2 * Note: this file was generated by the Gromacs avx_128_fma_double 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_double.h"
34 #include "kernelutil_x86_avx_128_fma_double.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_double
38 * Electrostatics interaction: GeneralizedBorn
39 * VdW interaction: CubicSplineTable
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_double
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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
61 int j_coord_offsetA,j_coord_offsetB;
62 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
64 real *shiftvec,*fshift,*x,*f;
65 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
67 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
68 int vdwjidx0A,vdwjidx0B;
69 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
70 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
71 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
74 __m128d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
75 __m128d minushalf = _mm_set1_pd(-0.5);
76 real *invsqrta,*dvda,*gbtab;
78 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
81 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
82 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
84 __m128i ifour = _mm_set1_epi32(4);
85 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
87 __m128d dummy_mask,cutoff_mask;
88 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
89 __m128d one = _mm_set1_pd(1.0);
90 __m128d two = _mm_set1_pd(2.0);
96 jindex = nlist->jindex;
98 shiftidx = nlist->shift;
100 shiftvec = fr->shift_vec[0];
101 fshift = fr->fshift[0];
102 facel = _mm_set1_pd(fr->epsfac);
103 charge = mdatoms->chargeA;
104 nvdwtype = fr->ntype;
106 vdwtype = mdatoms->typeA;
108 vftab = kernel_data->table_vdw->data;
109 vftabscale = _mm_set1_pd(kernel_data->table_vdw->scale);
111 invsqrta = fr->invsqrta;
113 gbtabscale = _mm_set1_pd(fr->gbtab.scale);
114 gbtab = fr->gbtab.data;
115 gbinvepsdiff = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
117 /* Avoid stupid compiler warnings */
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_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
142 fix0 = _mm_setzero_pd();
143 fiy0 = _mm_setzero_pd();
144 fiz0 = _mm_setzero_pd();
146 /* Load parameters for i particles */
147 iq0 = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
148 isai0 = _mm_load1_pd(invsqrta+inr+0);
149 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
151 /* Reset potential sums */
152 velecsum = _mm_setzero_pd();
153 vgbsum = _mm_setzero_pd();
154 vvdwsum = _mm_setzero_pd();
155 dvdasum = _mm_setzero_pd();
157 /* Start inner kernel loop */
158 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
161 /* Get j neighbor index, and coordinate index */
164 j_coord_offsetA = DIM*jnrA;
165 j_coord_offsetB = DIM*jnrB;
167 /* load j atom coordinates */
168 gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
171 /* Calculate displacement vector */
172 dx00 = _mm_sub_pd(ix0,jx0);
173 dy00 = _mm_sub_pd(iy0,jy0);
174 dz00 = _mm_sub_pd(iz0,jz0);
176 /* Calculate squared distance and things based on it */
177 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
179 rinv00 = gmx_mm_invsqrt_pd(rsq00);
181 /* Load parameters for j particles */
182 jq0 = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
183 isaj0 = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
184 vdwjidx0A = 2*vdwtype[jnrA+0];
185 vdwjidx0B = 2*vdwtype[jnrB+0];
187 /**************************
188 * CALCULATE INTERACTIONS *
189 **************************/
191 r00 = _mm_mul_pd(rsq00,rinv00);
193 /* Compute parameters for interactions between i and j atoms */
194 qq00 = _mm_mul_pd(iq0,jq0);
195 gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
196 vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
198 /* Calculate table index by multiplying r with table scale and truncate to integer */
199 rt = _mm_mul_pd(r00,vftabscale);
200 vfitab = _mm_cvttpd_epi32(rt);
202 vfeps = _mm_frcz_pd(rt);
204 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
206 twovfeps = _mm_add_pd(vfeps,vfeps);
207 vfitab = _mm_slli_epi32(vfitab,3);
209 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
210 isaprod = _mm_mul_pd(isai0,isaj0);
211 gbqqfactor = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
212 gbscale = _mm_mul_pd(isaprod,gbtabscale);
214 /* Calculate generalized born table index - this is a separate table from the normal one,
215 * but we use the same procedure by multiplying r with scale and truncating to integer.
