2 * Note: this file was generated by the Gromacs sse4_1_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_sse4_1_single.h"
34 #include "kernelutil_x86_sse4_1_single.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwCSTab_GeomP1P1_VF_sse4_1_single
38 * Electrostatics interaction: Ewald
39 * VdW interaction: CubicSplineTable
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_VF_sse4_1_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 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
80 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
81 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
83 __m128i ifour = _mm_set1_epi32(4);
84 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
87 __m128 ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
89 __m128 dummy_mask,cutoff_mask;
90 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
91 __m128 one = _mm_set1_ps(1.0);
92 __m128 two = _mm_set1_ps(2.0);
98 jindex = nlist->jindex;
100 shiftidx = nlist->shift;
102 shiftvec = fr->shift_vec[0];
103 fshift = fr->fshift[0];
104 facel = _mm_set1_ps(fr->epsfac);
105 charge = mdatoms->chargeA;
106 nvdwtype = fr->ntype;
108 vdwtype = mdatoms->typeA;
110 vftab = kernel_data->table_vdw->data;
111 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
113 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
114 ewtab = fr->ic->tabq_coul_FDV0;
115 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
116 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
118 /* Avoid stupid compiler warnings */
119 jnrA = jnrB = jnrC = jnrD = 0;
128 for(iidx=0;iidx<4*DIM;iidx++)
133 /* Start outer loop over neighborlists */
134 for(iidx=0; iidx<nri; iidx++)
136 /* Load shift vector for this list */
137 i_shift_offset = DIM*shiftidx[iidx];
139 /* Load limits for loop over neighbors */
140 j_index_start = jindex[iidx];
141 j_index_end = jindex[iidx+1];
143 /* Get outer coordinate index */
145 i_coord_offset = DIM*inr;
147 /* Load i particle coords and add shift vector */
148 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
150 fix0 = _mm_setzero_ps();
151 fiy0 = _mm_setzero_ps();
152 fiz0 = _mm_setzero_ps();
154 /* Load parameters for i particles */
155 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
156 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
158 /* Reset potential sums */
159 velecsum = _mm_setzero_ps();
160 vvdwsum = _mm_setzero_ps();
162 /* Start inner kernel loop */
163 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
166 /* Get j neighbor index, and coordinate index */
171 j_coord_offsetA = DIM*jnrA;
172 j_coord_offsetB = DIM*jnrB;
173 j_coord_offsetC = DIM*jnrC;
174 j_coord_offsetD = DIM*jnrD;
176 /* load j atom coordinates */
177 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
178 x+j_coord_offsetC,x+j_coord_offsetD,
181 /* Calculate displacement vector */
182 dx00 = _mm_sub_ps(ix0,jx0);
183 dy00 = _mm_sub_ps(iy0,jy0);
184 dz00 = _mm_sub_ps(iz0,jz0);
186 /* Calculate squared distance and things based on it */
187 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
189 rinv00 = gmx_mm_invsqrt_ps(rsq00);
191 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
193 /* Load parameters for j particles */
194 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
195 charge+jnrC+0,charge+jnrD+0);
196 vdwjidx0A = 2*vdwtype[jnrA+0];
197 vdwjidx0B = 2*vdwtype[jnrB+0];
198 vdwjidx0C = 2*vdwtype[jnrC+0];
199 vdwjidx0D = 2*vdwtype[jnrD+0];
201 /**************************
202 * CALCULATE INTERACTIONS *
203 **************************/
205 r00 = _mm_mul_ps(rsq00,rinv00);
