?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Removed include simple.h from nb_kernel_avx_128_fma_XX
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_avx_128_fma_single / nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_avx_128_fma_single.c
blob66ba31aef911a3a452f1795073fde6d2a4c1d68c
1 /*
2 * This file is part of the GROMACS molecular simulation package.
3 *
4 * Copyright (c) 2012,2013,2014,2015, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
8 *
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.
13 *
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.
18 *
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.
23 *
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.
31 *
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.
34 */
35 /*
36 * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
37 */
38 #include "gmxpre.h"
39
40 #include "config.h"
41
42 #include <math.h>
43
44 #include "../nb_kernel.h"
45 #include "gromacs/math/vec.h"
46 #include "gromacs/legacyheaders/nrnb.h"
47
48 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
49 #include "kernelutil_x86_avx_128_fma_single.h"
50
51 /*
52 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
53 * Electrostatics interaction: Coulomb
54 * VdW interaction: CubicSplineTable
55 * Geometry: Particle-Particle
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
60 (t_nblist * gmx_restrict nlist,
61 rvec * gmx_restrict xx,
62 rvec * gmx_restrict ff,
63 t_forcerec * gmx_restrict fr,
64 t_mdatoms * gmx_restrict mdatoms,
65 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
66 t_nrnb * gmx_restrict nrnb)
67 {
68 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
69 * just 0 for non-waters.
70 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
71 * jnr indices corresponding to data put in the four positions in the SIMD register.
72 */
73 int i_shift_offset,i_coord_offset,outeriter,inneriter;
74 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
75 int jnrA,jnrB,jnrC,jnrD;
76 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
77 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
78 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
79 real rcutoff_scalar;
80 real *shiftvec,*fshift,*x,*f;
81 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
82 real scratch[4*DIM];
83 __m128 fscal,rcutoff,rcutoff2,jidxall;
84 int vdwioffset0;
85 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
86 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
87 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
88 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
89 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
90 real *charge;
91 int nvdwtype;
92 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
93 int *vdwtype;
94 real *vdwparam;
95 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
96 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
97 __m128i vfitab;
98 __m128i ifour = _mm_set1_epi32(4);
99 __m128 rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
100 real *vftab;
101 __m128 dummy_mask,cutoff_mask;
102 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
103 __m128 one = _mm_set1_ps(1.0);
104 __m128 two = _mm_set1_ps(2.0);
105 x = xx[0];
106 f = ff[0];
107
108 nri = nlist->nri;
109 iinr = nlist->iinr;
110 jindex = nlist->jindex;
111 jjnr = nlist->jjnr;
112 shiftidx = nlist->shift;
113 gid = nlist->gid;
114 shiftvec = fr->shift_vec[0];
115 fshift = fr->fshift[0];
116 facel = _mm_set1_ps(fr->epsfac);
117 charge = mdatoms->chargeA;
118 nvdwtype = fr->ntype;
119 vdwparam = fr->nbfp;
120 vdwtype = mdatoms->typeA;
121
122 vftab = kernel_data->table_vdw->data;
123 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
124
125 /* Avoid stupid compiler warnings */
126 jnrA = jnrB = jnrC = jnrD = 0;
127 j_coord_offsetA = 0;
128 j_coord_offsetB = 0;
129 j_coord_offsetC = 0;
130 j_coord_offsetD = 0;
131
132 outeriter = 0;
133 inneriter = 0;
134
135 for(iidx=0;iidx<4*DIM;iidx++)
136 {
137 scratch[iidx] = 0.0;
138 }
139
140 /* Start outer loop over neighborlists */
141 for(iidx=0; iidx<nri; iidx++)
142 {
143 /* Load shift vector for this list */
144 i_shift_offset = DIM*shiftidx[iidx];
145
146 /* Load limits for loop over neighbors */
147 j_index_start = jindex[iidx];
148 j_index_end = jindex[iidx+1];
149
150 /* Get outer coordinate index */
151 inr = iinr[iidx];
152 i_coord_offset = DIM*inr;
153
154 /* Load i particle coords and add shift vector */
155 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
156
157 fix0 = _mm_setzero_ps();
158 fiy0 = _mm_setzero_ps();
159 fiz0 = _mm_setzero_ps();
