mm_sub_ps(iz2,jz0);
212 /* Calculate squared distance and things based on it */
213 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
214 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
215 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
217 rinv00 = gmx_mm_invsqrt_ps(rsq00);
218 rinv10 = gmx_mm_invsqrt_ps(rsq10);
219 rinv20 = gmx_mm_invsqrt_ps(rsq20);
221 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
222 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
223 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
225 /* Load parameters for j particles */
226 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
227 charge+jnrC+0,charge+jnrD+0);
228 vdwjidx0A = 2*vdwtype[jnrA+0];
229 vdwjidx0B = 2*vdwtype[jnrB+0];
230 vdwjidx0C = 2*vdwtype[jnrC+0];
231 vdwjidx0D = 2*vdwtype[jnrD+0];
233 fjx0 = _mm_setzero_ps();
234 fjy0 = _mm_setzero_ps();
235 fjz0 = _mm_setzero_ps();
237 /**************************
238 * CALCULATE INTERACTIONS *
239 **************************/
241 r00 = _mm_mul_ps(rsq00,rinv00);
243 /* Compute parameters for interactions between i and j atoms */
244 qq00 = _mm_mul_ps(iq0,jq0);
245 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
246 vdwparam+vdwioffset0+vdwjidx0B,
247 vdwparam+vdwioffset0+vdwjidx0C,
248 vdwparam+vdwioffset0+vdwjidx0D,
251 /* Calculate table index by multiplying r with table scale and truncate to integer */
252 rt = _mm_mul_ps(r00,vftabscale);
253 vfitab = _mm_cvttps_epi32(rt);
255 vfeps = _mm_frcz_ps(rt);
257 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
259 twovfeps = _mm_add_ps(vfeps,vfeps);
260 vfitab = _mm_slli_epi32(vfitab,3);
262 /* EWALD ELECTROSTATICS */
264 /* Analytical PME correction */
265 zeta2 = _mm_mul_ps(beta2,rsq00);
266 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
267 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
268 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
269 felec = _mm_mul_ps(qq00,felec);
270 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
271 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
272 velec = _mm_mul_ps(qq00,velec);
274 /* CUBIC SPLINE TABLE DISPERSION */
275 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
276 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
277 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
278 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
279 _MM_TRANSPOSE4_PS(Y,F,G,H);
280 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
281 VV = _mm_macc_ps(vfeps,Fp,Y);
282 vvdw6 = _mm_mul_ps(c6_00,VV);
283 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
284 fvdw6 = _mm_mul_ps(c6_00,FF);
286 /* CUBIC SPLINE TABLE REPULSION */
287 vfitab = _mm_add_epi32(vfitab,ifour);
288 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
289 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
290 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
291 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
292 _MM_TRANSPOSE4_PS(Y,F,G,H);
293 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
294 VV = _mm_macc_ps(vfeps,Fp,Y);
295 vvdw12 = _mm_mul_ps(c12_00,VV);
296 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
297 fvdw12 = _mm_mul_ps(c12_00,FF);
298 vvdw = _mm_add_ps(vvdw12,vvdw6);
299 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
301 /* Update potential sum for this i atom from the interaction with this j atom. */
302 velecsum = _mm_add_ps(velecsum,velec);
303 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
305 fscal = _mm_add_ps(felec,fvdw);
307 /* Update vectorial force */
308 fix0 = _mm_macc_ps(dx00,fscal,fix0);
309 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
310 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
312 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
313 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
314 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
316 /**************************
317 * CALCULATE INTERACTIONS *
318 **************************/
320 r10 = _mm_mul_ps(rsq10,rinv10);
322 /* Compute parameters for interactions between i and j atoms */
323 qq10 = _mm_mul_ps(iq1,jq0);
325 /* EWALD ELECTROSTATICS */
327 /* Analytical PME correction */
328 zeta2 = _mm_mul_ps(beta2,rsq10);
329 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
330 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
331 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
332 felec = _mm_mul_ps(qq10,felec);
333 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
334 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
335 velec = _mm_mul_ps(qq10,velec);
337 /* Update potential sum for this i atom from the interaction with this j atom. */
338 velecsum = _mm_add_ps(velecsum,velec);
342 /* Update vectorial force */
343 fix1 = _mm_macc_ps(dx10,fscal,fix1);
344 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
345 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
347 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
348 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
349 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
351 /**************************
352 * CALCULATE INTERACTIONS *
353 **************************/
355 r20 = _mm_mul_ps(rsq20,rinv20);
357 /* Compute parameters for interactions between i and j atoms */
358 qq20 = _mm_mul_ps(iq2,jq0);
