?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Removed simple.h from nb_kernel_sse4_1_XX
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sse4_1_single / nb_kernel_ElecGB_VdwCSTab_GeomP1P1_sse4_1_single.c
blob77e3dddf76c443ef692d6d5a39011fcc2492e3c8
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 sse4_1_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_sse4_1_single.h"
49 #include "kernelutil_x86_sse4_1_single.h"
50
51 /*
52 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_sse4_1_single
53 * Electrostatics interaction: GeneralizedBorn
54 * VdW interaction: CubicSplineTable
55 * Geometry: Particle-Particle
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_sse4_1_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 SSE, 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 tx,ty,tz,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 __m128i gbitab;
92 __m128 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
93 __m128 minushalf = _mm_set1_ps(-0.5);
94 real *invsqrta,*dvda,*gbtab;
95 int nvdwtype;
96 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
97 int *vdwtype;
98 real *vdwparam;
99 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
100 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
101 __m128i vfitab;
102 __m128i ifour = _mm_set1_epi32(4);
103 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
104 real *vftab;
105 __m128 dummy_mask,cutoff_mask;
106 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
107 __m128 one = _mm_set1_ps(1.0);
108 __m128 two = _mm_set1_ps(2.0);
109 x = xx[0];
110 f = ff[0];
111
112 nri = nlist->nri;
113 iinr = nlist->iinr;
114 jindex = nlist->jindex;
115 jjnr = nlist->jjnr;
116 shiftidx = nlist->shift;
117 gid = nlist->gid;
118 shiftvec = fr->shift_vec[0];
119 fshift = fr->fshift[0];
120 facel = _mm_set1_ps(fr->epsfac);
121 charge = mdatoms->chargeA;
122 nvdwtype = fr->ntype;
123 vdwparam = fr->nbfp;
124 vdwtype = mdatoms->typeA;
125
126 vftab = kernel_data->table_vdw->data;
127 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
128
129 invsqrta = fr->invsqrta;
130 dvda = fr->dvda;
131 gbtabscale = _mm_set1_ps(fr->gbtab.scale);
132 gbtab = fr->gbtab.data;
133 gbinvepsdiff = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
134
135 /* Avoid stupid compiler warnings */
136 jnrA = jnrB = jnrC = jnrD = 0;
137 j_coord_offsetA = 0;
138 j_coord_offsetB = 0;
139 j_coord_offsetC = 0;
140 j_coord_offsetD = 0;
141
142 outeriter = 0;
143 inneriter = 0;
144
145 for(iidx=0;iidx<4*DIM;iidx++)
146 {
147 scratch[iidx] = 0.0;
148 }
149
150 /* Start outer loop over neighborlists */
151 for(iidx=0; iidx<nri; iidx++)
152 {
153 /* Load shift vector for this list */
154 i_shift_offset = DIM*shiftidx[iidx];
155
156 /* Load limits for loop over neighbors */
157 j_index_start = jindex[iidx];
158 j_index_end = jindex[iidx+1];
159
160 /* Get outer coordinate index */
161 inr = iinr[iidx];
162 i_coord_offset = DIM*inr;
163
164 /* Load i particle coords and add shift vector */
165 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
166
167 fix0 = _mm_setzero_ps();
168 fiy0 = _mm_setzero_ps();
169 fiz0 = _mm_setzero_ps();
170
171 /* Load parameters for i particles */
172 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
173 isai0 = _mm_load1_ps(invsqrta+inr+0);
174 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
175
176 /* Reset potential sums */
177 velecsum = _mm_setzero_ps();
178 vgbsum = _mm_setzero_ps();
179 vvdwsum = _mm_setzero_ps();
180 dvdasum = _mm_setzero_ps();
181
182 /* Start inner kernel loop */
183 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
184 {
185
186 /* Get j neighbor index, and coordinate index */
187 jnrA = jjnr[jidx];
188 jnrB = jjnr[jidx+1];
189 jnrC = jjnr[jidx+2];
190 jnrD = jjnr[jidx+3];
191 j_coord_offsetA = DIM*jnrA;
192 j_coord_offsetB = DIM*jnrB;
193 j_coord_offsetC = DIM*jnrC;
194 j_coord_offsetD = DIM*jnrD;
195
196 /* load j atom coordinates */
