2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012, by the GROMACS development team, led by
5 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
6 * others, as listed in the AUTHORS file in the top-level source
7 * directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
36 * Note: this file was generated by the GROMACS sse2_double kernel generator.
44 #include "../nb_kernel.h"
45 #include "types/simple.h"
49 #include "gmx_math_x86_sse2_double.h"
50 #include "kernelutil_x86_sse2_double.h"
53 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_sse2_double
54 * Electrostatics interaction: GeneralizedBorn
55 * VdW interaction: CubicSplineTable
56 * Geometry: Particle-Particle
57 * Calculate force/pot: PotentialAndForce
60 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_sse2_double
61 (t_nblist * gmx_restrict nlist,
62 rvec * gmx_restrict xx,
63 rvec * gmx_restrict ff,
64 t_forcerec * gmx_restrict fr,
65 t_mdatoms * gmx_restrict mdatoms,
66 nb_kernel_data_t * gmx_restrict kernel_data,
67 t_nrnb * gmx_restrict nrnb)
69 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
70 * just 0 for non-waters.
71 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
72 * jnr indices corresponding to data put in the four positions in the SIMD register.
74 int i_shift_offset,i_coord_offset,outeriter,inneriter;
75 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
77 int j_coord_offsetA,j_coord_offsetB;
78 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
80 real *shiftvec,*fshift,*x,*f;
81 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
83 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
84 int vdwjidx0A,vdwjidx0B;
85 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
86 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
87 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
90 __m128d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,dvdatmp;
91 __m128d minushalf = _mm_set1_pd(-0.5);
92 real *invsqrta,*dvda,*gbtab;
94 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
97 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
98 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
100 __m128i ifour = _mm_set1_epi32(4);
101 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
103 __m128d dummy_mask,cutoff_mask;
104 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
105 __m128d one = _mm_set1_pd(1.0);
106 __m128d two = _mm_set1_pd(2.0);
112 jindex = nlist->jindex;
114 shiftidx = nlist->shift;
116 shiftvec = fr->shift_vec[0];
117 fshift = fr->fshift[0];
118 facel = _mm_set1_pd(fr->epsfac);
119 charge = mdatoms->chargeA;
120 nvdwtype = fr->ntype;
122 vdwtype = mdatoms->typeA;
124 vftab = kernel_data->table_vdw->data;
125 vftabscale = _mm_set1_pd(kernel_data->table_vdw->scale);
127 invsqrta = fr->invsqrta;
129 gbtabscale = _mm_set1_pd(fr->gbtab.scale);
130 gbtab = fr->gbtab.data;
131 gbinvepsdiff = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
133 /* Avoid stupid compiler warnings */
141 /* Start outer loop over neighborlists */
142 for(iidx=0; iidx<nri; iidx++)
144 /* Load shift vector for this list */
145 i_shift_offset = DIM*shiftidx[iidx];
147 /* Load limits for loop over neighbors */
148 j_index_start = jindex[iidx];
149 j_index_end = jindex[iidx+1];
151 /* Get outer coordinate index */
153 i_coord_offset = DIM*inr;
155 /* Load i particle coords and add shift vector */
156 gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
158 fix0 = _mm_setzero_pd();
159 fiy0 = _mm_setzero_pd();
160 fiz0 = _mm_setzero_pd();
162 /* Load parameters for i particles */
163 iq0 = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
164 isai0 = _mm_load1_pd(invsqrta+inr+0);
165 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
167 /* Reset potential sums */
