?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Removed simple.h from nb_kernel_sse2_XX
[gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sse2_single / nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_sse2_single.c
blob0c74c199c1407a4ffce9e74e8ae54ced75cb4413
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 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/math/vec.h"
46 #include "gromacs/legacyheaders/nrnb.h"
47
48 #include "gromacs/simd/math_x86_sse2_single.h"
49 #include "kernelutil_x86_sse2_single.h"
50
51 /*
52 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_sse2_single
53 * Electrostatics interaction: ReactionField
54 * VdW interaction: CubicSplineTable
55 * Geometry: Water3-Water3
56 * Calculate force/pot: PotentialAndForce
57 */
58 void
59 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_sse2_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 vdwioffset1;
87 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
88 int vdwioffset2;
89 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
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 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
97 __m128 dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
98 __m128 dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
99 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
100 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
101 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
102 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
103 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
104 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
105 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
106 real *charge;
107 int nvdwtype;
108 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
109 int *vdwtype;
110 real *vdwparam;
111 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
112 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
113 __m128i vfitab;
114 __m128i ifour = _mm_set1_epi32(4);
115 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
116 real *vftab;
117 __m128 dummy_mask,cutoff_mask;
118 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
119 __m128 one = _mm_set1_ps(1.0);
120 __m128 two = _mm_set1_ps(2.0);
121 x = xx[0];
122 f = ff[0];
123
124 nri = nlist->nri;
125 iinr = nlist->iinr;
126 jindex = nlist->jindex;
127 jjnr = nlist->jjnr;
128 shiftidx = nlist->shift;
129 gid = nlist->gid;
130 shiftvec = fr->shift_vec[0];
131 fshift = fr->fshift[0];
132 facel = _mm_set1_ps(fr->epsfac);
133 charge = mdatoms->chargeA;
134 krf = _mm_set1_ps(fr->ic->k_rf);
135 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
136 crf = _mm_set1_ps(fr->ic->c_rf);
137 nvdwtype = fr->ntype;
138 vdwparam = fr->nbfp;
139 vdwtype = mdatoms->typeA;
140
141 vftab = kernel_data->table_vdw->data;
142 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
143
144 /* Setup water-specific parameters */
145 inr = nlist->iinr[0];
146 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
147 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
148 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
149 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
150
151 jq0 = _mm_set1_ps(charge[inr+0]);
152 jq1 = _mm_set1_ps(charge[inr+1]);
153 jq2 = _mm_set1_ps(charge[inr+2]);
154 vdwjidx0A = 2*vdwtype[inr+0];
155 qq00 = _mm_mul_ps(iq0,jq0);
156 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
157 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
158 qq01 = _mm_mul_ps(iq0,jq1);
159 qq02 = _mm_mul_ps(iq0,jq2);
160 qq10 = _mm_mul_ps(iq1,jq0);
161 qq11 = _mm_mul_ps(iq1,jq1);
162 qq12 = _mm_mul_ps(iq1,jq2);
163 qq20 = _mm_mul_ps(iq2,jq0);
164 qq21 = _mm_mul_ps(iq2,jq1);
165 qq22 = _mm_mul_ps(iq2,jq2);
166
167 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
168 rcutoff_scalar = fr->rcoulomb;
169 rcutoff = _mm_set1_ps(rcutoff_scalar);
170 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
171
172 /* Avoid stupid compiler warnings */
173 jnrA = jnrB = jnrC = jnrD = 0;
174 j_coord_offsetA = 0;
175 j_coord_offsetB = 0;
176 j_coord_offsetC = 0;
177 j_coord_offsetD = 0;
178
179 outeriter = 0;
180 inneriter = 0;
181
182 for(iidx=0;iidx<4*DIM;iidx++)
183 {
184 scratch[iidx] = 0.0;
185 }
186
187 /* Start outer loop over neighborlists */
188 for(iidx=0; iidx<nri; iidx++)
189 {
190 /* Load shift vector for this list */
191 i_shift_offset = DIM*shiftidx[iidx];
192
193 /* Load limits for loop over neighbors */
194 j_index_start = jindex[iidx];
195 j_index_end = jindex[iidx+1];
196
197 /* Get outer coordinate index */
198 inr = iinr[iidx];
199 i_coord_offset = DIM*inr;
200
201 /* Load i particle coords and add shift vector */
202 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
203 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
204
205 fix0 = _mm_setzero_ps();
206 fiy0 = _mm_setzero_ps();
207 fiz0 = _mm_setzero_ps();
208 fix1 = _mm_setzero_ps();
209 fiy1 = _mm_setzero_ps();
210 fiz1 = _mm_setzero_ps();
211 fix2 = _mm_setzero_ps();
212 fiy2 = _mm_setzero_ps();
213 fiz2 = _mm_setzero_ps();
214
215 /* Reset potential sums */
216 velecsum = _mm_setzero_ps();
217 vvdwsum = _mm_setzero_ps();
218
219 /* Start inner kernel loop */
220 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
221 {
222
223 /* Get j neighbor index, and coordinate index */
224 jnrA = jjnr[jidx];
225 jnrB = jjnr[jidx+1];
226 jnrC = jjnr[jidx+2];
227 jnrD = jjnr[jidx+3];
228 j_coord_offsetA = DIM*jnrA;
229 j_coord_offsetB = DIM*jnrB;
230 j_coord_offsetC = DIM*jnrC;
231 j_coord_offsetD = DIM*jnrD;
232
233 /* load j atom coordinates */
234 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
235 x+j_coord_offsetC,x+j_coord_offsetD,
236 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
237
238 /* Calculate displacement vector */
239 dx00 = _mm_sub_ps(ix0,jx0);
240 dy00 = _mm_sub_ps(iy0,jy0);
241 dz00 = _mm_sub_ps(iz0,jz0);
242 dx01 = _mm_sub_ps(ix0,jx1);
243 dy01 = _mm_sub_ps(iy0,jy1);
244 dz01 = _mm_sub_ps(iz0,jz1);
245 dx02 = _mm_sub_ps(ix0,jx2);
246 dy02 = _mm_sub_ps(iy0,jy2);
247 dz02 = _mm_sub_ps(iz0,jz2);
248 dx10 = _mm_sub_ps(ix1,jx0);
249 dy10 = _mm_sub_ps(iy1,jy0);
250 dz10 = _mm_sub_ps(iz1,jz0);
251 dx11 = _mm_sub_ps(ix1,jx1);
252 dy11 = _mm_sub_ps(iy1,jy1);
253 dz11 = _mm_sub_ps(iz1,jz1);
254 dx12 = _mm_sub_ps(ix1,jx2);
255 dy12 = _mm_sub_ps(iy1,jy2);
256 dz12 = _mm_sub_ps(iz1,jz2);
257 dx20 = _mm_sub_ps(ix2,jx0);
258 dy20 = _mm_sub_ps(iy2,jy0);
259 dz20 = _mm_sub_ps(iz2,jz0);
260 dx21 = _mm_sub_ps(ix2,jx1);
261 dy21 = _mm_sub_ps(iy2,jy1);
262 dz21 = _mm_sub_ps(iz2,jz1);
263 dx22 = _mm_sub_ps(ix2,jx2);
264 dy22 = _mm_sub_ps(iy2,jy2);
265 dz22 = _mm_sub_ps(iz2,jz2);
266
267 /* Calculate squared distance and things based on it */
268 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
269 rsq01 = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
270 rsq02 = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
271 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
272 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
273 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
274 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
275 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
276 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
277
278 rinv00 = gmx_mm_invsqrt_ps(rsq00);
279 rinv01 = gmx_mm_invsqrt_ps(rsq01);
280 rinv02 = gmx_mm_invsqrt_ps(rsq02);
281 rinv10 = gmx_mm_invsqrt_ps(rsq10);
282 rinv11 = gmx_mm_invsqrt_ps(rsq11);
283 rinv12 = gmx_mm_invsqrt_ps(rsq12);
284 rinv20 = gmx_mm_invsqrt_ps(rsq20);
285 rinv21 = gmx_mm_invsqrt_ps(rsq21);
286 rinv22 = gmx_mm_invsqrt_ps(rsq22);
287
288 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
289 rinvsq01 = _mm_mul_ps(rinv01,rinv01);
290 rinvsq02 = _mm_mul_ps(rinv02,rinv02);
291 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
292 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
293 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
294 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
295 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
296 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
297
298 fjx0 = _mm_setzero_ps();
299 fjy0 = _mm_setzero_ps();
300 fjz0 = _mm_setzero_ps();
301 fjx1 = _mm_setzero_ps();
302 fjy1 = _mm_setzero_ps();
303 fjz1 = _mm_setzero_ps();
304 fjx2 = _mm_setzero_ps();
305 fjy2 = _mm_setzero_ps();
306 fjz2 = _mm_setzero_ps();
307
308 /**************************
309 * CALCULATE INTERACTIONS *
310 **************************/
311
312 if (gmx_mm_any_lt(rsq00,rcutoff2))
313 {
314
315 r00 = _mm_mul_ps(rsq00,rinv00);
316
317 /* Calculate table index by multiplying r with table scale and truncate to integer */
318 rt = _mm_mul_ps(r00,vftabscale);
319 vfitab = _mm_cvttps_epi32(rt);
320 vfeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
321 vfitab = _mm_slli_epi32(vfitab,3);
322
323 /* REACTION-FIELD ELECTROSTATICS */
324 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
325 felec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
326
327 /* CUBIC SPLINE TABLE DISPERSION */
328 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
329 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
330 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
331 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
332 _MM_TRANSPOSE4_PS(Y,F,G,H);
333 Heps = _mm_mul_ps(vfeps,H);
334 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
335 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
336 vvdw6 = _mm_mul_ps(c6_00,VV);
337 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
338 fvdw6 = _mm_mul_ps(c6_00,FF);
339
340 /* CUBIC SPLINE TABLE REPULSION */
341 vfitab = _mm_add_epi32(vfitab,ifour);
342 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
343 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
344 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
345 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
