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_ElecRFCut_VdwCSTab_GeomP1P1_VF_sse4_1_single
38 * Electrostatics interaction: ReactionField
39 * VdW interaction: CubicSplineTable
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecRFCut_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;
86 __m128 dummy_mask,cutoff_mask;
87 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
88 __m128 one = _mm_set1_ps(1.0);
89 __m128 two = _mm_set1_ps(2.0);
95 jindex = nlist->jindex;
97 shiftidx = nlist->shift;
99 shiftvec = fr->shift_vec[0];
100 fshift = fr->fshift[0];
101 facel = _mm_set1_ps(fr->epsfac);
102 charge = mdatoms->chargeA;
103 krf = _mm_set1_ps(fr->ic->k_rf);
104 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
105 crf = _mm_set1_ps(fr->ic->c_rf);
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 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
114 rcutoff_scalar = fr->rcoulomb;
115 rcutoff = _mm_set1_ps(rcutoff_scalar);
116 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
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 if (gmx_mm_any_lt(rsq00,rcutoff2))
208 r00 = _mm_mul_ps(rsq00,rinv00);
210 /* Compute parameters for interactions between i and j atoms */
211 qq00 = _mm_mul_ps(iq0,jq0);
212 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
213 vdwparam+vdwioffset0+vdwjidx0B,
214 vdwparam+vdwioffset0+vdwjidx0C,
215 vdwparam+vdwioffset0+vdwjidx0D,
218 /* Calculate table index by multiplying r with table scale and truncate to integer */
219 rt = _mm_mul_ps(r00,vftabscale);
220 vfitab = _mm_cvttps_epi32(rt);
221 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
222 vfitab = _mm_slli_epi32(vfitab,3);
224 /* REACTION-FIELD ELECTROSTATICS */
225 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
226 felec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
228 /* CUBIC SPLINE TABLE DISPERSION */
229 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
230 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
231 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
232 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
233 _MM_TRANSPOSE4_PS(Y,F,G,H);
234 Heps = _mm_mul_ps(vfeps,H);
235 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
236 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
237 vvdw6 = _mm_mul_ps(c6_00,VV);
238 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
239 fvdw6 = _mm_mul_ps(c6_00,FF);
241 /* CUBIC SPLINE TABLE REPULSION */
242 vfitab = _mm_add_epi32(vfitab,ifour);
243 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
244 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
245 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
246 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
247 _MM_TRANSPOSE4_PS(Y,F,G,H);
248 Heps = _mm_mul_ps(vfeps,H);
249 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
250 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
251 vvdw12 = _mm_mul_ps(c12_00,VV);
252 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
253 fvdw12 = _mm_mul_ps(c12_00,FF);
254 vvdw = _mm_add_ps(vvdw12,vvdw6);
255 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
257 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
259 /* Update potential sum for this i atom from the interaction with this j atom. */
260 velec = _mm_and_ps(velec,cutoff_mask);
261 velecsum = _mm_add_ps(velecsum,velec);
262 vvdw = _mm_and_ps(vvdw,cutoff_mask);
263 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
265 fscal = _mm_add_ps(felec,fvdw);
267 fscal = _mm_and_ps(fscal,cutoff_mask);
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);
287 /* Inner loop uses 72 flops */
293 /* Get j neighbor index, and coordinate index */
294 jnrlistA = jjnr[jidx];
295 jnrlistB = jjnr[jidx+1];
296 jnrlistC = jjnr[jidx+2];
297 jnrlistD = jjnr[jidx+3];
298 /* Sign of each element will be negative for non-real atoms.