207 /* Compute parameters for interactions between i and j atoms */
208 qq00 = _mm_mul_ps(iq0,jq0);
209 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
210 vdwparam+vdwioffset0+vdwjidx0B,
211 vdwparam+vdwioffset0+vdwjidx0C,
212 vdwparam+vdwioffset0+vdwjidx0D,
215 /* Calculate table index by multiplying r with table scale and truncate to integer */
216 rt = _mm_mul_ps(r00,vftabscale);
217 vfitab = _mm_cvttps_epi32(rt);
218 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
219 vfitab = _mm_slli_epi32(vfitab,3);
221 /* EWALD ELECTROSTATICS */
223 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
224 ewrt = _mm_mul_ps(r00,ewtabscale);
225 ewitab = _mm_cvttps_epi32(ewrt);
226 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
227 ewitab = _mm_slli_epi32(ewitab,2);
228 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
229 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
230 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
231 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
232 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
233 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
234 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
235 velec = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
236 felec = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
238 /* CUBIC SPLINE TABLE DISPERSION */
239 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
240 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
241 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
242 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
243 _MM_TRANSPOSE4_PS(Y,F,G,H);
244 Heps = _mm_mul_ps(vfeps,H);
245 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
246 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
247 vvdw6 = _mm_mul_ps(c6_00,VV);
248 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
249 fvdw6 = _mm_mul_ps(c6_00,FF);
251 /* CUBIC SPLINE TABLE REPULSION */
252 vfitab = _mm_add_epi32(vfitab,ifour);
253 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
254 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
255 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
256 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
257 _MM_TRANSPOSE4_PS(Y,F,G,H);
258 Heps = _mm_mul_ps(vfeps,H);
259 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
260 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
261 vvdw12 = _mm_mul_ps(c12_00,VV);
262 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
263 fvdw12 = _mm_mul_ps(c12_00,FF);
264 vvdw = _mm_add_ps(vvdw12,vvdw6);
265 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
267 /* Update potential sum for this i atom from the interaction with this j atom. */
268 velecsum = _mm_add_ps(velecsum,velec);
269 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
271 fscal = _mm_add_ps(felec,fvdw);
273 /* Calculate temporary vectorial force */
274 tx = _mm_mul_ps(fscal,dx00);
275 ty = _mm_mul_ps(fscal,dy00);
276 tz = _mm_mul_ps(fscal,dz00);
278 /* Update vectorial force */
279 fix0 = _mm_add_ps(fix0,tx);
280 fiy0 = _mm_add_ps(fiy0,ty);
281 fiz0 = _mm_add_ps(fiz0,tz);
283 fjptrA = f+j_coord_offsetA;
284 fjptrB = f+j_coord_offsetB;
285 fjptrC = f+j_coord_offsetC;
286 fjptrD = f+j_coord_offsetD;
287 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
289 /* Inner loop uses 75 flops */
295 /* Get j neighbor index, and coordinate index */
296 jnrlistA = jjnr[jidx];
297 jnrlistB = jjnr[jidx+1];
298 jnrlistC = jjnr[jidx+2];
299 jnrlistD = jjnr[jidx+3];
300 /* Sign of each element will be negative for non-real atoms.
