?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Remove nb-parameters from t_forcerec
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sse2_single / nb_kernel_ElecRF_VdwCSTab_GeomW4W4_sse2_single.c
blobec79b493a5ed422d8adfc0ec668b672e2685ce24
1 /*
2 * This file is part of the GROMACS molecular simulation package.
3 *
4 * Copyright (c) 2012,2013,2014,2015,2017, 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 sse2_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/gmxlib/nrnb.h"
46
47 #include "kernelutil_x86_sse2_single.h"
48
49 /*
50 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_sse2_single
51 * Electrostatics interaction: ReactionField
52 * VdW interaction: CubicSplineTable
53 * Geometry: Water4-Water4
54 * Calculate force/pot: PotentialAndForce
55 */
56 void
57 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_VF_sse2_single
58 (t_nblist * gmx_restrict nlist,
59 rvec * gmx_restrict xx,
60 rvec * gmx_restrict ff,
61 struct t_forcerec * gmx_restrict fr,
62 t_mdatoms * gmx_restrict mdatoms,
63 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
64 t_nrnb * gmx_restrict nrnb)
65 {
66 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
67 * just 0 for non-waters.
68 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
69 * jnr indices corresponding to data put in the four positions in the SIMD register.
70 */
71 int i_shift_offset,i_coord_offset,outeriter,inneriter;
72 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
73 int jnrA,jnrB,jnrC,jnrD;
74 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
75 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
76 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
77 real rcutoff_scalar;
78 real *shiftvec,*fshift,*x,*f;
79 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
80 real scratch[4*DIM];
81 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
82 int vdwioffset0;
83 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
84 int vdwioffset1;
85 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
86 int vdwioffset2;
87 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
88 int vdwioffset3;
89 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
90 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
91 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
92 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
93 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
94 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
95 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
96 int vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
97 __m128 jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
98 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
99 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
100 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
101 __m128 dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
102 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
103 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
104 __m128 dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
105 __m128 dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
106 __m128 dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
107 __m128 dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
108 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
109 real *charge;
110 int nvdwtype;
111 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
112 int *vdwtype;
113 real *vdwparam;
114 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
115 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
116 __m128i vfitab;
117 __m128i ifour = _mm_set1_epi32(4);
118 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
119 real *vftab;
120 __m128 dummy_mask,cutoff_mask;
121 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
122 __m128 one = _mm_set1_ps(1.0);
123 __m128 two = _mm_set1_ps(2.0);
124 x = xx[0];
125 f = ff[0];
126
127 nri = nlist->nri;
128 iinr = nlist->iinr;
129 jindex = nlist->jindex;
130 jjnr = nlist->jjnr;
131 shiftidx = nlist->shift;
132 gid = nlist->gid;
133 shiftvec = fr->shift_vec[0];
134 fshift = fr->fshift[0];
135 facel = _mm_set1_ps(fr->ic->epsfac);
136 charge = mdatoms->chargeA;
137 krf = _mm_set1_ps(fr->ic->k_rf);
138 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
139 crf = _mm_set1_ps(fr->ic->c_rf);
140 nvdwtype = fr->ntype;
141 vdwparam = fr->nbfp;
142 vdwtype = mdatoms->typeA;
143
144 vftab = kernel_data->table_vdw->data;
145 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
146
147 /* Setup water-specific parameters */
148 inr = nlist->iinr[0];
149 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
150 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
151 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
152 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
153
154 jq1 = _mm_set1_ps(charge[inr+1]);
155 jq2 = _mm_set1_ps(charge[inr+2]);
156 jq3 = _mm_set1_ps(charge[inr+3]);
157 vdwjidx0A = 2*vdwtype[inr+0];
158 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
159 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
160 qq11 = _mm_mul_ps(iq1,jq1);
161 qq12 = _mm_mul_ps(iq1,jq2);
162 qq13 = _mm_mul_ps(iq1,jq3);
163 qq21 = _mm_mul_ps(iq2,jq1);
164 qq22 = _mm_mul_ps(iq2,jq2);
165 qq23 = _mm_mul_ps(iq2,jq3);
166 qq31 = _mm_mul_ps(iq3,jq1);
167 qq32 = _mm_mul_ps(iq3,jq2);
168 qq33 = _mm_mul_ps(iq3,jq3);
169
170 /* Avoid stupid compiler warnings */
171 jnrA = jnrB = jnrC = jnrD = 0;
172 j_coord_offsetA = 0;
173 j_coord_offsetB = 0;
174 j_coord_offsetC = 0;
175 j_coord_offsetD = 0;
176
177 outeriter = 0;
178 inneriter = 0;
179
180 for(iidx=0;iidx<4*DIM;iidx++)
181 {
182 scratch[iidx] = 0.0;
183 }
184
185 /* Start outer loop over neighborlists */
186 for(iidx=0; iidx<nri; iidx++)
187 {
188 /* Load shift vector for this list */
189 i_shift_offset = DIM*shiftidx[iidx];
190
191 /* Load limits for loop over neighbors */
192 j_index_start = jindex[iidx];
193 j_index_end = jindex[iidx+1];
194
195 /* Get outer coordinate index */
196 inr = iinr[iidx];
197 i_coord_offset = DIM*inr;
198
199 /* Load i particle coords and add shift vector */
200 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
201 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
202
203 fix0 = _mm_setzero_ps();
204 fiy0 = _mm_setzero_ps();
205 fiz0 = _mm_setzero_ps();
206 fix1 = _mm_setzero_ps();
207 fiy1 = _mm_setzero_ps();
208 fiz1 = _mm_setzero_ps();
209 fix2 = _mm_setzero_ps();
210 fiy2 = _mm_setzero_ps();
211 fiz2 = _mm_setzero_ps();
212 fix3 = _mm_setzero_ps();
213 fiy3 = _mm_setzero_ps();
214 fiz3 = _mm_setzero_ps();
215
216 /* Reset potential sums */
217 velecsum = _mm_setzero_ps();
218 vvdwsum = _mm_setzero_ps();
219
220 /* Start inner kernel loop */
221 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
222 {
223
224 /* Get j neighbor index, and coordinate index */
225 jnrA = jjnr[jidx];
226 jnrB = jjnr[jidx+1];
227 jnrC = jjnr[jidx+2];
228 jnrD = jjnr[jidx+3];
229 j_coord_offsetA = DIM*jnrA;
230 j_coord_offsetB = DIM*jnrB;
231 j_coord_offsetC = DIM*jnrC;
232 j_coord_offsetD = DIM*jnrD;
233
234 /* load j atom coordinates */
235 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
236 x+j_coord_offsetC,x+j_coord_offsetD,
237 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
238 &jy2,&jz2,&jx3,&jy3,&jz3);
239
240 /* Calculate displacement vector */
241 dx00 = _mm_sub_ps(ix0,jx0);
242 dy00 = _mm_sub_ps(iy0,jy0);
243 dz00 = _mm_sub_ps(iz0,jz0);
244 dx11 = _mm_sub_ps(ix1,jx1);
245 dy11 = _mm_sub_ps(iy1,jy1);
246 dz11 = _mm_sub_ps(iz1,jz1);
247 dx12 = _mm_sub_ps(ix1,jx2);
248 dy12 = _mm_sub_ps(iy1,jy2);
249 dz12 = _mm_sub_ps(iz1,jz2);
250 dx13 = _mm_sub_ps(ix1,jx3);
251 dy13 = _mm_sub_ps(iy1,jy3);
252 dz13 = _mm_sub_ps(iz1,jz3);
253 dx21 = _mm_sub_ps(ix2,jx1);
254 dy21 = _mm_sub_ps(iy2,jy1);
255 dz21 = _mm_sub_ps(iz2,jz1);
256 dx22 = _mm_sub_ps(ix2,jx2);
257 dy22 = _mm_sub_ps(iy2,jy2);
258 dz22 = _mm_sub_ps(iz2,jz2);
259 dx23 = _mm_sub_ps(ix2,jx3);
260 dy23 = _mm_sub_ps(iy2,jy3);
261 dz23 = _mm_sub_ps(iz2,jz3);
262 dx31 = _mm_sub_ps(ix3,jx1);
263 dy31 = _mm_sub_ps(iy3,jy1);
264 dz31 = _mm_sub_ps(iz3,jz1);
265 dx32 = _mm_sub_ps(ix3,jx2);
266 dy32 = _mm_sub_ps(iy3,jy2);
267 dz32 = _mm_sub_ps(iz3,jz2);
268 dx33 = _mm_sub_ps(ix3,jx3);
269 dy33 = _mm_sub_ps(iy3,jy3);
270 dz33 = _mm_sub_ps(iz3,jz3);
271
272 /* Calculate squared distance and things based on it */
273 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
274 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
275 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
276 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
277 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
278 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
279 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
280 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
281 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
282 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