299 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
300 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
302 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
303 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
304 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
305 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
306 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
307 j_coord_offsetA = DIM*jnrA;
308 j_coord_offsetB = DIM*jnrB;
309 j_coord_offsetC = DIM*jnrC;
310 j_coord_offsetD = DIM*jnrD;
312 /* load j atom coordinates */
313 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
314 x+j_coord_offsetC,x+j_coord_offsetD,
317 /* Calculate displacement vector */
318 dx00 = _mm_sub_ps(ix0,jx0);
319 dy00 = _mm_sub_ps(iy0,jy0);
320 dz00 = _mm_sub_ps(iz0,jz0);
322 /* Calculate squared distance and things based on it */
323 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
325 rinv00 = gmx_mm_invsqrt_ps(rsq00);
327 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
329 /* Load parameters for j particles */
330 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
331 charge+jnrC+0,charge+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 if (gmx_mm_any_lt(rsq00,rcutoff2))
344 r00 = _mm_mul_ps(rsq00,rinv00);
345 r00 = _mm_andnot_ps(dummy_mask,r00);
347 /* Compute parameters for interactions between i and j atoms */
348 qq00 = _mm_mul_ps(iq0,jq0);
349 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
350 vdwparam+vdwioffset0+vdwjidx0B,
351 vdwparam+vdwioffset0+vdwjidx0C,
352 vdwparam+vdwioffset0+vdwjidx0D,
355 /* Calculate table index by multiplying r with table scale and truncate to integer */
356 rt = _mm_mul_ps(r00,vftabscale);
357 vfitab = _mm_cvttps_epi32(rt);
358 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
359 vfitab = _mm_slli_epi32(vfitab,3);
361 /* REACTION-FIELD ELECTROSTATICS */
362 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
363 felec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
365 /* CUBIC SPLINE TABLE DISPERSION */
366 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
367 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
368 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
369 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
370 _MM_TRANSPOSE4_PS(Y,F,G,H);
371 Heps = _mm_mul_ps(vfeps,H);
372 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
373 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
374 vvdw6 = _mm_mul_ps(c6_00,VV);
375 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
376 fvdw6 = _mm_mul_ps(c6_00,FF);
378 /* CUBIC SPLINE TABLE REPULSION */
379 vfitab = _mm_add_epi32(vfitab,ifour);
380 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
381 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
382 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
383 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
384 _MM_TRANSPOSE4_PS(Y,F,G,H);
385 Heps = _mm_mul_ps(vfeps,H);
386 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
387 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
388 vvdw12 = _mm_mul_ps(c12_00,VV);
389 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
390 fvdw12 = _mm_mul_ps(c12_00,FF);
391 vvdw = _mm_add_ps(vvdw12,vvdw6);
392 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
394 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
396 /* Update potential sum for this i atom from the interaction with this j atom. */
397 velec = _mm_and_ps(velec,cutoff_mask);
398 velec = _mm_andnot_ps(dummy_mask,velec);
399 velecsum = _mm_add_ps(velecsum,velec);
400 vvdw = _mm_and_ps(vvdw,cutoff_mask);
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_and_ps(fscal,cutoff_mask);
408 fscal = _mm_andnot_ps(dummy_mask,fscal);
410 /* Calculate temporary vectorial force */
411 tx = _mm_mul_ps(fscal,dx00);
412 ty = _mm_mul_ps(fscal,dy00);
413 tz = _mm_mul_ps(fscal,dz00);
415 /* Update vectorial force */
416 fix0 = _mm_add_ps(fix0,tx);
417 fiy0 = _mm_add_ps(fiy0,ty);
418 fiz0 = _mm_add_ps(fiz0,tz);
420 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
421 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
422 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
423 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
424 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
428 /* Inner loop uses 73 flops */
431 /* End of innermost loop */
433 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
434 f+i_coord_offset,fshift+i_shift_offset);
437 /* Update potential energies */
438 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
439 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
441 /* Increment number of inner iterations */
442 inneriter += j_index_end - j_index_start;
444 /* Outer loop uses 9 flops */
447 /* Increment number of outer iterations */
450 /* Update outer/inner flops */
452 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*73);
455 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_sse4_1_single
456 * Electrostatics interaction: ReactionField
457 * VdW interaction: CubicSplineTable
458 * Geometry: Particle-Particle
459 * Calculate force/pot: Force
462 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_sse4_1_single
463 (t_nblist * gmx_restrict nlist,
464 rvec * gmx_restrict xx,
465 rvec * gmx_restrict ff,
466 t_forcerec * gmx_restrict fr,
467 t_mdatoms * gmx_restrict mdatoms,
468 nb_kernel_data_t * gmx_restrict kernel_data,
469 t_nrnb * gmx_restrict nrnb)
471 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
472 * just 0 for non-waters.
473 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
474 * jnr indices corresponding to data put in the four positions in the SIMD register.