168 velecsum = _mm_setzero_pd();
169 vgbsum = _mm_setzero_pd();
170 vvdwsum = _mm_setzero_pd();
171 dvdasum = _mm_setzero_pd();
173 /* Start inner kernel loop */
174 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
177 /* Get j neighbor index, and coordinate index */
180 j_coord_offsetA = DIM*jnrA;
181 j_coord_offsetB = DIM*jnrB;
183 /* load j atom coordinates */
184 gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
187 /* Calculate displacement vector */
188 dx00 = _mm_sub_pd(ix0,jx0);
189 dy00 = _mm_sub_pd(iy0,jy0);
190 dz00 = _mm_sub_pd(iz0,jz0);
192 /* Calculate squared distance and things based on it */
193 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
195 rinv00 = gmx_mm_invsqrt_pd(rsq00);
197 /* Load parameters for j particles */
198 jq0 = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
199 isaj0 = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
200 vdwjidx0A = 2*vdwtype[jnrA+0];
201 vdwjidx0B = 2*vdwtype[jnrB+0];
203 /**************************
204 * CALCULATE INTERACTIONS *
205 **************************/
207 r00 = _mm_mul_pd(rsq00,rinv00);
209 /* Compute parameters for interactions between i and j atoms */
210 qq00 = _mm_mul_pd(iq0,jq0);
211 gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
212 vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
214 /* Calculate table index by multiplying r with table scale and truncate to integer */
215 rt = _mm_mul_pd(r00,vftabscale);
216 vfitab = _mm_cvttpd_epi32(rt);
217 vfeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
218 vfitab = _mm_slli_epi32(vfitab,3);
220 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
221 isaprod = _mm_mul_pd(isai0,isaj0);
222 gbqqfactor = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
223 gbscale = _mm_mul_pd(isaprod,gbtabscale);
225 /* Calculate generalized born table index - this is a separate table from the normal one,
226 * but we use the same procedure by multiplying r with scale and truncating to integer.
228 rt = _mm_mul_pd(r00,gbscale);
229 gbitab = _mm_cvttpd_epi32(rt);
230 gbeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(gbitab));
231 gbitab = _mm_slli_epi32(gbitab,2);
233 Y = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,0) );
234 F = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,1) );
235 GMX_MM_TRANSPOSE2_PD(Y,F);
236 G = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,0) +2);
237 H = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,1) +2);
238 GMX_MM_TRANSPOSE2_PD(G,H);
239 Heps = _mm_mul_pd(gbeps,H);
240 Fp = _mm_add_pd(F,_mm_mul_pd(gbeps,_mm_add_pd(G,Heps)));
241 VV = _mm_add_pd(Y,_mm_mul_pd(gbeps,Fp));
242 vgb = _mm_mul_pd(gbqqfactor,VV);
244 FF = _mm_add_pd(Fp,_mm_mul_pd(gbeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
245 fgb = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
246 dvdatmp = _mm_mul_pd(minushalf,_mm_add_pd(vgb,_mm_mul_pd(fgb,r00)));
247 dvdasum = _mm_add_pd(dvdasum,dvdatmp);
248 gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
249 velec = _mm_mul_pd(qq00,rinv00);
250 felec = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(velec,rinv00),fgb),rinv00);
252 /* CUBIC SPLINE TABLE DISPERSION */
253 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
254 F = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
255 GMX_MM_TRANSPOSE2_PD(Y,F);
256 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
257 H = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
258 GMX_MM_TRANSPOSE2_PD(G,H);
259 Heps = _mm_mul_pd(vfeps,H);
260 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
261 VV = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
262 vvdw6 = _mm_mul_pd(c6_00,VV);
263 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
264 fvdw6 = _mm_mul_pd(c6_00,FF);
266 /* CUBIC SPLINE TABLE REPULSION */