217 rt = _mm_mul_pd(r00,gbscale);
218 gbitab = _mm_cvttpd_epi32(rt);
220 gbeps = _mm_frcz_pd(rt);
222 gbeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
224 gbitab = _mm_slli_epi32(gbitab,2);
226 Y = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
227 F = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
228 GMX_MM_TRANSPOSE2_PD(Y,F);
229 G = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
230 H = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
231 GMX_MM_TRANSPOSE2_PD(G,H);
232 Fp = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
233 VV = _mm_macc_pd(gbeps,Fp,Y);
234 vgb = _mm_mul_pd(gbqqfactor,VV);
236 twogbeps = _mm_add_pd(gbeps,gbeps);
237 FF = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
238 fgb = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
239 dvdatmp = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
240 dvdasum = _mm_add_pd(dvdasum,dvdatmp);
241 gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
242 velec = _mm_mul_pd(qq00,rinv00);
243 felec = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
245 /* CUBIC SPLINE TABLE DISPERSION */
246 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
247 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
248 GMX_MM_TRANSPOSE2_PD(Y,F);
249 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
250 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
251 GMX_MM_TRANSPOSE2_PD(G,H);
252 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
253 VV = _mm_macc_pd(vfeps,Fp,Y);
254 vvdw6 = _mm_mul_pd(c6_00,VV);
255 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
256 fvdw6 = _mm_mul_pd(c6_00,FF);
258 /* CUBIC SPLINE TABLE REPULSION */
259 vfitab = _mm_add_epi32(vfitab,ifour);
260 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
261 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
262 GMX_MM_TRANSPOSE2_PD(Y,F);
263 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
264 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
265 GMX_MM_TRANSPOSE2_PD(G,H);
266 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
267 VV = _mm_macc_pd(vfeps,Fp,Y);
268 vvdw12 = _mm_mul_pd(c12_00,VV);
269 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
270 fvdw12 = _mm_mul_pd(c12_00,FF);
271 vvdw = _mm_add_pd(vvdw12,vvdw6);
272 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
274 /* Update potential sum for this i atom from the interaction with this j atom. */
275 velecsum = _mm_add_pd(velecsum,velec);
276 vgbsum = _mm_add_pd(vgbsum,vgb);
277 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
279 fscal = _mm_add_pd(felec,fvdw);
281 /* Update vectorial force */
282 fix0 = _mm_macc_pd(dx00,fscal,fix0);
283 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
284 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
286 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
287 _mm_mul_pd(dx00,fscal),
288 _mm_mul_pd(dy00,fscal),
289 _mm_mul_pd(dz00,fscal));
291 /* Inner loop uses 95 flops */
298 j_coord_offsetA = DIM*jnrA;
300 /* load j atom coordinates */
301 gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
304 /* Calculate displacement vector */
305 dx00 = _mm_sub_pd(ix0,jx0);
306 dy00 = _mm_sub_pd(iy0,jy0);
307 dz00 = _mm_sub_pd(iz0,jz0);
309 /* Calculate squared distance and things based on it */
310 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
312 rinv00 = gmx_mm_invsqrt_pd(rsq00);
314 /* Load parameters for j particles */
315 jq0 = _mm_load_sd(charge+jnrA+0);
316 isaj0 = _mm_load_sd(invsqrta+jnrA+0);
317 vdwjidx0A = 2*vdwtype[jnrA+0];
319 /**************************
320 * CALCULATE INTERACTIONS *
321 **************************/
323 r00 = _mm_mul_pd(rsq00,rinv00);
325 /* Compute parameters for interactions between i and j atoms */
326 qq00 = _mm_mul_pd(iq0,jq0);
327 gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
329 /* Calculate table index by multiplying r with table scale and truncate to integer */
330 rt = _mm_mul_pd(r00,vftabscale);
331 vfitab = _mm_cvttpd_epi32(rt);
333 vfeps = _mm_frcz_pd(rt);
335 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
337 twovfeps = _mm_add_pd(vfeps,vfeps);
338 vfitab = _mm_slli_epi32(vfitab,3);
340 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
341 isaprod = _mm_mul_pd(isai0,isaj0);
342 gbqqfactor = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
343 gbscale = _mm_mul_pd(isaprod,gbtabscale);
345 /* Calculate generalized born table index - this is a separate table from the normal one,
346 * but we use the same procedure by multiplying r with scale and truncating to integer.