476 int i_shift_offset,i_coord_offset,outeriter,inneriter;
477 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
478 int jnrA,jnrB,jnrC,jnrD;
479 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
480 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
481 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
483 real *shiftvec,*fshift,*x,*f;
484 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
486 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
488 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
489 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
490 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
491 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
492 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
495 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
498 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
499 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
501 __m128i ifour = _mm_set1_epi32(4);
502 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
504 __m128 dummy_mask,cutoff_mask;
505 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
506 __m128 one = _mm_set1_ps(1.0);
507 __m128 two = _mm_set1_ps(2.0);
513 jindex = nlist->jindex;
515 shiftidx = nlist->shift;
517 shiftvec = fr->shift_vec[0];
518 fshift = fr->fshift[0];
519 facel = _mm_set1_ps(fr->epsfac);
520 charge = mdatoms->chargeA;
521 krf = _mm_set1_ps(fr->ic->k_rf);
522 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
523 crf = _mm_set1_ps(fr->ic->c_rf);
524 nvdwtype = fr->ntype;
526 vdwtype = mdatoms->typeA;
528 vftab = kernel_data->table_vdw->data;
529 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
531 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
532 rcutoff_scalar = fr->rcoulomb;
533 rcutoff = _mm_set1_ps(rcutoff_scalar);
534 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
536 /* Avoid stupid compiler warnings */
537 jnrA = jnrB = jnrC = jnrD = 0;
546 for(iidx=0;iidx<4*DIM;iidx++)
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_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
568 fix0 = _mm_setzero_ps();
569 fiy0 = _mm_setzero_ps();
570 fiz0 = _mm_setzero_ps();
572 /* Load parameters for i particles */
573 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
574 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
576 /* Start inner kernel loop */
577 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
580 /* Get j neighbor index, and coordinate index */
585 j_coord_offsetA = DIM*jnrA;
586 j_coord_offsetB = DIM*jnrB;
587 j_coord_offsetC = DIM*jnrC;
588 j_coord_offsetD = DIM*jnrD;
590 /* load j atom coordinates */
591 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
592 x+j_coord_offsetC,x+j_coord_offsetD,
595 /* Calculate displacement vector */
596 dx00 = _mm_sub_ps(ix0,jx0);
597 dy00 = _mm_sub_ps(iy0,jy0);
598 dz00 = _mm_sub_ps(iz0,jz0);
600 /* Calculate squared distance and things based on it */
601 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
603 rinv00 = gmx_mm_invsqrt_ps(rsq00);
605 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
607 /* Load parameters for j particles */
608 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
609 charge+jnrC+0,charge+jnrD+0);
610 vdwjidx0A = 2*vdwtype[jnrA+0];
611 vdwjidx0B = 2*vdwtype[jnrB+0];
612 vdwjidx0C = 2*vdwtype[jnrC+0];
613 vdwjidx0D = 2*vdwtype[jnrD+0];
615 /**************************
616 * CALCULATE INTERACTIONS *
617 **************************/
619 if (gmx_mm_any_lt(rsq00,rcutoff2))
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 /* Calculate table index by multiplying r with table scale and truncate to integer */
633 rt = _mm_mul_ps(r00,vftabscale);
634 vfitab = _mm_cvttps_epi32(rt);
635 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
636 vfitab = _mm_slli_epi32(vfitab,3);
638 /* REACTION-FIELD ELECTROSTATICS */
639 felec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
641 /* CUBIC SPLINE TABLE DISPERSION */
642 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
643 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
644 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
645 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
646 _MM_TRANSPOSE4_PS(Y,F,G,H);
647 Heps = _mm_mul_ps(vfeps,H);
648 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
649 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
650 fvdw6 = _mm_mul_ps(c6_00,FF);
652 /* CUBIC SPLINE TABLE REPULSION */
653 vfitab = _mm_add_epi32(vfitab,ifour);
654 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
655 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
656 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
657 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
658 _MM_TRANSPOSE4_PS(Y,F,G,H);
659 Heps = _mm_mul_ps(vfeps,H);
660 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
661 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
662 fvdw12 = _mm_mul_ps(c12_00,FF);
663 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
665 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
667 fscal = _mm_add_ps(felec,fvdw);
669 fscal = _mm_and_ps(fscal,cutoff_mask);
671 /* Calculate temporary vectorial force */
672 tx = _mm_mul_ps(fscal,dx00);
673 ty = _mm_mul_ps(fscal,dy00);
674 tz = _mm_mul_ps(fscal,dz00);
676 /* Update vectorial force */
677 fix0 = _mm_add_ps(fix0,tx);
678 fiy0 = _mm_add_ps(fiy0,ty);
679 fiz0 = _mm_add_ps(fiz0,tz);
681 fjptrA = f+j_coord_offsetA;
682 fjptrB = f+j_coord_offsetB;
683 fjptrC = f+j_coord_offsetC;
684 fjptrD = f+j_coord_offsetD;
685 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
689 /* Inner loop uses 57 flops */
695 /* Get j neighbor index, and coordinate index */
696 jnrlistA = jjnr[jidx];
697 jnrlistB = jjnr[jidx+1];
698 jnrlistC = jjnr[jidx+2];
699 jnrlistD = jjnr[jidx+3];
700 /* Sign of each element will be negative for non-real atoms.