346 _MM_TRANSPOSE4_PS(Y,F,G,H);
347 Heps = _mm_mul_ps(vfeps,H);
348 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
349 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
350 vvdw12 = _mm_mul_ps(c12_00,VV);
351 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
352 fvdw12 = _mm_mul_ps(c12_00,FF);
353 vvdw = _mm_add_ps(vvdw12,vvdw6);
354 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
355
356 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
357
358 /* Update potential sum for this i atom from the interaction with this j atom. */
359 velec = _mm_and_ps(velec,cutoff_mask);
360 velecsum = _mm_add_ps(velecsum,velec);
361 vvdw = _mm_and_ps(vvdw,cutoff_mask);
362 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
363
364 fscal = _mm_add_ps(felec,fvdw);
365
366 fscal = _mm_and_ps(fscal,cutoff_mask);
367
368 /* Calculate temporary vectorial force */
369 tx = _mm_mul_ps(fscal,dx00);
370 ty = _mm_mul_ps(fscal,dy00);
371 tz = _mm_mul_ps(fscal,dz00);
372
373 /* Update vectorial force */
374 fix0 = _mm_add_ps(fix0,tx);
375 fiy0 = _mm_add_ps(fiy0,ty);
376 fiz0 = _mm_add_ps(fiz0,tz);
377
378 fjx0 = _mm_add_ps(fjx0,tx);
379 fjy0 = _mm_add_ps(fjy0,ty);
380 fjz0 = _mm_add_ps(fjz0,tz);
381
382 }
383
384 /**************************
385 * CALCULATE INTERACTIONS *
386 **************************/
387
388 if (gmx_mm_any_lt(rsq01,rcutoff2))
389 {
390
391 /* REACTION-FIELD ELECTROSTATICS */
392 velec = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
393 felec = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
394
395 cutoff_mask = _mm_cmplt_ps(rsq01,rcutoff2);
396
397 /* Update potential sum for this i atom from the interaction with this j atom. */
398 velec = _mm_and_ps(velec,cutoff_mask);
399 velecsum = _mm_add_ps(velecsum,velec);
400
401 fscal = felec;
402
403 fscal = _mm_and_ps(fscal,cutoff_mask);
404
405 /* Calculate temporary vectorial force */
406 tx = _mm_mul_ps(fscal,dx01);
407 ty = _mm_mul_ps(fscal,dy01);
408 tz = _mm_mul_ps(fscal,dz01);
409
410 /* Update vectorial force */
411 fix0 = _mm_add_ps(fix0,tx);
412 fiy0 = _mm_add_ps(fiy0,ty);
413 fiz0 = _mm_add_ps(fiz0,tz);
414
415 fjx1 = _mm_add_ps(fjx1,tx);
416 fjy1 = _mm_add_ps(fjy1,ty);
417 fjz1 = _mm_add_ps(fjz1,tz);
418
419 }
420
421 /**************************
422 * CALCULATE INTERACTIONS *
423 **************************/
424
425 if (gmx_mm_any_lt(rsq02,rcutoff2))
426 {
427
428 /* REACTION-FIELD ELECTROSTATICS */
429 velec = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
430 felec = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
431
432 cutoff_mask = _mm_cmplt_ps(rsq02,rcutoff2);
433
434 /* Update potential sum for this i atom from the interaction with this j atom. */
435 velec = _mm_and_ps(velec,cutoff_mask);
436 velecsum = _mm_add_ps(velecsum,velec);
437
438 fscal = felec;
439
440 fscal = _mm_and_ps(fscal,cutoff_mask);
441
442 /* Calculate temporary vectorial force */
443 tx = _mm_mul_ps(fscal,dx02);
444 ty = _mm_mul_ps(fscal,dy02);
445 tz = _mm_mul_ps(fscal,dz02);
446
447 /* Update vectorial force */
448 fix0 = _mm_add_ps(fix0,tx);
449 fiy0 = _mm_add_ps(fiy0,ty);
450 fiz0 = _mm_add_ps(fiz0,tz);
451
452 fjx2 = _mm_add_ps(fjx2,tx);
453 fjy2 = _mm_add_ps(fjy2,ty);
454 fjz2 = _mm_add_ps(fjz2,tz);
455
456 }
457
458 /**************************
459 * CALCULATE INTERACTIONS *
460 **************************/
461
462 if (gmx_mm_any_lt(rsq10,rcutoff2))
463 {
464
465 /* REACTION-FIELD ELECTROSTATICS */
466 velec = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
467 felec = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
468
469 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
470
471 /* Update potential sum for this i atom from the interaction with this j atom. */
472 velec = _mm_and_ps(velec,cutoff_mask);
473 velecsum = _mm_add_ps(velecsum,velec);
474
475 fscal = felec;
476
477 fscal = _mm_and_ps(fscal,cutoff_mask);
478
479 /* Calculate temporary vectorial force */
480 tx = _mm_mul_ps(fscal,dx10);
481 ty = _mm_mul_ps(fscal,dy10);
482 tz = _mm_mul_ps(fscal,dz10);
483
484 /* Update vectorial force */
485 fix1 = _mm_add_ps(fix1,tx);
486 fiy1 = _mm_add_ps(fiy1,ty);
487 fiz1 = _mm_add_ps(fiz1,tz);
488
489 fjx0 = _mm_add_ps(fjx0,tx);
490 fjy0 = _mm_add_ps(fjy0,ty);
491 fjz0 = _mm_add_ps(fjz0,tz);
492
493 }
494
495 /**************************
496 * CALCULATE INTERACTIONS *
497 **************************/
498
499 if (gmx_mm_any_lt(rsq11,rcutoff2))
500 {
501
502 /* REACTION-FIELD ELECTROSTATICS */
503 velec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
504 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
505
506 cutoff_mask = _mm_cmplt_ps(rsq11,rcutoff2);
507
508 /* Update potential sum for this i atom from the interaction with this j atom. */
509 velec = _mm_and_ps(velec,cutoff_mask);
510 velecsum = _mm_add_ps(velecsum,velec);
511
512 fscal = felec;
513
514 fscal = _mm_and_ps(fscal,cutoff_mask);
515
516 /* Calculate temporary vectorial force */
517 tx = _mm_mul_ps(fscal,dx11);
518 ty = _mm_mul_ps(fscal,dy11);
519 tz = _mm_mul_ps(fscal,dz11);
520
521 /* Update vectorial force */
522 fix1 = _mm_add_ps(fix1,tx);
523 fiy1 = _mm_add_ps(fiy1,ty);
524 fiz1 = _mm_add_ps(fiz1,tz);
525
526 fjx1 = _mm_add_ps(fjx1,tx);
527 fjy1 = _mm_add_ps(fjy1,ty);
528 fjz1 = _mm_add_ps(fjz1,tz);
529
530 }
531
532 /**************************
533 * CALCULATE INTERACTIONS *
534 **************************/
535
536 if (gmx_mm_any_lt(rsq12,rcutoff2))
537 {
538
539 /* REACTION-FIELD ELECTROSTATICS */
540 velec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
541 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
542
543 cutoff_mask = _mm_cmplt_ps(rsq12,rcutoff2);
544
545 /* Update potential sum for this i atom from the interaction with this j atom. */
546 velec = _mm_and_ps(velec,cutoff_mask);
547 velecsum = _mm_add_ps(velecsum,velec);
548
549 fscal = felec;
550
551 fscal = _mm_and_ps(fscal,cutoff_mask);
552
553 /* Calculate temporary vectorial force */
554 tx = _mm_mul_ps(fscal,dx12);
555 ty = _mm_mul_ps(fscal,dy12);
556 tz = _mm_mul_ps(fscal,dz12);
557
558 /* Update vectorial force */
559 fix1 = _mm_add_ps(fix1,tx);
560 fiy1 = _mm_add_ps(fiy1,ty);
561 fiz1 = _mm_add_ps(fiz1,tz);
562
563 fjx2 = _mm_add_ps(fjx2,tx);
564 fjy2 = _mm_add_ps(fjy2,ty);
565 fjz2 = _mm_add_ps(fjz2,tz);
566
567 }
568
569 /**************************
570 * CALCULATE INTERACTIONS *
571 **************************/
572
573 if (gmx_mm_any_lt(rsq20,rcutoff2))
574 {
575
576 /* REACTION-FIELD ELECTROSTATICS */
577 velec = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
578 felec = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
579
580 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
581
582 /* Update potential sum for this i atom from the interaction with this j atom. */
583 velec = _mm_and_ps(velec,cutoff_mask);
584 velecsum = _mm_add_ps(velecsum,velec);
585
586 fscal = felec;
587
588 fscal = _mm_and_ps(fscal,cutoff_mask);
589
590 /* Calculate temporary vectorial force */
591 tx = _mm_mul_ps(fscal,dx20);
592 ty = _mm_mul_ps(fscal,dy20);
593 tz = _mm_mul_ps(fscal,dz20);
594
595 /* Update vectorial force */
596 fix2 = _mm_add_ps(fix2,tx);
597 fiy2 = _mm_add_ps(fiy2,ty);
598 fiz2 = _mm_add_ps(fiz2,tz);
599
600 fjx0 = _mm_add_ps(fjx0,tx);
601 fjy0 = _mm_add_ps(fjy0,ty);
602 fjz0 = _mm_add_ps(fjz0,tz);
603
604 }
605
606 /**************************
607 * CALCULATE INTERACTIONS *
608 **************************/
609
610 if (gmx_mm_any_lt(rsq21,rcutoff2))
611 {
612
613 /* REACTION-FIELD ELECTROSTATICS */
614 velec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
615 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
616
617 cutoff_mask = _mm_cmplt_ps(rsq21,rcutoff2);
618
619 /* Update potential sum for this i atom from the interaction with this j atom. */
620 velec = _mm_and_ps(velec,cutoff_mask);
621 velecsum = _mm_add_ps(velecsum,velec);
622
623 fscal = felec;
624
625 fscal = _mm_and_ps(fscal,cutoff_mask);
626
627 /* Calculate temporary vectorial force */
628 tx = _mm_mul_ps(fscal,dx21);
629 ty = _mm_mul_ps(fscal,dy21);
630 tz = _mm_mul_ps(fscal,dz21);
631
632 /* Update vectorial force */
633 fix2 = _mm_add_ps(fix2,tx);
634 fiy2 = _mm_add_ps(fiy2,ty);
635 fiz2 = _mm_add_ps(fiz2,tz);
636
637 fjx1 = _mm_add_ps(fjx1,tx);
638 fjy1 = _mm_add_ps(fjy1,ty);
639 fjz1 = _mm_add_ps(fjz1,tz);
640
641 }
642
643 /**************************
644 * CALCULATE INTERACTIONS *
645 **************************/
646
647 if (gmx_mm_any_lt(rsq22,rcutoff2))
648 {
649
650 /* REACTION-FIELD ELECTROSTATICS */
651 velec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
652 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
653
654 cutoff_mask = _mm_cmplt_ps(rsq22,rcutoff2);
655
656 /* Update potential sum for this i atom from the interaction with this j atom. */
657 velec = _mm_and_ps(velec,cutoff_mask);
658 velecsum = _mm_add_ps(velecsum,velec);
659
660 fscal = felec;
661
662 fscal = _mm_and_ps(fscal,cutoff_mask);
663
664 /* Calculate temporary vectorial force */
665 tx = _mm_mul_ps(fscal,dx22);
666 ty = _mm_mul_ps(fscal,dy22);
667 tz = _mm_mul_ps(fscal,dz22);
668
669 /* Update vectorial force */
670 fix2 = _mm_add_ps(fix2,tx);
671 fiy2 = _mm_add_ps(fiy2,ty);
672 fiz2 = _mm_add_ps(fiz2,tz);
673
674 fjx2 = _mm_add_ps(fjx2,tx);
675 fjy2 = _mm_add_ps(fjy2,ty);
676 fjz2 = _mm_add_ps(fjz2,tz);
677
678 }
679
680 fjptrA = f+j_coord_offsetA;
681 fjptrB = f+j_coord_offsetB;
682 fjptrC = f+j_coord_offsetC;
683 fjptrD = f+j_coord_offsetD;
684
685 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
686 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
687
688 /* Inner loop uses 360 flops */
689 }
690
691 if(jidx<j_index_end)
692 {
693
694 /* Get j neighbor index, and coordinate index */
695 jnrlistA = jjnr[jidx];
696 jnrlistB = jjnr[jidx+1];
697 jnrlistC = jjnr[jidx+2];
698 jnrlistD = jjnr[jidx+3];
699 /* Sign of each element will be negative for non-real atoms.