283
284 rinv00 = sse2_invsqrt_f(rsq00);
285 rinv11 = sse2_invsqrt_f(rsq11);
286 rinv12 = sse2_invsqrt_f(rsq12);
287 rinv13 = sse2_invsqrt_f(rsq13);
288 rinv21 = sse2_invsqrt_f(rsq21);
289 rinv22 = sse2_invsqrt_f(rsq22);
290 rinv23 = sse2_invsqrt_f(rsq23);
291 rinv31 = sse2_invsqrt_f(rsq31);
292 rinv32 = sse2_invsqrt_f(rsq32);
293 rinv33 = sse2_invsqrt_f(rsq33);
294
295 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
296 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
297 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
298 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
299 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
300 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
301 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
302 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
303 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
304
305 fjx0 = _mm_setzero_ps();
306 fjy0 = _mm_setzero_ps();
307 fjz0 = _mm_setzero_ps();
308 fjx1 = _mm_setzero_ps();
309 fjy1 = _mm_setzero_ps();
310 fjz1 = _mm_setzero_ps();
311 fjx2 = _mm_setzero_ps();
312 fjy2 = _mm_setzero_ps();
313 fjz2 = _mm_setzero_ps();
314 fjx3 = _mm_setzero_ps();
315 fjy3 = _mm_setzero_ps();
316 fjz3 = _mm_setzero_ps();
317
318 /**************************
319 * CALCULATE INTERACTIONS *
320 **************************/
321
322 r00 = _mm_mul_ps(rsq00,rinv00);
323
324 /* Calculate table index by multiplying r with table scale and truncate to integer */
325 rt = _mm_mul_ps(r00,vftabscale);
326 vfitab = _mm_cvttps_epi32(rt);
327 vfeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
328 vfitab = _mm_slli_epi32(vfitab,3);
329
330 /* CUBIC SPLINE TABLE DISPERSION */
331 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
332 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
333 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
334 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
335 _MM_TRANSPOSE4_PS(Y,F,G,H);
336 Heps = _mm_mul_ps(vfeps,H);
337 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
338 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
339 vvdw6 = _mm_mul_ps(c6_00,VV);
340 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
341 fvdw6 = _mm_mul_ps(c6_00,FF);
342
343 /* CUBIC SPLINE TABLE REPULSION */
344 vfitab = _mm_add_epi32(vfitab,ifour);
345 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
346 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
347 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
348 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
349 _MM_TRANSPOSE4_PS(Y,F,G,H);
350 Heps = _mm_mul_ps(vfeps,H);
351 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
352 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
353 vvdw12 = _mm_mul_ps(c12_00,VV);
354 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
355 fvdw12 = _mm_mul_ps(c12_00,FF);
356 vvdw = _mm_add_ps(vvdw12,vvdw6);
357 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
358
359 /* Update potential sum for this i atom from the interaction with this j atom. */
360 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
361
362 fscal = fvdw;
363
364 /* Calculate temporary vectorial force */
365 tx = _mm_mul_ps(fscal,dx00);
366 ty = _mm_mul_ps(fscal,dy00);
367 tz = _mm_mul_ps(fscal,dz00);
368
369 /* Update vectorial force */
370 fix0 = _mm_add_ps(fix0,tx);
371 fiy0 = _mm_add_ps(fiy0,ty);
372 fiz0 = _mm_add_ps(fiz0,tz);
373
374 fjx0 = _mm_add_ps(fjx0,tx);
375 fjy0 = _mm_add_ps(fjy0,ty);
376 fjz0 = _mm_add_ps(fjz0,tz);
377
378 /**************************
379 * CALCULATE INTERACTIONS *
380 **************************/
381
382 /* REACTION-FIELD ELECTROSTATICS */
383 velec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
384 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
385
386 /* Update potential sum for this i atom from the interaction with this j atom. */
387 velecsum = _mm_add_ps(velecsum,velec);
388
389 fscal = felec;
390
391 /* Calculate temporary vectorial force */
392 tx = _mm_mul_ps(fscal,dx11);
393 ty = _mm_mul_ps(fscal,dy11);
394 tz = _mm_mul_ps(fscal,dz11);
395
396 /* Update vectorial force */
397 fix1 = _mm_add_ps(fix1,tx);
398 fiy1 = _mm_add_ps(fiy1,ty);
399 fiz1 = _mm_add_ps(fiz1,tz);
400
401 fjx1 = _mm_add_ps(fjx1,tx);
402 fjy1 = _mm_add_ps(fjy1,ty);
403 fjz1 = _mm_add_ps(fjz1,tz);
404
405 /**************************
406 * CALCULATE INTERACTIONS *
407 **************************/
408
409 /* REACTION-FIELD ELECTROSTATICS */
410 velec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
411 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
412
413 /* Update potential sum for this i atom from the interaction with this j atom. */
414 velecsum = _mm_add_ps(velecsum,velec);
415
416 fscal = felec;
417
418 /* Calculate temporary vectorial force */
419 tx = _mm_mul_ps(fscal,dx12);
420 ty = _mm_mul_ps(fscal,dy12);
421 tz = _mm_mul_ps(fscal,dz12);
422
423 /* Update vectorial force */
424 fix1 = _mm_add_ps(fix1,tx);
425 fiy1 = _mm_add_ps(fiy1,ty);
426 fiz1 = _mm_add_ps(fiz1,tz);
427
428 fjx2 = _mm_add_ps(fjx2,tx);
429 fjy2 = _mm_add_ps(fjy2,ty);
430 fjz2 = _mm_add_ps(fjz2,tz);
431
432 /**************************
433 * CALCULATE INTERACTIONS *
434 **************************/
435
436 /* REACTION-FIELD ELECTROSTATICS */
437 velec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
438 felec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
439
440 /* Update potential sum for this i atom from the interaction with this j atom. */
441 velecsum = _mm_add_ps(velecsum,velec);
442
443 fscal = felec;
444
445 /* Calculate temporary vectorial force */
446 tx = _mm_mul_ps(fscal,dx13);
447 ty = _mm_mul_ps(fscal,dy13);
448 tz = _mm_mul_ps(fscal,dz13);
449
450 /* Update vectorial force */
451 fix1 = _mm_add_ps(fix1,tx);
452 fiy1 = _mm_add_ps(fiy1,ty);
453 fiz1 = _mm_add_ps(fiz1,tz);
454
455 fjx3 = _mm_add_ps(fjx3,tx);
456 fjy3 = _mm_add_ps(fjy3,ty);
457 fjz3 = _mm_add_ps(fjz3,tz);
458
459 /**************************
460 * CALCULATE INTERACTIONS *
461 **************************/
462
463 /* REACTION-FIELD ELECTROSTATICS */
464 velec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
465 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
466
467 /* Update potential sum for this i atom from the interaction with this j atom. */
468 velecsum = _mm_add_ps(velecsum,velec);
469
470 fscal = felec;
471
472 /* Calculate temporary vectorial force */
473 tx = _mm_mul_ps(fscal,dx21);
474 ty = _mm_mul_ps(fscal,dy21);
475 tz = _mm_mul_ps(fscal,dz21);
476
477 /* Update vectorial force */
478 fix2 = _mm_add_ps(fix2,tx);
479 fiy2 = _mm_add_ps(fiy2,ty);
480 fiz2 = _mm_add_ps(fiz2,tz);
481
482 fjx1 = _mm_add_ps(fjx1,tx);
483 fjy1 = _mm_add_ps(fjy1,ty);
484 fjz1 = _mm_add_ps(fjz1,tz);
485
486 /**************************
487 * CALCULATE INTERACTIONS *
488 **************************/
489
490 /* REACTION-FIELD ELECTROSTATICS */
491 velec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
492 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
493
494 /* Update potential sum for this i atom from the interaction with this j atom. */
495 velecsum = _mm_add_ps(velecsum,velec);
496
497 fscal = felec;
498
499 /* Calculate temporary vectorial force */
500 tx = _mm_mul_ps(fscal,dx22);
501 ty = _mm_mul_ps(fscal,dy22);
502 tz = _mm_mul_ps(fscal,dz22);
503
504 /* Update vectorial force */
505 fix2 = _mm_add_ps(fix2,tx);
506 fiy2 = _mm_add_ps(fiy2,ty);
507 fiz2 = _mm_add_ps(fiz2,tz);
508
509 fjx2 = _mm_add_ps(fjx2,tx);
510 fjy2 = _mm_add_ps(fjy2,ty);
511 fjz2 = _mm_add_ps(fjz2,tz);
512
513 /**************************
514 * CALCULATE INTERACTIONS *
515 **************************/
516
517 /* REACTION-FIELD ELECTROSTATICS */
518 velec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
519 felec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
520
521 /* Update potential sum for this i atom from the interaction with this j atom. */
522 velecsum = _mm_add_ps(velecsum,velec);
523
524 fscal = felec;
525
526 /* Calculate temporary vectorial force */
527 tx = _mm_mul_ps(fscal,dx23);
528 ty = _mm_mul_ps(fscal,dy23);
529 tz = _mm_mul_ps(fscal,dz23);
530
531 /* Update vectorial force */
532 fix2 = _mm_add_ps(fix2,tx);
533 fiy2 = _mm_add_ps(fiy2,ty);
534 fiz2 = _mm_add_ps(fiz2,tz);
535
536 fjx3 = _mm_add_ps(fjx3,tx);
537 fjy3 = _mm_add_ps(fjy3,ty);
538 fjz3 = _mm_add_ps(fjz3,tz);
539
540 /**************************
541 * CALCULATE INTERACTIONS *
542 **************************/
543
544 /* REACTION-FIELD ELECTROSTATICS */
545 velec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
546 felec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
547
548 /* Update potential sum for this i atom from the interaction with this j atom. */
549 velecsum = _mm_add_ps(velecsum,velec);
550
551 fscal = felec;
552
553 /* Calculate temporary vectorial force */
554 tx = _mm_mul_ps(fscal,dx31);
555 ty = _mm_mul_ps(fscal,dy31);
556 tz = _mm_mul_ps(fscal,dz31);
557
558 /* Update vectorial force */