267 vfitab = _mm_add_epi32(vfitab,ifour);
268 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
269 F = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
270 GMX_MM_TRANSPOSE2_PD(Y,F);
271 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
272 H = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
273 GMX_MM_TRANSPOSE2_PD(G,H);
274 Heps = _mm_mul_pd(vfeps,H);
275 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
276 VV = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
277 vvdw12 = _mm_mul_pd(c12_00,VV);
278 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
279 fvdw12 = _mm_mul_pd(c12_00,FF);
280 vvdw = _mm_add_pd(vvdw12,vvdw6);
281 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
283 /* Update potential sum for this i atom from the interaction with this j atom. */
284 velecsum = _mm_add_pd(velecsum,velec);
285 vgbsum = _mm_add_pd(vgbsum,vgb);
286 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
288 fscal = _mm_add_pd(felec,fvdw);
290 /* Calculate temporary vectorial force */
291 tx = _mm_mul_pd(fscal,dx00);
292 ty = _mm_mul_pd(fscal,dy00);
293 tz = _mm_mul_pd(fscal,dz00);
295 /* Update vectorial force */
296 fix0 = _mm_add_pd(fix0,tx);
297 fiy0 = _mm_add_pd(fiy0,ty);
298 fiz0 = _mm_add_pd(fiz0,tz);
300 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,tx,ty,tz);
302 /* Inner loop uses 92 flops */
309 j_coord_offsetA = DIM*jnrA;
311 /* load j atom coordinates */
312 gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
315 /* Calculate displacement vector */
316 dx00 = _mm_sub_pd(ix0,jx0);
317 dy00 = _mm_sub_pd(iy0,jy0);
318 dz00 = _mm_sub_pd(iz0,jz0);
320 /* Calculate squared distance and things based on it */
321 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
323 rinv00 = gmx_mm_invsqrt_pd(rsq00);
325 /* Load parameters for j particles */
326 jq0 = _mm_load_sd(charge+jnrA+0);
327 isaj0 = _mm_load_sd(invsqrta+jnrA+0);
328 vdwjidx0A = 2*vdwtype[jnrA+0];
330 /**************************
331 * CALCULATE INTERACTIONS *
332 **************************/
334 r00 = _mm_mul_pd(rsq00,rinv00);
336 /* Compute parameters for interactions between i and j atoms */
337 qq00 = _mm_mul_pd(iq0,jq0);
338 gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
340 /* Calculate table index by multiplying r with table scale and truncate to integer */
341 rt = _mm_mul_pd(r00,vftabscale);
342 vfitab = _mm_cvttpd_epi32(rt);
343 vfeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
344 vfitab = _mm_slli_epi32(vfitab,3);
346 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
347 isaprod = _mm_mul_pd(isai0,isaj0);
348 gbqqfactor = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
349 gbscale = _mm_mul_pd(isaprod,gbtabscale);
351 /* Calculate generalized born table index - this is a separate table from the normal one,
352 * but we use the same procedure by multiplying r with scale and truncating to integer.
354 rt = _mm_mul_pd(r00,gbscale);
355 gbitab = _mm_cvttpd_epi32(rt);
356 gbeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(gbitab));
357 gbitab = _mm_slli_epi32(gbitab,2);
359 Y = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,0) );
360 F = _mm_setzero_pd();
361 GMX_MM_TRANSPOSE2_PD(Y,F);
362 G = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,0) +2);
363 H = _mm_setzero_pd();
364 GMX_MM_TRANSPOSE2_PD(G,H);
365 Heps = _mm_mul_pd(gbeps,H);
366 Fp = _mm_add_pd(F,_mm_mul_pd(gbeps,_mm_add_pd(G,Heps)));
367 VV = _mm_add_pd(Y,_mm_mul_pd(gbeps,Fp));
368 vgb = _mm_mul_pd(gbqqfactor,VV);
370 FF = _mm_add_pd(Fp,_mm_mul_pd(gbeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
371 fgb = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
372 dvdatmp = _mm_mul_pd(minushalf,_mm_add_pd(vgb,_mm_mul_pd(fgb,r00)));
373 dvdasum = _mm_add_pd(dvdasum,dvdatmp);