700 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
701 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
702 */
703 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
704 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
705 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
706 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
707 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
708 j_coord_offsetA = DIM*jnrA;
709 j_coord_offsetB = DIM*jnrB;
710 j_coord_offsetC = DIM*jnrC;
711 j_coord_offsetD = DIM*jnrD;
712
713 /* load j atom coordinates */
714 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
715 x+j_coord_offsetC,x+j_coord_offsetD,
716 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
717
718 /* Calculate displacement vector */
719 dx00 = _mm_sub_ps(ix0,jx0);
720 dy00 = _mm_sub_ps(iy0,jy0);
721 dz00 = _mm_sub_ps(iz0,jz0);
722 dx01 = _mm_sub_ps(ix0,jx1);
723 dy01 = _mm_sub_ps(iy0,jy1);
724 dz01 = _mm_sub_ps(iz0,jz1);
725 dx02 = _mm_sub_ps(ix0,jx2);
726 dy02 = _mm_sub_ps(iy0,jy2);
727 dz02 = _mm_sub_ps(iz0,jz2);
728 dx10 = _mm_sub_ps(ix1,jx0);
729 dy10 = _mm_sub_ps(iy1,jy0);
730 dz10 = _mm_sub_ps(iz1,jz0);
731 dx11 = _mm_sub_ps(ix1,jx1);
732 dy11 = _mm_sub_ps(iy1,jy1);
733 dz11 = _mm_sub_ps(iz1,jz1);
734 dx12 = _mm_sub_ps(ix1,jx2);
735 dy12 = _mm_sub_ps(iy1,jy2);
736 dz12 = _mm_sub_ps(iz1,jz2);
737 dx20 = _mm_sub_ps(ix2,jx0);
738 dy20 = _mm_sub_ps(iy2,jy0);
739 dz20 = _mm_sub_ps(iz2,jz0);
740 dx21 = _mm_sub_ps(ix2,jx1);
741 dy21 = _mm_sub_ps(iy2,jy1);
742 dz21 = _mm_sub_ps(iz2,jz1);
743 dx22 = _mm_sub_ps(ix2,jx2);
744 dy22 = _mm_sub_ps(iy2,jy2);
745 dz22 = _mm_sub_ps(iz2,jz2);
746
747 /* Calculate squared distance and things based on it */
748 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
749 rsq01 = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
750 rsq02 = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
751 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
752 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
753 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
754 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
755 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
756 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
757
758 rinv00 = gmx_mm_invsqrt_ps(rsq00);
759 rinv01 = gmx_mm_invsqrt_ps(rsq01);
760 rinv02 = gmx_mm_invsqrt_ps(rsq02);
761 rinv10 = gmx_mm_invsqrt_ps(rsq10);
762 rinv11 = gmx_mm_invsqrt_ps(rsq11);
763 rinv12 = gmx_mm_invsqrt_ps(rsq12);
764 rinv20 = gmx_mm_invsqrt_ps(rsq20);
765 rinv21 = gmx_mm_invsqrt_ps(rsq21);
766 rinv22 = gmx_mm_invsqrt_ps(rsq22);
767
768 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
769 rinvsq01 = _mm_mul_ps(rinv01,rinv01);
770 rinvsq02 = _mm_mul_ps(rinv02,rinv02);
771 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
772 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
773 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
774 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
775 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
776 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
777
778 fjx0 = _mm_setzero_ps();
779 fjy0 = _mm_setzero_ps();
780 fjz0 = _mm_setzero_ps();
781 fjx1 = _mm_setzero_ps();
782 fjy1 = _mm_setzero_ps();
783 fjz1 = _mm_setzero_ps();
784 fjx2 = _mm_setzero_ps();
785 fjy2 = _mm_setzero_ps();
786 fjz2 = _mm_setzero_ps();
787
788 /**************************
789 * CALCULATE INTERACTIONS *
790 **************************/
791
792 if (gmx_mm_any_lt(rsq00,rcutoff2))
793 {
794
795 r00 = _mm_mul_ps(rsq00,rinv00);
796 r00 = _mm_andnot_ps(dummy_mask,r00);
797
798 /* Calculate table index by multiplying r with table scale and truncate to integer */
799 rt = _mm_mul_ps(r00,vftabscale);
800 vfitab = _mm_cvttps_epi32(rt);
801 vfeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
802 vfitab = _mm_slli_epi32(vfitab,3);
803
804 /* REACTION-FIELD ELECTROSTATICS */
805 velec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
806 felec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
807
808 /* CUBIC SPLINE TABLE DISPERSION */
809 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
810 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
811 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
812 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
813 _MM_TRANSPOSE4_PS(Y,F,G,H);
814 Heps = _mm_mul_ps(vfeps,H);
815 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
816 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
817 vvdw6 = _mm_mul_ps(c6_00,VV);
818 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
819 fvdw6 = _mm_mul_ps(c6_00,FF);
820
821 /* CUBIC SPLINE TABLE REPULSION */
822 vfitab = _mm_add_epi32(vfitab,ifour);
823 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
824 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
825 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
826 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
827 _MM_TRANSPOSE4_PS(Y,F,G,H);
828 Heps = _mm_mul_ps(vfeps,H);
829 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
830 VV = _mm_add_ps(Y,_mm_mul_ps(vfeps,Fp));
831 vvdw12 = _mm_mul_ps(c12_00,VV);
832 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
833 fvdw12 = _mm_mul_ps(c12_00,FF);
834 vvdw = _mm_add_ps(vvdw12,vvdw6);
835 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
836
837 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
838
839 /* Update potential sum for this i atom from the interaction with this j atom. */
840 velec = _mm_and_ps(velec,cutoff_mask);
841 velec = _mm_andnot_ps(dummy_mask,velec);
842 velecsum = _mm_add_ps(velecsum,velec);
843 vvdw = _mm_and_ps(vvdw,cutoff_mask);
844 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
845 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
846
847 fscal = _mm_add_ps(felec,fvdw);
848
849 fscal = _mm_and_ps(fscal,cutoff_mask);
850
851 fscal = _mm_andnot_ps(dummy_mask,fscal);
852
853 /* Calculate temporary vectorial force */
854 tx = _mm_mul_ps(fscal,dx00);
855 ty = _mm_mul_ps(fscal,dy00);
856 tz = _mm_mul_ps(fscal,dz00);
857
858 /* Update vectorial force */
859 fix0 = _mm_add_ps(fix0,tx);
860 fiy0 = _mm_add_ps(fiy0,ty);
861 fiz0 = _mm_add_ps(fiz0,tz);
862
863 fjx0 = _mm_add_ps(fjx0,tx);
864 fjy0 = _mm_add_ps(fjy0,ty);
865 fjz0 = _mm_add_ps(fjz0,tz);
866
867 }
868
869 /**************************
870 * CALCULATE INTERACTIONS *
871 **************************/
872
873 if (gmx_mm_any_lt(rsq01,rcutoff2))
874 {
875
876 /* REACTION-FIELD ELECTROSTATICS */
877 velec = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
878 felec = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
879
880 cutoff_mask = _mm_cmplt_ps(rsq01,rcutoff2);
881
882 /* Update potential sum for this i atom from the interaction with this j atom. */
883 velec = _mm_and_ps(velec,cutoff_mask);
884 velec = _mm_andnot_ps(dummy_mask,velec);
885 velecsum = _mm_add_ps(velecsum,velec);
886
887 fscal = felec;
888
889 fscal = _mm_and_ps(fscal,cutoff_mask);
890
891 fscal = _mm_andnot_ps(dummy_mask,fscal);
892
893 /* Calculate temporary vectorial force */
894 tx = _mm_mul_ps(fscal,dx01);
895 ty = _mm_mul_ps(fscal,dy01);
896 tz = _mm_mul_ps(fscal,dz01);
897
898 /* Update vectorial force */
899 fix0 = _mm_add_ps(fix0,tx);
900 fiy0 = _mm_add_ps(fiy0,ty);
901 fiz0 = _mm_add_ps(fiz0,tz);
902
903 fjx1 = _mm_add_ps(fjx1,tx);
904 fjy1 = _mm_add_ps(fjy1,ty);
905 fjz1 = _mm_add_ps(fjz1,tz);
906
907 }
908
909 /**************************
910 * CALCULATE INTERACTIONS *
911 **************************/
912
913 if (gmx_mm_any_lt(rsq02,rcutoff2))
914 {
915
916 /* REACTION-FIELD ELECTROSTATICS */
917 velec = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
918 felec = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
919
920 cutoff_mask = _mm_cmplt_ps(rsq02,rcutoff2);
921
922 /* Update potential sum for this i atom from the interaction with this j atom. */
923 velec = _mm_and_ps(velec,cutoff_mask);
924 velec = _mm_andnot_ps(dummy_mask,velec);
925 velecsum = _mm_add_ps(velecsum,velec);
926
927 fscal = felec;
928
929 fscal = _mm_and_ps(fscal,cutoff_mask);
930
931 fscal = _mm_andnot_ps(dummy_mask,fscal);
932
933 /* Calculate temporary vectorial force */
934 tx = _mm_mul_ps(fscal,dx02);
935 ty = _mm_mul_ps(fscal,dy02);
936 tz = _mm_mul_ps(fscal,dz02);
937
938 /* Update vectorial force */
939 fix0 = _mm_add_ps(fix0,tx);
940 fiy0 = _mm_add_ps(fiy0,ty);
941 fiz0 = _mm_add_ps(fiz0,tz);
942
943 fjx2 = _mm_add_ps(fjx2,tx);
944 fjy2 = _mm_add_ps(fjy2,ty);
945 fjz2 = _mm_add_ps(fjz2,tz);
946
947 }
948
949 /**************************