197 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
198 x+j_coord_offsetC,x+j_coord_offsetD,
199 &jx0,&jy0,&jz0);
200
201 /* Calculate displacement vector */
202 dx00 = _mm_sub_ps(ix0,jx0);
203 dy00 = _mm_sub_ps(iy0,jy0);
204 dz00 = _mm_sub_ps(iz0,jz0);
205
206 /* Calculate squared distance and things based on it */
207 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
208
209 rinv00 = gmx_mm_invsqrt_ps(rsq00);
210
211 /* Load parameters for j particles */
212 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
213 charge+jnrC+0,charge+jnrD+0);
214 isaj0 = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
215 invsqrta+jnrC+0,invsqrta+jnrD+0);
216 vdwjidx0A = 2*vdwtype[jnrA+0];
217 vdwjidx0B = 2*vdwtype[jnrB+0];
218 vdwjidx0C = 2*vdwtype[jnrC+0];
219 vdwjidx0D = 2*vdwtype[jnrD+0];
220
221 /**************************
222 * CALCULATE INTERACTIONS *
223 **************************/
224
225 r00 = _mm_mul_ps(rsq00,rinv00);
226
227 /* Compute parameters for interactions between i and j atoms */
228 qq00 = _mm_mul_ps(iq0,jq0);
229 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
230 vdwparam+vdwioffset0+vdwjidx0B,
231 vdwparam+vdwioffset0+vdwjidx0C,
232 vdwparam+vdwioffset0+vdwjidx0D,
233 &c6_00,&c12_00);
234
235 /* Calculate table index by multiplying r with table scale and truncate to integer */
236 rt = _mm_mul_ps(r00,vftabscale);
237 vfitab = _mm_cvttps_epi32(rt);
238 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
239 vfitab = _mm_slli_epi32(vfitab,3);
240
241 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
242 isaprod = _mm_mul_ps(isai0,isaj0);
243 gbqqfactor = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
244 gbscale = _mm_mul_ps(isaprod,gbtabscale);
245
246 /* Calculate generalized born table index - this is a separate table from the normal one,
247 * but we use the same procedure by multiplying r with scale and truncating to integer.
248 */
249 rt = _mm_mul_ps(r00,gbscale);
250 gbitab = _mm_cvttps_epi32(rt);
251 gbeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
252 gbitab = _mm_slli_epi32(gbitab,2);
253 Y = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
254 F = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
255 G = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
256 H = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
257 _MM_TRANSPOSE4_PS(Y,F,G,H);
258 Heps = _mm_mul_ps(gbeps,H);
259 Fp = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
260 VV = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
261 vgb = _mm_mul_ps(gbqqfactor,VV);
262
263 FF = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
264 fgb = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
265 dvdatmp = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
266 dvdasum = _mm_add_ps(dvdasum,dvdatmp);
267 fjptrA = dvda+jnrA;
268 fjptrB = dvda+jnrB;
269 fjptrC = dvda+jnrC;
270 fjptrD = dvda+jnrD;
271 gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
272 velec = _mm_mul_ps(qq00,rinv00);
273 felec = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
274
275 /* CUBIC SPLINE TABLE DISPERSION */
276 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
277 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
278 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
279 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
280 _MM_TRANSPOSE4_PS(Y,F,G,H);
281 Heps = _mm_mul_ps(vfeps,H);
282 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
283 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
284 vvdw6 = _mm_mul_ps(c6_00,VV);
285 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
286 fvdw6 = _mm_mul_ps(c6_00,FF);
287
288 /* CUBIC SPLINE TABLE REPULSION */
289 vfitab = _mm_add_epi32(vfitab,ifour);
290 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
291 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
292 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
293 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
294 _MM_TRANSPOSE4_PS(Y,F,G,H);