348 rt = _mm_mul_pd(r00,gbscale);
349 gbitab = _mm_cvttpd_epi32(rt);
351 gbeps = _mm_frcz_pd(rt);
353 gbeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
355 gbitab = _mm_slli_epi32(gbitab,2);
357 Y = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
358 F = _mm_setzero_pd();
359 GMX_MM_TRANSPOSE2_PD(Y,F);
360 G = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
361 H = _mm_setzero_pd();
362 GMX_MM_TRANSPOSE2_PD(G,H);
363 Fp = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
364 VV = _mm_macc_pd(gbeps,Fp,Y);
365 vgb = _mm_mul_pd(gbqqfactor,VV);
367 twogbeps = _mm_add_pd(gbeps,gbeps);
368 FF = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
369 fgb = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
370 dvdatmp = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
371 dvdasum = _mm_add_pd(dvdasum,dvdatmp);
372 gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
373 velec = _mm_mul_pd(qq00,rinv00);
374 felec = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
376 /* CUBIC SPLINE TABLE DISPERSION */
377 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
378 F = _mm_setzero_pd();
379 GMX_MM_TRANSPOSE2_PD(Y,F);
380 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
381 H = _mm_setzero_pd();
382 GMX_MM_TRANSPOSE2_PD(G,H);
383 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
384 VV = _mm_macc_pd(vfeps,Fp,Y);
385 vvdw6 = _mm_mul_pd(c6_00,VV);
386 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
387 fvdw6 = _mm_mul_pd(c6_00,FF);
389 /* CUBIC SPLINE TABLE REPULSION */
390 vfitab = _mm_add_epi32(vfitab,ifour);
391 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
392 F = _mm_setzero_pd();
393 GMX_MM_TRANSPOSE2_PD(Y,F);
394 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
395 H = _mm_setzero_pd();
396 GMX_MM_TRANSPOSE2_PD(G,H);
397 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
398 VV = _mm_macc_pd(vfeps,Fp,Y);
399 vvdw12 = _mm_mul_pd(c12_00,VV);
400 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
401 fvdw12 = _mm_mul_pd(c12_00,FF);
402 vvdw = _mm_add_pd(vvdw12,vvdw6);
403 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
405 /* Update potential sum for this i atom from the interaction with this j atom. */
406 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
407 velecsum = _mm_add_pd(velecsum,velec);
408 vgb = _mm_unpacklo_pd(vgb,_mm_setzero_pd());
409 vgbsum = _mm_add_pd(vgbsum,vgb);
410 vvdw = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
411 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
413 fscal = _mm_add_pd(felec,fvdw);
415 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
417 /* Update vectorial force */
418 fix0 = _mm_macc_pd(dx00,fscal,fix0);
419 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
420 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
422 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
423 _mm_mul_pd(dx00,fscal),
424 _mm_mul_pd(dy00,fscal),
425 _mm_mul_pd(dz00,fscal));
427 /* Inner loop uses 95 flops */
430 /* End of innermost loop */
432 gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
433 f+i_coord_offset,fshift+i_shift_offset);
436 /* Update potential energies */
437 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
438 gmx_mm_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
439 gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
440 dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
441 gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
443 /* Increment number of inner iterations */
444 inneriter += j_index_end - j_index_start;
446 /* Outer loop uses 10 flops */
449 /* Increment number of outer iterations */
452 /* Update outer/inner flops */
454 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*95);
457 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_double
458 * Electrostatics interaction: GeneralizedBorn
459 * VdW interaction: CubicSplineTable
460 * Geometry: Particle-Particle
461 * Calculate force/pot: Force
464 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_double
465 (t_nblist * gmx_restrict nlist,
466 rvec * gmx_restrict xx,
467 rvec * gmx_restrict ff,
468 t_forcerec * gmx_restrict fr,
469 t_mdatoms * gmx_restrict mdatoms,
470 nb_kernel_data_t * gmx_restrict kernel_data,
471 t_nrnb * gmx_restrict nrnb)
473 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
474 * just 0 for non-waters.