701 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
702 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
704 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
705 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
706 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
707 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
708 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
709 j_coord_offsetA = DIM*jnrA;
710 j_coord_offsetB = DIM*jnrB;
711 j_coord_offsetC = DIM*jnrC;
712 j_coord_offsetD = DIM*jnrD;
714 /* load j atom coordinates */
715 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
716 x+j_coord_offsetC,x+j_coord_offsetD,
719 /* Calculate displacement vector */
720 dx00 = _mm_sub_ps(ix0,jx0);
721 dy00 = _mm_sub_ps(iy0,jy0);
722 dz00 = _mm_sub_ps(iz0,jz0);
724 /* Calculate squared distance and things based on it */
725 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
727 rinv00 = gmx_mm_invsqrt_ps(rsq00);
729 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
731 /* Load parameters for j particles */
732 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
733 charge+jnrC+0,charge+jnrD+0);
734 vdwjidx0A = 2*vdwtype[jnrA+0];
735 vdwjidx0B = 2*vdwtype[jnrB+0];
736 vdwjidx0C = 2*vdwtype[jnrC+0];
737 vdwjidx0D = 2*vdwtype[jnrD+0];
739 /**************************
740 * CALCULATE INTERACTIONS *
741 **************************/
743 if (gmx_mm_any_lt(rsq00,rcutoff2))
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 /* Calculate table index by multiplying r with table scale and truncate to integer */
758 rt = _mm_mul_ps(r00,vftabscale);
759 vfitab = _mm_cvttps_epi32(rt);
760 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
761 vfitab = _mm_slli_epi32(vfitab,3);
763 /* REACTION-FIELD ELECTROSTATICS */
764 felec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
766 /* CUBIC SPLINE TABLE DISPERSION */
767 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
768 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
769 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
770 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
771 _MM_TRANSPOSE4_PS(Y,F,G,H);
772 Heps = _mm_mul_ps(vfeps,H);
773 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
774 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
775 fvdw6 = _mm_mul_ps(c6_00,FF);
777 /* CUBIC SPLINE TABLE REPULSION */
778 vfitab = _mm_add_epi32(vfitab,ifour);
779 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
780 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
781 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
782 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
783 _MM_TRANSPOSE4_PS(Y,F,G,H);
784 Heps = _mm_mul_ps(vfeps,H);
785 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
786 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
787 fvdw12 = _mm_mul_ps(c12_00,FF);
788 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
790 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
792 fscal = _mm_add_ps(felec,fvdw);
794 fscal = _mm_and_ps(fscal,cutoff_mask);
796 fscal = _mm_andnot_ps(dummy_mask,fscal);
798 /* Calculate temporary vectorial force */
799 tx = _mm_mul_ps(fscal,dx00);
800 ty = _mm_mul_ps(fscal,dy00);
801 tz = _mm_mul_ps(fscal,dz00);
803 /* Update vectorial force */
804 fix0 = _mm_add_ps(fix0,tx);
805 fiy0 = _mm_add_ps(fiy0,ty);
806 fiz0 = _mm_add_ps(fiz0,tz);
808 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
809 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
810 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
811 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
812 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
816 /* Inner loop uses 58 flops */
819 /* End of innermost loop */
821 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
822 f+i_coord_offset,fshift+i_shift_offset);
824 /* Increment number of inner iterations */
825 inneriter += j_index_end - j_index_start;
827 /* Outer loop uses 7 flops */
830 /* Increment number of outer iterations */
833 /* Update outer/inner flops */
835 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*58);