295 Heps = _mm_mul_ps(vfeps,H);
296 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
297 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
298 vvdw12 = _mm_mul_ps(c12_00,VV);
299 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
300 fvdw12 = _mm_mul_ps(c12_00,FF);
301 vvdw = _mm_add_ps(vvdw12,vvdw6);
302 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
303
304 /* Update potential sum for this i atom from the interaction with this j atom. */
305 velecsum = _mm_add_ps(velecsum,velec);
306 vgbsum = _mm_add_ps(vgbsum,vgb);
307 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
308
309 fscal = _mm_add_ps(felec,fvdw);
310
311 /* Calculate temporary vectorial force */
312 tx = _mm_mul_ps(fscal,dx00);
313 ty = _mm_mul_ps(fscal,dy00);
314 tz = _mm_mul_ps(fscal,dz00);
315
316 /* Update vectorial force */
317 fix0 = _mm_add_ps(fix0,tx);
318 fiy0 = _mm_add_ps(fiy0,ty);
319 fiz0 = _mm_add_ps(fiz0,tz);
320
321 fjptrA = f+j_coord_offsetA;
322 fjptrB = f+j_coord_offsetB;
323 fjptrC = f+j_coord_offsetC;
324 fjptrD = f+j_coord_offsetD;
325 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
326
327 /* Inner loop uses 92 flops */
328 }
329
330 if(jidx<j_index_end)
331 {
332
333 /* Get j neighbor index, and coordinate index */
334 jnrlistA = jjnr[jidx];
335 jnrlistB = jjnr[jidx+1];
336 jnrlistC = jjnr[jidx+2];
337 jnrlistD = jjnr[jidx+3];
338 /* Sign of each element will be negative for non-real atoms.
339 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
340 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
341 */
342 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
343 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
344 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
345 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
346 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
347 j_coord_offsetA = DIM*jnrA;
348 j_coord_offsetB = DIM*jnrB;
349 j_coord_offsetC = DIM*jnrC;
350 j_coord_offsetD = DIM*jnrD;
351
352 /* load j atom coordinates */
353 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
354 x+j_coord_offsetC,x+j_coord_offsetD,
355 &jx0,&jy0,&jz0);
356
357 /* Calculate displacement vector */
358 dx00 = _mm_sub_ps(ix0,jx0);
359 dy00 = _mm_sub_ps(iy0,jy0);
360 dz00 = _mm_sub_ps(iz0,jz0);
361
362 /* Calculate squared distance and things based on it */
363 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
364
365 rinv00 = gmx_mm_invsqrt_ps(rsq00);
366
367 /* Load parameters for j particles */
368 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
369 charge+jnrC+0,charge+jnrD+0);
370 isaj0 = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
371 invsqrta+jnrC+0,invsqrta+jnrD+0);
372 vdwjidx0A = 2*vdwtype[jnrA+0];
373 vdwjidx0B = 2*vdwtype[jnrB+0];
374 vdwjidx0C = 2*vdwtype[jnrC+0];
375 vdwjidx0D = 2*vdwtype[jnrD+0];
376
377 /**************************
378 * CALCULATE INTERACTIONS *
379 **************************/
380
381 r00 = _mm_mul_ps(rsq00,rinv00);
382 r00 = _mm_andnot_ps(dummy_mask,r00);
383
384 /* Compute parameters for interactions between i and j atoms */
385 qq00 = _mm_mul_ps(iq0,jq0);
386 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
387 vdwparam+vdwioffset0+vdwjidx0B,
388 vdwparam+vdwioffset0+vdwjidx0C,
389 vdwparam+vdwioffset0+vdwjidx0D,
390 &c6_00,&c12_00);
391
392 /* Calculate table index by multiplying r with table scale and truncate to integer */
393 rt = _mm_mul_ps(r00,vftabscale);
394 vfitab = _mm_cvttps_epi32(rt);
395 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
396 vfitab = _mm_slli_epi32(vfitab,3);
397
398 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
399 isaprod = _mm_mul_ps(isai0,isaj0);
400 gbqqfactor = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
401 gbscale = _mm_mul_ps(isaprod,gbtabscale);
402
403 /* Calculate generalized born table index - this is a separate table from the normal one,
404 * but we use the same procedure by multiplying r with scale and truncating to integer.