360 /* EWALD ELECTROSTATICS */
362 /* Analytical PME correction */
363 zeta2 = _mm_mul_ps(beta2,rsq20);
364 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
365 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
366 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
367 felec = _mm_mul_ps(qq20,felec);
368 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
369 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
370 velec = _mm_mul_ps(qq20,velec);
372 /* Update potential sum for this i atom from the interaction with this j atom. */
373 velecsum = _mm_add_ps(velecsum,velec);
377 /* Update vectorial force */
378 fix2 = _mm_macc_ps(dx20,fscal,fix2);
379 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
380 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
382 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
383 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
384 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
386 fjptrA = f+j_coord_offsetA;
387 fjptrB = f+j_coord_offsetB;
388 fjptrC = f+j_coord_offsetC;
389 fjptrD = f+j_coord_offsetD;
391 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
393 /* Inner loop uses 121 flops */
399 /* Get j neighbor index, and coordinate index */
400 jnrlistA = jjnr[jidx];
401 jnrlistB = jjnr[jidx+1];
402 jnrlistC = jjnr[jidx+2];
403 jnrlistD = jjnr[jidx+3];
404 /* Sign of each element will be negative for non-real atoms.
405 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
406 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
408 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
409 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
410 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
411 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
412 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
413 j_coord_offsetA = DIM*jnrA;
414 j_coord_offsetB = DIM*jnrB;
415 j_coord_offsetC = DIM*jnrC;
416 j_coord_offsetD = DIM*jnrD;
418 /* load j atom coordinates */
419 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
420 x+j_coord_offsetC,x+j_coord_offsetD,
423 /* Calculate displacement vector */
424 dx00 = _mm_sub_ps(ix0,jx0);
425 dy00 = _mm_sub_ps(iy0,jy0);
426 dz00 = _mm_sub_ps(iz0,jz0);
427 dx10 = _mm_sub_ps(ix1,jx0);
428 dy10 = _mm_sub_ps(iy1,jy0);
429 dz10 = _mm_sub_ps(iz1,jz0);
430 dx20 = _mm_sub_ps(ix2,jx0);
431 dy20 = _mm_sub_ps(iy2,jy0);
432 dz20 = _mm_sub_ps(iz2,jz0);
434 /* Calculate squared distance and things based on it */
435 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
436 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
437 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
439 rinv00 = gmx_mm_invsqrt_ps(rsq00);
440 rinv10 = gmx_mm_invsqrt_ps(rsq10);
441 rinv20 = gmx_mm_invsqrt_ps(rsq20);
443 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
444 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
445 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
447 /* Load parameters for j particles */
448 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
449 charge+jnrC+0,charge+jnrD+0);
450 vdwjidx0A = 2*vdwtype[jnrA+0];
451 vdwjidx0B = 2*vdwtype[jnrB+0];
452 vdwjidx0C = 2*vdwtype[jnrC+0];
453 vdwjidx0D = 2*vdwtype[jnrD+0];
455 fjx0 = _mm_setzero_ps();
456 fjy0 = _mm_setzero_ps();
457 fjz0 = _mm_setzero_ps();
459 /**************************
460 * CALCULATE INTERACTIONS *
461 **************************/
463 r00 = _mm_mul_ps(rsq00,rinv00);
464 r00 = _mm_andnot_ps(dummy_mask,r00);
466 /* Compute parameters for interactions between i and j atoms */
467 qq00 = _mm_mul_ps(iq0,jq0);
468 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
469 vdwparam+vdwioffset0+vdwjidx0B,
470 vdwparam+vdwioffset0+vdwjidx0C,
471 vdwparam+vdwioffset0+vdwjidx0D,
474 /* Calculate table index by multiplying r with table scale and truncate to integer */
475 rt = _mm_mul_ps(r00,vftabscale);
476 vfitab = _mm_cvttps_epi32(rt);
478 vfeps = _mm_frcz_ps(rt);
480 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
482 twovfeps = _mm_add_ps(vfeps,vfeps);
483 vfitab = _mm_slli_epi32(vfitab,3);
485 /* EWALD ELECTROSTATICS */
487 /* Analytical PME correction */
488 zeta2 = _mm_mul_ps(beta2,rsq00);
489 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
490 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
491 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
492 felec = _mm_mul_ps(qq00,felec);
493 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
494 velec = _mm_nmacc_ps(pmecorrV,beta,rinv00);
495 velec = _mm_mul_ps(qq00,velec);
497 /* CUBIC SPLINE TABLE DISPERSION */
498 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
499 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
500 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
501 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
502 _MM_TRANSPOSE4_PS(Y,F,G,H);
503 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
504 VV = _mm_macc_ps(vfeps,Fp,Y);
505 vvdw6 = _mm_mul_ps(c6_00,VV);
506 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
507 fvdw6 = _mm_mul_ps(c6_00,FF);
509 /* CUBIC SPLINE TABLE REPULSION */
510 vfitab = _mm_add_epi32(vfitab,ifour);
511 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
512 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
513 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