301 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
302 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
304 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
305 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
306 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
307 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
308 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
309 j_coord_offsetA = DIM*jnrA;
310 j_coord_offsetB = DIM*jnrB;
311 j_coord_offsetC = DIM*jnrC;
312 j_coord_offsetD = DIM*jnrD;
314 /* load j atom coordinates */
315 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
316 x+j_coord_offsetC,x+j_coord_offsetD,
319 /* Calculate displacement vector */
320 dx00 = _mm_sub_ps(ix0,jx0);
321 dy00 = _mm_sub_ps(iy0,jy0);
322 dz00 = _mm_sub_ps(iz0,jz0);
324 /* Calculate squared distance and things based on it */
325 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
327 rinv00 = gmx_mm_invsqrt_ps(rsq00);
329 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
331 /* Load parameters for j particles */
332 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
333 charge+jnrC+0,charge+jnrD+0);
334 vdwjidx0A = 2*vdwtype[jnrA+0];
335 vdwjidx0B = 2*vdwtype[jnrB+0];
336 vdwjidx0C = 2*vdwtype[jnrC+0];
337 vdwjidx0D = 2*vdwtype[jnrD+0];
339 /**************************
340 * CALCULATE INTERACTIONS *
341 **************************/
343 r00 = _mm_mul_ps(rsq00,rinv00);
344 r00 = _mm_andnot_ps(dummy_mask,r00);
346 /* Compute parameters for interactions between i and j atoms */
347 qq00 = _mm_mul_ps(iq0,jq0);
348 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
349 vdwparam+vdwioffset0+vdwjidx0B,
350 vdwparam+vdwioffset0+vdwjidx0C,
351 vdwparam+vdwioffset0+vdwjidx0D,
354 /* Calculate table index by multiplying r with table scale and truncate to integer */
355 rt = _mm_mul_ps(r00,vftabscale);
356 vfitab = _mm_cvttps_epi32(rt);
357 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
358 vfitab = _mm_slli_epi32(vfitab,3);
360 /* EWALD ELECTROSTATICS */
362 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
363 ewrt = _mm_mul_ps(r00,ewtabscale);
364 ewitab = _mm_cvttps_epi32(ewrt);
365 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
366 ewitab = _mm_slli_epi32(ewitab,2);
367 ewtabF = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
368 ewtabD = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
369 ewtabV = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
370 ewtabFn = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
371 _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
372 felec = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
373 velec = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
374 velec = _mm_mul_ps(qq00,_mm_sub_ps(rinv00,velec));
375 felec = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
377 /* CUBIC SPLINE TABLE DISPERSION */
378 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
379 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
380 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
381 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
382 _MM_TRANSPOSE4_PS(Y,F,G,H);
383 Heps = _mm_mul_ps(vfeps,H);
384 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
385 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
386 vvdw6 = _mm_mul_ps(c6_00,VV);
387 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
388 fvdw6 = _mm_mul_ps(c6_00,FF);
390 /* CUBIC SPLINE TABLE REPULSION */
391 vfitab = _mm_add_epi32(vfitab,ifour);
392 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
393 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
394 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
395 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
396 _MM_TRANSPOSE4_PS(Y,F,G,H);
397 Heps = _mm_mul_ps(vfeps,H);
398 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
399 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
400 vvdw12 = _mm_mul_ps(c12_00,VV);
401 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
402 fvdw12 = _mm_mul_ps(c12_00,FF);
403 vvdw = _mm_add_ps(vvdw12,vvdw6);
404 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
406 /* Update potential sum for this i atom from the interaction with this j atom. */
407 velec = _mm_andnot_ps(dummy_mask,velec);
408 velecsum = _mm_add_ps(velecsum,velec);
409 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
410 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
412 fscal = _mm_add_ps(felec,fvdw);
414 fscal = _mm_andnot_ps(dummy_mask,fscal);
416 /* Calculate temporary vectorial force */
417 tx = _mm_mul_ps(fscal,dx00);
418 ty = _mm_mul_ps(fscal,dy00);
419 tz = _mm_mul_ps(fscal,dz00);
421 /* Update vectorial force */
422 fix0 = _mm_add_ps(fix0,tx);
423 fiy0 = _mm_add_ps(fiy0,ty);
424 fiz0 = _mm_add_ps(fiz0,tz);
426 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
427 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
428 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
429 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
430 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
432 /* Inner loop uses 76 flops */
435 /* End of innermost loop */
437 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
438 f+i_coord_offset,fshift+i_shift_offset);
441 /* Update potential energies */
442 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
443 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
445 /* Increment number of inner iterations */
446 inneriter += j_index_end - j_index_start;
448 /* Outer loop uses 9 flops */
451 /* Increment number of outer iterations */
454 /* Update outer/inner flops */
456 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*76);
459 * Gromacs nonbonded kernel: nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_sse4_1_single
460 * Electrostatics interaction: Ewald
461 * VdW interaction: CubicSplineTable
462 * Geometry: Particle-Particle
463 * Calculate force/pot: Force
466 nb_kernel_ElecEw_VdwCSTab_GeomP1P1_F_sse4_1_single
467 (t_nblist * gmx_restrict nlist,
468 rvec * gmx_restrict xx,
469 rvec * gmx_restrict ff,
470 t_forcerec * gmx_restrict fr,
471 t_mdatoms * gmx_restrict mdatoms,
472 nb_kernel_data_t * gmx_restrict kernel_data,
473 t_nrnb * gmx_restrict nrnb)
475 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
476 * just 0 for non-waters.
