2 * Note: this file was generated by the Gromacs sse2_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_sse2_single.h"
34 #include "kernelutil_x86_sse2_single.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_sse2_single
38 * Electrostatics interaction: GeneralizedBorn
39 * VdW interaction: LennardJones
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_sse2_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 SSE, 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 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
70 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
71 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
72 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
73 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
74 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
77 __m128 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
78 __m128 minushalf = _mm_set1_ps(-0.5);
79 real *invsqrta,*dvda,*gbtab;
81 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
84 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
85 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
87 __m128i ifour = _mm_set1_epi32(4);
88 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
90 __m128 dummy_mask,cutoff_mask;
91 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
92 __m128 one = _mm_set1_ps(1.0);
93 __m128 two = _mm_set1_ps(2.0);
99 jindex = nlist->jindex;
101 shiftidx = nlist->shift;
103 shiftvec = fr->shift_vec[0];
104 fshift = fr->fshift[0];
105 facel = _mm_set1_ps(fr->epsfac);
106 charge = mdatoms->chargeA;
107 nvdwtype = fr->ntype;
109 vdwtype = mdatoms->typeA;
111 invsqrta = fr->invsqrta;
113 gbtabscale = _mm_set1_ps(fr->gbtab.scale);
114 gbtab = fr->gbtab.data;
115 gbinvepsdiff = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
117 /* Avoid stupid compiler warnings */
118 jnrA = jnrB = jnrC = jnrD = 0;
127 for(iidx=0;iidx<4*DIM;iidx++)
132 /* Start outer loop over neighborlists */
133 for(iidx=0; iidx<nri; iidx++)
135 /* Load shift vector for this list */
136 i_shift_offset = DIM*shiftidx[iidx];
138 /* Load limits for loop over neighbors */
139 j_index_start = jindex[iidx];
140 j_index_end = jindex[iidx+1];
142 /* Get outer coordinate index */
144 i_coord_offset = DIM*inr;
146 /* Load i particle coords and add shift vector */
147 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
149 fix0 = _mm_setzero_ps();
150 fiy0 = _mm_setzero_ps();
151 fiz0 = _mm_setzero_ps();
153 /* Load parameters for i particles */
154 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
155 isai0 = _mm_load1_ps(invsqrta+inr+0);
156 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
158 /* Reset potential sums */
159 velecsum = _mm_setzero_ps();
160 vgbsum = _mm_setzero_ps();
161 vvdwsum = _mm_setzero_ps();
162 dvdasum = _mm_setzero_ps();
164 /* Start inner kernel loop */
165 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
168 /* Get j neighbor index, and coordinate index */
173 j_coord_offsetA = DIM*jnrA;
174 j_coord_offsetB = DIM*jnrB;
175 j_coord_offsetC = DIM*jnrC;
176 j_coord_offsetD = DIM*jnrD;
178 /* load j atom coordinates */
179 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
180 x+j_coord_offsetC,x+j_coord_offsetD,
183 /* Calculate displacement vector */
184 dx00 = _mm_sub_ps(ix0,jx0);
185 dy00 = _mm_sub_ps(iy0,jy0);
186 dz00 = _mm_sub_ps(iz0,jz0);
188 /* Calculate squared distance and things based on it */
189 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
191 rinv00 = gmx_mm_invsqrt_ps(rsq00);
193 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
195 /* Load parameters for j particles */
196 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
197 charge+jnrC+0,charge+jnrD+0);
198 isaj0 = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
199 invsqrta+jnrC+0,invsqrta+jnrD+0);
200 vdwjidx0A = 2*vdwtype[jnrA+0];
201 vdwjidx0B = 2*vdwtype[jnrB+0];
202 vdwjidx0C = 2*vdwtype[jnrC+0];
203 vdwjidx0D = 2*vdwtype[jnrD+0];
205 /**************************
206 * CALCULATE INTERACTIONS *
207 **************************/
209 r00 = _mm_mul_ps(rsq00,rinv00);
211 /* Compute parameters for interactions between i and j atoms */
212 qq00 = _mm_mul_ps(iq0,jq0);
213 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
214 vdwparam+vdwioffset0+vdwjidx0B,
215 vdwparam+vdwioffset0+vdwjidx0C,
216 vdwparam+vdwioffset0+vdwjidx0D,
219 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
220 isaprod = _mm_mul_ps(isai0,isaj0);
221 gbqqfactor = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
222 gbscale = _mm_mul_ps(isaprod,gbtabscale);
224 /* Calculate generalized born table index - this is a separate table from the normal one,
225 * but we use the same procedure by multiplying r with scale and truncating to integer.