559 fix3 = _mm_add_ps(fix3,tx);
560 fiy3 = _mm_add_ps(fiy3,ty);
561 fiz3 = _mm_add_ps(fiz3,tz);
562
563 fjx1 = _mm_add_ps(fjx1,tx);
564 fjy1 = _mm_add_ps(fjy1,ty);
565 fjz1 = _mm_add_ps(fjz1,tz);
566
567 /**************************
568 * CALCULATE INTERACTIONS *
569 **************************/
570
571 /* REACTION-FIELD ELECTROSTATICS */
572 velec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
573 felec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
574
575 /* Update potential sum for this i atom from the interaction with this j atom. */
576 velecsum = _mm_add_ps(velecsum,velec);
577
578 fscal = felec;
579
580 /* Calculate temporary vectorial force */
581 tx = _mm_mul_ps(fscal,dx32);
582 ty = _mm_mul_ps(fscal,dy32);
583 tz = _mm_mul_ps(fscal,dz32);
584
585 /* Update vectorial force */
586 fix3 = _mm_add_ps(fix3,tx);
587 fiy3 = _mm_add_ps(fiy3,ty);
588 fiz3 = _mm_add_ps(fiz3,tz);
589
590 fjx2 = _mm_add_ps(fjx2,tx);
591 fjy2 = _mm_add_ps(fjy2,ty);
592 fjz2 = _mm_add_ps(fjz2,tz);
593
594 /**************************
595 * CALCULATE INTERACTIONS *
596 **************************/
597
598 /* REACTION-FIELD ELECTROSTATICS */
599 velec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
600 felec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
601
602 /* Update potential sum for this i atom from the interaction with this j atom. */
603 velecsum = _mm_add_ps(velecsum,velec);
604
605 fscal = felec;
606
607 /* Calculate temporary vectorial force */
608 tx = _mm_mul_ps(fscal,dx33);
609 ty = _mm_mul_ps(fscal,dy33);
610 tz = _mm_mul_ps(fscal,dz33);
611
612 /* Update vectorial force */
613 fix3 = _mm_add_ps(fix3,tx);
614 fiy3 = _mm_add_ps(fiy3,ty);
615 fiz3 = _mm_add_ps(fiz3,tz);
616
617 fjx3 = _mm_add_ps(fjx3,tx);
618 fjy3 = _mm_add_ps(fjy3,ty);
619 fjz3 = _mm_add_ps(fjz3,tz);
620
621 fjptrA = f+j_coord_offsetA;
622 fjptrB = f+j_coord_offsetB;
623 fjptrC = f+j_coord_offsetC;
624 fjptrD = f+j_coord_offsetD;
625
626 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
627 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
628 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
629
630 /* Inner loop uses 347 flops */
631 }
632
633 if(jidx<j_index_end)
634 {
635
636 /* Get j neighbor index, and coordinate index */
637 jnrlistA = jjnr[jidx];
638 jnrlistB = jjnr[jidx+1];
639 jnrlistC = jjnr[jidx+2];
640 jnrlistD = jjnr[jidx+3];
641 /* Sign of each element will be negative for non-real atoms.
642 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
643 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
644 */
645 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
646 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
647 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
648 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
649 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
650 j_coord_offsetA = DIM*jnrA;
651 j_coord_offsetB = DIM*jnrB;
652 j_coord_offsetC = DIM*jnrC;
653 j_coord_offsetD = DIM*jnrD;
654
655 /* load j atom coordinates */
656 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
657 x+j_coord_offsetC,x+j_coord_offsetD,
658 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
659 &jy2,&jz2,&jx3,&jy3,&jz3);
660
661 /* Calculate displacement vector */
662 dx00 = _mm_sub_ps(ix0,jx0);
663 dy00 = _mm_sub_ps(iy0,jy0);
664 dz00 = _mm_sub_ps(iz0,jz0);
665 dx11 = _mm_sub_ps(ix1,jx1);
666 dy11 = _mm_sub_ps(iy1,jy1);
667 dz11 = _mm_sub_ps(iz1,jz1);
668 dx12 = _mm_sub_ps(ix1,jx2);
669 dy12 = _mm_sub_ps(iy1,jy2);
670 dz12 = _mm_sub_ps(iz1,jz2);
671 dx13 = _mm_sub_ps(ix1,jx3);
672 dy13 = _mm_sub_ps(iy1,jy3);
673 dz13 = _mm_sub_ps(iz1,jz3);
674 dx21 = _mm_sub_ps(ix2,jx1);
675 dy21 = _mm_sub_ps(iy2,jy1);
676 dz21 = _mm_sub_ps(iz2,jz1);
677 dx22 = _mm_sub_ps(ix2,jx2);
678 dy22 = _mm_sub_ps(iy2,jy2);
679 dz22 = _mm_sub_ps(iz2,jz2);
680 dx23 = _mm_sub_ps(ix2,jx3);
681 dy23 = _mm_sub_ps(iy2,jy3);
682 dz23 = _mm_sub_ps(iz2,jz3);
683 dx31 = _mm_sub_ps(ix3,jx1);
684 dy31 = _mm_sub_ps(iy3,jy1);
685 dz31 = _mm_sub_ps(iz3,jz1);
686 dx32 = _mm_sub_ps(ix3,jx2);
687 dy32 = _mm_sub_ps(iy3,jy2);
688 dz32 = _mm_sub_ps(iz3,jz2);
689 dx33 = _mm_sub_ps(ix3,jx3);
690 dy33 = _mm_sub_ps(iy3,jy3);
691 dz33 = _mm_sub_ps(iz3,jz3);
692
693 /* Calculate squared distance and things based on it */
694 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
695 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
696 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
697 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
698 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
699 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
700 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
701 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
702 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
703 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
704
705 rinv00 = sse2_invsqrt_f(rsq00);
706 rinv11 = sse2_invsqrt_f(rsq11);
707 rinv12 = sse2_invsqrt_f(rsq12);
708 rinv13 = sse2_invsqrt_f(rsq13);
709 rinv21 = sse2_invsqrt_f(rsq21);
710 rinv22 = sse2_invsqrt_f(rsq22);
711 rinv23 = sse2_invsqrt_f(rsq23);
712 rinv31 = sse2_invsqrt_f(rsq31);
713 rinv32 = sse2_invsqrt_f(rsq32);
714 rinv33 = sse2_invsqrt_f(rsq33);
715
716 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
717 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
718 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
719 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
720 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
721 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
722 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
723 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
724 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
725
726 fjx0 = _mm_setzero_ps();
727 fjy0 = _mm_setzero_ps();
728 fjz0 = _mm_setzero_ps();
729 fjx1 = _mm_setzero_ps();
730 fjy1 = _mm_setzero_ps();
731 fjz1 = _mm_setzero_ps();
732 fjx2 = _mm_setzero_ps();
733 fjy2 = _mm_setzero_ps();
734 fjz2 = _mm_setzero_ps();
735 fjx3 = _mm_setzero_ps();
736 fjy3 = _mm_setzero_ps();
737 fjz3 = _mm_setzero_ps();
738
739 /**************************
740 * CALCULATE INTERACTIONS *
741 **************************/
742
743 r00 = _mm_mul_ps(rsq00,rinv00);
744 r00 = _mm_andnot_ps(dummy_mask,r00);
745
746 /* Calculate table index by multiplying r with table scale and truncate to integer */
747 rt = _mm_mul_ps(r00,vftabscale);
748 vfitab = _mm_cvttps_epi32(rt);
749 vfeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
750 vfitab = _mm_slli_epi32(vfitab,3);
751
752 /* CUBIC SPLINE TABLE DISPERSION */
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 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
761 vvdw6 = _mm_mul_ps(c6_00,VV);
762 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
763 fvdw6 = _mm_mul_ps(c6_00,FF);
764
765 /* CUBIC SPLINE TABLE REPULSION */
766 vfitab = _mm_add_epi32(vfitab,ifour);
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 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
775 vvdw12 = _mm_mul_ps(c12_00,VV);
776 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
777 fvdw12 = _mm_mul_ps(c12_00,FF);
778 vvdw = _mm_add_ps(vvdw12,vvdw6);
779 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
780
781 /* Update potential sum for this i atom from the interaction with this j atom. */
782 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
783 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
784
785 fscal = fvdw;
786
787 fscal = _mm_andnot_ps(dummy_mask,fscal);
788
789 /* Calculate temporary vectorial force */
790 tx = _mm_mul_ps(fscal,dx00);
791 ty = _mm_mul_ps(fscal,dy00);
792 tz = _mm_mul_ps(fscal,dz00);
793
794 /* Update vectorial force */
795 fix0 = _mm_add_ps(fix0,tx);
796 fiy0 = _mm_add_ps(fiy0,ty);
797 fiz0 = _mm_add_ps(fiz0,tz);
798
799 fjx0 = _mm_add_ps(fjx0,tx);
800 fjy0 = _mm_add_ps(fjy0,ty);
801 fjz0 = _mm_add_ps(fjz0,tz);
802
803 /**************************
804 * CALCULATE INTERACTIONS *
805 **************************/
806
807 /* REACTION-FIELD ELECTROSTATICS */
808 velec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
809 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
810
811 /* Update potential sum for this i atom from the interaction with this j atom. */
812 velec = _mm_andnot_ps(dummy_mask,velec);
813 velecsum = _mm_add_ps(velecsum,velec);
814
815 fscal = felec;
816
817 fscal = _mm_andnot_ps(dummy_mask,fscal);
818
819 /* Calculate temporary vectorial force */
820 tx = _mm_mul_ps(fscal,dx11);
821 ty = _mm_mul_ps(fscal,dy11);
822 tz = _mm_mul_ps(fscal,dz11);
823
824 /* Update vectorial force */
825 fix1 = _mm_add_ps(fix1,tx);
826 fiy1 = _mm_add_ps(fiy1,ty);
827 fiz1 = _mm_add_ps(fiz1,tz);
828
829 fjx1 = _mm_add_ps(fjx1,tx);
830 fjy1 = _mm_add_ps(fjy1,ty);
831 fjz1 = _mm_add_ps(fjz1,tz);
832
833 /**************************
834 * CALCULATE INTERACTIONS *
835 **************************/
836
837 /* REACTION-FIELD ELECTROSTATICS */