477 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
478 * jnr indices corresponding to data put in the four positions in the SIMD register.
480 int i_shift_offset,i_coord_offset,outeriter,inneriter;
481 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
482 int jnrA,jnrB,jnrC,jnrD;
483 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
484 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
485 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
487 real *shiftvec,*fshift,*x,*f;
488 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
490 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
492 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
493 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
494 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
495 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
496 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
499 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
502 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
503 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
505 __m128i ifour = _mm_set1_epi32(4);
506 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
509 __m128 ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
511 __m128 dummy_mask,cutoff_mask;
512 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
513 __m128 one = _mm_set1_ps(1.0);
514 __m128 two = _mm_set1_ps(2.0);
520 jindex = nlist->jindex;
522 shiftidx = nlist->shift;
524 shiftvec = fr->shift_vec[0];
525 fshift = fr->fshift[0];
526 facel = _mm_set1_ps(fr->epsfac);
527 charge = mdatoms->chargeA;
528 nvdwtype = fr->ntype;
530 vdwtype = mdatoms->typeA;
532 vftab = kernel_data->table_vdw->data;
533 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
535 sh_ewald = _mm_set1_ps(fr->ic->sh_ewald);
536 ewtab = fr->ic->tabq_coul_F;
537 ewtabscale = _mm_set1_ps(fr->ic->tabq_scale);
538 ewtabhalfspace = _mm_set1_ps(0.5/fr->ic->tabq_scale);
540 /* Avoid stupid compiler warnings */
541 jnrA = jnrB = jnrC = jnrD = 0;
550 for(iidx=0;iidx<4*DIM;iidx++)
555 /* Start outer loop over neighborlists */
556 for(iidx=0; iidx<nri; iidx++)
558 /* Load shift vector for this list */
559 i_shift_offset = DIM*shiftidx[iidx];
561 /* Load limits for loop over neighbors */
562 j_index_start = jindex[iidx];
563 j_index_end = jindex[iidx+1];
565 /* Get outer coordinate index */
567 i_coord_offset = DIM*inr;
569 /* Load i particle coords and add shift vector */
570 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
572 fix0 = _mm_setzero_ps();
573 fiy0 = _mm_setzero_ps();
574 fiz0 = _mm_setzero_ps();
576 /* Load parameters for i particles */
577 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
578 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
580 /* Start inner kernel loop */
581 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
584 /* Get j neighbor index, and coordinate index */
589 j_coord_offsetA = DIM*jnrA;
590 j_coord_offsetB = DIM*jnrB;
591 j_coord_offsetC = DIM*jnrC;
592 j_coord_offsetD = DIM*jnrD;
594 /* load j atom coordinates */
595 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
596 x+j_coord_offsetC,x+j_coord_offsetD,
599 /* Calculate displacement vector */
600 dx00 = _mm_sub_ps(ix0,jx0);
601 dy00 = _mm_sub_ps(iy0,jy0);
602 dz00 = _mm_sub_ps(iz0,jz0);
604 /* Calculate squared distance and things based on it */
605 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