227 rt = _mm_mul_ps(r00,gbscale);
228 gbitab = _mm_cvttps_epi32(rt);
229 gbeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
230 gbitab = _mm_slli_epi32(gbitab,2);
232 Y = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
233 F = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
234 G = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
235 H = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
236 _MM_TRANSPOSE4_PS(Y,F,G,H);
237 Heps = _mm_mul_ps(gbeps,H);
238 Fp = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
239 VV = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
240 vgb = _mm_mul_ps(gbqqfactor,VV);
242 FF = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
243 fgb = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
244 dvdatmp = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
245 dvdasum = _mm_add_ps(dvdasum,dvdatmp);
250 gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
251 velec = _mm_mul_ps(qq00,rinv00);
252 felec = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
254 /* LENNARD-JONES DISPERSION/REPULSION */
256 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
257 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
258 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
259 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
260 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
262 /* Update potential sum for this i atom from the interaction with this j atom. */
263 velecsum = _mm_add_ps(velecsum,velec);
264 vgbsum = _mm_add_ps(vgbsum,vgb);
265 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
267 fscal = _mm_add_ps(felec,fvdw);
269 /* Calculate temporary vectorial force */
270 tx = _mm_mul_ps(fscal,dx00);
271 ty = _mm_mul_ps(fscal,dy00);
272 tz = _mm_mul_ps(fscal,dz00);
274 /* Update vectorial force */
275 fix0 = _mm_add_ps(fix0,tx);
276 fiy0 = _mm_add_ps(fiy0,ty);
277 fiz0 = _mm_add_ps(fiz0,tz);
279 fjptrA = f+j_coord_offsetA;
280 fjptrB = f+j_coord_offsetB;
281 fjptrC = f+j_coord_offsetC;
282 fjptrD = f+j_coord_offsetD;
283 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
285 /* Inner loop uses 71 flops */
291 /* Get j neighbor index, and coordinate index */
292 jnrlistA = jjnr[jidx];
293 jnrlistB = jjnr[jidx+1];
294 jnrlistC = jjnr[jidx+2];
295 jnrlistD = jjnr[jidx+3];
296 /* Sign of each element will be negative for non-real atoms.
297 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
298 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
300 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
301 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
302 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
303 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
304 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
305 j_coord_offsetA = DIM*jnrA;
306 j_coord_offsetB = DIM*jnrB;
307 j_coord_offsetC = DIM*jnrC;
308 j_coord_offsetD = DIM*jnrD;
310 /* load j atom coordinates */
311 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
312 x+j_coord_offsetC,x+j_coord_offsetD,
315 /* Calculate displacement vector */
316 dx00 = _mm_sub_ps(ix0,jx0);
317 dy00 = _mm_sub_ps(iy0,jy0);
318 dz00 = _mm_sub_ps(iz0,jz0);
320 /* Calculate squared distance and things based on it */
321 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
323 rinv00 = gmx_mm_invsqrt_ps(rsq00);
325 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
327 /* Load parameters for j particles */
328 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
329 charge+jnrC+0,charge+jnrD+0);
330 isaj0 = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
331 invsqrta+jnrC+0,invsqrta+jnrD+0);
332 vdwjidx0A = 2*vdwtype[jnrA+0];
333 vdwjidx0B = 2*vdwtype[jnrB+0];
334 vdwjidx0C = 2*vdwtype[jnrC+0];
335 vdwjidx0D = 2*vdwtype[jnrD+0];
337 /**************************
338 * CALCULATE INTERACTIONS *
339 **************************/
341 r00 = _mm_mul_ps(rsq00,rinv00);
342 r00 = _mm_andnot_ps(dummy_mask,r00);
344 /* Compute parameters for interactions between i and j atoms */
345 qq00 = _mm_mul_ps(iq0,jq0);
346 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
347 vdwparam+vdwioffset0+vdwjidx0B,
348 vdwparam+vdwioffset0+vdwjidx0C,
349 vdwparam+vdwioffset0+vdwjidx0D,
352 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
353 isaprod = _mm_mul_ps(isai0,isaj0);
354 gbqqfactor = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
355 gbscale = _mm_mul_ps(isaprod,gbtabscale);
357 /* Calculate generalized born table index - this is a separate table from the normal one,
358 * but we use the same procedure by multiplying r with scale and truncating to integer.