475 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
476 * jnr indices corresponding to data put in the four positions in the SIMD register.
478 int i_shift_offset,i_coord_offset,outeriter,inneriter;
479 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
481 int j_coord_offsetA,j_coord_offsetB;
482 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
484 real *shiftvec,*fshift,*x,*f;
485 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
487 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
488 int vdwjidx0A,vdwjidx0B;
489 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
490 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
491 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
494 __m128d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,twogbeps,dvdatmp;
495 __m128d minushalf = _mm_set1_pd(-0.5);
496 real *invsqrta,*dvda,*gbtab;
498 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
501 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
502 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
504 __m128i ifour = _mm_set1_epi32(4);
505 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
507 __m128d dummy_mask,cutoff_mask;
508 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
509 __m128d one = _mm_set1_pd(1.0);
510 __m128d two = _mm_set1_pd(2.0);
516 jindex = nlist->jindex;
518 shiftidx = nlist->shift;
520 shiftvec = fr->shift_vec[0];
521 fshift = fr->fshift[0];
522 facel = _mm_set1_pd(fr->epsfac);
523 charge = mdatoms->chargeA;
524 nvdwtype = fr->ntype;
526 vdwtype = mdatoms->typeA;
528 vftab = kernel_data->table_vdw->data;
529 vftabscale = _mm_set1_pd(kernel_data->table_vdw->scale);
531 invsqrta = fr->invsqrta;
533 gbtabscale = _mm_set1_pd(fr->gbtab.scale);
534 gbtab = fr->gbtab.data;
535 gbinvepsdiff = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
537 /* Avoid stupid compiler warnings */
545 /* Start outer loop over neighborlists */
546 for(iidx=0; iidx<nri; iidx++)
548 /* Load shift vector for this list */
549 i_shift_offset = DIM*shiftidx[iidx];
551 /* Load limits for loop over neighbors */
552 j_index_start = jindex[iidx];
553 j_index_end = jindex[iidx+1];
555 /* Get outer coordinate index */
557 i_coord_offset = DIM*inr;
559 /* Load i particle coords and add shift vector */
560 gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
562 fix0 = _mm_setzero_pd();
563 fiy0 = _mm_setzero_pd();
564 fiz0 = _mm_setzero_pd();
566 /* Load parameters for i particles */
567 iq0 = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
568 isai0 = _mm_load1_pd(invsqrta+inr+0);
569 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
571 dvdasum = _mm_setzero_pd();
573 /* Start inner kernel loop */
574 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
577 /* Get j neighbor index, and coordinate index */
580 j_coord_offsetA = DIM*jnrA;
581 j_coord_offsetB = DIM*jnrB;
583 /* load j atom coordinates */
584 gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
587 /* Calculate displacement vector */
588 dx00 = _mm_sub_pd(ix0,jx0);
589 dy00 = _mm_sub_pd(iy0,jy0);
590 dz00 = _mm_sub_pd(iz0,jz0);
592 /* Calculate squared distance and things based on it */
593 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
595 rinv00 = gmx_mm_invsqrt_pd(rsq00);
597 /* Load parameters for j particles */
598 jq0 = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
599 isaj0 = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
600 vdwjidx0A = 2*vdwtype[jnrA+0];
601 vdwjidx0B = 2*vdwtype[jnrB+0];
603 /**************************
604 * CALCULATE INTERACTIONS *
605 **************************/
607 r00 = _mm_mul_pd(rsq00,rinv00);
609 /* Compute parameters for interactions between i and j atoms */
610 qq00 = _mm_mul_pd(iq0,jq0);
611 gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
612 vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
614 /* Calculate table index by multiplying r with table scale and truncate to integer */
615 rt = _mm_mul_pd(r00,vftabscale);
616 vfitab = _mm_cvttpd_epi32(rt);
618 vfeps = _mm_frcz_pd(rt);
620 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
622 twovfeps = _mm_add_pd(vfeps,vfeps);
623 vfitab = _mm_slli_epi32(vfitab,3);
625 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
626 isaprod = _mm_mul_pd(isai0,isaj0);
627 gbqqfactor = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
628 gbscale = _mm_mul_pd(isaprod,gbtabscale);
630 /* Calculate generalized born table index - this is a separate table from the normal one,
631 * but we use the same procedure by multiplying r with scale and truncating to integer.