405 */
406 rt = _mm_mul_ps(r00,gbscale);
407 gbitab = _mm_cvttps_epi32(rt);
408 gbeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
409 gbitab = _mm_slli_epi32(gbitab,2);
410 Y = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
411 F = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
412 G = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
413 H = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
414 _MM_TRANSPOSE4_PS(Y,F,G,H);
415 Heps = _mm_mul_ps(gbeps,H);
416 Fp = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
417 VV = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
418 vgb = _mm_mul_ps(gbqqfactor,VV);
419
420 FF = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
421 fgb = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
422 dvdatmp = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
423 dvdatmp = _mm_andnot_ps(dummy_mask,dvdatmp);
424 dvdasum = _mm_add_ps(dvdasum,dvdatmp);
425 /* 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. */
426 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
427 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
428 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
429 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
430 gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
431 velec = _mm_mul_ps(qq00,rinv00);
432 felec = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
433
434 /* CUBIC SPLINE TABLE DISPERSION */
435 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
436 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
437 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
438 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
439 _MM_TRANSPOSE4_PS(Y,F,G,H);
440 Heps = _mm_mul_ps(vfeps,H);
441 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
442 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
443 vvdw6 = _mm_mul_ps(c6_00,VV);
444 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
445 fvdw6 = _mm_mul_ps(c6_00,FF);
446
447 /* CUBIC SPLINE TABLE REPULSION */
448 vfitab = _mm_add_epi32(vfitab,ifour);
449 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
450 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
451 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
452 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
453 _MM_TRANSPOSE4_PS(Y,F,G,H);
454 Heps = _mm_mul_ps(vfeps,H);
455 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
456 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
457 vvdw12 = _mm_mul_ps(c12_00,VV);
458 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
459 fvdw12 = _mm_mul_ps(c12_00,FF);
460 vvdw = _mm_add_ps(vvdw12,vvdw6);
461 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
462
463 /* Update potential sum for this i atom from the interaction with this j atom. */
464 velec = _mm_andnot_ps(dummy_mask,velec);
465 velecsum = _mm_add_ps(velecsum,velec);
466 vgb = _mm_andnot_ps(dummy_mask,vgb);
467 vgbsum = _mm_add_ps(vgbsum,vgb);
468 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
469 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
470
471 fscal = _mm_add_ps(felec,fvdw);
472
473 fscal = _mm_andnot_ps(dummy_mask,fscal);
474
475 /* Calculate temporary vectorial force */
476 tx = _mm_mul_ps(fscal,dx00);
477 ty = _mm_mul_ps(fscal,dy00);
478 tz = _mm_mul_ps(fscal,dz00);
479
480 /* Update vectorial force */
481 fix0 = _mm_add_ps(fix0,tx);
482 fiy0 = _mm_add_ps(fiy0,ty);
483 fiz0 = _mm_add_ps(fiz0,tz);
484
485 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
486 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
487 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
488 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
489 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
490
491 /* Inner loop uses 93 flops */
492 }
493
494 /* End of innermost loop */
495
496 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
497 f+i_coord_offset,fshift+i_shift_offset);
498
499 ggid = gid[iidx];
500 /* Update potential energies */
501 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
502 gmx_mm_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
503 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
504 dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
505 gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
506
507 /* Increment number of inner iterations */
508 inneriter += j_index_end - j_index_start;
509
510 /* Outer loop uses 10 flops */
511 }
512
513 /* Increment number of outer iterations */
514 outeriter += nri;
515
516 /* Update outer/inner flops */
517
518 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*10 + inneriter*93);
519 }
520 /*
521 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_sse4_1_single
522 * Electrostatics interaction: GeneralizedBorn
523 * VdW interaction: CubicSplineTable
524 * Geometry: Particle-Particle
525 * Calculate force/pot: Force
526 */
527 void
528 nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_sse4_1_single
529 (t_nblist * gmx_restrict nlist,
530 rvec * gmx_restrict xx,
531 rvec * gmx_restrict ff,
532 t_forcerec * gmx_restrict fr,
533 t_mdatoms * gmx_restrict mdatoms,
534 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
535 t_nrnb * gmx_restrict nrnb)
536 {
537 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
538 * just 0 for non-waters.