360 rt = _mm_mul_ps(r00,gbscale);
361 gbitab = _mm_cvttps_epi32(rt);
362 gbeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
363 gbitab = _mm_slli_epi32(gbitab,2);
365 Y = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
366 F = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
367 G = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
368 H = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
369 _MM_TRANSPOSE4_PS(Y,F,G,H);
370 Heps = _mm_mul_ps(gbeps,H);
371 Fp = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
372 VV = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
373 vgb = _mm_mul_ps(gbqqfactor,VV);
375 FF = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
376 fgb = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
377 dvdatmp = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
378 dvdasum = _mm_add_ps(dvdasum,dvdatmp);
379 /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
380 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
381 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
382 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
383 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
384 gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
385 velec = _mm_mul_ps(qq00,rinv00);
386 felec = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
388 /* LENNARD-JONES DISPERSION/REPULSION */
390 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
391 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
392 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
393 vvdw = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
394 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
396 /* Update potential sum for this i atom from the interaction with this j atom. */
397 velec = _mm_andnot_ps(dummy_mask,velec);
398 velecsum = _mm_add_ps(velecsum,velec);
399 vgb = _mm_andnot_ps(dummy_mask,vgb);
400 vgbsum = _mm_add_ps(vgbsum,vgb);
401 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
402 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
404 fscal = _mm_add_ps(felec,fvdw);
406 fscal = _mm_andnot_ps(dummy_mask,fscal);
408 /* Calculate temporary vectorial force */
409 tx = _mm_mul_ps(fscal,dx00);
410 ty = _mm_mul_ps(fscal,dy00);
411 tz = _mm_mul_ps(fscal,dz00);
413 /* Update vectorial force */
414 fix0 = _mm_add_ps(fix0,tx);
415 fiy0 = _mm_add_ps(fiy0,ty);
416 fiz0 = _mm_add_ps(fiz0,tz);
418 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
419 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
420 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
421 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
422 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
424 /* Inner loop uses 72 flops */
427 /* End of innermost loop */
429 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
430 f+i_coord_offset,fshift+i_shift_offset);
433 /* Update potential energies */
434 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
435 gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
436 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
437 dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
438 gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
440 /* Increment number of inner iterations */
441 inneriter += j_index_end - j_index_start;
443 /* Outer loop uses 10 flops */
446 /* Increment number of outer iterations */
449 /* Update outer/inner flops */
451 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*72);
454 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_sse2_single
455 * Electrostatics interaction: GeneralizedBorn
456 * VdW interaction: LennardJones
457 * Geometry: Particle-Particle
458 * Calculate force/pot: Force
461 nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_sse2_single
462 (t_nblist * gmx_restrict nlist,
463 rvec * gmx_restrict xx,
464 rvec * gmx_restrict ff,
465 t_forcerec * gmx_restrict fr,
466 t_mdatoms * gmx_restrict mdatoms,
467 nb_kernel_data_t * gmx_restrict kernel_data,
468 t_nrnb * gmx_restrict nrnb)
470 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
471 * just 0 for non-waters.
472 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
473 * jnr indices corresponding to data put in the four positions in the SIMD register.