633 rt = _mm_mul_pd(r00,gbscale);
634 gbitab = _mm_cvttpd_epi32(rt);
636 gbeps = _mm_frcz_pd(rt);
638 gbeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
640 gbitab = _mm_slli_epi32(gbitab,2);
642 Y = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
643 F = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
644 GMX_MM_TRANSPOSE2_PD(Y,F);
645 G = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
646 H = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,1) +2);
647 GMX_MM_TRANSPOSE2_PD(G,H);
648 Fp = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
649 VV = _mm_macc_pd(gbeps,Fp,Y);
650 vgb = _mm_mul_pd(gbqqfactor,VV);
652 twogbeps = _mm_add_pd(gbeps,gbeps);
653 FF = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
654 fgb = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
655 dvdatmp = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
656 dvdasum = _mm_add_pd(dvdasum,dvdatmp);
657 gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
658 velec = _mm_mul_pd(qq00,rinv00);
659 felec = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
661 /* CUBIC SPLINE TABLE DISPERSION */
662 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
663 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
664 GMX_MM_TRANSPOSE2_PD(Y,F);
665 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
666 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
667 GMX_MM_TRANSPOSE2_PD(G,H);
668 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
669 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
670 fvdw6 = _mm_mul_pd(c6_00,FF);
672 /* CUBIC SPLINE TABLE REPULSION */
673 vfitab = _mm_add_epi32(vfitab,ifour);
674 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
675 F = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
676 GMX_MM_TRANSPOSE2_PD(Y,F);
677 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
678 H = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
679 GMX_MM_TRANSPOSE2_PD(G,H);
680 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
681 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
682 fvdw12 = _mm_mul_pd(c12_00,FF);
683 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
685 fscal = _mm_add_pd(felec,fvdw);
687 /* Update vectorial force */
688 fix0 = _mm_macc_pd(dx00,fscal,fix0);
689 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
690 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
692 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
693 _mm_mul_pd(dx00,fscal),
694 _mm_mul_pd(dy00,fscal),
695 _mm_mul_pd(dz00,fscal));
697 /* Inner loop uses 85 flops */
704 j_coord_offsetA = DIM*jnrA;
706 /* load j atom coordinates */
707 gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
710 /* Calculate displacement vector */
711 dx00 = _mm_sub_pd(ix0,jx0);
712 dy00 = _mm_sub_pd(iy0,jy0);
713 dz00 = _mm_sub_pd(iz0,jz0);
715 /* Calculate squared distance and things based on it */
716 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
718 rinv00 = gmx_mm_invsqrt_pd(rsq00);
720 /* Load parameters for j particles */
721 jq0 = _mm_load_sd(charge+jnrA+0);
722 isaj0 = _mm_load_sd(invsqrta+jnrA+0);
723 vdwjidx0A = 2*vdwtype[jnrA+0];
725 /**************************
726 * CALCULATE INTERACTIONS *
727 **************************/
729 r00 = _mm_mul_pd(rsq00,rinv00);
731 /* Compute parameters for interactions between i and j atoms */
732 qq00 = _mm_mul_pd(iq0,jq0);
733 gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
735 /* Calculate table index by multiplying r with table scale and truncate to integer */
736 rt = _mm_mul_pd(r00,vftabscale);
737 vfitab = _mm_cvttpd_epi32(rt);
739 vfeps = _mm_frcz_pd(rt);
741 vfeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
743 twovfeps = _mm_add_pd(vfeps,vfeps);
744 vfitab = _mm_slli_epi32(vfitab,3);
746 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
747 isaprod = _mm_mul_pd(isai0,isaj0);
748 gbqqfactor = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
749 gbscale = _mm_mul_pd(isaprod,gbtabscale);
751 /* Calculate generalized born table index - this is a separate table from the normal one,
752 * but we use the same procedure by multiplying r with scale and truncating to integer.