539 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
540 * jnr indices corresponding to data put in the four positions in the SIMD register.
541 */
542 int i_shift_offset,i_coord_offset,outeriter,inneriter;
543 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
544 int jnrA,jnrB,jnrC,jnrD;
545 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
546 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
547 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
548 real rcutoff_scalar;
549 real *shiftvec,*fshift,*x,*f;
550 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
551 real scratch[4*DIM];
552 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
553 int vdwioffset0;
554 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
555 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
556 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
557 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
558 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
559 real *charge;
560 __m128i gbitab;
561 __m128 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
562 __m128 minushalf = _mm_set1_ps(-0.5);
563 real *invsqrta,*dvda,*gbtab;
564 int nvdwtype;
565 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
566 int *vdwtype;
567 real *vdwparam;
568 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
569 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
570 __m128i vfitab;
571 __m128i ifour = _mm_set1_epi32(4);
572 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
573 real *vftab;
574 __m128 dummy_mask,cutoff_mask;
575 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
576 __m128 one = _mm_set1_ps(1.0);
577 __m128 two = _mm_set1_ps(2.0);
578 x = xx[0];
579 f = ff[0];
580
581 nri = nlist->nri;
582 iinr = nlist->iinr;
583 jindex = nlist->jindex;
584 jjnr = nlist->jjnr;
585 shiftidx = nlist->shift;
586 gid = nlist->gid;
587 shiftvec = fr->shift_vec[0];
588 fshift = fr->fshift[0];
589 facel = _mm_set1_ps(fr->epsfac);
590 charge = mdatoms->chargeA;
591 nvdwtype = fr->ntype;
592 vdwparam = fr->nbfp;
593 vdwtype = mdatoms->typeA;
594
595 vftab = kernel_data->table_vdw->data;
596 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
597
598 invsqrta = fr->invsqrta;
599 dvda = fr->dvda;
600 gbtabscale = _mm_set1_ps(fr->gbtab.scale);
601 gbtab = fr->gbtab.data;
602 gbinvepsdiff = _mm_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
603
604 /* Avoid stupid compiler warnings */
605 jnrA = jnrB = jnrC = jnrD = 0;
606 j_coord_offsetA = 0;
607 j_coord_offsetB = 0;
608 j_coord_offsetC = 0;
609 j_coord_offsetD = 0;
610
611 outeriter = 0;
612 inneriter = 0;
613
614 for(iidx=0;iidx<4*DIM;iidx++)
615 {
616 scratch[iidx] = 0.0;
617 }
618
619 /* Start outer loop over neighborlists */
620 for(iidx=0; iidx<nri; iidx++)
621 {
622 /* Load shift vector for this list */
623 i_shift_offset = DIM*shiftidx[iidx];
624
625 /* Load limits for loop over neighbors */