514 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
515 _MM_TRANSPOSE4_PS(Y,F,G,H);
516 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
517 VV = _mm_macc_ps(vfeps,Fp,Y);
518 vvdw12 = _mm_mul_ps(c12_00,VV);
519 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
520 fvdw12 = _mm_mul_ps(c12_00,FF);
521 vvdw = _mm_add_ps(vvdw12,vvdw6);
522 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
524 /* Update potential sum for this i atom from the interaction with this j atom. */
525 velec = _mm_andnot_ps(dummy_mask,velec);
526 velecsum = _mm_add_ps(velecsum,velec);
527 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
528 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
530 fscal = _mm_add_ps(felec,fvdw);
532 fscal = _mm_andnot_ps(dummy_mask,fscal);
534 /* Update vectorial force */
535 fix0 = _mm_macc_ps(dx00,fscal,fix0);
536 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
537 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
539 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
540 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
541 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
543 /**************************
544 * CALCULATE INTERACTIONS *
545 **************************/
547 r10 = _mm_mul_ps(rsq10,rinv10);
548 r10 = _mm_andnot_ps(dummy_mask,r10);
550 /* Compute parameters for interactions between i and j atoms */
551 qq10 = _mm_mul_ps(iq1,jq0);
553 /* EWALD ELECTROSTATICS */
555 /* Analytical PME correction */
556 zeta2 = _mm_mul_ps(beta2,rsq10);
557 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
558 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
559 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
560 felec = _mm_mul_ps(qq10,felec);
561 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
562 velec = _mm_nmacc_ps(pmecorrV,beta,rinv10);
563 velec = _mm_mul_ps(qq10,velec);
565 /* Update potential sum for this i atom from the interaction with this j atom. */
566 velec = _mm_andnot_ps(dummy_mask,velec);
567 velecsum = _mm_add_ps(velecsum,velec);
571 fscal = _mm_andnot_ps(dummy_mask,fscal);
573 /* Update vectorial force */
574 fix1 = _mm_macc_ps(dx10,fscal,fix1);
575 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
576 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
578 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
579 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
580 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
582 /**************************
583 * CALCULATE INTERACTIONS *
584 **************************/
586 r20 = _mm_mul_ps(rsq20,rinv20);
587 r20 = _mm_andnot_ps(dummy_mask,r20);
589 /* Compute parameters for interactions between i and j atoms */
590 qq20 = _mm_mul_ps(iq2,jq0);
592 /* EWALD ELECTROSTATICS */
594 /* Analytical PME correction */
595 zeta2 = _mm_mul_ps(beta2,rsq20);
596 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
597 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
598 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
599 felec = _mm_mul_ps(qq20,felec);
600 pmecorrV = gmx_mm_pmecorrV_ps(zeta2);
601 velec = _mm_nmacc_ps(pmecorrV,beta,rinv20);
602 velec = _mm_mul_ps(qq20,velec);
604 /* Update potential sum for this i atom from the interaction with this j atom. */
605 velec = _mm_andnot_ps(dummy_mask,velec);
606 velecsum = _mm_add_ps(velecsum,velec);
610 fscal = _mm_andnot_ps(dummy_mask,fscal);
612 /* Update vectorial force */
613 fix2 = _mm_macc_ps(dx20,fscal,fix2);
614 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
615 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
617 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
618 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
619 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
621 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
622 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
623 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
624 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
626 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
628 /* Inner loop uses 124 flops */
631 /* End of innermost loop */
633 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
634 f+i_coord_offset,fshift+i_shift_offset);
637 /* Update potential energies */
638 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
639 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
641 /* Increment number of inner iterations */
642 inneriter += j_index_end - j_index_start;
644 /* Outer loop uses 20 flops */
647 /* Increment number of outer iterations */
650 /* Update outer/inner flops */
652 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_VF,outeriter*20 + inneriter*124);
655 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwCSTab_GeomW3P1_F_avx_128_fma_single
656 * Electrostatics interaction: Ewald
657 * VdW interaction: CubicSplineTable
658 * Geometry: Water3-Particle
659 * Calculate force/pot: Force
662 nb_kernel_ElecEw_VdwCSTab_GeomW3P1_F_avx_128_fma_single
663 (t_nblist * gmx_restrict nlist,
664 rvec * gmx_restrict xx,
665 rvec * gmx_restrict ff,
666 t_forcerec * gmx_restrict fr,
667 t_mdatoms * gmx_restrict mdatoms,
668 nb_kernel_data_t * gmx_restrict kernel_data,
669 t_nrnb * gmx_restrict nrnb)
671 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
672 * just 0 for non-waters.