838 velec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
839 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
840
841 /* Update potential sum for this i atom from the interaction with this j atom. */
842 velec = _mm_andnot_ps(dummy_mask,velec);
843 velecsum = _mm_add_ps(velecsum,velec);
844
845 fscal = felec;
846
847 fscal = _mm_andnot_ps(dummy_mask,fscal);
848
849 /* Calculate temporary vectorial force */
850 tx = _mm_mul_ps(fscal,dx12);
851 ty = _mm_mul_ps(fscal,dy12);
852 tz = _mm_mul_ps(fscal,dz12);
853
854 /* Update vectorial force */
855 fix1 = _mm_add_ps(fix1,tx);
856 fiy1 = _mm_add_ps(fiy1,ty);
857 fiz1 = _mm_add_ps(fiz1,tz);
858
859 fjx2 = _mm_add_ps(fjx2,tx);
860 fjy2 = _mm_add_ps(fjy2,ty);
861 fjz2 = _mm_add_ps(fjz2,tz);
862
863 /**************************
864 * CALCULATE INTERACTIONS *
865 **************************/
866
867 /* REACTION-FIELD ELECTROSTATICS */
868 velec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
869 felec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
870
871 /* Update potential sum for this i atom from the interaction with this j atom. */
872 velec = _mm_andnot_ps(dummy_mask,velec);
873 velecsum = _mm_add_ps(velecsum,velec);
874
875 fscal = felec;
876
877 fscal = _mm_andnot_ps(dummy_mask,fscal);
878
879 /* Calculate temporary vectorial force */
880 tx = _mm_mul_ps(fscal,dx13);
881 ty = _mm_mul_ps(fscal,dy13);
882 tz = _mm_mul_ps(fscal,dz13);
883
884 /* Update vectorial force */
885 fix1 = _mm_add_ps(fix1,tx);
886 fiy1 = _mm_add_ps(fiy1,ty);
887 fiz1 = _mm_add_ps(fiz1,tz);
888
889 fjx3 = _mm_add_ps(fjx3,tx);
890 fjy3 = _mm_add_ps(fjy3,ty);
891 fjz3 = _mm_add_ps(fjz3,tz);
892
893 /**************************
894 * CALCULATE INTERACTIONS *
895 **************************/
896
897 /* REACTION-FIELD ELECTROSTATICS */
898 velec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
899 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
900
901 /* Update potential sum for this i atom from the interaction with this j atom. */
902 velec = _mm_andnot_ps(dummy_mask,velec);
903 velecsum = _mm_add_ps(velecsum,velec);
904
905 fscal = felec;
906
907 fscal = _mm_andnot_ps(dummy_mask,fscal);
908
909 /* Calculate temporary vectorial force */
910 tx = _mm_mul_ps(fscal,dx21);
911 ty = _mm_mul_ps(fscal,dy21);
912 tz = _mm_mul_ps(fscal,dz21);
913
914 /* Update vectorial force */
915 fix2 = _mm_add_ps(fix2,tx);
916 fiy2 = _mm_add_ps(fiy2,ty);
917 fiz2 = _mm_add_ps(fiz2,tz);
918
919 fjx1 = _mm_add_ps(fjx1,tx);
920 fjy1 = _mm_add_ps(fjy1,ty);
921 fjz1 = _mm_add_ps(fjz1,tz);
922
923 /**************************
924 * CALCULATE INTERACTIONS *
925 **************************/
926
927 /* REACTION-FIELD ELECTROSTATICS */
928 velec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
929 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
930
931 /* Update potential sum for this i atom from the interaction with this j atom. */
932 velec = _mm_andnot_ps(dummy_mask,velec);
933 velecsum = _mm_add_ps(velecsum,velec);
934
935 fscal = felec;
936
937 fscal = _mm_andnot_ps(dummy_mask,fscal);
938
939 /* Calculate temporary vectorial force */
940 tx = _mm_mul_ps(fscal,dx22);
941 ty = _mm_mul_ps(fscal,dy22);
942 tz = _mm_mul_ps(fscal,dz22);
943
944 /* Update vectorial force */
945 fix2 = _mm_add_ps(fix2,tx);
946 fiy2 = _mm_add_ps(fiy2,ty);
947 fiz2 = _mm_add_ps(fiz2,tz);
948
949 fjx2 = _mm_add_ps(fjx2,tx);
950 fjy2 = _mm_add_ps(fjy2,ty);
951 fjz2 = _mm_add_ps(fjz2,tz);
952
953 /**************************
954 * CALCULATE INTERACTIONS *
955 **************************/
956
957 /* REACTION-FIELD ELECTROSTATICS */
958 velec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
959 felec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
960
961 /* Update potential sum for this i atom from the interaction with this j atom. */
962 velec = _mm_andnot_ps(dummy_mask,velec);
963 velecsum = _mm_add_ps(velecsum,velec);
964
965 fscal = felec;
966
967 fscal = _mm_andnot_ps(dummy_mask,fscal);
968
969 /* Calculate temporary vectorial force */
970 tx = _mm_mul_ps(fscal,dx23);
971 ty = _mm_mul_ps(fscal,dy23);
972 tz = _mm_mul_ps(fscal,dz23);
973
974 /* Update vectorial force */
975 fix2 = _mm_add_ps(fix2,tx);
976 fiy2 = _mm_add_ps(fiy2,ty);
977 fiz2 = _mm_add_ps(fiz2,tz);
978
979 fjx3 = _mm_add_ps(fjx3,tx);
980 fjy3 = _mm_add_ps(fjy3,ty);
981 fjz3 = _mm_add_ps(fjz3,tz);
982
983 /**************************
984 * CALCULATE INTERACTIONS *
985 **************************/
986
987 /* REACTION-FIELD ELECTROSTATICS */
988 velec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
989 felec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
990
991 /* Update potential sum for this i atom from the interaction with this j atom. */
992 velec = _mm_andnot_ps(dummy_mask,velec);
993 velecsum = _mm_add_ps(velecsum,velec);
994
995 fscal = felec;
996
997 fscal = _mm_andnot_ps(dummy_mask,fscal);
998
999 /* Calculate temporary vectorial force */
1000 tx = _mm_mul_ps(fscal,dx31);
1001 ty = _mm_mul_ps(fscal,dy31);
1002 tz = _mm_mul_ps(fscal,dz31);
1003
1004 /* Update vectorial force */
1005 fix3 = _mm_add_ps(fix3,tx);
1006 fiy3 = _mm_add_ps(fiy3,ty);
1007 fiz3 = _mm_add_ps(fiz3,tz);
1008
1009 fjx1 = _mm_add_ps(fjx1,tx);
1010 fjy1 = _mm_add_ps(fjy1,ty);
1011 fjz1 = _mm_add_ps(fjz1,tz);
1012
1013 /**************************
1014 * CALCULATE INTERACTIONS *
1015 **************************/
1016
1017 /* REACTION-FIELD ELECTROSTATICS */
1018 velec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1019 felec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1020
1021 /* Update potential sum for this i atom from the interaction with this j atom. */
1022 velec = _mm_andnot_ps(dummy_mask,velec);
1023 velecsum = _mm_add_ps(velecsum,velec);
1024
1025 fscal = felec;
1026
1027 fscal = _mm_andnot_ps(dummy_mask,fscal);
1028
1029 /* Calculate temporary vectorial force */
1030 tx = _mm_mul_ps(fscal,dx32);
1031 ty = _mm_mul_ps(fscal,dy32);
1032 tz = _mm_mul_ps(fscal,dz32);
1033
1034 /* Update vectorial force */
1035 fix3 = _mm_add_ps(fix3,tx);
1036 fiy3 = _mm_add_ps(fiy3,ty);
1037 fiz3 = _mm_add_ps(fiz3,tz);
1038
1039 fjx2 = _mm_add_ps(fjx2,tx);
1040 fjy2 = _mm_add_ps(fjy2,ty);
1041 fjz2 = _mm_add_ps(fjz2,tz);
1042
1043 /**************************
1044 * CALCULATE INTERACTIONS *
1045 **************************/
1046
1047 /* REACTION-FIELD ELECTROSTATICS */
1048 velec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1049 felec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1050
1051 /* Update potential sum for this i atom from the interaction with this j atom. */
1052 velec = _mm_andnot_ps(dummy_mask,velec);
1053 velecsum = _mm_add_ps(velecsum,velec);
1054
1055 fscal = felec;
1056
1057 fscal = _mm_andnot_ps(dummy_mask,fscal);
1058
1059 /* Calculate temporary vectorial force */
1060 tx = _mm_mul_ps(fscal,dx33);
1061 ty = _mm_mul_ps(fscal,dy33);
1062 tz = _mm_mul_ps(fscal,dz33);
1063
1064 /* Update vectorial force */
1065 fix3 = _mm_add_ps(fix3,tx);
1066 fiy3 = _mm_add_ps(fiy3,ty);
1067 fiz3 = _mm_add_ps(fiz3,tz);
1068
1069 fjx3 = _mm_add_ps(fjx3,tx);
1070 fjy3 = _mm_add_ps(fjy3,ty);
1071 fjz3 = _mm_add_ps(fjz3,tz);
1072
1073 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1074 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1075 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1076 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1077
1078 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1079 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1080 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1081
1082 /* Inner loop uses 348 flops */
1083 }
1084
1085 /* End of innermost loop */
1086
1087 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1088 f+i_coord_offset,fshift+i_shift_offset);
1089
1090 ggid = gid[iidx];
1091 /* Update potential energies */
1092 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1093 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1094
1095 /* Increment number of inner iterations */
1096 inneriter += j_index_end - j_index_start;
1097
1098 /* Outer loop uses 26 flops */
1099 }
1100
1101 /* Increment number of outer iterations */
1102 outeriter += nri;
1103
1104 /* Update outer/inner flops */
1105
1106 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*348);
1107 }
1108 /*
1109 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse2_single
1110 * Electrostatics interaction: ReactionField
1111 * VdW interaction: CubicSplineTable
1112 * Geometry: Water4-Water4
1113 * Calculate force/pot: Force
1114 */
1115 void
1116 nb_kernel_ElecRF_VdwCSTab_GeomW4W4_F_sse2_single
1117 (t_nblist * gmx_restrict nlist,
1118 rvec * gmx_restrict xx,
1119 rvec * gmx_restrict ff,
1120 struct t_forcerec * gmx_restrict fr,
1121 t_mdatoms * gmx_restrict mdatoms,
1122 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1123 t_nrnb * gmx_restrict nrnb)
1124 {
1125 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1126 * just 0 for non-waters.