374 gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
375 velec = _mm_mul_pd(qq00,rinv00);
376 felec = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(velec,rinv00),fgb),rinv00);
378 /* CUBIC SPLINE TABLE DISPERSION */
379 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
380 F = _mm_setzero_pd();
381 GMX_MM_TRANSPOSE2_PD(Y,F);
382 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
383 H = _mm_setzero_pd();
384 GMX_MM_TRANSPOSE2_PD(G,H);
385 Heps = _mm_mul_pd(vfeps,H);
386 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
387 VV = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
388 vvdw6 = _mm_mul_pd(c6_00,VV);
389 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
390 fvdw6 = _mm_mul_pd(c6_00,FF);
392 /* CUBIC SPLINE TABLE REPULSION */
393 vfitab = _mm_add_epi32(vfitab,ifour);
394 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
395 F = _mm_setzero_pd();
396 GMX_MM_TRANSPOSE2_PD(Y,F);
397 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
398 H = _mm_setzero_pd();
399 GMX_MM_TRANSPOSE2_PD(G,H);
400 Heps = _mm_mul_pd(vfeps,H);
401 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
402 VV = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
403 vvdw12 = _mm_mul_pd(c12_00,VV);
404 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
405 fvdw12 = _mm_mul_pd(c12_00,FF);
406 vvdw = _mm_add_pd(vvdw12,vvdw6);
407 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
409 /* Update potential sum for this i atom from the interaction with this j atom. */
410 velec = _mm_unpacklo_pd(velec,_mm_setzero_pd());
411 velecsum = _mm_add_pd(velecsum,velec);
412 vgb = _mm_unpacklo_pd(vgb,_mm_setzero_pd());
413 vgbsum = _mm_add_pd(vgbsum,vgb);
414 vvdw = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
415 vvdwsum = _mm_add_pd(vvdwsum,vvdw);
417 fscal = _mm_add_pd(felec,fvdw);
419 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
421 /* Calculate temporary vectorial force */
422 tx = _mm_mul_pd(fscal,dx00);
423 ty = _mm_mul_pd(fscal,dy00);
424 tz = _mm_mul_pd(fscal,dz00);
426 /* Update vectorial force */
427 fix0 = _mm_add_pd(fix0,tx);
428 fiy0 = _mm_add_pd(fiy0,ty);
429 fiz0 = _mm_add_pd(fiz0,tz);
431 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,tx,ty,tz);
433 /* Inner loop uses 92 flops */
436 /* End of innermost loop */
438 gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
439 f+i_coord_offset,fshift+i_shift_offset);
442 /* Update potential energies */
443 gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
444 gmx_mm_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
445 gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
446 dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
447 gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
449 /* Increment number of inner iterations */
450 inneriter += j_index_end - j_index_start;
452 /* Outer loop uses 10 flops */
455 /* Increment number of outer iterations */
458 /* Update outer/inner flops */
460 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*92);
463 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_sse2_double
464 * Electrostatics interaction: GeneralizedBorn
465 * VdW interaction: CubicSplineTable
466 * Geometry: Particle-Particle
467 * Calculate force/pot: Force
470 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_sse2_double
471 (t_nblist * gmx_restrict nlist,
472 rvec * gmx_restrict xx,
473 rvec * gmx_restrict ff,
474 t_forcerec * gmx_restrict fr,
475 t_mdatoms * gmx_restrict mdatoms,
476 nb_kernel_data_t * gmx_restrict kernel_data,
477 t_nrnb * gmx_restrict nrnb)
479 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
480 * just 0 for non-waters.
481 * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
482 * jnr indices corresponding to data put in the four positions in the SIMD register.