673 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
674 * jnr indices corresponding to data put in the four positions in the SIMD register.
676 int i_shift_offset,i_coord_offset,outeriter,inneriter;
677 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
678 int jnrA,jnrB,jnrC,jnrD;
679 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
680 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
681 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
683 real *shiftvec,*fshift,*x,*f;
684 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
686 __m128 fscal,rcutoff,rcutoff2,jidxall;
688 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
690 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
692 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
693 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
694 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
695 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
696 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
697 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
698 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
701 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
704 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
705 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
707 __m128i ifour = _mm_set1_epi32(4);
708 __m128 rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
711 __m128 ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
712 __m128 beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
714 __m128 dummy_mask,cutoff_mask;
715 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
716 __m128 one = _mm_set1_ps(1.0);
717 __m128 two = _mm_set1_ps(2.0);
723 jindex = nlist->jindex;
725 shiftidx = nlist->shift;
727 shiftvec = fr->shift_vec[0];
728 fshift = fr->fshift[0];
729 facel = _mm_set1_ps(fr->epsfac);
730 charge = mdatoms->chargeA;
731 nvdwtype = fr->ntype;
733 vdwtype = mdatoms->typeA;
735 vftab = kernel_data->table_vdw->data;
736 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
738 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
739 beta = _mm_set1_ps(fr->ic->ewaldcoeff);
740 beta2 = _mm_mul_ps(beta,beta);
741 beta3 = _mm_mul_ps(beta,beta2);
742 ewtab = fr->ic->tabq_coul_F;
743 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
744 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
746 /* Setup water-specific parameters */
747 inr = nlist->iinr[0];
748 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
749 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
750 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
751 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
753 /* Avoid stupid compiler warnings */
754 jnrA = jnrB = jnrC = jnrD = 0;
763 for(iidx=0;iidx<4*DIM;iidx++)
768 /* Start outer loop over neighborlists */
769 for(iidx=0; iidx<nri; iidx++)
771 /* Load shift vector for this list */
772 i_shift_offset = DIM*shiftidx[iidx];
774 /* Load limits for loop over neighbors */
775 j_index_start = jindex[iidx];
776 j_index_end = jindex[iidx+1];
778 /* Get outer coordinate index */
780 i_coord_offset = DIM*inr;
782 /* Load i particle coords and add shift vector */
783 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
784 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
786 fix0 = _mm_setzero_ps();
787 fiy0 = _mm_setzero_ps();
788 fiz0 = _mm_setzero_ps();
789 fix1 = _mm_setzero_ps();
790 fiy1 = _mm_setzero_ps();
791 fiz1 = _mm_setzero_ps();
792 fix2 = _mm_setzero_ps();
793 fiy2 = _mm_setzero_ps();
794 fiz2 = _mm_setzero_ps();
796 /* Start inner kernel loop */
797 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
800 /* Get j neighbor index, and coordinate index */
805 j_coord_offsetA = DIM*jnrA;
806 j_coord_offsetB = DIM*jnrB;
807 j_coord_offsetC = DIM*jnrC;
808 j_coord_offsetD = DIM*jnrD;
810 /* load j atom coordinates */
811 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
812 x+j_coord_offsetC,x+j_coord_offsetD,
815 /* Calculate displacement vector */
816 dx00 = _mm_sub_ps(ix0,jx0);
817 dy00 = _mm_sub_ps(iy0,jy0);
818 dz00 = _mm_sub_ps(iz0,jz0);
819 dx10 = _mm_sub_ps(ix1,jx0);
820 dy10 = _mm_sub_ps(iy1,jy0);
821 dz10 = _mm_sub_ps(iz1,jz0);
822 dx20 = _mm_sub_ps(ix2,jx0);
823 dy20 = _mm_sub_ps(iy2,jy0);
824 dz20 = _mm_sub_ps(iz2,jz0);
826 /* Calculate squared distance and things based on it */
827 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
828 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
829 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
831 rinv00 = gmx_mm_invsqrt_ps(rsq00);
832 rinv10 = gmx_mm_invsqrt_ps(rsq10);
833 rinv20 = gmx_mm_invsqrt_ps(rsq20);
835 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
836 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
837 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
839 /* Load parameters for j particles */
840 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
841 charge+jnrC+0,charge+jnrD+0);