160
161 /* Load parameters for i particles */
162 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
163 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
164
165 /* Reset potential sums */
166 velecsum = _mm_setzero_ps();
167 vvdwsum = _mm_setzero_ps();
168
169 /* Start inner kernel loop */
170 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
171 {
172
173 /* Get j neighbor index, and coordinate index */
174 jnrA = jjnr[jidx];
175 jnrB = jjnr[jidx+1];
176 jnrC = jjnr[jidx+2];
177 jnrD = jjnr[jidx+3];
178 j_coord_offsetA = DIM*jnrA;
179 j_coord_offsetB = DIM*jnrB;
180 j_coord_offsetC = DIM*jnrC;
181 j_coord_offsetD = DIM*jnrD;
182
183 /* load j atom coordinates */
184 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
185 x+j_coord_offsetC,x+j_coord_offsetD,
186 &jx0,&jy0,&jz0);
187
188 /* Calculate displacement vector */
189 dx00 = _mm_sub_ps(ix0,jx0);
190 dy00 = _mm_sub_ps(iy0,jy0);
191 dz00 = _mm_sub_ps(iz0,jz0);
192
193 /* Calculate squared distance and things based on it */
194 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
195
196 rinv00 = gmx_mm_invsqrt_ps(rsq00);
197
198 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
199
200 /* Load parameters for j particles */
201 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
202 charge+jnrC+0,charge+jnrD+0);
203 vdwjidx0A = 2*vdwtype[jnrA+0];
204 vdwjidx0B = 2*vdwtype[jnrB+0];
205 vdwjidx0C = 2*vdwtype[jnrC+0];
206 vdwjidx0D = 2*vdwtype[jnrD+0];
207
208 /**************************
209 * CALCULATE INTERACTIONS *
210 **************************/
211
212 r00 = _mm_mul_ps(rsq00,rinv00);
213
214 /* Compute parameters for interactions between i and j atoms */
215 qq00 = _mm_mul_ps(iq0,jq0);
216 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
217 vdwparam+vdwioffset0+vdwjidx0B,
218 vdwparam+vdwioffset0+vdwjidx0C,
219 vdwparam+vdwioffset0+vdwjidx0D,
220 &c6_00,&c12_00);
221
222 /* Calculate table index by multiplying r with table scale and truncate to integer */
223 rt = _mm_mul_ps(r00,vftabscale);
224 vfitab = _mm_cvttps_epi32(rt);
225 #ifdef __XOP__
226 vfeps = _mm_frcz_ps(rt);
227 #else
228 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
229 #endif
230 twovfeps = _mm_add_ps(vfeps,vfeps);
231 vfitab = _mm_slli_epi32(vfitab,3);
232
233 /* COULOMB ELECTROSTATICS */
234 velec = _mm_mul_ps(qq00,rinv00);
235 felec = _mm_mul_ps(velec,rinvsq00);
236
237 /* CUBIC SPLINE TABLE DISPERSION */
238 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
239 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
240 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
241 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
242 _MM_TRANSPOSE4_PS(Y,F,G,H);
243 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
244 VV = _mm_macc_ps(vfeps,Fp,Y);
245 vvdw6 = _mm_mul_ps(c6_00,VV);
246 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
247 fvdw6 = _mm_mul_ps(c6_00,FF);
248
249 /* CUBIC SPLINE TABLE REPULSION */
250 vfitab = _mm_add_epi32(vfitab,ifour);
251 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
252 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
253 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
254 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
255 _MM_TRANSPOSE4_PS(Y,F,G,H);
256 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
257 VV = _mm_macc_ps(vfeps,Fp,Y);
258 vvdw12 = _mm_mul_ps(c12_00,VV);
259 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
260 fvdw12 = _mm_mul_ps(c12_00,FF);
261 vvdw = _mm_add_ps(vvdw12,vvdw6);
262 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
263
264 /* Update potential sum for this i atom from the interaction with this j atom. */
265 velecsum = _mm_add_ps(velecsum,velec);
266 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
267
268 fscal = _mm_add_ps(felec,fvdw);
269
270 /* Update vectorial force */
271 fix0 = _mm_macc_ps(dx00,fscal,fix0);
272 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
273 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
274
275 fjptrA = f+j_coord_offsetA;
276 fjptrB = f+j_coord_offsetB;
277 fjptrC = f+j_coord_offsetC;
278 fjptrD = f+j_coord_offsetD;
279 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
280 _mm_mul_ps(dx00,fscal),
281 _mm_mul_ps(dy00,fscal),
282 _mm_mul_ps(dz00,fscal));
283
284 /* Inner loop uses 66 flops */
285 }
286
287 if(jidx<j_index_end)
288 {
289
290 /* Get j neighbor index, and coordinate index */
291 jnrlistA = jjnr[jidx];
292 jnrlistB = jjnr[jidx+1];
293 jnrlistC = jjnr[jidx+2];
294 jnrlistD = jjnr[jidx+3];
295 /* Sign of each element will be negative for non-real atoms.