950 * CALCULATE INTERACTIONS *
951 **************************/
952
953 if (gmx_mm_any_lt(rsq10,rcutoff2))
954 {
955
956 /* REACTION-FIELD ELECTROSTATICS */
957 velec = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
958 felec = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
959
960 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
961
962 /* Update potential sum for this i atom from the interaction with this j atom. */
963 velec = _mm_and_ps(velec,cutoff_mask);
964 velec = _mm_andnot_ps(dummy_mask,velec);
965 velecsum = _mm_add_ps(velecsum,velec);
966
967 fscal = felec;
968
969 fscal = _mm_and_ps(fscal,cutoff_mask);
970
971 fscal = _mm_andnot_ps(dummy_mask,fscal);
972
973 /* Calculate temporary vectorial force */
974 tx = _mm_mul_ps(fscal,dx10);
975 ty = _mm_mul_ps(fscal,dy10);
976 tz = _mm_mul_ps(fscal,dz10);
977
978 /* Update vectorial force */
979 fix1 = _mm_add_ps(fix1,tx);
980 fiy1 = _mm_add_ps(fiy1,ty);
981 fiz1 = _mm_add_ps(fiz1,tz);
982
983 fjx0 = _mm_add_ps(fjx0,tx);
984 fjy0 = _mm_add_ps(fjy0,ty);
985 fjz0 = _mm_add_ps(fjz0,tz);
986
987 }
988
989 /**************************
990 * CALCULATE INTERACTIONS *
991 **************************/
992
993 if (gmx_mm_any_lt(rsq11,rcutoff2))
994 {
995
996 /* REACTION-FIELD ELECTROSTATICS */
997 velec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
998 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
999
1000 cutoff_mask = _mm_cmplt_ps(rsq11,rcutoff2);
1001
1002 /* Update potential sum for this i atom from the interaction with this j atom. */
1003 velec = _mm_and_ps(velec,cutoff_mask);
1004 velec = _mm_andnot_ps(dummy_mask,velec);
1005 velecsum = _mm_add_ps(velecsum,velec);
1006
1007 fscal = felec;
1008
1009 fscal = _mm_and_ps(fscal,cutoff_mask);
1010
1011 fscal = _mm_andnot_ps(dummy_mask,fscal);
1012
1013 /* Calculate temporary vectorial force */
1014 tx = _mm_mul_ps(fscal,dx11);
1015 ty = _mm_mul_ps(fscal,dy11);
1016 tz = _mm_mul_ps(fscal,dz11);
1017
1018 /* Update vectorial force */
1019 fix1 = _mm_add_ps(fix1,tx);
1020 fiy1 = _mm_add_ps(fiy1,ty);
1021 fiz1 = _mm_add_ps(fiz1,tz);
1022
1023 fjx1 = _mm_add_ps(fjx1,tx);
1024 fjy1 = _mm_add_ps(fjy1,ty);
1025 fjz1 = _mm_add_ps(fjz1,tz);
1026
1027 }
1028
1029 /**************************
1030 * CALCULATE INTERACTIONS *
1031 **************************/
1032
1033 if (gmx_mm_any_lt(rsq12,rcutoff2))
1034 {
1035
1036 /* REACTION-FIELD ELECTROSTATICS */
1037 velec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
1038 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1039
1040 cutoff_mask = _mm_cmplt_ps(rsq12,rcutoff2);
1041
1042 /* Update potential sum for this i atom from the interaction with this j atom. */
1043 velec = _mm_and_ps(velec,cutoff_mask);
1044 velec = _mm_andnot_ps(dummy_mask,velec);
1045 velecsum = _mm_add_ps(velecsum,velec);
1046
1047 fscal = felec;
1048
1049 fscal = _mm_and_ps(fscal,cutoff_mask);
1050
1051 fscal = _mm_andnot_ps(dummy_mask,fscal);
1052
1053 /* Calculate temporary vectorial force */
1054 tx = _mm_mul_ps(fscal,dx12);
1055 ty = _mm_mul_ps(fscal,dy12);
1056 tz = _mm_mul_ps(fscal,dz12);
1057
1058 /* Update vectorial force */
1059 fix1 = _mm_add_ps(fix1,tx);
1060 fiy1 = _mm_add_ps(fiy1,ty);
1061 fiz1 = _mm_add_ps(fiz1,tz);
1062
1063 fjx2 = _mm_add_ps(fjx2,tx);
1064 fjy2 = _mm_add_ps(fjy2,ty);
1065 fjz2 = _mm_add_ps(fjz2,tz);
1066
1067 }
1068
1069 /**************************
1070 * CALCULATE INTERACTIONS *
1071 **************************/
1072
1073 if (gmx_mm_any_lt(rsq20,rcutoff2))
1074 {
1075
1076 /* REACTION-FIELD ELECTROSTATICS */
1077 velec = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
1078 felec = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1079
1080 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
1081
1082 /* Update potential sum for this i atom from the interaction with this j atom. */
1083 velec = _mm_and_ps(velec,cutoff_mask);
1084 velec = _mm_andnot_ps(dummy_mask,velec);
1085 velecsum = _mm_add_ps(velecsum,velec);
1086
1087 fscal = felec;
1088
1089 fscal = _mm_and_ps(fscal,cutoff_mask);
1090
1091 fscal = _mm_andnot_ps(dummy_mask,fscal);
1092
1093 /* Calculate temporary vectorial force */
1094 tx = _mm_mul_ps(fscal,dx20);
1095 ty = _mm_mul_ps(fscal,dy20);
1096 tz = _mm_mul_ps(fscal,dz20);
1097
1098 /* Update vectorial force */
1099 fix2 = _mm_add_ps(fix2,tx);
1100 fiy2 = _mm_add_ps(fiy2,ty);
1101 fiz2 = _mm_add_ps(fiz2,tz);
1102
1103 fjx0 = _mm_add_ps(fjx0,tx);
1104 fjy0 = _mm_add_ps(fjy0,ty);
1105 fjz0 = _mm_add_ps(fjz0,tz);
1106
1107 }
1108
1109 /**************************
1110 * CALCULATE INTERACTIONS *
1111 **************************/
1112
1113 if (gmx_mm_any_lt(rsq21,rcutoff2))
1114 {
1115
1116 /* REACTION-FIELD ELECTROSTATICS */
1117 velec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
1118 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1119
1120 cutoff_mask = _mm_cmplt_ps(rsq21,rcutoff2);
1121
1122 /* Update potential sum for this i atom from the interaction with this j atom. */
1123 velec = _mm_and_ps(velec,cutoff_mask);
1124 velec = _mm_andnot_ps(dummy_mask,velec);
1125 velecsum = _mm_add_ps(velecsum,velec);
1126
1127 fscal = felec;
1128
1129 fscal = _mm_and_ps(fscal,cutoff_mask);
1130
1131 fscal = _mm_andnot_ps(dummy_mask,fscal);
1132
1133 /* Calculate temporary vectorial force */
1134 tx = _mm_mul_ps(fscal,dx21);
1135 ty = _mm_mul_ps(fscal,dy21);
1136 tz = _mm_mul_ps(fscal,dz21);
1137
1138 /* Update vectorial force */
1139 fix2 = _mm_add_ps(fix2,tx);
1140 fiy2 = _mm_add_ps(fiy2,ty);
1141 fiz2 = _mm_add_ps(fiz2,tz);
1142
1143 fjx1 = _mm_add_ps(fjx1,tx);
1144 fjy1 = _mm_add_ps(fjy1,ty);
1145 fjz1 = _mm_add_ps(fjz1,tz);
1146
1147 }
1148
1149 /**************************
1150 * CALCULATE INTERACTIONS *
1151 **************************/
1152
1153 if (gmx_mm_any_lt(rsq22,rcutoff2))
1154 {
1155
1156 /* REACTION-FIELD ELECTROSTATICS */
1157 velec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
1158 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1159
1160 cutoff_mask = _mm_cmplt_ps(rsq22,rcutoff2);
1161
1162 /* Update potential sum for this i atom from the interaction with this j atom. */
1163 velec = _mm_and_ps(velec,cutoff_mask);
1164 velec = _mm_andnot_ps(dummy_mask,velec);
1165 velecsum = _mm_add_ps(velecsum,velec);
1166
1167 fscal = felec;
1168
1169 fscal = _mm_and_ps(fscal,cutoff_mask);
1170
1171 fscal = _mm_andnot_ps(dummy_mask,fscal);
1172
1173 /* Calculate temporary vectorial force */
1174 tx = _mm_mul_ps(fscal,dx22);
1175 ty = _mm_mul_ps(fscal,dy22);
1176 tz = _mm_mul_ps(fscal,dz22);
1177
1178 /* Update vectorial force */
1179 fix2 = _mm_add_ps(fix2,tx);
1180 fiy2 = _mm_add_ps(fiy2,ty);
1181 fiz2 = _mm_add_ps(fiz2,tz);
1182
1183 fjx2 = _mm_add_ps(fjx2,tx);
1184 fjy2 = _mm_add_ps(fjy2,ty);
1185 fjz2 = _mm_add_ps(fjz2,tz);
1186
1187 }
1188
1189 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1190 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1191 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1192 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1193
1194 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1195 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1196
1197 /* Inner loop uses 361 flops */
1198 }
1199
1200 /* End of innermost loop */
1201
1202 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1203 f+i_coord_offset,fshift+i_shift_offset);
1204
1205 ggid = gid[iidx];
1206 /* Update potential energies */
1207 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1208 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1209
1210 /* Increment number of inner iterations */
1211 inneriter += j_index_end - j_index_start;
1212
1213 /* Outer loop uses 20 flops */
1214 }
1215
1216 /* Increment number of outer iterations */
1217 outeriter += nri;
1218
1219 /* Update outer/inner flops */
1220
1221 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*361);
1222 }
1223 /*
1224 * Gromacs nonbonded kernel: nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sse2_single
1225 * Electrostatics interaction: ReactionField
1226 * VdW interaction: CubicSplineTable
1227 * Geometry: Water3-Water3
1228 * Calculate force/pot: Force
1229 */
1230 void
1231 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sse2_single
1232 (t_nblist * gmx_restrict nlist,
1233 rvec * gmx_restrict xx,
1234 rvec * gmx_restrict ff,
1235 t_forcerec * gmx_restrict fr,
1236 t_mdatoms * gmx_restrict mdatoms,
1237 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1238 t_nrnb * gmx_restrict nrnb)
1239 {
1240 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1241 * just 0 for non-waters.