842 vdwjidx0A = 2*vdwtype[jnrA+0];
843 vdwjidx0B = 2*vdwtype[jnrB+0];
844 vdwjidx0C = 2*vdwtype[jnrC+0];
845 vdwjidx0D = 2*vdwtype[jnrD+0];
847 fjx0 = _mm_setzero_ps();
848 fjy0 = _mm_setzero_ps();
849 fjz0 = _mm_setzero_ps();
851 /**************************
852 * CALCULATE INTERACTIONS *
853 **************************/
855 r00 = _mm_mul_ps(rsq00,rinv00);
857 /* Compute parameters for interactions between i and j atoms */
858 qq00 = _mm_mul_ps(iq0,jq0);
859 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
860 vdwparam+vdwioffset0+vdwjidx0B,
861 vdwparam+vdwioffset0+vdwjidx0C,
862 vdwparam+vdwioffset0+vdwjidx0D,
865 /* Calculate table index by multiplying r with table scale and truncate to integer */
866 rt = _mm_mul_ps(r00,vftabscale);
867 vfitab = _mm_cvttps_epi32(rt);
869 vfeps = _mm_frcz_ps(rt);
871 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
873 twovfeps = _mm_add_ps(vfeps,vfeps);
874 vfitab = _mm_slli_epi32(vfitab,3);
876 /* EWALD ELECTROSTATICS */
878 /* Analytical PME correction */
879 zeta2 = _mm_mul_ps(beta2,rsq00);
880 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
881 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
882 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
883 felec = _mm_mul_ps(qq00,felec);
885 /* CUBIC SPLINE TABLE DISPERSION */
886 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
887 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
888 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
889 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
890 _MM_TRANSPOSE4_PS(Y,F,G,H);
891 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
892 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
893 fvdw6 = _mm_mul_ps(c6_00,FF);
895 /* CUBIC SPLINE TABLE REPULSION */
896 vfitab = _mm_add_epi32(vfitab,ifour);
897 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
898 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
899 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
900 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
901 _MM_TRANSPOSE4_PS(Y,F,G,H);
902 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
903 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
904 fvdw12 = _mm_mul_ps(c12_00,FF);
905 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
907 fscal = _mm_add_ps(felec,fvdw);
909 /* Update vectorial force */
910 fix0 = _mm_macc_ps(dx00,fscal,fix0);
911 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
912 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
914 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
915 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
916 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
918 /**************************
919 * CALCULATE INTERACTIONS *
920 **************************/
922 r10 = _mm_mul_ps(rsq10,rinv10);
924 /* Compute parameters for interactions between i and j atoms */
925 qq10 = _mm_mul_ps(iq1,jq0);
927 /* EWALD ELECTROSTATICS */
929 /* Analytical PME correction */
930 zeta2 = _mm_mul_ps(beta2,rsq10);
931 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
932 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
933 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
934 felec = _mm_mul_ps(qq10,felec);
938 /* Update vectorial force */
939 fix1 = _mm_macc_ps(dx10,fscal,fix1);
940 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
941 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
943 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
944 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
945 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
947 /**************************
948 * CALCULATE INTERACTIONS *
949 **************************/
951 r20 = _mm_mul_ps(rsq20,rinv20);
953 /* Compute parameters for interactions between i and j atoms */
954 qq20 = _mm_mul_ps(iq2,jq0);
956 /* EWALD ELECTROSTATICS */
958 /* Analytical PME correction */
959 zeta2 = _mm_mul_ps(beta2,rsq20);
960 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
961 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
962 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
963 felec = _mm_mul_ps(qq20,felec);
967 /* Update vectorial force */
968 fix2 = _mm_macc_ps(dx20,fscal,fix2);
969 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
970 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
972 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
973 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
974 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
976 fjptrA = f+j_coord_offsetA;
977 fjptrB = f+j_coord_offsetB;
978 fjptrC = f+j_coord_offsetC;
979 fjptrD = f+j_coord_offsetD;
981 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
983 /* Inner loop uses 110 flops */
989 /* Get j neighbor index, and coordinate index */
990 jnrlistA = jjnr[jidx];
991 jnrlistB = jjnr[jidx+1];
992 jnrlistC = jjnr[jidx+2];
993 jnrlistD = jjnr[jidx+3];
994 /* Sign of each element will be negative for non-real atoms.