296 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
297 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
298 */
299 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
300 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
301 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
302 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
303 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
304 j_coord_offsetA = DIM*jnrA;
305 j_coord_offsetB = DIM*jnrB;
306 j_coord_offsetC = DIM*jnrC;
307 j_coord_offsetD = DIM*jnrD;
308
309 /* load j atom coordinates */
310 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
311 x+j_coord_offsetC,x+j_coord_offsetD,
312 &jx0,&jy0,&jz0);
313
314 /* Calculate displacement vector */
315 dx00 = _mm_sub_ps(ix0,jx0);
316 dy00 = _mm_sub_ps(iy0,jy0);
317 dz00 = _mm_sub_ps(iz0,jz0);
318
319 /* Calculate squared distance and things based on it */
320 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
321
322 rinv00 = gmx_mm_invsqrt_ps(rsq00);
323
324 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
325
326 /* Load parameters for j particles */
327 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
328 charge+jnrC+0,charge+jnrD+0);
329 vdwjidx0A = 2*vdwtype[jnrA+0];
330 vdwjidx0B = 2*vdwtype[jnrB+0];
331 vdwjidx0C = 2*vdwtype[jnrC+0];
332 vdwjidx0D = 2*vdwtype[jnrD+0];
333
334 /**************************
335 * CALCULATE INTERACTIONS *
336 **************************/
337
338 r00 = _mm_mul_ps(rsq00,rinv00);
339 r00 = _mm_andnot_ps(dummy_mask,r00);
340
341 /* Compute parameters for interactions between i and j atoms */
342 qq00 = _mm_mul_ps(iq0,jq0);
343 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
344 vdwparam+vdwioffset0+vdwjidx0B,
345 vdwparam+vdwioffset0+vdwjidx0C,
346 vdwparam+vdwioffset0+vdwjidx0D,
347 &c6_00,&c12_00);
348
349 /* Calculate table index by multiplying r with table scale and truncate to integer */
350 rt = _mm_mul_ps(r00,vftabscale);
351 vfitab = _mm_cvttps_epi32(rt);
352 #ifdef __XOP__
353 vfeps = _mm_frcz_ps(rt);
354 #else
355 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
356 #endif
357 twovfeps = _mm_add_ps(vfeps,vfeps);
358 vfitab = _mm_slli_epi32(vfitab,3);
359
360 /* COULOMB ELECTROSTATICS */
361 velec = _mm_mul_ps(qq00,rinv00);
362 felec = _mm_mul_ps(velec,rinvsq00);
363
364 /* CUBIC SPLINE TABLE DISPERSION */
365 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
366 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
367 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
368 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
369 _MM_TRANSPOSE4_PS(Y,F,G,H);
370 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
371 VV = _mm_macc_ps(vfeps,Fp,Y);
372 vvdw6 = _mm_mul_ps(c6_00,VV);
373 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
374 fvdw6 = _mm_mul_ps(c6_00,FF);
375
376 /* CUBIC SPLINE TABLE REPULSION */
377 vfitab = _mm_add_epi32(vfitab,ifour);
378 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
379 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
380 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
381 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
382 _MM_TRANSPOSE4_PS(Y,F,G,H);
383 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
384 VV = _mm_macc_ps(vfeps,Fp,Y);
385 vvdw12 = _mm_mul_ps(c12_00,VV);
386 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
387 fvdw12 = _mm_mul_ps(c12_00,FF);
388 vvdw = _mm_add_ps(vvdw12,vvdw6);
389 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
390
391 /* Update potential sum for this i atom from the interaction with this j atom. */
392 velec = _mm_andnot_ps(dummy_mask,velec);
393 velecsum = _mm_add_ps(velecsum,velec);
394 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
395 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
396
397 fscal = _mm_add_ps(felec,fvdw);
398
399 fscal = _mm_andnot_ps(dummy_mask,fscal);
400
401 /* Update vectorial force */
402 fix0 = _mm_macc_ps(dx00,fscal,fix0);
403 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
404 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
405
406 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
407 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
408 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
409 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