1127 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1128 * jnr indices corresponding to data put in the four positions in the SIMD register.
1129 */
1130 int i_shift_offset,i_coord_offset,outeriter,inneriter;
1131 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1132 int jnrA,jnrB,jnrC,jnrD;
1133 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1134 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1135 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
1136 real rcutoff_scalar;
1137 real *shiftvec,*fshift,*x,*f;
1138 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1139 real scratch[4*DIM];
1140 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1141 int vdwioffset0;
1142 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1143 int vdwioffset1;
1144 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1145 int vdwioffset2;
1146 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1147 int vdwioffset3;
1148 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1149 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1150 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1151 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1152 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1153 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1154 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1155 int vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1156 __m128 jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1157 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1158 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1159 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1160 __m128 dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1161 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1162 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1163 __m128 dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1164 __m128 dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1165 __m128 dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1166 __m128 dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1167 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
1168 real *charge;
1169 int nvdwtype;
1170 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1171 int *vdwtype;
1172 real *vdwparam;
1173 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
1174 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
1175 __m128i vfitab;
1176 __m128i ifour = _mm_set1_epi32(4);
1177 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1178 real *vftab;
1179 __m128 dummy_mask,cutoff_mask;
1180 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1181 __m128 one = _mm_set1_ps(1.0);
1182 __m128 two = _mm_set1_ps(2.0);
1183 x = xx[0];
1184 f = ff[0];
1185
1186 nri = nlist->nri;
1187 iinr = nlist->iinr;
1188 jindex = nlist->jindex;
1189 jjnr = nlist->jjnr;
1190 shiftidx = nlist->shift;
1191 gid = nlist->gid;
1192 shiftvec = fr->shift_vec[0];
1193 fshift = fr->fshift[0];
1194 facel = _mm_set1_ps(fr->ic->epsfac);
1195 charge = mdatoms->chargeA;
1196 krf = _mm_set1_ps(fr->ic->k_rf);
1197 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
1198 crf = _mm_set1_ps(fr->ic->c_rf);
1199 nvdwtype = fr->ntype;
1200 vdwparam = fr->nbfp;
1201 vdwtype = mdatoms->typeA;
1202
1203 vftab = kernel_data->table_vdw->data;
1204 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
1205
1206 /* Setup water-specific parameters */
1207 inr = nlist->iinr[0];
1208 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1209 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1210 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1211 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
1212
1213 jq1 = _mm_set1_ps(charge[inr+1]);
1214 jq2 = _mm_set1_ps(charge[inr+2]);
1215 jq3 = _mm_set1_ps(charge[inr+3]);
1216 vdwjidx0A = 2*vdwtype[inr+0];
1217 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1218 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1219 qq11 = _mm_mul_ps(iq1,jq1);
1220 qq12 = _mm_mul_ps(iq1,jq2);
1221 qq13 = _mm_mul_ps(iq1,jq3);
1222 qq21 = _mm_mul_ps(iq2,jq1);
1223 qq22 = _mm_mul_ps(iq2,jq2);
1224 qq23 = _mm_mul_ps(iq2,jq3);
1225 qq31 = _mm_mul_ps(iq3,jq1);
1226 qq32 = _mm_mul_ps(iq3,jq2);
1227 qq33 = _mm_mul_ps(iq3,jq3);
1228
1229 /* Avoid stupid compiler warnings */
1230 jnrA = jnrB = jnrC = jnrD = 0;
1231 j_coord_offsetA = 0;
1232 j_coord_offsetB = 0;
1233 j_coord_offsetC = 0;
1234 j_coord_offsetD = 0;
1235
1236 outeriter = 0;
1237 inneriter = 0;
1238
1239 for(iidx=0;iidx<4*DIM;iidx++)
1240 {
1241 scratch[iidx] = 0.0;
1242 }
1243
1244 /* Start outer loop over neighborlists */
1245 for(iidx=0; iidx<nri; iidx++)
1246 {
1247 /* Load shift vector for this list */
1248 i_shift_offset = DIM*shiftidx[iidx];
1249
1250 /* Load limits for loop over neighbors */
1251 j_index_start = jindex[iidx];
1252 j_index_end = jindex[iidx+1];
1253
1254 /* Get outer coordinate index */
1255 inr = iinr[iidx];
1256 i_coord_offset = DIM*inr;
1257
1258 /* Load i particle coords and add shift vector */
1259 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1260 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1261
1262 fix0 = _mm_setzero_ps();
1263 fiy0 = _mm_setzero_ps();
1264 fiz0 = _mm_setzero_ps();
1265 fix1 = _mm_setzero_ps();
1266 fiy1 = _mm_setzero_ps();
1267 fiz1 = _mm_setzero_ps();
1268 fix2 = _mm_setzero_ps();
1269 fiy2 = _mm_setzero_ps();
1270 fiz2 = _mm_setzero_ps();
1271 fix3 = _mm_setzero_ps();
1272 fiy3 = _mm_setzero_ps();
1273 fiz3 = _mm_setzero_ps();
1274
1275 /* Start inner kernel loop */
1276 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1277 {
1278
1279 /* Get j neighbor index, and coordinate index */
1280 jnrA = jjnr[jidx];
1281 jnrB = jjnr[jidx+1];
1282 jnrC = jjnr[jidx+2];
1283 jnrD = jjnr[jidx+3];
1284 j_coord_offsetA = DIM*jnrA;
1285 j_coord_offsetB = DIM*jnrB;
1286 j_coord_offsetC = DIM*jnrC;
1287 j_coord_offsetD = DIM*jnrD;
1288
1289 /* load j atom coordinates */
1290 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1291 x+j_coord_offsetC,x+j_coord_offsetD,
1292 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1293 &jy2,&jz2,&jx3,&jy3,&jz3);
1294
1295 /* Calculate displacement vector */
1296 dx00 = _mm_sub_ps(ix0,jx0);
1297 dy00 = _mm_sub_ps(iy0,jy0);
1298 dz00 = _mm_sub_ps(iz0,jz0);
1299 dx11 = _mm_sub_ps(ix1,jx1);
1300 dy11 = _mm_sub_ps(iy1,jy1);
1301 dz11 = _mm_sub_ps(iz1,jz1);
1302 dx12 = _mm_sub_ps(ix1,jx2);
1303 dy12 = _mm_sub_ps(iy1,jy2);
1304 dz12 = _mm_sub_ps(iz1,jz2);
1305 dx13 = _mm_sub_ps(ix1,jx3);
1306 dy13 = _mm_sub_ps(iy1,jy3);
1307 dz13 = _mm_sub_ps(iz1,jz3);
1308 dx21 = _mm_sub_ps(ix2,jx1);
1309 dy21 = _mm_sub_ps(iy2,jy1);
1310 dz21 = _mm_sub_ps(iz2,jz1);
1311 dx22 = _mm_sub_ps(ix2,jx2);
1312 dy22 = _mm_sub_ps(iy2,jy2);
1313 dz22 = _mm_sub_ps(iz2,jz2);
1314 dx23 = _mm_sub_ps(ix2,jx3);
1315 dy23 = _mm_sub_ps(iy2,jy3);
1316 dz23 = _mm_sub_ps(iz2,jz3);
1317 dx31 = _mm_sub_ps(ix3,jx1);
1318 dy31 = _mm_sub_ps(iy3,jy1);
1319 dz31 = _mm_sub_ps(iz3,jz1);
1320 dx32 = _mm_sub_ps(ix3,jx2);
1321 dy32 = _mm_sub_ps(iy3,jy2);
1322 dz32 = _mm_sub_ps(iz3,jz2);
1323 dx33 = _mm_sub_ps(ix3,jx3);
1324 dy33 = _mm_sub_ps(iy3,jy3);
1325 dz33 = _mm_sub_ps(iz3,jz3);
1326
1327 /* Calculate squared distance and things based on it */
1328 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1329 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1330 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1331 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1332 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1333 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1334 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1335 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1336 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1337 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1338