475 int i_shift_offset,i_coord_offset,outeriter,inneriter;
476 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
477 int jnrA,jnrB,jnrC,jnrD;
478 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
479 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
480 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
482 real *shiftvec,*fshift,*x,*f;
483 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
485 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
487 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
488 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
489 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
490 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
491 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
494 __m128 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
495 __m128 minushalf = _mm_set1_ps(-0.5);
496 real *invsqrta,*dvda,*gbtab;
498 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
501 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
502 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
504 __m128i ifour = _mm_set1_epi32(4);
505 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
507 __m128 dummy_mask,cutoff_mask;
508 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
509 __m128 one = _mm_set1_ps(1.0);
510 __m128 two = _mm_set1_ps(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_ps(fr->epsfac);
523 charge = mdatoms->chargeA;
524 nvdwtype = fr->ntype;
526 vdwtype = mdatoms->typeA;
528 invsqrta = fr->invsqrta;
530 gbtabscale = _mm_set1_ps(fr->gbtab.scale);
531 gbtab = fr->gbtab.data;
532 gbinvepsdiff = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
534 /* Avoid stupid compiler warnings */
535 jnrA = jnrB = jnrC = jnrD = 0;
544 for(iidx=0;iidx<4*DIM;iidx++)
549 /* Start outer loop over neighborlists */
550 for(iidx=0; iidx<nri; iidx++)
552 /* Load shift vector for this list */
553 i_shift_offset = DIM*shiftidx[iidx];
555 /* Load limits for loop over neighbors */
556 j_index_start = jindex[iidx];
557 j_index_end = jindex[iidx+1];
559 /* Get outer coordinate index */
561 i_coord_offset = DIM*inr;
563 /* Load i particle coords and add shift vector */
564 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
566 fix0 = _mm_setzero_ps();
567 fiy0 = _mm_setzero_ps();
568 fiz0 = _mm_setzero_ps();
570 /* Load parameters for i particles */
571 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
572 isai0 = _mm_load1_ps(invsqrta+inr+0);
573 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
575 dvdasum = _mm_setzero_ps();
577 /* Start inner kernel loop */
578 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
581 /* Get j neighbor index, and coordinate index */
586 j_coord_offsetA = DIM*jnrA;
587 j_coord_offsetB = DIM*jnrB;
588 j_coord_offsetC = DIM*jnrC;
589 j_coord_offsetD = DIM*jnrD;
591 /* load j atom coordinates */
592 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
593 x+j_coord_offsetC,x+j_coord_offsetD,
596 /* Calculate displacement vector */
597 dx00 = _mm_sub_ps(ix0,jx0);
598 dy00 = _mm_sub_ps(iy0,jy0);
599 dz00 = _mm_sub_ps(iz0,jz0);
601 /* Calculate squared distance and things based on it */
602 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
604 rinv00 = gmx_mm_invsqrt_ps(rsq00);
606 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
608 /* Load parameters for j particles */
609 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
610 charge+jnrC+0,charge+jnrD+0);
611 isaj0 = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
612 invsqrta+jnrC+0,invsqrta+jnrD+0);
613 vdwjidx0A = 2*vdwtype[jnrA+0];
614 vdwjidx0B = 2*vdwtype[jnrB+0];
615 vdwjidx0C = 2*vdwtype[jnrC+0];
616 vdwjidx0D = 2*vdwtype[jnrD+0];
618 /**************************
619 * CALCULATE INTERACTIONS *
620 **************************/
622 r00 = _mm_mul_ps(rsq00,rinv00);
624 /* Compute parameters for interactions between i and j atoms */
625 qq00 = _mm_mul_ps(iq0,jq0);
626 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
627 vdwparam+vdwioffset0+vdwjidx0B,
628 vdwparam+vdwioffset0+vdwjidx0C,
629 vdwparam+vdwioffset0+vdwjidx0D,
632 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
633 isaprod = _mm_mul_ps(isai0,isaj0);
634 gbqqfactor = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
635 gbscale = _mm_mul_ps(isaprod,gbtabscale);
637 /* Calculate generalized born table index - this is a separate table from the normal one,
638 * but we use the same procedure by multiplying r with scale and truncating to integer.
640 rt = _mm_mul_ps(r00,gbscale);
641 gbitab = _mm_cvttps_epi32(rt);
642 gbeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
643 gbitab = _mm_slli_epi32(gbitab,2);
645 Y = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
646 F = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
647 G = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
648 H = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
649 _MM_TRANSPOSE4_PS(Y,F,G,H);
650 Heps = _mm_mul_ps(gbeps,H);
651 Fp = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
652 VV = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
653 vgb = _mm_mul_ps(gbqqfactor,VV);
655 FF = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
656 fgb = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
657 dvdatmp = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
658 dvdasum = _mm_add_ps(dvdasum,dvdatmp);
663 gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
664 velec = _mm_mul_ps(qq00,rinv00);
665 felec = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
667 /* LENNARD-JONES DISPERSION/REPULSION */
669 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
670 fvdw = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
672 fscal = _mm_add_ps(felec,fvdw);
674 /* Calculate temporary vectorial force */
675 tx = _mm_mul_ps(fscal,dx00);
676 ty = _mm_mul_ps(fscal,dy00);
677 tz = _mm_mul_ps(fscal,dz00);
679 /* Update vectorial force */
680 fix0 = _mm_add_ps(fix0,tx);
681 fiy0 = _mm_add_ps(fiy0,ty);
682 fiz0 = _mm_add_ps(fiz0,tz);
684 fjptrA = f+j_coord_offsetA;
685 fjptrB = f+j_coord_offsetB;
686 fjptrC = f+j_coord_offsetC;
687 fjptrD = f+j_coord_offsetD;
688 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
690 /* Inner loop uses 64 flops */
696 /* Get j neighbor index, and coordinate index */
697 jnrlistA = jjnr[jidx];
698 jnrlistB = jjnr[jidx+1];
699 jnrlistC = jjnr[jidx+2];
700 jnrlistD = jjnr[jidx+3];
701 /* Sign of each element will be negative for non-real atoms.