626 j_index_start = jindex[iidx];
627 j_index_end = jindex[iidx+1];
628
629 /* Get outer coordinate index */
630 inr = iinr[iidx];
631 i_coord_offset = DIM*inr;
632
633 /* Load i particle coords and add shift vector */
634 gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
635
636 fix0 = _mm_setzero_ps();
637 fiy0 = _mm_setzero_ps();
638 fiz0 = _mm_setzero_ps();
639
640 /* Load parameters for i particles */
641 iq0 = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
642 isai0 = _mm_load1_ps(invsqrta+inr+0);
643 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
644
645 dvdasum = _mm_setzero_ps();
646
647 /* Start inner kernel loop */
648 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
649 {
650
651 /* Get j neighbor index, and coordinate index */
652 jnrA = jjnr[jidx];
653 jnrB = jjnr[jidx+1];
654 jnrC = jjnr[jidx+2];
655 jnrD = jjnr[jidx+3];
656 j_coord_offsetA = DIM*jnrA;
657 j_coord_offsetB = DIM*jnrB;
658 j_coord_offsetC = DIM*jnrC;
659 j_coord_offsetD = DIM*jnrD;
660
661 /* load j atom coordinates */
662 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
663 x+j_coord_offsetC,x+j_coord_offsetD,
664 &jx0,&jy0,&jz0);
665
666 /* Calculate displacement vector */
667 dx00 = _mm_sub_ps(ix0,jx0);
668 dy00 = _mm_sub_ps(iy0,jy0);
669 dz00 = _mm_sub_ps(iz0,jz0);
670
671 /* Calculate squared distance and things based on it */
672 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
673
674 rinv00 = gmx_mm_invsqrt_ps(rsq00);
675
676 /* Load parameters for j particles */
677 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
678 charge+jnrC+0,charge+jnrD+0);
679 isaj0 = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
680 invsqrta+jnrC+0,invsqrta+jnrD+0);
681 vdwjidx0A = 2*vdwtype[jnrA+0];
682 vdwjidx0B = 2*vdwtype[jnrB+0];
683 vdwjidx0C = 2*vdwtype[jnrC+0];
684 vdwjidx0D = 2*vdwtype[jnrD+0];
685
686 /**************************
687 * CALCULATE INTERACTIONS *
688 **************************/
689
690 r00 = _mm_mul_ps(rsq00,rinv00);
691
692 /* Compute parameters for interactions between i and j atoms */
693 qq00 = _mm_mul_ps(iq0,jq0);
694 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
695 vdwparam+vdwioffset0+vdwjidx0B,
696 vdwparam+vdwioffset0+vdwjidx0C,
697 vdwparam+vdwioffset0+vdwjidx0D,
698 &c6_00,&c12_00);
699
700 /* Calculate table index by multiplying r with table scale and truncate to integer */
701 rt = _mm_mul_ps(r00,vftabscale);
702 vfitab = _mm_cvttps_epi32(rt);
703 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
704 vfitab = _mm_slli_epi32(vfitab,3);
705
706 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
707 isaprod = _mm_mul_ps(isai0,isaj0);
708 gbqqfactor = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
709 gbscale = _mm_mul_ps(isaprod,gbtabscale);
710
711 /* Calculate generalized born table index - this is a separate table from the normal one,
712 * but we use the same procedure by multiplying r with scale and truncating to integer.