1242 * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1243 * jnr indices corresponding to data put in the four positions in the SIMD register.
1244 */
1245 int i_shift_offset,i_coord_offset,outeriter,inneriter;
1246 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1247 int jnrA,jnrB,jnrC,jnrD;
1248 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1249 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1250 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
1251 real rcutoff_scalar;
1252 real *shiftvec,*fshift,*x,*f;
1253 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1254 real scratch[4*DIM];
1255 __m128 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1256 int vdwioffset0;
1257 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1258 int vdwioffset1;
1259 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1260 int vdwioffset2;
1261 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1262 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1263 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1264 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1265 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1266 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1267 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1268 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1269 __m128 dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1270 __m128 dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1271 __m128 dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1272 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1273 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1274 __m128 dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1275 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1276 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1277 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
1278 real *charge;
1279 int nvdwtype;
1280 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1281 int *vdwtype;
1282 real *vdwparam;
1283 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
1284 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
1285 __m128i vfitab;
1286 __m128i ifour = _mm_set1_epi32(4);
1287 __m128 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1288 real *vftab;
1289 __m128 dummy_mask,cutoff_mask;
1290 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1291 __m128 one = _mm_set1_ps(1.0);
1292 __m128 two = _mm_set1_ps(2.0);
1293 x = xx[0];
1294 f = ff[0];
1295
1296 nri = nlist->nri;
1297 iinr = nlist->iinr;
1298 jindex = nlist->jindex;
1299 jjnr = nlist->jjnr;
1300 shiftidx = nlist->shift;
1301 gid = nlist->gid;
1302 shiftvec = fr->shift_vec[0];
1303 fshift = fr->fshift[0];
1304 facel = _mm_set1_ps(fr->epsfac);
1305 charge = mdatoms->chargeA;
1306 krf = _mm_set1_ps(fr->ic->k_rf);
1307 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
1308 crf = _mm_set1_ps(fr->ic->c_rf);
1309 nvdwtype = fr->ntype;
1310 vdwparam = fr->nbfp;
1311 vdwtype = mdatoms->typeA;
1312
1313 vftab = kernel_data->table_vdw->data;
1314 vftabscale = _mm_set1_ps(kernel_data->table_vdw->scale);
1315
1316 /* Setup water-specific parameters */
1317 inr = nlist->iinr[0];
1318 iq0 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1319 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1320 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1321 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
1322
1323 jq0 = _mm_set1_ps(charge[inr+0]);
1324 jq1 = _mm_set1_ps(charge[inr+1]);
1325 jq2 = _mm_set1_ps(charge[inr+2]);
1326 vdwjidx0A = 2*vdwtype[inr+0];
1327 qq00 = _mm_mul_ps(iq0,jq0);
1328 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1329 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1330 qq01 = _mm_mul_ps(iq0,jq1);
1331 qq02 = _mm_mul_ps(iq0,jq2);
1332 qq10 = _mm_mul_ps(iq1,jq0);
1333 qq11 = _mm_mul_ps(iq1,jq1);
1334 qq12 = _mm_mul_ps(iq1,jq2);
1335 qq20 = _mm_mul_ps(iq2,jq0);
1336 qq21 = _mm_mul_ps(iq2,jq1);
1337 qq22 = _mm_mul_ps(iq2,jq2);
1338
1339 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1340 rcutoff_scalar = fr->rcoulomb;
1341 rcutoff = _mm_set1_ps(rcutoff_scalar);
1342 rcutoff2 = _mm_mul_ps(rcutoff,rcutoff);
1343
1344 /* Avoid stupid compiler warnings */
1345 jnrA = jnrB = jnrC = jnrD = 0;
1346 j_coord_offsetA = 0;
1347 j_coord_offsetB = 0;
1348 j_coord_offsetC = 0;
1349 j_coord_offsetD = 0;
1350
1351 outeriter = 0;
1352 inneriter = 0;
1353
1354 for(iidx=0;iidx<4*DIM;iidx++)
1355 {
1356 scratch[iidx] = 0.0;
1357 }
1358
1359 /* Start outer loop over neighborlists */
1360 for(iidx=0; iidx<nri; iidx++)
1361 {
1362 /* Load shift vector for this list */
1363 i_shift_offset = DIM*shiftidx[iidx];
1364
1365 /* Load limits for loop over neighbors */
1366 j_index_start = jindex[iidx];
1367 j_index_end = jindex[iidx+1];
1368
1369 /* Get outer coordinate index */
1370 inr = iinr[iidx];
1371 i_coord_offset = DIM*inr;
1372
1373 /* Load i particle coords and add shift vector */
1374 gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1375 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1376
1377 fix0 = _mm_setzero_ps();
1378 fiy0 = _mm_setzero_ps();
1379 fiz0 = _mm_setzero_ps();
1380 fix1 = _mm_setzero_ps();
1381 fiy1 = _mm_setzero_ps();
1382 fiz1 = _mm_setzero_ps();
1383 fix2 = _mm_setzero_ps();
1384 fiy2 = _mm_setzero_ps();
1385 fiz2 = _mm_setzero_ps();
1386
1387 /* Start inner kernel loop */
1388 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1389 {
1390
1391 /* Get j neighbor index, and coordinate index */
1392 jnrA = jjnr[jidx];
1393 jnrB = jjnr[jidx+1];
1394 jnrC = jjnr[jidx+2];
1395 jnrD = jjnr[jidx+3];
1396 j_coord_offsetA = DIM*jnrA;
1397 j_coord_offsetB = DIM*jnrB;
1398 j_coord_offsetC = DIM*jnrC;
1399 j_coord_offsetD = DIM*jnrD;
1400
1401 /* load j atom coordinates */
1402 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1403 x+j_coord_offsetC,x+j_coord_offsetD,
1404 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1405
1406 /* Calculate displacement vector */
1407 dx00 = _mm_sub_ps(ix0,jx0);
1408 dy00 = _mm_sub_ps(iy0,jy0);
1409 dz00 = _mm_sub_ps(iz0,jz0);
1410 dx01 = _mm_sub_ps(ix0,jx1);
1411 dy01 = _mm_sub_ps(iy0,jy1);
1412 dz01 = _mm_sub_ps(iz0,jz1);
1413 dx02 = _mm_sub_ps(ix0,jx2);
1414 dy02 = _mm_sub_ps(iy0,jy2);
1415 dz02 = _mm_sub_ps(iz0,jz2);
1416 dx10 = _mm_sub_ps(ix1,jx0);
1417 dy10 = _mm_sub_ps(iy1,jy0);
1418 dz10 = _mm_sub_ps(iz1,jz0);
1419 dx11 = _mm_sub_ps(ix1,jx1);
1420 dy11 = _mm_sub_ps(iy1,jy1);
1421 dz11 = _mm_sub_ps(iz1,jz1);
1422 dx12 = _mm_sub_ps(ix1,jx2);
1423 dy12 = _mm_sub_ps(iy1,jy2);
1424 dz12 = _mm_sub_ps(iz1,jz2);
1425 dx20 = _mm_sub_ps(ix2,jx0);
1426 dy20 = _mm_sub_ps(iy2,jy0);
1427 dz20 = _mm_sub_ps(iz2,jz0);
1428 dx21 = _mm_sub_ps(ix2,jx1);
1429 dy21 = _mm_sub_ps(iy2,jy1);
1430 dz21 = _mm_sub_ps(iz2,jz1);
1431 dx22 = _mm_sub_ps(ix2,jx2);
1432 dy22 = _mm_sub_ps(iy2,jy2);
1433 dz22 = _mm_sub_ps(iz2,jz2);
1434
1435 /* Calculate squared distance and things based on it */
1436 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1437 rsq01 = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1438 rsq02 = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1439 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1440 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1441 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1442 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1443 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1444 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1445
1446 rinv00 = gmx_mm_invsqrt_ps(rsq00);
1447 rinv01 = gmx_mm_invsqrt_ps(rsq01);
1448 rinv02 = gmx_mm_invsqrt_ps(rsq02);
1449 rinv10 = gmx_mm_invsqrt_ps(rsq10);
1450 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1451 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1452 rinv20 = gmx_mm_invsqrt_ps(rsq20);
1453 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1454 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1455
1456 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
1457 rinvsq01 = _mm_mul_ps(rinv01,rinv01);
1458 rinvsq02 = _mm_mul_ps(rinv02,rinv02);
1459 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
1460 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1461 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1462 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
1463 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1464 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1465
1466 fjx0 = _mm_setzero_ps();
1467 fjy0 = _mm_setzero_ps();
1468 fjz0 = _mm_setzero_ps();
1469 fjx1 = _mm_setzero_ps();
1470 fjy1 = _mm_setzero_ps();
1471 fjz1 = _mm_setzero_ps();
1472 fjx2 = _mm_setzero_ps();
1473 fjy2 = _mm_setzero_ps();
1474 fjz2 = _mm_setzero_ps();
1475
1476 /**************************
1477 * CALCULATE INTERACTIONS *
1478 **************************/
1479
1480 if (gmx_mm_any_lt(rsq00,rcutoff2))
1481 {
1482
1483 r00 = _mm_mul_ps(rsq00,rinv00);
1484
1485 /* Calculate table index by multiplying r with table scale and truncate to integer */