754 rt = _mm_mul_pd(r00,gbscale);
755 gbitab = _mm_cvttpd_epi32(rt);
757 gbeps = _mm_frcz_pd(rt);
759 gbeps = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
761 gbitab = _mm_slli_epi32(gbitab,2);
763 Y = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
764 F = _mm_setzero_pd();
765 GMX_MM_TRANSPOSE2_PD(Y,F);
766 G = _mm_load_pd( gbtab + _mm_extract_epi32(gbitab,0) +2);
767 H = _mm_setzero_pd();
768 GMX_MM_TRANSPOSE2_PD(G,H);
769 Fp = _mm_macc_pd(gbeps,_mm_macc_pd(gbeps,H,G),F);
770 VV = _mm_macc_pd(gbeps,Fp,Y);
771 vgb = _mm_mul_pd(gbqqfactor,VV);
773 twogbeps = _mm_add_pd(gbeps,gbeps);
774 FF = _mm_macc_pd(_mm_macc_pd(twogbeps,H,G),gbeps,Fp);
775 fgb = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
776 dvdatmp = _mm_mul_pd(minushalf,_mm_macc_pd(fgb,r00,vgb));
777 dvdasum = _mm_add_pd(dvdasum,dvdatmp);
778 gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
779 velec = _mm_mul_pd(qq00,rinv00);
780 felec = _mm_mul_pd(_mm_msub_pd(velec,rinv00,fgb),rinv00);
782 /* CUBIC SPLINE TABLE DISPERSION */
783 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
784 F = _mm_setzero_pd();
785 GMX_MM_TRANSPOSE2_PD(Y,F);
786 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
787 H = _mm_setzero_pd();
788 GMX_MM_TRANSPOSE2_PD(G,H);
789 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
790 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
791 fvdw6 = _mm_mul_pd(c6_00,FF);
793 /* CUBIC SPLINE TABLE REPULSION */
794 vfitab = _mm_add_epi32(vfitab,ifour);
795 Y = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
796 F = _mm_setzero_pd();
797 GMX_MM_TRANSPOSE2_PD(Y,F);
798 G = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
799 H = _mm_setzero_pd();
800 GMX_MM_TRANSPOSE2_PD(G,H);
801 Fp = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
802 FF = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
803 fvdw12 = _mm_mul_pd(c12_00,FF);
804 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
806 fscal = _mm_add_pd(felec,fvdw);
808 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
810 /* Update vectorial force */
811 fix0 = _mm_macc_pd(dx00,fscal,fix0);
812 fiy0 = _mm_macc_pd(dy00,fscal,fiy0);
813 fiz0 = _mm_macc_pd(dz00,fscal,fiz0);
815 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
816 _mm_mul_pd(dx00,fscal),
817 _mm_mul_pd(dy00,fscal),
818 _mm_mul_pd(dz00,fscal));
820 /* Inner loop uses 85 flops */
823 /* End of innermost loop */
825 gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
826 f+i_coord_offset,fshift+i_shift_offset);
828 dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
829 gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
831 /* Increment number of inner iterations */
832 inneriter += j_index_end - j_index_start;
834 /* Outer loop uses 7 flops */
837 /* Increment number of outer iterations */
840 /* Update outer/inner flops */
842 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*85);