995 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
996 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
998 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
999 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1000 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1001 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1002 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1003 j_coord_offsetA = DIM*jnrA;
1004 j_coord_offsetB = DIM*jnrB;
1005 j_coord_offsetC = DIM*jnrC;
1006 j_coord_offsetD = DIM*jnrD;
1008 /* load j atom coordinates */
1009 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1010 x+j_coord_offsetC,x+j_coord_offsetD,
1013 /* Calculate displacement vector */
1014 dx00 = _mm_sub_ps(ix0,jx0);
1015 dy00 = _mm_sub_ps(iy0,jy0);
1016 dz00 = _mm_sub_ps(iz0,jz0);
1017 dx10 = _mm_sub_ps(ix1,jx0);
1018 dy10 = _mm_sub_ps(iy1,jy0);
1019 dz10 = _mm_sub_ps(iz1,jz0);
1020 dx20 = _mm_sub_ps(ix2,jx0);
1021 dy20 = _mm_sub_ps(iy2,jy0);
1022 dz20 = _mm_sub_ps(iz2,jz0);
1024 /* Calculate squared distance and things based on it */
1025 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1026 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1027 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1029 rinv00 = gmx_mm_invsqrt_ps(rsq00);
1030 rinv10 = gmx_mm_invsqrt_ps(rsq10);
1031 rinv20 = gmx_mm_invsqrt_ps(rsq20);
1033 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
1034 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
1035 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
1037 /* Load parameters for j particles */
1038 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
1039 charge+jnrC+0,charge+jnrD+0);
1040 vdwjidx0A = 2*vdwtype[jnrA+0];
1041 vdwjidx0B = 2*vdwtype[jnrB+0];
1042 vdwjidx0C = 2*vdwtype[jnrC+0];
1043 vdwjidx0D = 2*vdwtype[jnrD+0];
1045 fjx0 = _mm_setzero_ps();
1046 fjy0 = _mm_setzero_ps();
1047 fjz0 = _mm_setzero_ps();
1049 /**************************
1050 * CALCULATE INTERACTIONS *
1051 **************************/
1053 r00 = _mm_mul_ps(rsq00,rinv00);
1054 r00 = _mm_andnot_ps(dummy_mask,r00);
1056 /* Compute parameters for interactions between i and j atoms */
1057 qq00 = _mm_mul_ps(iq0,jq0);
1058 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
1059 vdwparam+vdwioffset0+vdwjidx0B,
1060 vdwparam+vdwioffset0+vdwjidx0C,
1061 vdwparam+vdwioffset0+vdwjidx0D,
1064 /* Calculate table index by multiplying r with table scale and truncate to integer */
1065 rt = _mm_mul_ps(r00,vftabscale);
1066 vfitab = _mm_cvttps_epi32(rt);
1068 vfeps = _mm_frcz_ps(rt);
1070 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
1072 twovfeps = _mm_add_ps(vfeps,vfeps);
1073 vfitab = _mm_slli_epi32(vfitab,3);
1075 /* EWALD ELECTROSTATICS */
1077 /* Analytical PME correction */
1078 zeta2 = _mm_mul_ps(beta2,rsq00);
1079 rinv3 = _mm_mul_ps(rinvsq00,rinv00);
1080 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1081 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1082 felec = _mm_mul_ps(qq00,felec);
1084 /* CUBIC SPLINE TABLE DISPERSION */
1085 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1086 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1087 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1088 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1089 _MM_TRANSPOSE4_PS(Y,F,G,H);
1090 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1091 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1092 fvdw6 = _mm_mul_ps(c6_00,FF);
1094 /* CUBIC SPLINE TABLE REPULSION */
1095 vfitab = _mm_add_epi32(vfitab,ifour);