607 rinv00 = gmx_mm_invsqrt_ps(rsq00);
609 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
611 /* Load parameters for j particles */
612 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
613 charge+jnrC+0,charge+jnrD+0);
614 vdwjidx0A = 2*vdwtype[jnrA+0];
615 vdwjidx0B = 2*vdwtype[jnrB+0];
616 vdwjidx0C = 2*vdwtype[jnrC+0];
617 vdwjidx0D = 2*vdwtype[jnrD+0];
619 /**************************
620 * CALCULATE INTERACTIONS *
621 **************************/
623 r00 = _mm_mul_ps(rsq00,rinv00);
625 /* Compute parameters for interactions between i and j atoms */
626 qq00 = _mm_mul_ps(iq0,jq0);
627 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
628 vdwparam+vdwioffset0+vdwjidx0B,
629 vdwparam+vdwioffset0+vdwjidx0C,
630 vdwparam+vdwioffset0+vdwjidx0D,
633 /* Calculate table index by multiplying r with table scale and truncate to integer */
634 rt = _mm_mul_ps(r00,vftabscale);
635 vfitab = _mm_cvttps_epi32(rt);
636 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
637 vfitab = _mm_slli_epi32(vfitab,3);
639 /* EWALD ELECTROSTATICS */
641 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
642 ewrt = _mm_mul_ps(r00,ewtabscale);
643 ewitab = _mm_cvttps_epi32(ewrt);
644 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
645 gmx_mm_load_4pair_swizzle_ps(ewtab + gmx_mm_extract_epi32(ewitab,0),ewtab + gmx_mm_extract_epi32(ewitab,1),
646 ewtab + gmx_mm_extract_epi32(ewitab,2),ewtab + gmx_mm_extract_epi32(ewitab,3),
648 felec = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
649 felec = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
651 /* CUBIC SPLINE TABLE DISPERSION */
652 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
653 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
654 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
655 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
656 _MM_TRANSPOSE4_PS(Y,F,G,H);
657 Heps = _mm_mul_ps(vfeps,H);
658 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
659 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
660 fvdw6 = _mm_mul_ps(c6_00,FF);
662 /* CUBIC SPLINE TABLE REPULSION */
663 vfitab = _mm_add_epi32(vfitab,ifour);
664 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
665 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
666 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
667 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
668 _MM_TRANSPOSE4_PS(Y,F,G,H);
669 Heps = _mm_mul_ps(vfeps,H);
670 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
671 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
672 fvdw12 = _mm_mul_ps(c12_00,FF);
673 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
675 fscal = _mm_add_ps(felec,fvdw);
677 /* Calculate temporary vectorial force */
678 tx = _mm_mul_ps(fscal,dx00);
679 ty = _mm_mul_ps(fscal,dy00);
680 tz = _mm_mul_ps(fscal,dz00);
682 /* Update vectorial force */
683 fix0 = _mm_add_ps(fix0,tx);
684 fiy0 = _mm_add_ps(fiy0,ty);
685 fiz0 = _mm_add_ps(fiz0,tz);
687 fjptrA = f+j_coord_offsetA;
688 fjptrB = f+j_coord_offsetB;
689 fjptrC = f+j_coord_offsetC;
690 fjptrD = f+j_coord_offsetD;
691 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
693 /* Inner loop uses 62 flops */
699 /* Get j neighbor index, and coordinate index */
700 jnrlistA = jjnr[jidx];
701 jnrlistB = jjnr[jidx+1];
702 jnrlistC = jjnr[jidx+2];
703 jnrlistD = jjnr[jidx+3];
704 /* Sign of each element will be negative for non-real atoms.