1486 rt = _mm_mul_ps(r00,vftabscale);
1487 vfitab = _mm_cvttps_epi32(rt);
1488 vfeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1489 vfitab = _mm_slli_epi32(vfitab,3);
1490
1491 /* REACTION-FIELD ELECTROSTATICS */
1492 felec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1493
1494 /* CUBIC SPLINE TABLE DISPERSION */
1495 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1496 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1497 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1498 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1499 _MM_TRANSPOSE4_PS(Y,F,G,H);
1500 Heps = _mm_mul_ps(vfeps,H);
1501 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1502 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1503 fvdw6 = _mm_mul_ps(c6_00,FF);
1504
1505 /* CUBIC SPLINE TABLE REPULSION */
1506 vfitab = _mm_add_epi32(vfitab,ifour);
1507 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1508 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1509 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1510 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1511 _MM_TRANSPOSE4_PS(Y,F,G,H);
1512 Heps = _mm_mul_ps(vfeps,H);
1513 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1514 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1515 fvdw12 = _mm_mul_ps(c12_00,FF);
1516 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1517
1518 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
1519
1520 fscal = _mm_add_ps(felec,fvdw);
1521
1522 fscal = _mm_and_ps(fscal,cutoff_mask);
1523
1524 /* Calculate temporary vectorial force */
1525 tx = _mm_mul_ps(fscal,dx00);
1526 ty = _mm_mul_ps(fscal,dy00);
1527 tz = _mm_mul_ps(fscal,dz00);
1528
1529 /* Update vectorial force */
1530 fix0 = _mm_add_ps(fix0,tx);
1531 fiy0 = _mm_add_ps(fiy0,ty);
1532 fiz0 = _mm_add_ps(fiz0,tz);
1533
1534 fjx0 = _mm_add_ps(fjx0,tx);
1535 fjy0 = _mm_add_ps(fjy0,ty);
1536 fjz0 = _mm_add_ps(fjz0,tz);
1537
1538 }
1539
1540 /**************************
1541 * CALCULATE INTERACTIONS *
1542 **************************/
1543
1544 if (gmx_mm_any_lt(rsq01,rcutoff2))
1545 {
1546
1547 /* REACTION-FIELD ELECTROSTATICS */
1548 felec = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1549
1550 cutoff_mask = _mm_cmplt_ps(rsq01,rcutoff2);
1551
1552 fscal = felec;
1553
1554 fscal = _mm_and_ps(fscal,cutoff_mask);
1555
1556 /* Calculate temporary vectorial force */
1557 tx = _mm_mul_ps(fscal,dx01);
1558 ty = _mm_mul_ps(fscal,dy01);
1559 tz = _mm_mul_ps(fscal,dz01);
1560
1561 /* Update vectorial force */
1562 fix0 = _mm_add_ps(fix0,tx);
1563 fiy0 = _mm_add_ps(fiy0,ty);
1564 fiz0 = _mm_add_ps(fiz0,tz);
1565
1566 fjx1 = _mm_add_ps(fjx1,tx);
1567 fjy1 = _mm_add_ps(fjy1,ty);
1568 fjz1 = _mm_add_ps(fjz1,tz);
1569
1570 }
1571
1572 /**************************
1573 * CALCULATE INTERACTIONS *
1574 **************************/
1575
1576 if (gmx_mm_any_lt(rsq02,rcutoff2))
1577 {
1578
1579 /* REACTION-FIELD ELECTROSTATICS */
1580 felec = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1581
1582 cutoff_mask = _mm_cmplt_ps(rsq02,rcutoff2);
1583
1584 fscal = felec;
1585
1586 fscal = _mm_and_ps(fscal,cutoff_mask);
1587
1588 /* Calculate temporary vectorial force */
1589 tx = _mm_mul_ps(fscal,dx02);
1590 ty = _mm_mul_ps(fscal,dy02);
1591 tz = _mm_mul_ps(fscal,dz02);
1592
1593 /* Update vectorial force */
1594 fix0 = _mm_add_ps(fix0,tx);
1595 fiy0 = _mm_add_ps(fiy0,ty);
1596 fiz0 = _mm_add_ps(fiz0,tz);
1597
1598 fjx2 = _mm_add_ps(fjx2,tx);
1599 fjy2 = _mm_add_ps(fjy2,ty);
1600 fjz2 = _mm_add_ps(fjz2,tz);
1601
1602 }
1603
1604 /**************************
1605 * CALCULATE INTERACTIONS *
1606 **************************/
1607
1608 if (gmx_mm_any_lt(rsq10,rcutoff2))
1609 {
1610
1611 /* REACTION-FIELD ELECTROSTATICS */
1612 felec = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1613
1614 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
1615
1616 fscal = felec;
1617
1618 fscal = _mm_and_ps(fscal,cutoff_mask);
1619
1620 /* Calculate temporary vectorial force */
1621 tx = _mm_mul_ps(fscal,dx10);
1622 ty = _mm_mul_ps(fscal,dy10);
1623 tz = _mm_mul_ps(fscal,dz10);
1624
1625 /* Update vectorial force */
1626 fix1 = _mm_add_ps(fix1,tx);
1627 fiy1 = _mm_add_ps(fiy1,ty);
1628 fiz1 = _mm_add_ps(fiz1,tz);
1629
1630 fjx0 = _mm_add_ps(fjx0,tx);
1631 fjy0 = _mm_add_ps(fjy0,ty);
1632 fjz0 = _mm_add_ps(fjz0,tz);
1633
1634 }
1635
1636 /**************************
1637 * CALCULATE INTERACTIONS *
1638 **************************/
1639
1640 if (gmx_mm_any_lt(rsq11,rcutoff2))
1641 {
1642
1643 /* REACTION-FIELD ELECTROSTATICS */
1644 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1645
1646 cutoff_mask = _mm_cmplt_ps(rsq11,rcutoff2);
1647
1648 fscal = felec;
1649
1650 fscal = _mm_and_ps(fscal,cutoff_mask);
1651
1652 /* Calculate temporary vectorial force */
1653 tx = _mm_mul_ps(fscal,dx11);
1654 ty = _mm_mul_ps(fscal,dy11);
1655 tz = _mm_mul_ps(fscal,dz11);
1656
1657 /* Update vectorial force */
1658 fix1 = _mm_add_ps(fix1,tx);
1659 fiy1 = _mm_add_ps(fiy1,ty);
1660 fiz1 = _mm_add_ps(fiz1,tz);
1661
1662 fjx1 = _mm_add_ps(fjx1,tx);
1663 fjy1 = _mm_add_ps(fjy1,ty);
1664 fjz1 = _mm_add_ps(fjz1,tz);
1665
1666 }
1667
1668 /**************************
1669 * CALCULATE INTERACTIONS *
1670 **************************/
1671
1672 if (gmx_mm_any_lt(rsq12,rcutoff2))
1673 {
1674
1675 /* REACTION-FIELD ELECTROSTATICS */
1676 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1677
1678 cutoff_mask = _mm_cmplt_ps(rsq12,rcutoff2);
1679
1680 fscal = felec;
1681
1682 fscal = _mm_and_ps(fscal,cutoff_mask);
1683
1684 /* Calculate temporary vectorial force */
1685 tx = _mm_mul_ps(fscal,dx12);
1686 ty = _mm_mul_ps(fscal,dy12);
1687 tz = _mm_mul_ps(fscal,dz12);
1688
1689 /* Update vectorial force */
1690 fix1 = _mm_add_ps(fix1,tx);
1691 fiy1 = _mm_add_ps(fiy1,ty);
1692 fiz1 = _mm_add_ps(fiz1,tz);
1693
1694 fjx2 = _mm_add_ps(fjx2,tx);
1695 fjy2 = _mm_add_ps(fjy2,ty);
1696 fjz2 = _mm_add_ps(fjz2,tz);
1697
1698 }
1699
1700 /**************************
1701 * CALCULATE INTERACTIONS *
1702 **************************/
1703
1704 if (gmx_mm_any_lt(rsq20,rcutoff2))
1705 {
1706
1707 /* REACTION-FIELD ELECTROSTATICS */
1708 felec = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1709
1710 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
1711
1712 fscal = felec;
1713
1714 fscal = _mm_and_ps(fscal,cutoff_mask);
1715
1716 /* Calculate temporary vectorial force */
1717 tx = _mm_mul_ps(fscal,dx20);
1718 ty = _mm_mul_ps(fscal,dy20);
1719 tz = _mm_mul_ps(fscal,dz20);
1720
1721 /* Update vectorial force */
1722 fix2 = _mm_add_ps(fix2,tx);
1723 fiy2 = _mm_add_ps(fiy2,ty);
1724 fiz2 = _mm_add_ps(fiz2,tz);
1725
1726 fjx0 = _mm_add_ps(fjx0,tx);
1727 fjy0 = _mm_add_ps(fjy0,ty);
1728 fjz0 = _mm_add_ps(fjz0,tz);
1729
1730 }
1731
1732 /**************************
1733 * CALCULATE INTERACTIONS *
1734 **************************/
1735
1736 if (gmx_mm_any_lt(rsq21,rcutoff2))
1737 {
1738
1739 /* REACTION-FIELD ELECTROSTATICS */
1740 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1741
1742 cutoff_mask = _mm_cmplt_ps(rsq21,rcutoff2);
1743
1744 fscal = felec;
1745
1746 fscal = _mm_and_ps(fscal,cutoff_mask);
1747
1748 /* Calculate temporary vectorial force */
1749 tx = _mm_mul_ps(fscal,dx21);
1750 ty = _mm_mul_ps(fscal,dy21);
1751 tz = _mm_mul_ps(fscal,dz21);
1752
1753 /* Update vectorial force */
1754 fix2 = _mm_add_ps(fix2,tx);
1755 fiy2 = _mm_add_ps(fiy2,ty);
1756 fiz2 = _mm_add_ps(fiz2,tz);
1757
1758 fjx1 = _mm_add_ps(fjx1,tx);
1759 fjy1 = _mm_add_ps(fjy1,ty);
1760 fjz1 = _mm_add_ps(fjz1,tz);
1761
1762 }
1763
1764 /**************************
1765 * CALCULATE INTERACTIONS *
1766 **************************/
1767
1768 if (gmx_mm_any_lt(rsq22,rcutoff2))
1769 {
1770
1771 /* REACTION-FIELD ELECTROSTATICS */
1772 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1773
1774 cutoff_mask = _mm_cmplt_ps(rsq22,rcutoff2);
1775
1776 fscal = felec;
1777
1778 fscal = _mm_and_ps(fscal,cutoff_mask);
1779
1780 /* Calculate temporary vectorial force */
1781 tx = _mm_mul_ps(fscal,dx22);
1782 ty = _mm_mul_ps(fscal,dy22);
1783 tz = _mm_mul_ps(fscal,dz22);
1784
1785 /* Update vectorial force */
1786 fix2 = _mm_add_ps(fix2,tx);
1787 fiy2 = _mm_add_ps(fiy2,ty);
1788 fiz2 = _mm_add_ps(fiz2,tz);
1789
1790 fjx2 = _mm_add_ps(fjx2,tx);
1791 fjy2 = _mm_add_ps(fjy2,ty);
1792 fjz2 = _mm_add_ps(fjz2,tz);
1793
1794 }
1795
1796 fjptrA = f+j_coord_offsetA;
1797 fjptrB = f+j_coord_offsetB;
1798 fjptrC = f+j_coord_offsetC;
1799 fjptrD = f+j_coord_offsetD;
1800
1801 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1802 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1803
1804 /* Inner loop uses 297 flops */
1805 }
1806
1807 if(jidx<j_index_end)
1808 {
1809
1810 /* Get j neighbor index, and coordinate index */
1811 jnrlistA = jjnr[jidx];
1812 jnrlistB = jjnr[jidx+1];
1813 jnrlistC = jjnr[jidx+2];
1814 jnrlistD = jjnr[jidx+3];
1815 /* Sign of each element will be negative for non-real atoms.