1339 rinv00 = sse2_invsqrt_f(rsq00);
1340 rinv11 = sse2_invsqrt_f(rsq11);
1341 rinv12 = sse2_invsqrt_f(rsq12);
1342 rinv13 = sse2_invsqrt_f(rsq13);
1343 rinv21 = sse2_invsqrt_f(rsq21);
1344 rinv22 = sse2_invsqrt_f(rsq22);
1345 rinv23 = sse2_invsqrt_f(rsq23);
1346 rinv31 = sse2_invsqrt_f(rsq31);
1347 rinv32 = sse2_invsqrt_f(rsq32);
1348 rinv33 = sse2_invsqrt_f(rsq33);
1349
1350 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1351 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1352 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1353 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1354 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1355 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1356 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1357 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1358 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1359
1360 fjx0 = _mm_setzero_ps();
1361 fjy0 = _mm_setzero_ps();
1362 fjz0 = _mm_setzero_ps();
1363 fjx1 = _mm_setzero_ps();
1364 fjy1 = _mm_setzero_ps();
1365 fjz1 = _mm_setzero_ps();
1366 fjx2 = _mm_setzero_ps();
1367 fjy2 = _mm_setzero_ps();
1368 fjz2 = _mm_setzero_ps();
1369 fjx3 = _mm_setzero_ps();
1370 fjy3 = _mm_setzero_ps();
1371 fjz3 = _mm_setzero_ps();
1372
1373 /**************************
1374 * CALCULATE INTERACTIONS *
1375 **************************/
1376
1377 r00 = _mm_mul_ps(rsq00,rinv00);
1378
1379 /* Calculate table index by multiplying r with table scale and truncate to integer */
1380 rt = _mm_mul_ps(r00,vftabscale);
1381 vfitab = _mm_cvttps_epi32(rt);
1382 vfeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1383 vfitab = _mm_slli_epi32(vfitab,3);
1384
1385 /* CUBIC SPLINE TABLE DISPERSION */
1386 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1387 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1388 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1389 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1390 _MM_TRANSPOSE4_PS(Y,F,G,H);
1391 Heps = _mm_mul_ps(vfeps,H);
1392 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1393 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1394 fvdw6 = _mm_mul_ps(c6_00,FF);
1395
1396 /* CUBIC SPLINE TABLE REPULSION */
1397 vfitab = _mm_add_epi32(vfitab,ifour);
1398 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1399 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1400 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1401 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1402 _MM_TRANSPOSE4_PS(Y,F,G,H);
1403 Heps = _mm_mul_ps(vfeps,H);
1404 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1405 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1406 fvdw12 = _mm_mul_ps(c12_00,FF);
1407 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1408
1409 fscal = fvdw;
1410
1411 /* Calculate temporary vectorial force */
1412 tx = _mm_mul_ps(fscal,dx00);
1413 ty = _mm_mul_ps(fscal,dy00);
1414 tz = _mm_mul_ps(fscal,dz00);
1415
1416 /* Update vectorial force */
1417 fix0 = _mm_add_ps(fix0,tx);
1418 fiy0 = _mm_add_ps(fiy0,ty);
1419 fiz0 = _mm_add_ps(fiz0,tz);
1420
1421 fjx0 = _mm_add_ps(fjx0,tx);
1422 fjy0 = _mm_add_ps(fjy0,ty);
1423 fjz0 = _mm_add_ps(fjz0,tz);
1424
1425 /**************************
1426 * CALCULATE INTERACTIONS *
1427 **************************/
1428
1429 /* REACTION-FIELD ELECTROSTATICS */
1430 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1431
1432 fscal = felec;
1433
1434 /* Calculate temporary vectorial force */
1435 tx = _mm_mul_ps(fscal,dx11);
1436 ty = _mm_mul_ps(fscal,dy11);
1437 tz = _mm_mul_ps(fscal,dz11);
1438
1439 /* Update vectorial force */
1440 fix1 = _mm_add_ps(fix1,tx);
1441 fiy1 = _mm_add_ps(fiy1,ty);
1442 fiz1 = _mm_add_ps(fiz1,tz);
1443
1444 fjx1 = _mm_add_ps(fjx1,tx);
1445 fjy1 = _mm_add_ps(fjy1,ty);
1446 fjz1 = _mm_add_ps(fjz1,tz);
1447
1448 /**************************
1449 * CALCULATE INTERACTIONS *
1450 **************************/
1451
1452 /* REACTION-FIELD ELECTROSTATICS */
1453 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1454
1455 fscal = felec;
1456
1457 /* Calculate temporary vectorial force */
1458 tx = _mm_mul_ps(fscal,dx12);
1459 ty = _mm_mul_ps(fscal,dy12);
1460 tz = _mm_mul_ps(fscal,dz12);
1461
1462 /* Update vectorial force */
1463 fix1 = _mm_add_ps(fix1,tx);
1464 fiy1 = _mm_add_ps(fiy1,ty);
1465 fiz1 = _mm_add_ps(fiz1,tz);
1466
1467 fjx2 = _mm_add_ps(fjx2,tx);
1468 fjy2 = _mm_add_ps(fjy2,ty);
1469 fjz2 = _mm_add_ps(fjz2,tz);
1470
1471 /**************************
1472 * CALCULATE INTERACTIONS *
1473 **************************/
1474
1475 /* REACTION-FIELD ELECTROSTATICS */
1476 felec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1477
1478 fscal = felec;
1479
1480 /* Calculate temporary vectorial force */
1481 tx = _mm_mul_ps(fscal,dx13);
1482 ty = _mm_mul_ps(fscal,dy13);
1483 tz = _mm_mul_ps(fscal,dz13);
1484
1485 /* Update vectorial force */
1486 fix1 = _mm_add_ps(fix1,tx);
1487 fiy1 = _mm_add_ps(fiy1,ty);
1488 fiz1 = _mm_add_ps(fiz1,tz);
1489
1490 fjx3 = _mm_add_ps(fjx3,tx);
1491 fjy3 = _mm_add_ps(fjy3,ty);
1492 fjz3 = _mm_add_ps(fjz3,tz);
1493
1494 /**************************
1495 * CALCULATE INTERACTIONS *
1496 **************************/
1497
1498 /* REACTION-FIELD ELECTROSTATICS */
1499 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1500
1501 fscal = felec;
1502
1503 /* Calculate temporary vectorial force */
1504 tx = _mm_mul_ps(fscal,dx21);
1505 ty = _mm_mul_ps(fscal,dy21);
1506 tz = _mm_mul_ps(fscal,dz21);
1507
1508 /* Update vectorial force */
1509 fix2 = _mm_add_ps(fix2,tx);
1510 fiy2 = _mm_add_ps(fiy2,ty);
1511 fiz2 = _mm_add_ps(fiz2,tz);
1512
1513 fjx1 = _mm_add_ps(fjx1,tx);
1514 fjy1 = _mm_add_ps(fjy1,ty);
1515 fjz1 = _mm_add_ps(fjz1,tz);
1516
1517 /**************************
1518 * CALCULATE INTERACTIONS *
1519 **************************/
1520
1521 /* REACTION-FIELD ELECTROSTATICS */
1522 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1523
1524 fscal = felec;
1525
1526 /* Calculate temporary vectorial force */
1527 tx = _mm_mul_ps(fscal,dx22);
1528 ty = _mm_mul_ps(fscal,dy22);
1529 tz = _mm_mul_ps(fscal,dz22);
1530
1531 /* Update vectorial force */
1532 fix2 = _mm_add_ps(fix2,tx);
1533 fiy2 = _mm_add_ps(fiy2,ty);
1534 fiz2 = _mm_add_ps(fiz2,tz);
1535
1536 fjx2 = _mm_add_ps(fjx2,tx);
1537 fjy2 = _mm_add_ps(fjy2,ty);
1538 fjz2 = _mm_add_ps(fjz2,tz);
1539
1540 /**************************
1541 * CALCULATE INTERACTIONS *
1542 **************************/
1543
1544 /* REACTION-FIELD ELECTROSTATICS */
1545 felec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1546
1547 fscal = felec;
1548
1549 /* Calculate temporary vectorial force */
1550 tx = _mm_mul_ps(fscal,dx23);
1551 ty = _mm_mul_ps(fscal,dy23);
1552 tz = _mm_mul_ps(fscal,dz23);
1553
1554 /* Update vectorial force */
1555 fix2 = _mm_add_ps(fix2,tx);
1556 fiy2 = _mm_add_ps(fiy2,ty);
1557 fiz2 = _mm_add_ps(fiz2,tz);
1558
1559 fjx3 = _mm_add_ps(fjx3,tx);
1560 fjy3 = _mm_add_ps(fjy3,ty);
1561 fjz3 = _mm_add_ps(fjz3,tz);
1562
1563 /**************************
1564 * CALCULATE INTERACTIONS *
1565 **************************/
1566
1567 /* REACTION-FIELD ELECTROSTATICS */
1568 felec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1569
1570 fscal = felec;
1571
1572 /* Calculate temporary vectorial force */
1573 tx = _mm_mul_ps(fscal,dx31);