705 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
706 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
708 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
709 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
710 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
711 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
712 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
713 j_coord_offsetA = DIM*jnrA;
714 j_coord_offsetB = DIM*jnrB;
715 j_coord_offsetC = DIM*jnrC;
716 j_coord_offsetD = DIM*jnrD;
718 /* load j atom coordinates */
719 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
720 x+j_coord_offsetC,x+j_coord_offsetD,
723 /* Calculate displacement vector */
724 dx00 = _mm_sub_ps(ix0,jx0);
725 dy00 = _mm_sub_ps(iy0,jy0);
726 dz00 = _mm_sub_ps(iz0,jz0);
728 /* Calculate squared distance and things based on it */
729 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
731 rinv00 = gmx_mm_invsqrt_ps(rsq00);
733 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
735 /* Load parameters for j particles */
736 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
737 charge+jnrC+0,charge+jnrD+0);
738 vdwjidx0A = 2*vdwtype[jnrA+0];
739 vdwjidx0B = 2*vdwtype[jnrB+0];
740 vdwjidx0C = 2*vdwtype[jnrC+0];
741 vdwjidx0D = 2*vdwtype[jnrD+0];
743 /**************************
744 * CALCULATE INTERACTIONS *
745 **************************/
747 r00 = _mm_mul_ps(rsq00,rinv00);
748 r00 = _mm_andnot_ps(dummy_mask,r00);
750 /* Compute parameters for interactions between i and j atoms */
751 qq00 = _mm_mul_ps(iq0,jq0);
752 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
753 vdwparam+vdwioffset0+vdwjidx0B,
754 vdwparam+vdwioffset0+vdwjidx0C,
755 vdwparam+vdwioffset0+vdwjidx0D,
758 /* Calculate table index by multiplying r with table scale and truncate to integer */
759 rt = _mm_mul_ps(r00,vftabscale);
760 vfitab = _mm_cvttps_epi32(rt);
761 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
762 vfitab = _mm_slli_epi32(vfitab,3);
764 /* EWALD ELECTROSTATICS */
766 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
767 ewrt = _mm_mul_ps(r00,ewtabscale);
768 ewitab = _mm_cvttps_epi32(ewrt);
769 eweps = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
770 gmx_mm_load_4pair_swizzle_ps(ewtab + gmx_mm_extract_epi32(ewitab,0),ewtab + gmx_mm_extract_epi32(ewitab,1),
771 ewtab + gmx_mm_extract_epi32(ewitab,2),ewtab + gmx_mm_extract_epi32(ewitab,3),
773 felec = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
774 felec = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
776 /* CUBIC SPLINE TABLE DISPERSION */
777 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
778 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
779 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
780 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
781 _MM_TRANSPOSE4_PS(Y,F,G,H);
782 Heps = _mm_mul_ps(vfeps,H);
783 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
784 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
785 fvdw6 = _mm_mul_ps(c6_00,FF);
787 /* CUBIC SPLINE TABLE REPULSION */
788 vfitab = _mm_add_epi32(vfitab,ifour);
789 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
790 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
791 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
792 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
793 _MM_TRANSPOSE4_PS(Y,F,G,H);
794 Heps = _mm_mul_ps(vfeps,H);
795 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
796 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
797 fvdw12 = _mm_mul_ps(c12_00,FF);
798 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
800 fscal = _mm_add_ps(felec,fvdw);
802 fscal = _mm_andnot_ps(dummy_mask,fscal);
804 /* Calculate temporary vectorial force */
805 tx = _mm_mul_ps(fscal,dx00);
806 ty = _mm_mul_ps(fscal,dy00);
807 tz = _mm_mul_ps(fscal,dz00);
809 /* Update vectorial force */
810 fix0 = _mm_add_ps(fix0,tx);
811 fiy0 = _mm_add_ps(fiy0,ty);
812 fiz0 = _mm_add_ps(fiz0,tz);
814 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
815 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
816 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
817 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
818 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
820 /* Inner loop uses 63 flops */
823 /* End of innermost loop */
825 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
826 f+i_coord_offset,fshift+i_shift_offset);
828 /* Increment number of inner iterations */
829 inneriter += j_index_end - j_index_start;
831 /* Outer loop uses 7 flops */
834 /* Increment number of outer iterations */
837 /* Update outer/inner flops */
839 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*63);