1816 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1817 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1818 */
1819 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1820 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1821 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1822 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1823 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1824 j_coord_offsetA = DIM*jnrA;
1825 j_coord_offsetB = DIM*jnrB;
1826 j_coord_offsetC = DIM*jnrC;
1827 j_coord_offsetD = DIM*jnrD;
1828
1829 /* load j atom coordinates */
1830 gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1831 x+j_coord_offsetC,x+j_coord_offsetD,
1832 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1833
1834 /* Calculate displacement vector */
1835 dx00 = _mm_sub_ps(ix0,jx0);
1836 dy00 = _mm_sub_ps(iy0,jy0);
1837 dz00 = _mm_sub_ps(iz0,jz0);
1838 dx01 = _mm_sub_ps(ix0,jx1);
1839 dy01 = _mm_sub_ps(iy0,jy1);
1840 dz01 = _mm_sub_ps(iz0,jz1);
1841 dx02 = _mm_sub_ps(ix0,jx2);
1842 dy02 = _mm_sub_ps(iy0,jy2);
1843 dz02 = _mm_sub_ps(iz0,jz2);
1844 dx10 = _mm_sub_ps(ix1,jx0);
1845 dy10 = _mm_sub_ps(iy1,jy0);
1846 dz10 = _mm_sub_ps(iz1,jz0);
1847 dx11 = _mm_sub_ps(ix1,jx1);
1848 dy11 = _mm_sub_ps(iy1,jy1);
1849 dz11 = _mm_sub_ps(iz1,jz1);
1850 dx12 = _mm_sub_ps(ix1,jx2);
1851 dy12 = _mm_sub_ps(iy1,jy2);
1852 dz12 = _mm_sub_ps(iz1,jz2);
1853 dx20 = _mm_sub_ps(ix2,jx0);
1854 dy20 = _mm_sub_ps(iy2,jy0);
1855 dz20 = _mm_sub_ps(iz2,jz0);
1856 dx21 = _mm_sub_ps(ix2,jx1);
1857 dy21 = _mm_sub_ps(iy2,jy1);
1858 dz21 = _mm_sub_ps(iz2,jz1);
1859 dx22 = _mm_sub_ps(ix2,jx2);
1860 dy22 = _mm_sub_ps(iy2,jy2);
1861 dz22 = _mm_sub_ps(iz2,jz2);
1862
1863 /* Calculate squared distance and things based on it */
1864 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1865 rsq01 = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1866 rsq02 = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1867 rsq10 = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1868 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1869 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1870 rsq20 = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1871 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1872 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1873
1874 rinv00 = gmx_mm_invsqrt_ps(rsq00);
1875 rinv01 = gmx_mm_invsqrt_ps(rsq01);
1876 rinv02 = gmx_mm_invsqrt_ps(rsq02);
1877 rinv10 = gmx_mm_invsqrt_ps(rsq10);
1878 rinv11 = gmx_mm_invsqrt_ps(rsq11);
1879 rinv12 = gmx_mm_invsqrt_ps(rsq12);
1880 rinv20 = gmx_mm_invsqrt_ps(rsq20);
1881 rinv21 = gmx_mm_invsqrt_ps(rsq21);
1882 rinv22 = gmx_mm_invsqrt_ps(rsq22);
1883
1884 rinvsq00 = _mm_mul_ps(rinv00,rinv00);
1885 rinvsq01 = _mm_mul_ps(rinv01,rinv01);
1886 rinvsq02 = _mm_mul_ps(rinv02,rinv02);
1887 rinvsq10 = _mm_mul_ps(rinv10,rinv10);
1888 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1889 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1890 rinvsq20 = _mm_mul_ps(rinv20,rinv20);
1891 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1892 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1893
1894 fjx0 = _mm_setzero_ps();
1895 fjy0 = _mm_setzero_ps();
1896 fjz0 = _mm_setzero_ps();
1897 fjx1 = _mm_setzero_ps();
1898 fjy1 = _mm_setzero_ps();
1899 fjz1 = _mm_setzero_ps();
1900 fjx2 = _mm_setzero_ps();
1901 fjy2 = _mm_setzero_ps();
1902 fjz2 = _mm_setzero_ps();
1903
1904 /**************************
1905 * CALCULATE INTERACTIONS *
1906 **************************/
1907
1908 if (gmx_mm_any_lt(rsq00,rcutoff2))
1909 {
1910
1911 r00 = _mm_mul_ps(rsq00,rinv00);
1912 r00 = _mm_andnot_ps(dummy_mask,r00);
1913
1914 /* Calculate table index by multiplying r with table scale and truncate to integer */
1915 rt = _mm_mul_ps(r00,vftabscale);
1916 vfitab = _mm_cvttps_epi32(rt);
1917 vfeps = _mm_sub_ps(rt,_mm_cvtepi32_ps(vfitab));
1918 vfitab = _mm_slli_epi32(vfitab,3);
1919
1920 /* REACTION-FIELD ELECTROSTATICS */
1921 felec = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1922
1923 /* CUBIC SPLINE TABLE DISPERSION */
1924 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1925 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1926 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1927 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1928 _MM_TRANSPOSE4_PS(Y,F,G,H);
1929 Heps = _mm_mul_ps(vfeps,H);
1930 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1931 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1932 fvdw6 = _mm_mul_ps(c6_00,FF);
1933
1934 /* CUBIC SPLINE TABLE REPULSION */
1935 vfitab = _mm_add_epi32(vfitab,ifour);
1936 Y = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,0) );
1937 F = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,1) );
1938 G = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,2) );
1939 H = _mm_load_ps( vftab + gmx_mm_extract_epi32(vfitab,3) );
1940 _MM_TRANSPOSE4_PS(Y,F,G,H);
1941 Heps = _mm_mul_ps(vfeps,H);
1942 Fp = _mm_add_ps(F,_mm_mul_ps(vfeps,_mm_add_ps(G,Heps)));
1943 FF = _mm_add_ps(Fp,_mm_mul_ps(vfeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
1944 fvdw12 = _mm_mul_ps(c12_00,FF);
1945 fvdw = _mm_xor_ps(signbit,_mm_mul_ps(_mm_add_ps(fvdw6,fvdw12),_mm_mul_ps(vftabscale,rinv00)));
1946
1947 cutoff_mask = _mm_cmplt_ps(rsq00,rcutoff2);
1948
1949 fscal = _mm_add_ps(felec,fvdw);
1950
1951 fscal = _mm_and_ps(fscal,cutoff_mask);
1952
1953 fscal = _mm_andnot_ps(dummy_mask,fscal);
1954
1955 /* Calculate temporary vectorial force */
1956 tx = _mm_mul_ps(fscal,dx00);
1957 ty = _mm_mul_ps(fscal,dy00);
1958 tz = _mm_mul_ps(fscal,dz00);
1959
1960 /* Update vectorial force */
1961 fix0 = _mm_add_ps(fix0,tx);
1962 fiy0 = _mm_add_ps(fiy0,ty);
1963 fiz0 = _mm_add_ps(fiz0,tz);
1964
1965 fjx0 = _mm_add_ps(fjx0,tx);
1966 fjy0 = _mm_add_ps(fjy0,ty);
1967 fjz0 = _mm_add_ps(fjz0,tz);
1968
1969 }
1970
1971 /**************************
1972 * CALCULATE INTERACTIONS *
1973 **************************/
1974
1975 if (gmx_mm_any_lt(rsq01,rcutoff2))
1976 {
1977
1978 /* REACTION-FIELD ELECTROSTATICS */
1979 felec = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1980
1981 cutoff_mask = _mm_cmplt_ps(rsq01,rcutoff2);
1982
1983 fscal = felec;
1984
1985 fscal = _mm_and_ps(fscal,cutoff_mask);
1986
1987 fscal = _mm_andnot_ps(dummy_mask,fscal);
1988
1989 /* Calculate temporary vectorial force */
1990 tx = _mm_mul_ps(fscal,dx01);
1991 ty = _mm_mul_ps(fscal,dy01);
1992 tz = _mm_mul_ps(fscal,dz01);
1993
1994 /* Update vectorial force */
1995 fix0 = _mm_add_ps(fix0,tx);
1996 fiy0 = _mm_add_ps(fiy0,ty);
1997 fiz0 = _mm_add_ps(fiz0,tz);
1998
1999 fjx1 = _mm_add_ps(fjx1,tx);
2000 fjy1 = _mm_add_ps(fjy1,ty);
2001 fjz1 = _mm_add_ps(fjz1,tz);
2002
2003 }
2004
2005 /**************************
2006 * CALCULATE INTERACTIONS *
2007 **************************/
2008