1574 ty = _mm_mul_ps(fscal,dy31);
1575 tz = _mm_mul_ps(fscal,dz31);
1576
1577 /* Update vectorial force */
1578 fix3 = _mm_add_ps(fix3,tx);
1579 fiy3 = _mm_add_ps(fiy3,ty);
1580 fiz3 = _mm_add_ps(fiz3,tz);
1581
1582 fjx1 = _mm_add_ps(fjx1,tx);
1583 fjy1 = _mm_add_ps(fjy1,ty);
1584 fjz1 = _mm_add_ps(fjz1,tz);
1585
1586 /**************************
1587 * CALCULATE INTERACTIONS *
1588 **************************/
1589
1590 /* REACTION-FIELD ELECTROSTATICS */
1591 felec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1592
1593 fscal = felec;
1594
1595 /* Calculate temporary vectorial force */
1596 tx = _mm_mul_ps(fscal,dx32);
1597 ty = _mm_mul_ps(fscal,dy32);
1598 tz = _mm_mul_ps(fscal,dz32);
1599
1600 /* Update vectorial force */
1601 fix3 = _mm_add_ps(fix3,tx);
1602 fiy3 = _mm_add_ps(fiy3,ty);
1603 fiz3 = _mm_add_ps(fiz3,tz);
1604
1605 fjx2 = _mm_add_ps(fjx2,tx);
1606 fjy2 = _mm_add_ps(fjy2,ty);
1607 fjz2 = _mm_add_ps(fjz2,tz);
1608
1609 /**************************
1610 * CALCULATE INTERACTIONS *
1611 **************************/
1612
1613 /* REACTION-FIELD ELECTROSTATICS */
1614 felec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1615
1616 fscal = felec;
1617
1618 /* Calculate temporary vectorial force */
1619 tx = _mm_mul_ps(fscal,dx33);
1620 ty = _mm_mul_ps(fscal,dy33);
1621 tz = _mm_mul_ps(fscal,dz33);
1622
1623 /* Update vectorial force */
1624 fix3 = _mm_add_ps(fix3,tx);
1625 fiy3 = _mm_add_ps(fiy3,ty);
1626 fiz3 = _mm_add_ps(fiz3,tz);
1627
1628 fjx3 = _mm_add_ps(fjx3,tx);
1629 fjy3 = _mm_add_ps(fjy3,ty);
1630 fjz3 = _mm_add_ps(fjz3,tz);
1631
1632 fjptrA = f+j_coord_offsetA;
1633 fjptrB = f+j_coord_offsetB;
1634 fjptrC = f+j_coord_offsetC;
1635 fjptrD = f+j_coord_offsetD;
1636
1637 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1638 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1639 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1640
1641 /* Inner loop uses 294 flops */
1642 }
1643
1644 if(jidx<j_index_end)
1645 {
1646
1647 /* Get j neighbor index, and coordinate index */
1648 jnrlistA = jjnr[jidx];
1649 jnrlistB = jjnr[jidx+1];
1650 jnrlistC = jjnr[jidx+2];
1651 jnrlistD = jjnr[jidx+3];
1652 /* Sign of each element will be negative for non-real atoms.
1653 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1654 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1655 */
1656 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1657 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1658 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1659 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1660 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1661 j_coord_offsetA = DIM*jnrA;
1662 j_coord_offsetB = DIM*jnrB;
1663 j_coord_offsetC = DIM*jnrC;
1664 j_coord_offsetD = DIM*jnrD;
1665
1666 /* load j atom coordinates */
1667 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1668 x+j_coord_offsetC,x+j_coord_offsetD,
1669 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1670 &jy2,&jz2,&jx3,&jy3,&jz3);
1671
1672 /* Calculate displacement vector */
1673 dx00 = _mm_sub_ps(ix0,jx0);
1674 dy00 = _mm_sub_ps(iy0,jy0);
1675 dz00 = _mm_sub_ps(iz0,jz0);
1676 dx11 = _mm_sub_ps(ix1,jx1);
1677 dy11 = _mm_sub_ps(iy1,jy1);
1678 dz11 = _mm_sub_ps(iz1,jz1);
1679 dx12 = _mm_sub_ps(ix1,jx2);
1680 dy12 = _mm_sub_ps(iy1,jy2);
1681 dz12 = _mm_sub_ps(iz1,jz2);
1682 dx13 = _mm_sub_ps(ix1,jx3);
1683 dy13 = _mm_sub_ps(iy1,jy3);
1684 dz13 = _mm_sub_ps(iz1,jz3);
1685 dx21 = _mm_sub_ps(ix2,jx1);
1686 dy21 = _mm_sub_ps(iy2,jy1);
1687 dz21 = _mm_sub_ps(iz2,jz1);
1688 dx22 = _mm_sub_ps(ix2,jx2);
1689 dy22 = _mm_sub_ps(iy2,jy2);
1690 dz22 = _mm_sub_ps(iz2,jz2);
1691 dx23 = _mm_sub_ps(ix2,jx3);
1692 dy23 = _mm_sub_ps(iy2,jy3);
1693 dz23 = _mm_sub_ps(iz2,jz3);
1694 dx31 = _mm_sub_ps(ix3,jx1);
1695 dy31 = _mm_sub_ps(iy3,jy1);
1696 dz31 = _mm_sub_ps(iz3,jz1);
1697 dx32 = _mm_sub_ps(ix3,jx2);
1698 dy32 = _mm_sub_ps(iy3,jy2);
1699 dz32 = _mm_sub_ps(iz3,jz2);
1700 dx33 = _mm_sub_ps(ix3,jx3);
1701 dy33 = _mm_sub_ps(iy3,jy3);
1702 dz33 = _mm_sub_ps(iz3,jz3);
1703
1704 /* Calculate squared distance and things based on it */
1705 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1706 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1707 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1708 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1709 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1710 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1711 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1712 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1713 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1714 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1715
1716 rinv00 = sse2_invsqrt_f(rsq00);
1717 rinv11 = sse2_invsqrt_f(rsq11);
1718 rinv12 = sse2_invsqrt_f(rsq12);
1719 rinv13 = sse2_invsqrt_f(rsq13);
1720 rinv21 = sse2_invsqrt_f(rsq21);
1721 rinv22 = sse2_invsqrt_f(rsq22);
1722 rinv23 = sse2_invsqrt_f(rsq23);
1723 rinv31 = sse2_invsqrt_f(rsq31);
1724 rinv32 = sse2_invsqrt_f(rsq32);
1725 rinv33 = sse2_invsqrt_f(rsq33);
1726
1727 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1728 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1729 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1730 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1731 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1732 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1733 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1734 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1735 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1736
1737 fjx0 = _mm_setzero_ps();
1738 fjy0 = _mm_setzero_ps();
1739 fjz0 = _mm_setzero_ps();
1740 fjx1 = _mm_setzero_ps();
1741 fjy1 = _mm_setzero_ps();
1742 fjz1 = _mm_setzero_ps();
1743 fjx2 = _mm_setzero_ps();
1744 fjy2 = _mm_setzero_ps();
1745 fjz2 = _mm_setzero_ps();
1746 fjx3 = _mm_setzero_ps();
1747 fjy3 = _mm_setzero_ps();
1748 fjz3 = _mm_setzero_ps();
1749
1750 /**************************
1751 * CALCULATE INTERACTIONS *
1752 **************************/
1753
1754 r00 = _mm_mul_ps(rsq00,rinv00);
1755 r00 = _mm_andnot_ps(dummy_mask,r00);
1756
1757 /* Calculate table index by multiplying r with table scale and truncate to integer */
1758 rt = _mm_mul_ps(r00,vftabscale);
1759 vfitab = _mm_cvttps_epi32(rt);
1760 vfeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1761 vfitab = _mm_slli_epi32(vfitab,3);
1762
1763 /* CUBIC SPLINE TABLE DISPERSION */
1764 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1765 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1766 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1767 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1768 _MM_TRANSPOSE4_PS(Y,F,G,H);
1769 Heps = _mm_mul_ps(vfeps,H);
1770 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1771 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1772 fvdw6 = _mm_mul_ps(c6_00,FF);
1773
1774 /* CUBIC SPLINE TABLE REPULSION */
1775 vfitab = _mm_add_epi32(vfitab,ifour);
1776 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1777 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1778 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1779 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1780 _MM_TRANSPOSE4_PS(Y,F,G,H);
1781 Heps = _mm_mul_ps(vfeps,H);
1782 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1783 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1784 fvdw12 = _mm_mul_ps(c12_00,FF);
1785 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1786
1787 fscal = fvdw;
1788
1789 fscal = _mm_andnot_ps(dummy_mask,fscal);