484 int i_shift_offset,i_coord_offset,outeriter,inneriter;
485 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
487 int j_coord_offsetA,j_coord_offsetB;
488 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
490 real *shiftvec,*fshift,*x,*f;
491 __m128d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
493 __m128d ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
494 int vdwjidx0A,vdwjidx0B;
495 __m128d jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
496 __m128d dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
497 __m128d velec,felec,velecsum,facel,crf,krf,krf2;
500 __m128d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,dvdaj,gbeps,dvdatmp;
501 __m128d minushalf = _mm_set1_pd(-0.5);
502 real *invsqrta,*dvda,*gbtab;
504 __m128d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
507 __m128d one_sixth = _mm_set1_pd(1.0/6.0);
508 __m128d one_twelfth = _mm_set1_pd(1.0/12.0);
510 __m128i ifour = _mm_set1_epi32(4);
511 __m128d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
513 __m128d dummy_mask,cutoff_mask;
514 __m128d signbit = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
515 __m128d one = _mm_set1_pd(1.0);
516 __m128d two = _mm_set1_pd(2.0);
522 jindex = nlist->jindex;
524 shiftidx = nlist->shift;
526 shiftvec = fr->shift_vec[0];
527 fshift = fr->fshift[0];
528 facel = _mm_set1_pd(fr->epsfac);
529 charge = mdatoms->chargeA;
530 nvdwtype = fr->ntype;
532 vdwtype = mdatoms->typeA;
534 vftab = kernel_data->table_vdw->data;
535 vftabscale = _mm_set1_pd(kernel_data->table_vdw->scale);
537 invsqrta = fr->invsqrta;
539 gbtabscale = _mm_set1_pd(fr->gbtab.scale);
540 gbtab = fr->gbtab.data;
541 gbinvepsdiff = _mm_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
543 /* Avoid stupid compiler warnings */
551 /* Start outer loop over neighborlists */
552 for(iidx=0; iidx<nri; iidx++)
554 /* Load shift vector for this list */
555 i_shift_offset = DIM*shiftidx[iidx];
557 /* Load limits for loop over neighbors */
558 j_index_start = jindex[iidx];
559 j_index_end = jindex[iidx+1];
561 /* Get outer coordinate index */
563 i_coord_offset = DIM*inr;
565 /* Load i particle coords and add shift vector */
566 gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
568 fix0 = _mm_setzero_pd();
569 fiy0 = _mm_setzero_pd();
570 fiz0 = _mm_setzero_pd();
572 /* Load parameters for i particles */
573 iq0 = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
574 isai0 = _mm_load1_pd(invsqrta+inr+0);
575 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
577 dvdasum = _mm_setzero_pd();
579 /* Start inner kernel loop */
580 for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
583 /* Get j neighbor index, and coordinate index */
586 j_coord_offsetA = DIM*jnrA;
587 j_coord_offsetB = DIM*jnrB;
589 /* load j atom coordinates */
590 gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
593 /* Calculate displacement vector */
594 dx00 = _mm_sub_pd(ix0,jx0);
595 dy00 = _mm_sub_pd(iy0,jy0);
596 dz00 = _mm_sub_pd(iz0,jz0);
598 /* Calculate squared distance and things based on it */
599 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
601 rinv00 = gmx_mm_invsqrt_pd(rsq00);
603 /* Load parameters for j particles */
604 jq0 = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
605 isaj0 = gmx_mm_load_2real_swizzle_pd(invsqrta+jnrA+0,invsqrta+jnrB+0);
606 vdwjidx0A = 2*vdwtype[jnrA+0];
607 vdwjidx0B = 2*vdwtype[jnrB+0];
609 /**************************
610 * CALCULATE INTERACTIONS *
611 **************************/
613 r00 = _mm_mul_pd(rsq00,rinv00);
615 /* Compute parameters for interactions between i and j atoms */
616 qq00 = _mm_mul_pd(iq0,jq0);
617 gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
618 vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
620 /* Calculate table index by multiplying r with table scale and truncate to integer */
621 rt = _mm_mul_pd(r00,vftabscale);
622 vfitab = _mm_cvttpd_epi32(rt);
623 vfeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
624 vfitab = _mm_slli_epi32(vfitab,3);
626 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
627 isaprod = _mm_mul_pd(isai0,isaj0);
628 gbqqfactor = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
629 gbscale = _mm_mul_pd(isaprod,gbtabscale);
631 /* Calculate generalized born table index - this is a separate table from the normal one,
632 * but we use the same procedure by multiplying r with scale and truncating to integer.