410 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
411 _mm_mul_ps(dx00,fscal),
412 _mm_mul_ps(dy00,fscal),
413 _mm_mul_ps(dz00,fscal));
414
415 /* Inner loop uses 67 flops */
416 }
417
418 /* End of innermost loop */
419
420 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
421 f+i_coord_offset,fshift+i_shift_offset);
422
423 ggid = gid[iidx];
424 /* Update potential energies */
425 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
426 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
427
428 /* Increment number of inner iterations */
429 inneriter += j_index_end - j_index_start;
430
431 /* Outer loop uses 9 flops */
432 }
433
434 /* Increment number of outer iterations */
435 outeriter += nri;
436
437 /* Update outer/inner flops */
438
439 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*67);
440 }
441 /*
442 * Gromacs nonbonded kernel: nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_128_fma_single
443 * Electrostatics interaction: Coulomb
444 * VdW interaction: CubicSplineTable
445 * Geometry: Particle-Particle
446 * Calculate force/pot: Force
447 */
448 void
449 nb_kernel_ElecCoul_VdwCSTab_GeomP1P1_F_avx_128_fma_single
450 (t_nblist * gmx_restrict nlist,
451 rvec * gmx_restrict xx,
452 rvec * gmx_restrict ff,
453 t_forcerec * gmx_restrict fr,
454 t_mdatoms * gmx_restrict mdatoms,
455 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
456 t_nrnb * gmx_restrict nrnb)
457 {
458 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
459 * just 0 for non-waters.
460 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
461 * jnr indices corresponding to data put in the four positions in the SIMD register.
462 */
463 int i_shift_offset,i_coord_offset,outeriter,inneriter;
464 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
465 int jnrA,jnrB,jnrC,jnrD;
466 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
467 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
468 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
469 real rcutoff_scalar;
470 real *shiftvec,*fshift,*x,*f;
471 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
472 real scratch[4*DIM];
473 __m128 fscal,rcutoff,rcutoff2,jidxall;
474 int vdwioffset0;
475 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
476 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
477 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
478 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
479 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
480 real *charge;
481 int nvdwtype;
482 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
483 int *vdwtype;
484 real *vdwparam;
485 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
486 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
487 __m128i vfitab;
488 __m128i ifour = _mm_set1_epi32(4);
489 __m128 rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
490 real *vftab;
491 __m128 dummy_mask,cutoff_mask;
492 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
493 __m128 one = _mm_set1_ps(1.0);
494 __m128 two = _mm_set1_ps(2.0);
495 x = xx[0];
496 f = ff[0];
497
498 nri = nlist->nri;
499 iinr = nlist->iinr;
500 jindex = nlist->jindex;
501 jjnr = nlist->jjnr;
502 shiftidx = nlist->shift;
503 gid = nlist->gid;
504 shiftvec = fr->shift_vec[0];
505 fshift = fr->fshift[0];
506 facel = _mm_set1_ps(fr->epsfac);
507 charge = mdatoms->chargeA;
508 nvdwtype = fr->ntype;
509 vdwparam = fr->nbfp;
510 vdwtype = mdatoms->typeA;
511
512 vftab = kernel_data->table_vdw->data;
513 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
514
515 /* Avoid stupid compiler warnings */
516 jnrA = jnrB = jnrC = jnrD = 0;
517 j_coord_offsetA = 0;
518 j_coord_offsetB = 0;
519 j_coord_offsetC = 0;
520 j_coord_offsetD = 0;
521
522 outeriter = 0;
523 inneriter = 0;
524
525 for(iidx=0;iidx<4*DIM;iidx++)
526 {
527 scratch[iidx] = 0.0;
528 }
529
530 /* Start outer loop over neighborlists */
531 for(iidx=0; iidx<nri; iidx++)
532 {
533 /* Load shift vector for this list */
534 i_shift_offset = DIM*shiftidx[iidx];
535
536 /* Load limits for loop over neighbors */
537 j_index_start = jindex[iidx];
538 j_index_end = jindex[iidx+1];
539
540 /* Get outer coordinate index */
541 inr = iinr[iidx];
542 i_coord_offset = DIM*inr;
543
544 /* Load i particle coords and add shift vector */