1790
1791 /* Calculate temporary vectorial force */
1792 tx = _mm_mul_ps(fscal,dx00);
1793 ty = _mm_mul_ps(fscal,dy00);
1794 tz = _mm_mul_ps(fscal,dz00);
1795
1796 /* Update vectorial force */
1797 fix0 = _mm_add_ps(fix0,tx);
1798 fiy0 = _mm_add_ps(fiy0,ty);
1799 fiz0 = _mm_add_ps(fiz0,tz);
1800
1801 fjx0 = _mm_add_ps(fjx0,tx);
1802 fjy0 = _mm_add_ps(fjy0,ty);
1803 fjz0 = _mm_add_ps(fjz0,tz);
1804
1805 /**************************
1806 * CALCULATE INTERACTIONS *
1807 **************************/
1808
1809 /* REACTION-FIELD ELECTROSTATICS */
1810 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1811
1812 fscal = felec;
1813
1814 fscal = _mm_andnot_ps(dummy_mask,fscal);
1815
1816 /* Calculate temporary vectorial force */
1817 tx = _mm_mul_ps(fscal,dx11);
1818 ty = _mm_mul_ps(fscal,dy11);
1819 tz = _mm_mul_ps(fscal,dz11);
1820
1821 /* Update vectorial force */
1822 fix1 = _mm_add_ps(fix1,tx);
1823 fiy1 = _mm_add_ps(fiy1,ty);
1824 fiz1 = _mm_add_ps(fiz1,tz);
1825
1826 fjx1 = _mm_add_ps(fjx1,tx);
1827 fjy1 = _mm_add_ps(fjy1,ty);
1828 fjz1 = _mm_add_ps(fjz1,tz);
1829
1830 /**************************
1831 * CALCULATE INTERACTIONS *
1832 **************************/
1833
1834 /* REACTION-FIELD ELECTROSTATICS */
1835 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1836
1837 fscal = felec;
1838
1839 fscal = _mm_andnot_ps(dummy_mask,fscal);
1840
1841 /* Calculate temporary vectorial force */
1842 tx = _mm_mul_ps(fscal,dx12);
1843 ty = _mm_mul_ps(fscal,dy12);
1844 tz = _mm_mul_ps(fscal,dz12);
1845
1846 /* Update vectorial force */
1847 fix1 = _mm_add_ps(fix1,tx);
1848 fiy1 = _mm_add_ps(fiy1,ty);
1849 fiz1 = _mm_add_ps(fiz1,tz);
1850
1851 fjx2 = _mm_add_ps(fjx2,tx);
1852 fjy2 = _mm_add_ps(fjy2,ty);
1853 fjz2 = _mm_add_ps(fjz2,tz);
1854
1855 /**************************
1856 * CALCULATE INTERACTIONS *
1857 **************************/
1858
1859 /* REACTION-FIELD ELECTROSTATICS */
1860 felec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1861
1862 fscal = felec;
1863
1864 fscal = _mm_andnot_ps(dummy_mask,fscal);
1865
1866 /* Calculate temporary vectorial force */
1867 tx = _mm_mul_ps(fscal,dx13);
1868 ty = _mm_mul_ps(fscal,dy13);
1869 tz = _mm_mul_ps(fscal,dz13);
1870
1871 /* Update vectorial force */
1872 fix1 = _mm_add_ps(fix1,tx);
1873 fiy1 = _mm_add_ps(fiy1,ty);
1874 fiz1 = _mm_add_ps(fiz1,tz);
1875
1876 fjx3 = _mm_add_ps(fjx3,tx);
1877 fjy3 = _mm_add_ps(fjy3,ty);
1878 fjz3 = _mm_add_ps(fjz3,tz);
1879
1880 /**************************
1881 * CALCULATE INTERACTIONS *
1882 **************************/
1883
1884 /* REACTION-FIELD ELECTROSTATICS */
1885 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1886
1887 fscal = felec;
1888
1889 fscal = _mm_andnot_ps(dummy_mask,fscal);
1890
1891 /* Calculate temporary vectorial force */
1892 tx = _mm_mul_ps(fscal,dx21);
1893 ty = _mm_mul_ps(fscal,dy21);
1894 tz = _mm_mul_ps(fscal,dz21);
1895
1896 /* Update vectorial force */
1897 fix2 = _mm_add_ps(fix2,tx);
1898 fiy2 = _mm_add_ps(fiy2,ty);
1899 fiz2 = _mm_add_ps(fiz2,tz);
1900
1901 fjx1 = _mm_add_ps(fjx1,tx);
1902 fjy1 = _mm_add_ps(fjy1,ty);
1903 fjz1 = _mm_add_ps(fjz1,tz);
1904
1905 /**************************
1906 * CALCULATE INTERACTIONS *
1907 **************************/
1908
1909 /* REACTION-FIELD ELECTROSTATICS */
1910 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1911
1912 fscal = felec;
1913
1914 fscal = _mm_andnot_ps(dummy_mask,fscal);
1915
1916 /* Calculate temporary vectorial force */
1917 tx = _mm_mul_ps(fscal,dx22);
1918 ty = _mm_mul_ps(fscal,dy22);
1919 tz = _mm_mul_ps(fscal,dz22);
1920
1921 /* Update vectorial force */
1922 fix2 = _mm_add_ps(fix2,tx);
1923 fiy2 = _mm_add_ps(fiy2,ty);
1924 fiz2 = _mm_add_ps(fiz2,tz);
1925
1926 fjx2 = _mm_add_ps(fjx2,tx);
1927 fjy2 = _mm_add_ps(fjy2,ty);
1928 fjz2 = _mm_add_ps(fjz2,tz);
1929
1930 /**************************
1931 * CALCULATE INTERACTIONS *
1932 **************************/
1933
1934 /* REACTION-FIELD ELECTROSTATICS */
1935 felec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1936
1937 fscal = felec;
1938
1939 fscal = _mm_andnot_ps(dummy_mask,fscal);
1940
1941 /* Calculate temporary vectorial force */
1942 tx = _mm_mul_ps(fscal,dx23);
1943 ty = _mm_mul_ps(fscal,dy23);
1944 tz = _mm_mul_ps(fscal,dz23);
1945
1946 /* Update vectorial force */
1947 fix2 = _mm_add_ps(fix2,tx);
1948 fiy2 = _mm_add_ps(fiy2,ty);
1949 fiz2 = _mm_add_ps(fiz2,tz);
1950
1951 fjx3 = _mm_add_ps(fjx3,tx);
1952 fjy3 = _mm_add_ps(fjy3,ty);
1953 fjz3 = _mm_add_ps(fjz3,tz);
1954
1955 /**************************
1956 * CALCULATE INTERACTIONS *
1957 **************************/
1958
1959 /* REACTION-FIELD ELECTROSTATICS */
1960 felec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1961
1962 fscal = felec;
1963
1964 fscal = _mm_andnot_ps(dummy_mask,fscal);
1965
1966 /* Calculate temporary vectorial force */
1967 tx = _mm_mul_ps(fscal,dx31);
1968 ty = _mm_mul_ps(fscal,dy31);
1969 tz = _mm_mul_ps(fscal,dz31);
1970
1971 /* Update vectorial force */
1972 fix3 = _mm_add_ps(fix3,tx);
1973 fiy3 = _mm_add_ps(fiy3,ty);
1974 fiz3 = _mm_add_ps(fiz3,tz);
1975
1976 fjx1 = _mm_add_ps(fjx1,tx);
1977 fjy1 = _mm_add_ps(fjy1,ty);
1978 fjz1 = _mm_add_ps(fjz1,tz);
1979
1980 /**************************
1981 * CALCULATE INTERACTIONS *
1982 **************************/
1983
1984 /* REACTION-FIELD ELECTROSTATICS */
1985 felec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1986
1987 fscal = felec;
1988
1989 fscal = _mm_andnot_ps(dummy_mask,fscal);
1990
1991 /* Calculate temporary vectorial force */
1992 tx = _mm_mul_ps(fscal,dx32);
1993 ty = _mm_mul_ps(fscal,dy32);
1994 tz = _mm_mul_ps(fscal,dz32);
1995
1996 /* Update vectorial force */
1997 fix3 = _mm_add_ps(fix3,tx);
1998 fiy3 = _mm_add_ps(fiy3,ty);
1999 fiz3 = _mm_add_ps(fiz3,tz);
2000
2001 fjx2 = _mm_add_ps(fjx2,tx);
2002 fjy2 = _mm_add_ps(fjy2,ty);
2003 fjz2 = _mm_add_ps(fjz2,tz);
2004
2005 /**************************
2006 * CALCULATE INTERACTIONS *
2007 **************************/
2008
2009 /* REACTION-FIELD ELECTROSTATICS */
2010 felec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2011
2012 fscal = felec;
2013
2014 fscal = _mm_andnot_ps(dummy_mask,fscal);
2015
2016 /* Calculate temporary vectorial force */
2017 tx = _mm_mul_ps(fscal,dx33);
2018 ty = _mm_mul_ps(fscal,dy33);
2019 tz = _mm_mul_ps(fscal,dz33);
2020
2021 /* Update vectorial force */
2022 fix3 = _mm_add_ps(fix3,tx);
2023 fiy3 = _mm_add_ps(fiy3,ty);
2024 fiz3 = _mm_add_ps(fiz3,tz);
2025
2026 fjx3 = _mm_add_ps(fjx3,tx);
2027 fjy3 = _mm_add_ps(fjy3,ty);
2028 fjz3 = _mm_add_ps(fjz3,tz);
2029
2030 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2031 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2032 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2033 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2034
2035 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2036 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2037 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2038
2039 /* Inner loop uses 295 flops */
2040 }
2041
2042 /* End of innermost loop */
2043
2044 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2045 f+i_coord_offset,fshift+i_shift_offset);
2046
2047 /* Increment number of inner iterations */
2048 inneriter += j_index_end - j_index_start;
2049
2050 /* Outer loop uses 24 flops */
2051 }
2052
2053 /* Increment number of outer iterations */
2054 outeriter += nri;
2055
2056 /* Update outer/inner flops */
2057
2058 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*295);
2059 }
The ultimate molecular dynamics simulation package