702 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
703 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
705 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
706 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
707 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
708 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
709 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
710 j_coord_offsetA = DIM*jnrA;
711 j_coord_offsetB = DIM*jnrB;
712 j_coord_offsetC = DIM*jnrC;
713 j_coord_offsetD = DIM*jnrD;
715 /* load j atom coordinates */
716 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
717 x+j_coord_offsetC,x+j_coord_offsetD,
720 /* Calculate displacement vector */
721 dx00 = _mm_sub_ps(ix0,jx0);
722 dy00 = _mm_sub_ps(iy0,jy0);
723 dz00 = _mm_sub_ps(iz0,jz0);
725 /* Calculate squared distance and things based on it */
726 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
728 rinv00 = gmx_mm_invsqrt_ps(rsq00);
730 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
732 /* Load parameters for j particles */
733 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
734 charge+jnrC+0,charge+jnrD+0);
735 isaj0 = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
736 invsqrta+jnrC+0,invsqrta+jnrD+0);
737 vdwjidx0A = 2*vdwtype[jnrA+0];
738 vdwjidx0B = 2*vdwtype[jnrB+0];
739 vdwjidx0C = 2*vdwtype[jnrC+0];
740 vdwjidx0D = 2*vdwtype[jnrD+0];
742 /**************************
743 * CALCULATE INTERACTIONS *
744 **************************/
746 r00 = _mm_mul_ps(rsq00,rinv00);
747 r00 = _mm_andnot_ps(dummy_mask,r00);
749 /* Compute parameters for interactions between i and j atoms */
750 qq00 = _mm_mul_ps(iq0,jq0);
751 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
752 vdwparam+vdwioffset0+vdwjidx0B,
753 vdwparam+vdwioffset0+vdwjidx0C,
754 vdwparam+vdwioffset0+vdwjidx0D,
757 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
758 isaprod = _mm_mul_ps(isai0,isaj0);
759 gbqqfactor = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
760 gbscale = _mm_mul_ps(isaprod,gbtabscale);
762 /* Calculate generalized born table index - this is a separate table from the normal one,
763 * but we use the same procedure by multiplying r with scale and truncating to integer.
765 rt = _mm_mul_ps(r00,gbscale);
766 gbitab = _mm_cvttps_epi32(rt);
767 gbeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(gbitab));
768 gbitab = _mm_slli_epi32(gbitab,2);
770 Y = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
771 F = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
772 G = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
773 H = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
774 _MM_TRANSPOSE4_PS(Y,F,G,H);
775 Heps = _mm_mul_ps(gbeps,H);
776 Fp = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
777 VV = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
778 vgb = _mm_mul_ps(gbqqfactor,VV);
780 FF = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
781 fgb = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
782 dvdatmp = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
783 dvdasum = _mm_add_ps(dvdasum,dvdatmp);
784 /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
785 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
786 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
787 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
788 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
789 gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
790 velec = _mm_mul_ps(qq00,rinv00);
791 felec = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
793 /* LENNARD-JONES DISPERSION/REPULSION */
795 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
796 fvdw = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
798 fscal = _mm_add_ps(felec,fvdw);
800 fscal = _mm_andnot_ps(dummy_mask,fscal);
802 /* Calculate temporary vectorial force */
803 tx = _mm_mul_ps(fscal,dx00);
804 ty = _mm_mul_ps(fscal,dy00);
805 tz = _mm_mul_ps(fscal,dz00);
807 /* Update vectorial force */
808 fix0 = _mm_add_ps(fix0,tx);
809 fiy0 = _mm_add_ps(fiy0,ty);
810 fiz0 = _mm_add_ps(fiz0,tz);
812 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
813 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
814 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
815 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
816 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
818 /* Inner loop uses 65 flops */
821 /* End of innermost loop */
823 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
824 f+i_coord_offset,fshift+i_shift_offset);
826 dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
827 gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
829 /* Increment number of inner iterations */
830 inneriter += j_index_end - j_index_start;
832 /* Outer loop uses 7 flops */
835 /* Increment number of outer iterations */
838 /* Update outer/inner flops */
840 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*65);