634 rt = _mm_mul_pd(r00,gbscale);
635 gbitab = _mm_cvttpd_epi32(rt);
636 gbeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(gbitab));
637 gbitab = _mm_slli_epi32(gbitab,2);
639 Y = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,0) );
640 F = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,1) );
641 GMX_MM_TRANSPOSE2_PD(Y,F);
642 G = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,0) +2);
643 H = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,1) +2);
644 GMX_MM_TRANSPOSE2_PD(G,H);
645 Heps = _mm_mul_pd(gbeps,H);
646 Fp = _mm_add_pd(F,_mm_mul_pd(gbeps,_mm_add_pd(G,Heps)));
647 VV = _mm_add_pd(Y,_mm_mul_pd(gbeps,Fp));
648 vgb = _mm_mul_pd(gbqqfactor,VV);
650 FF = _mm_add_pd(Fp,_mm_mul_pd(gbeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
651 fgb = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
652 dvdatmp = _mm_mul_pd(minushalf,_mm_add_pd(vgb,_mm_mul_pd(fgb,r00)));
653 dvdasum = _mm_add_pd(dvdasum,dvdatmp);
654 gmx_mm_increment_2real_swizzle_pd(dvda+jnrA,dvda+jnrB,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
655 velec = _mm_mul_pd(qq00,rinv00);
656 felec = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(velec,rinv00),fgb),rinv00);
658 /* CUBIC SPLINE TABLE DISPERSION */
659 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
660 F = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
661 GMX_MM_TRANSPOSE2_PD(Y,F);
662 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
663 H = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
664 GMX_MM_TRANSPOSE2_PD(G,H);
665 Heps = _mm_mul_pd(vfeps,H);
666 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
667 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
668 fvdw6 = _mm_mul_pd(c6_00,FF);
670 /* CUBIC SPLINE TABLE REPULSION */
671 vfitab = _mm_add_epi32(vfitab,ifour);
672 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
673 F = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
674 GMX_MM_TRANSPOSE2_PD(Y,F);
675 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
676 H = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
677 GMX_MM_TRANSPOSE2_PD(G,H);
678 Heps = _mm_mul_pd(vfeps,H);
679 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
680 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
681 fvdw12 = _mm_mul_pd(c12_00,FF);
682 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
684 fscal = _mm_add_pd(felec,fvdw);
686 /* Calculate temporary vectorial force */
687 tx = _mm_mul_pd(fscal,dx00);
688 ty = _mm_mul_pd(fscal,dy00);
689 tz = _mm_mul_pd(fscal,dz00);
691 /* Update vectorial force */
692 fix0 = _mm_add_pd(fix0,tx);
693 fiy0 = _mm_add_pd(fiy0,ty);
694 fiz0 = _mm_add_pd(fiz0,tz);
696 gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,tx,ty,tz);
698 /* Inner loop uses 82 flops */
705 j_coord_offsetA = DIM*jnrA;
707 /* load j atom coordinates */
708 gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
711 /* Calculate displacement vector */
712 dx00 = _mm_sub_pd(ix0,jx0);
713 dy00 = _mm_sub_pd(iy0,jy0);
714 dz00 = _mm_sub_pd(iz0,jz0);
716 /* Calculate squared distance and things based on it */
717 rsq00 = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
719 rinv00 = gmx_mm_invsqrt_pd(rsq00);
721 /* Load parameters for j particles */
722 jq0 = _mm_load_sd(charge+jnrA+0);
723 isaj0 = _mm_load_sd(invsqrta+jnrA+0);
724 vdwjidx0A = 2*vdwtype[jnrA+0];
726 /**************************
727 * CALCULATE INTERACTIONS *
728 **************************/
730 r00 = _mm_mul_pd(rsq00,rinv00);
732 /* Compute parameters for interactions between i and j atoms */
733 qq00 = _mm_mul_pd(iq0,jq0);
734 gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
736 /* Calculate table index by multiplying r with table scale and truncate to integer */
737 rt = _mm_mul_pd(r00,vftabscale);
738 vfitab = _mm_cvttpd_epi32(rt);
739 vfeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
740 vfitab = _mm_slli_epi32(vfitab,3);
742 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
743 isaprod = _mm_mul_pd(isai0,isaj0);
744 gbqqfactor = _mm_xor_pd(signbit,_mm_mul_pd(qq00,_mm_mul_pd(isaprod,gbinvepsdiff)));
745 gbscale = _mm_mul_pd(isaprod,gbtabscale);
747 /* Calculate generalized born table index - this is a separate table from the normal one,
748 * but we use the same procedure by multiplying r with scale and truncating to integer.