1096 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
1097 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
1098 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
1099 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
1100 _MM_TRANSPOSE4_PS(Y,F,G,H);
1101 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
1102 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
1103 fvdw12 = _mm_mul_ps(c12_00,FF);
1104 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1106 fscal = _mm_add_ps(felec,fvdw);
1108 fscal = _mm_andnot_ps(dummy_mask,fscal);
1110 /* Update vectorial force */
1111 fix0 = _mm_macc_ps(dx00,fscal,fix0);
1112 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
1113 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
1115 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
1116 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
1117 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
1119 /**************************
1120 * CALCULATE INTERACTIONS *
1121 **************************/
1123 r10 = _mm_mul_ps(rsq10,rinv10);
1124 r10 = _mm_andnot_ps(dummy_mask,r10);
1126 /* Compute parameters for interactions between i and j atoms */
1127 qq10 = _mm_mul_ps(iq1,jq0);
1129 /* EWALD ELECTROSTATICS */
1131 /* Analytical PME correction */
1132 zeta2 = _mm_mul_ps(beta2,rsq10);
1133 rinv3 = _mm_mul_ps(rinvsq10,rinv10);
1134 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1135 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1136 felec = _mm_mul_ps(qq10,felec);
1140 fscal = _mm_andnot_ps(dummy_mask,fscal);
1142 /* Update vectorial force */
1143 fix1 = _mm_macc_ps(dx10,fscal,fix1);
1144 fiy1 = _mm_macc_ps(dy10,fscal,fiy1);
1145 fiz1 = _mm_macc_ps(dz10,fscal,fiz1);
1147 fjx0 = _mm_macc_ps(dx10,fscal,fjx0);
1148 fjy0 = _mm_macc_ps(dy10,fscal,fjy0);
1149 fjz0 = _mm_macc_ps(dz10,fscal,fjz0);
1151 /**************************
1152 * CALCULATE INTERACTIONS *
1153 **************************/
1155 r20 = _mm_mul_ps(rsq20,rinv20);
1156 r20 = _mm_andnot_ps(dummy_mask,r20);
1158 /* Compute parameters for interactions between i and j atoms */
1159 qq20 = _mm_mul_ps(iq2,jq0);
1161 /* EWALD ELECTROSTATICS */
1163 /* Analytical PME correction */
1164 zeta2 = _mm_mul_ps(beta2,rsq20);
1165 rinv3 = _mm_mul_ps(rinvsq20,rinv20);
1166 pmecorrF = gmx_mm_pmecorrF_ps(zeta2);
1167 felec = _mm_macc_ps(pmecorrF,beta3,rinv3);
1168 felec = _mm_mul_ps(qq20,felec);
1172 fscal = _mm_andnot_ps(dummy_mask,fscal);
1174 /* Update vectorial force */
1175 fix2 = _mm_macc_ps(dx20,fscal,fix2);
1176 fiy2 = _mm_macc_ps(dy20,fscal,fiy2);
1177 fiz2 = _mm_macc_ps(dz20,fscal,fiz2);
1179 fjx0 = _mm_macc_ps(dx20,fscal,fjx0);
1180 fjy0 = _mm_macc_ps(dy20,fscal,fjy0);
1181 fjz0 = _mm_macc_ps(dz20,fscal,fjz0);
1183 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1184 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1185 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1186 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1188 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
1190 /* Inner loop uses 113 flops */
1193 /* End of innermost loop */
1195 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1196 f+i_coord_offset,fshift+i_shift_offset);
1198 /* Increment number of inner iterations */
1199 inneriter += j_index_end - j_index_start;
1201 /* Outer loop uses 18 flops */
1204 /* Increment number of outer iterations */
1207 /* Update outer/inner flops */
1209 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3_F,outeriter*18 + inneriter*113);