713 */
714 rt = _mm_mul_ps(r00,gbscale);
715 gbitab = _mm_cvttps_epi32(rt);
716 gbeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
717 gbitab = _mm_slli_epi32(gbitab,2);
718 Y = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
719 F = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
720 G = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
721 H = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
722 _MM_TRANSPOSE4_PS(Y,F,G,H);
723 Heps = _mm_mul_ps(gbeps,H);
724 Fp = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
725 VV = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
726 vgb = _mm_mul_ps(gbqqfactor,VV);
727
728 FF = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
729 fgb = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
730 dvdatmp = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
731 dvdasum = _mm_add_ps(dvdasum,dvdatmp);
732 fjptrA = dvda+jnrA;
733 fjptrB = dvda+jnrB;
734 fjptrC = dvda+jnrC;
735 fjptrD = dvda+jnrD;
736 gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
737 velec = _mm_mul_ps(qq00,rinv00);
738 felec = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
739
740 /* CUBIC SPLINE TABLE DISPERSION */
741 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
742 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
743 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
744 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
745 _MM_TRANSPOSE4_PS(Y,F,G,H);
746 Heps = _mm_mul_ps(vfeps,H);
747 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
748 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
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 + gmx_mm_extract_epi32(vfitab,0) );
754 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
755 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
756 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
757 _MM_TRANSPOSE4_PS(Y,F,G,H);
758 Heps = _mm_mul_ps(vfeps,H);
759 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
760 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
761 fvdw12 = _mm_mul_ps(c12_00,FF);
762 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
763
764 fscal = _mm_add_ps(felec,fvdw);
765
766 /* Calculate temporary vectorial force */
767 tx = _mm_mul_ps(fscal,dx00);
768 ty = _mm_mul_ps(fscal,dy00);
769 tz = _mm_mul_ps(fscal,dz00);
770
771 /* Update vectorial force */
772 fix0 = _mm_add_ps(fix0,tx);
773 fiy0 = _mm_add_ps(fiy0,ty);
774 fiz0 = _mm_add_ps(fiz0,tz);
775
776 fjptrA = f+j_coord_offsetA;
777 fjptrB = f+j_coord_offsetB;
778 fjptrC = f+j_coord_offsetC;
779 fjptrD = f+j_coord_offsetD;
780 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
781
782 /* Inner loop uses 82 flops */
783 }
784
785 if(jidx<j_index_end)
786 {
787
788 /* Get j neighbor index, and coordinate index */
789 jnrlistA = jjnr[jidx];
790 jnrlistB = jjnr[jidx+1];
791 jnrlistC = jjnr[jidx+2];
792 jnrlistD = jjnr[jidx+3];
793 /* Sign of each element will be negative for non-real atoms.
794 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
795 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
796 */
797 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
798 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
799 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
800 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
801 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
802 j_coord_offsetA = DIM*jnrA;
803 j_coord_offsetB = DIM*jnrB;
804 j_coord_offsetC = DIM*jnrC;
805 j_coord_offsetD = DIM*jnrD;
806
807 /* load j atom coordinates */
808 gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
809 x+j_coord_offsetC,x+j_coord_offsetD,
810 &jx0,&jy0,&jz0);
811
812 /* Calculate displacement vector */
813 dx00 = _mm_sub_ps(ix0,jx0);
814 dy00 = _mm_sub_ps(iy0,jy0);
815 dz00 = _mm_sub_ps(iz0,jz0);
816
817 /* Calculate squared distance and things based on it */
818 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
819
820 rinv00 = gmx_mm_invsqrt_ps(rsq00);
821
822 /* Load parameters for j particles */
823 jq0 = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
824 charge+jnrC+0,charge+jnrD+0);
825 isaj0 = gmx_mm_load_4real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
826 invsqrta+jnrC+0,invsqrta+jnrD+0);
827 vdwjidx0A = 2*vdwtype[jnrA+0];
828 vdwjidx0B = 2*vdwtype[jnrB+0];
829 vdwjidx0C = 2*vdwtype[jnrC+0];
830 vdwjidx0D = 2*vdwtype[jnrD+0];
831
832 /**************************
833 * CALCULATE INTERACTIONS *
834 **************************/
835
836 r00 = _mm_mul_ps(rsq00,rinv00);
837 r00 = _mm_andnot_ps(dummy_mask,r00);
838
839 /* Compute parameters for interactions between i and j atoms */
840 qq00 = _mm_mul_ps(iq0,jq0);
841 gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
842 vdwparam+vdwioffset0+vdwjidx0B,
843 vdwparam+vdwioffset0+vdwjidx0C,
844 vdwparam+vdwioffset0+vdwjidx0D,
845 &c6_00,&c12_00);
846
847 /* Calculate table index by multiplying r with table scale and truncate to integer */
848 rt = _mm_mul_ps(r00,vftabscale);
849 vfitab = _mm_cvttps_epi32(rt);
850 vfeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
851 vfitab = _mm_slli_epi32(vfitab,3);
852
853 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
854 isaprod = _mm_mul_ps(isai0,isaj0);
855 gbqqfactor = _mm_xor_ps(signbit,_mm_mul_ps(qq00,_mm_mul_ps(isaprod,gbinvepsdiff)));
856 gbscale = _mm_mul_ps(isaprod,gbtabscale);
857
858 /* Calculate generalized born table index - this is a separate table from the normal one,
859 * but we use the same procedure by multiplying r with scale and truncating to integer.