750 rt = _mm_mul_pd(r00,gbscale);
751 gbitab = _mm_cvttpd_epi32(rt);
752 gbeps = _mm_sub_pd(rt,_mm_cvtepi32_pd(gbitab));
753 gbitab = _mm_slli_epi32(gbitab,2);
755 Y = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,0) );
756 F = _mm_setzero_pd();
757 GMX_MM_TRANSPOSE2_PD(Y,F);
758 G = _mm_load_pd( gbtab + gmx_mm_extract_epi32(gbitab,0) +2);
759 H = _mm_setzero_pd();
760 GMX_MM_TRANSPOSE2_PD(G,H);
761 Heps = _mm_mul_pd(gbeps,H);
762 Fp = _mm_add_pd(F,_mm_mul_pd(gbeps,_mm_add_pd(G,Heps)));
763 VV = _mm_add_pd(Y,_mm_mul_pd(gbeps,Fp));
764 vgb = _mm_mul_pd(gbqqfactor,VV);
766 FF = _mm_add_pd(Fp,_mm_mul_pd(gbeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
767 fgb = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
768 dvdatmp = _mm_mul_pd(minushalf,_mm_add_pd(vgb,_mm_mul_pd(fgb,r00)));
769 dvdasum = _mm_add_pd(dvdasum,dvdatmp);
770 gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
771 velec = _mm_mul_pd(qq00,rinv00);
772 felec = _mm_mul_pd(_mm_sub_pd(_mm_mul_pd(velec,rinv00),fgb),rinv00);
774 /* CUBIC SPLINE TABLE DISPERSION */
775 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
776 F = _mm_setzero_pd();
777 GMX_MM_TRANSPOSE2_PD(Y,F);
778 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
779 H = _mm_setzero_pd();
780 GMX_MM_TRANSPOSE2_PD(G,H);
781 Heps = _mm_mul_pd(vfeps,H);
782 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
783 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
784 fvdw6 = _mm_mul_pd(c6_00,FF);
786 /* CUBIC SPLINE TABLE REPULSION */
787 vfitab = _mm_add_epi32(vfitab,ifour);
788 Y = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
789 F = _mm_setzero_pd();
790 GMX_MM_TRANSPOSE2_PD(Y,F);
791 G = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
792 H = _mm_setzero_pd();
793 GMX_MM_TRANSPOSE2_PD(G,H);
794 Heps = _mm_mul_pd(vfeps,H);
795 Fp = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
796 FF = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
797 fvdw12 = _mm_mul_pd(c12_00,FF);
798 fvdw = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
800 fscal = _mm_add_pd(felec,fvdw);
802 fscal = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
804 /* Calculate temporary vectorial force */
805 tx = _mm_mul_pd(fscal,dx00);
806 ty = _mm_mul_pd(fscal,dy00);
807 tz = _mm_mul_pd(fscal,dz00);
809 /* Update vectorial force */
810 fix0 = _mm_add_pd(fix0,tx);
811 fiy0 = _mm_add_pd(fiy0,ty);
812 fiz0 = _mm_add_pd(fiz0,tz);
814 gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,tx,ty,tz);
816 /* Inner loop uses 82 flops */
819 /* End of innermost loop */
821 gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
822 f+i_coord_offset,fshift+i_shift_offset);
824 dvdasum = _mm_mul_pd(dvdasum, _mm_mul_pd(isai0,isai0));
825 gmx_mm_update_1pot_pd(dvdasum,dvda+inr);
827 /* Increment number of inner iterations */
828 inneriter += j_index_end - j_index_start;
830 /* Outer loop uses 7 flops */
833 /* Increment number of outer iterations */
836 /* Update outer/inner flops */
838 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*82);