545 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
546
547 fix0 = _mm_setzero_ps();
548 fiy0 = _mm_setzero_ps();
549 fiz0 = _mm_setzero_ps();
550
551 /* Load parameters for i particles */
552 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
553 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
554
555 /* Start inner kernel loop */
556 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
557 {
558
559 /* Get j neighbor index, and coordinate index */
560 jnrA = jjnr[jidx];
561 jnrB = jjnr[jidx+1];
562 jnrC = jjnr[jidx+2];
563 jnrD = jjnr[jidx+3];
564 j_coord_offsetA = DIM*jnrA;
565 j_coord_offsetB = DIM*jnrB;
566 j_coord_offsetC = DIM*jnrC;
567 j_coord_offsetD = DIM*jnrD;
568
569 /* load j atom coordinates */
570 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
571 x+j_coord_offsetC,x+j_coord_offsetD,
572 &jx0,&jy0,&jz0);
573
574 /* Calculate displacement vector */
575 dx00 = _mm_sub_ps(ix0,jx0);
576 dy00 = _mm_sub_ps(iy0,jy0);
577 dz00 = _mm_sub_ps(iz0,jz0);
578
579 /* Calculate squared distance and things based on it */
580 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
581
582 rinv00 = gmx_mm_invsqrt_ps(rsq00);
583
584 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
585
586 /* Load parameters for j particles */
587 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
588 charge+jnrC+0,charge+jnrD+0);
589 vdwjidx0A = 2*vdwtype[jnrA+0];
590 vdwjidx0B = 2*vdwtype[jnrB+0];
591 vdwjidx0C = 2*vdwtype[jnrC+0];
592 vdwjidx0D = 2*vdwtype[jnrD+0];
593
594 /**************************
595 * CALCULATE INTERACTIONS *
596 **************************/
597
598 r00 = _mm_mul_ps(rsq00,rinv00);
599
600 /* Compute parameters for interactions between i and j atoms */
601 qq00 = _mm_mul_ps(iq0,jq0);
602 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
603 vdwparam+vdwioffset0+vdwjidx0B,
604 vdwparam+vdwioffset0+vdwjidx0C,
605 vdwparam+vdwioffset0+vdwjidx0D,
606 &c6_00,&c12_00);
607
608 /* Calculate table index by multiplying r with table scale and truncate to integer */
609 rt = _mm_mul_ps(r00,vftabscale);
610 vfitab = _mm_cvttps_epi32(rt);
611 #ifdef __XOP__
612 vfeps = _mm_frcz_ps(rt);
613 #else
614 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
615 #endif
616 twovfeps = _mm_add_ps(vfeps,vfeps);
617 vfitab = _mm_slli_epi32(vfitab,3);
618
619 /* COULOMB ELECTROSTATICS */
620 velec = _mm_mul_ps(qq00,rinv00);
621 felec = _mm_mul_ps(velec,rinvsq00);
622
623 /* CUBIC SPLINE TABLE DISPERSION */
624 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
625 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
626 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
627 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
628 _MM_TRANSPOSE4_PS(Y,F,G,H);
629 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
630 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
631 fvdw6 = _mm_mul_ps(c6_00,FF);
632
633 /* CUBIC SPLINE TABLE REPULSION */
634 vfitab = _mm_add_epi32(vfitab,ifour);
635 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
636 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
637 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
638 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
639 _MM_TRANSPOSE4_PS(Y,F,G,H);
640 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
641 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
642 fvdw12 = _mm_mul_ps(c12_00,FF);
643 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
644
645 fscal = _mm_add_ps(felec,fvdw);
646
647 /* Update vectorial force */
648 fix0 = _mm_macc_ps(dx00,fscal,fix0);
649 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
650 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
651
652 fjptrA = f+j_coord_offsetA;
653 fjptrB = f+j_coord_offsetB;
654 fjptrC = f+j_coord_offsetC;
655 fjptrD = f+j_coord_offsetD;
656 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
657 _mm_mul_ps(dx00,fscal),
658 _mm_mul_ps(dy00,fscal),
659 _mm_mul_ps(dz00,fscal));
660
661 /* Inner loop uses 57 flops */
662 }
663
664 if(jidx<j_index_end)
665 {
666
667 /* Get j neighbor index, and coordinate index */
668 jnrlistA = jjnr[jidx];
669 jnrlistB = jjnr[jidx+1];
670 jnrlistC = jjnr[jidx+2];
671 jnrlistD = jjnr[jidx+3];
672 /* Sign of each element will be negative for non-real atoms.