860 */
861 rt = _mm_mul_ps(r00,gbscale);
862 gbitab = _mm_cvttps_epi32(rt);
863 gbeps = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
864 gbitab = _mm_slli_epi32(gbitab,2);
865 Y = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,0) );
866 F = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,1) );
867 G = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,2) );
868 H = _mm_load_ps( gbtab + gmx_mm_extract_epi32(gbitab,3) );
869 _MM_TRANSPOSE4_PS(Y,F,G,H);
870 Heps = _mm_mul_ps(gbeps,H);
871 Fp = _mm_add_ps(F,_mm_mul_ps(gbeps,_mm_add_ps(G,Heps)));
872 VV = _mm_add_ps(Y,_mm_mul_ps(gbeps,Fp));
873 vgb = _mm_mul_ps(gbqqfactor,VV);
874
875 FF = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
876 fgb = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
877 dvdatmp = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
878 dvdatmp = _mm_andnot_ps(dummy_mask,dvdatmp);
879 dvdasum = _mm_add_ps(dvdasum,dvdatmp);
880 /* 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. */
881 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
882 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
883 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
884 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
885 gmx_mm_increment_4real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,_mm_mul_ps(dvdatmp,_mm_mul_ps(isaj0,isaj0)));
886 velec = _mm_mul_ps(qq00,rinv00);
887 felec = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(velec,rinv00),fgb),rinv00);
888
889 /* CUBIC SPLINE TABLE DISPERSION */
890 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
891 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
892 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
893 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
894 _MM_TRANSPOSE4_PS(Y,F,G,H);
895 Heps = _mm_mul_ps(vfeps,H);
896 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
897 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
898 fvdw6 = _mm_mul_ps(c6_00,FF);
899
900 /* CUBIC SPLINE TABLE REPULSION */
901 vfitab = _mm_add_epi32(vfitab,ifour);
902 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
903 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
904 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
905 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
906 _MM_TRANSPOSE4_PS(Y,F,G,H);
907 Heps = _mm_mul_ps(vfeps,H);
908 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
909 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
910 fvdw12 = _mm_mul_ps(c12_00,FF);
911 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
912
913 fscal = _mm_add_ps(felec,fvdw);
914
915 fscal = _mm_andnot_ps(dummy_mask,fscal);
916
917 /* Calculate temporary vectorial force */
918 tx = _mm_mul_ps(fscal,dx00);
919 ty = _mm_mul_ps(fscal,dy00);
920 tz = _mm_mul_ps(fscal,dz00);
921
922 /* Update vectorial force */
923 fix0 = _mm_add_ps(fix0,tx);
924 fiy0 = _mm_add_ps(fiy0,ty);
925 fiz0 = _mm_add_ps(fiz0,tz);
926
927 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
928 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
929 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
930 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
931 gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
932
933 /* Inner loop uses 83 flops */
934 }
935
936 /* End of innermost loop */
937
938 gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
939 f+i_coord_offset,fshift+i_shift_offset);
940
941 dvdasum = _mm_mul_ps(dvdasum, _mm_mul_ps(isai0,isai0));
942 gmx_mm_update_1pot_ps(dvdasum,dvda+inr);
943
944 /* Increment number of inner iterations */
945 inneriter += j_index_end - j_index_start;
946
947 /* Outer loop uses 7 flops */
948 }
949
950 /* Increment number of outer iterations */
951 outeriter += nri;
952
953 /* Update outer/inner flops */
954
955 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*83);
956 }
The ultimate molecular dynamics simulation package