2009 if (gmx_mm_any_lt(rsq02,rcutoff2))
2010 {
2011
2012 /* REACTION-FIELD ELECTROSTATICS */
2013 felec = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
2014
2015 cutoff_mask = _mm_cmplt_ps(rsq02,rcutoff2);
2016
2017 fscal = felec;
2018
2019 fscal = _mm_and_ps(fscal,cutoff_mask);
2020
2021 fscal = _mm_andnot_ps(dummy_mask,fscal);
2022
2023 /* Calculate temporary vectorial force */
2024 tx = _mm_mul_ps(fscal,dx02);
2025 ty = _mm_mul_ps(fscal,dy02);
2026 tz = _mm_mul_ps(fscal,dz02);
2027
2028 /* Update vectorial force */
2029 fix0 = _mm_add_ps(fix0,tx);
2030 fiy0 = _mm_add_ps(fiy0,ty);
2031 fiz0 = _mm_add_ps(fiz0,tz);
2032
2033 fjx2 = _mm_add_ps(fjx2,tx);
2034 fjy2 = _mm_add_ps(fjy2,ty);
2035 fjz2 = _mm_add_ps(fjz2,tz);
2036
2037 }
2038
2039 /**************************
2040 * CALCULATE INTERACTIONS *
2041 **************************/
2042
2043 if (gmx_mm_any_lt(rsq10,rcutoff2))
2044 {
2045
2046 /* REACTION-FIELD ELECTROSTATICS */
2047 felec = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
2048
2049 cutoff_mask = _mm_cmplt_ps(rsq10,rcutoff2);
2050
2051 fscal = felec;
2052
2053 fscal = _mm_and_ps(fscal,cutoff_mask);
2054
2055 fscal = _mm_andnot_ps(dummy_mask,fscal);
2056
2057 /* Calculate temporary vectorial force */
2058 tx = _mm_mul_ps(fscal,dx10);
2059 ty = _mm_mul_ps(fscal,dy10);
2060 tz = _mm_mul_ps(fscal,dz10);
2061
2062 /* Update vectorial force */
2063 fix1 = _mm_add_ps(fix1,tx);
2064 fiy1 = _mm_add_ps(fiy1,ty);
2065 fiz1 = _mm_add_ps(fiz1,tz);
2066
2067 fjx0 = _mm_add_ps(fjx0,tx);
2068 fjy0 = _mm_add_ps(fjy0,ty);
2069 fjz0 = _mm_add_ps(fjz0,tz);
2070
2071 }
2072
2073 /**************************
2074 * CALCULATE INTERACTIONS *
2075 **************************/
2076
2077 if (gmx_mm_any_lt(rsq11,rcutoff2))
2078 {
2079
2080 /* REACTION-FIELD ELECTROSTATICS */
2081 felec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
2082
2083 cutoff_mask = _mm_cmplt_ps(rsq11,rcutoff2);
2084
2085 fscal = felec;
2086
2087 fscal = _mm_and_ps(fscal,cutoff_mask);
2088
2089 fscal = _mm_andnot_ps(dummy_mask,fscal);
2090
2091 /* Calculate temporary vectorial force */
2092 tx = _mm_mul_ps(fscal,dx11);
2093 ty = _mm_mul_ps(fscal,dy11);
2094 tz = _mm_mul_ps(fscal,dz11);
2095
2096 /* Update vectorial force */
2097 fix1 = _mm_add_ps(fix1,tx);
2098 fiy1 = _mm_add_ps(fiy1,ty);
2099 fiz1 = _mm_add_ps(fiz1,tz);
2100
2101 fjx1 = _mm_add_ps(fjx1,tx);
2102 fjy1 = _mm_add_ps(fjy1,ty);
2103 fjz1 = _mm_add_ps(fjz1,tz);
2104
2105 }
2106
2107 /**************************
2108 * CALCULATE INTERACTIONS *
2109 **************************/
2110
2111 if (gmx_mm_any_lt(rsq12,rcutoff2))
2112 {
2113
2114 /* REACTION-FIELD ELECTROSTATICS */
2115 felec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
2116
2117 cutoff_mask = _mm_cmplt_ps(rsq12,rcutoff2);
2118
2119 fscal = felec;
2120
2121 fscal = _mm_and_ps(fscal,cutoff_mask);
2122
2123 fscal = _mm_andnot_ps(dummy_mask,fscal);
2124
2125 /* Calculate temporary vectorial force */
2126 tx = _mm_mul_ps(fscal,dx12);
2127 ty = _mm_mul_ps(fscal,dy12);
2128 tz = _mm_mul_ps(fscal,dz12);
2129
2130 /* Update vectorial force */
2131 fix1 = _mm_add_ps(fix1,tx);
2132 fiy1 = _mm_add_ps(fiy1,ty);
2133 fiz1 = _mm_add_ps(fiz1,tz);
2134
2135 fjx2 = _mm_add_ps(fjx2,tx);
2136 fjy2 = _mm_add_ps(fjy2,ty);
2137 fjz2 = _mm_add_ps(fjz2,tz);
2138
2139 }
2140
2141 /**************************
2142 * CALCULATE INTERACTIONS *
2143 **************************/
2144
2145 if (gmx_mm_any_lt(rsq20,rcutoff2))
2146 {
2147
2148 /* REACTION-FIELD ELECTROSTATICS */
2149 felec = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
2150
2151 cutoff_mask = _mm_cmplt_ps(rsq20,rcutoff2);
2152
2153 fscal = felec;
2154
2155 fscal = _mm_and_ps(fscal,cutoff_mask);
2156
2157 fscal = _mm_andnot_ps(dummy_mask,fscal);
2158
2159 /* Calculate temporary vectorial force */
2160 tx = _mm_mul_ps(fscal,dx20);
2161 ty = _mm_mul_ps(fscal,dy20);
2162 tz = _mm_mul_ps(fscal,dz20);
2163
2164 /* Update vectorial force */
2165 fix2 = _mm_add_ps(fix2,tx);
2166 fiy2 = _mm_add_ps(fiy2,ty);
2167 fiz2 = _mm_add_ps(fiz2,tz);
2168
2169 fjx0 = _mm_add_ps(fjx0,tx);
2170 fjy0 = _mm_add_ps(fjy0,ty);
2171 fjz0 = _mm_add_ps(fjz0,tz);
2172
2173 }
2174
2175 /**************************
2176 * CALCULATE INTERACTIONS *
2177 **************************/
2178
2179 if (gmx_mm_any_lt(rsq21,rcutoff2))
2180 {
2181
2182 /* REACTION-FIELD ELECTROSTATICS */
2183 felec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
2184
2185 cutoff_mask = _mm_cmplt_ps(rsq21,rcutoff2);
2186
2187 fscal = felec;
2188
2189 fscal = _mm_and_ps(fscal,cutoff_mask);
2190
2191 fscal = _mm_andnot_ps(dummy_mask,fscal);
2192
2193 /* Calculate temporary vectorial force */
2194 tx = _mm_mul_ps(fscal,dx21);
2195 ty = _mm_mul_ps(fscal,dy21);
2196 tz = _mm_mul_ps(fscal,dz21);
2197
2198 /* Update vectorial force */
2199 fix2 = _mm_add_ps(fix2,tx);
2200 fiy2 = _mm_add_ps(fiy2,ty);
2201 fiz2 = _mm_add_ps(fiz2,tz);
2202
2203 fjx1 = _mm_add_ps(fjx1,tx);
2204 fjy1 = _mm_add_ps(fjy1,ty);
2205 fjz1 = _mm_add_ps(fjz1,tz);
2206
2207 }
2208
2209 /**************************
2210 * CALCULATE INTERACTIONS *
2211 **************************/
2212
2213 if (gmx_mm_any_lt(rsq22,rcutoff2))
2214 {
2215
2216 /* REACTION-FIELD ELECTROSTATICS */
2217 felec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
2218
2219 cutoff_mask = _mm_cmplt_ps(rsq22,rcutoff2);
2220
2221 fscal = felec;
2222
2223 fscal = _mm_and_ps(fscal,cutoff_mask);
2224
2225 fscal = _mm_andnot_ps(dummy_mask,fscal);
2226
2227 /* Calculate temporary vectorial force */
2228 tx = _mm_mul_ps(fscal,dx22);
2229 ty = _mm_mul_ps(fscal,dy22);
2230 tz = _mm_mul_ps(fscal,dz22);
2231
2232 /* Update vectorial force */
2233 fix2 = _mm_add_ps(fix2,tx);
2234 fiy2 = _mm_add_ps(fiy2,ty);
2235 fiz2 = _mm_add_ps(fiz2,tz);
2236
2237 fjx2 = _mm_add_ps(fjx2,tx);
2238 fjy2 = _mm_add_ps(fjy2,ty);
2239 fjz2 = _mm_add_ps(fjz2,tz);
2240
2241 }
2242
2243 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2244 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2245 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2246 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2247
2248 gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2249 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2250
2251 /* Inner loop uses 298 flops */
2252 }
2253
2254 /* End of innermost loop */
2255
2256 gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2257 f+i_coord_offset,fshift+i_shift_offset);
2258
2259 /* Increment number of inner iterations */
2260 inneriter += j_index_end - j_index_start;
2261
2262 /* Outer loop uses 18 flops */
2263 }
2264
2265 /* Increment number of outer iterations */
2266 outeriter += nri;
2267
2268 /* Update outer/inner flops */
2269
2270 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*298);
2271 }
The ultimate molecular dynamics simulation package