673 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
674 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
675 */
676 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
677 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
678 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
679 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
680 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
681 j_coord_offsetA = DIM*jnrA;
682 j_coord_offsetB = DIM*jnrB;
683 j_coord_offsetC = DIM*jnrC;
684 j_coord_offsetD = DIM*jnrD;
685
686 /* load j atom coordinates */
687 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
688 x+j_coord_offsetC,x+j_coord_offsetD,
689 &jx0,&jy0,&jz0);
690
691 /* Calculate displacement vector */
692 dx00 = _mm_sub_ps(ix0,jx0);
693 dy00 = _mm_sub_ps(iy0,jy0);
694 dz00 = _mm_sub_ps(iz0,jz0);
695
696 /* Calculate squared distance and things based on it */
697 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
698
699 rinv00 = gmx_mm_invsqrt_ps(rsq00);
700
701 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
702
703 /* Load parameters for j particles */
704 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
705 charge+jnrC+0,charge+jnrD+0);
706 vdwjidx0A = 2*vdwtype[jnrA+0];
707 vdwjidx0B = 2*vdwtype[jnrB+0];
708 vdwjidx0C = 2*vdwtype[jnrC+0];
709 vdwjidx0D = 2*vdwtype[jnrD+0];
710
711 /**************************
712 * CALCULATE INTERACTIONS *
713 **************************/
714
715 r00 = _mm_mul_ps(rsq00,rinv00);
716 r00 = _mm_andnot_ps(dummy_mask,r00);
717
718 /* Compute parameters for interactions between i and j atoms */
719 qq00 = _mm_mul_ps(iq0,jq0);
720 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
721 vdwparam+vdwioffset0+vdwjidx0B,
722 vdwparam+vdwioffset0+vdwjidx0C,
723 vdwparam+vdwioffset0+vdwjidx0D,
724 &c6_00,&c12_00);
725
726 /* Calculate table index by multiplying r with table scale and truncate to integer */
727 rt = _mm_mul_ps(r00,vftabscale);
728 vfitab = _mm_cvttps_epi32(rt);
729 #ifdef __XOP__
730 vfeps = _mm_frcz_ps(rt);
731 #else
732 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
733 #endif
734 twovfeps = _mm_add_ps(vfeps,vfeps);
735 vfitab = _mm_slli_epi32(vfitab,3);
736
737 /* COULOMB ELECTROSTATICS */
738 velec = _mm_mul_ps(qq00,rinv00);
739 felec = _mm_mul_ps(velec,rinvsq00);
740
741 /* CUBIC SPLINE TABLE DISPERSION */
742 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
743 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
744 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
745 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
746 _MM_TRANSPOSE4_PS(Y,F,G,H);
747 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
748 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
749 fvdw6 = _mm_mul_ps(c6_00,FF);
750
751 /* CUBIC SPLINE TABLE REPULSION */
752 vfitab = _mm_add_epi32(vfitab,ifour);
753 Y = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
754 F = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
755 G = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
756 H = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
757 _MM_TRANSPOSE4_PS(Y,F,G,H);
758 Fp = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
759 FF = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
760 fvdw12 = _mm_mul_ps(c12_00,FF);
761 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
762
763 fscal = _mm_add_ps(felec,fvdw);
764
765 fscal = _mm_andnot_ps(dummy_mask,fscal);
766
767 /* Update vectorial force */
768 fix0 = _mm_macc_ps(dx00,fscal,fix0);
769 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
770 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
771
772 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
773 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
774 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
775 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
776 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
777 _mm_mul_ps(dx00,fscal),
778 _mm_mul_ps(dy00,fscal),
779 _mm_mul_ps(dz00,fscal));
780
781 /* Inner loop uses 58 flops */
782 }
783
784 /* End of innermost loop */
785
786 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
787 f+i_coord_offset,fshift+i_shift_offset);
788
789 /* Increment number of inner iterations */
790 inneriter += j_index_end - j_index_start;
791
792 /* Outer loop uses 7 flops */
793 }
794
795 /* Increment number of outer iterations */
796 outeriter += nri;
797
798 /* Update outer/inner flops */
799
800 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*58);
801 }
The ultimate molecular dynamics simulation package