?  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_avx_128_fma_single / nb_kernel_ElecRF_VdwLJ_GeomW4W4_avx_128_fma_single.c
blob1edc2e75ab2b25d9e3b1034227aea124dfdc360c
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 avx_128_fma_single kernel generator.
37 */
38 #include "gmxpre.h"
39
40 #include "config.h"
41
42 #include <math.h>
43
44 #include "../nb_kernel.h"
45 #include "gromacs/gmxlib/nrnb.h"
46
47 #include "kernelutil_x86_avx_128_fma_single.h"
48
49 /*
50 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_avx_128_fma_single
51 * Electrostatics interaction: ReactionField
52 * VdW interaction: LennardJones
53 * Geometry: Water4-Water4
54 * Calculate force/pot: PotentialAndForce
55 */
56 void
57 nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_avx_128_fma_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 AVX_128, 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 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 __m128 dummy_mask,cutoff_mask;
117 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
118 __m128 one = _mm_set1_ps(1.0);
119 __m128 two = _mm_set1_ps(2.0);
120 x = xx[0];
121 f = ff[0];
122
123 nri = nlist->nri;
124 iinr = nlist->iinr;
125 jindex = nlist->jindex;
126 jjnr = nlist->jjnr;
127 shiftidx = nlist->shift;
128 gid = nlist->gid;
129 shiftvec = fr->shift_vec[0];
130 fshift = fr->fshift[0];
131 facel = _mm_set1_ps(fr->ic->epsfac);
132 charge = mdatoms->chargeA;
133 krf = _mm_set1_ps(fr->ic->k_rf);
134 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
135 crf = _mm_set1_ps(fr->ic->c_rf);
136 nvdwtype = fr->ntype;
137 vdwparam = fr->nbfp;
138 vdwtype = mdatoms->typeA;
139
140 /* Setup water-specific parameters */
141 inr = nlist->iinr[0];
142 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
143 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
144 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
145 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
146
147 jq1 = _mm_set1_ps(charge[inr+1]);
148 jq2 = _mm_set1_ps(charge[inr+2]);
149 jq3 = _mm_set1_ps(charge[inr+3]);
150 vdwjidx0A = 2*vdwtype[inr+0];
151 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
152 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
153 qq11 = _mm_mul_ps(iq1,jq1);
154 qq12 = _mm_mul_ps(iq1,jq2);
155 qq13 = _mm_mul_ps(iq1,jq3);
156 qq21 = _mm_mul_ps(iq2,jq1);
157 qq22 = _mm_mul_ps(iq2,jq2);
158 qq23 = _mm_mul_ps(iq2,jq3);
159 qq31 = _mm_mul_ps(iq3,jq1);
160 qq32 = _mm_mul_ps(iq3,jq2);
161 qq33 = _mm_mul_ps(iq3,jq3);
162
163 /* Avoid stupid compiler warnings */
164 jnrA = jnrB = jnrC = jnrD = 0;
165 j_coord_offsetA = 0;
166 j_coord_offsetB = 0;
167 j_coord_offsetC = 0;
168 j_coord_offsetD = 0;
169
170 outeriter = 0;
171 inneriter = 0;
172
173 for(iidx=0;iidx<4*DIM;iidx++)
174 {
175 scratch[iidx] = 0.0;
176 }
177
178 /* Start outer loop over neighborlists */
179 for(iidx=0; iidx<nri; iidx++)
180 {
181 /* Load shift vector for this list */
182 i_shift_offset = DIM*shiftidx[iidx];
183
184 /* Load limits for loop over neighbors */
185 j_index_start = jindex[iidx];
186 j_index_end = jindex[iidx+1];
187
188 /* Get outer coordinate index */
189 inr = iinr[iidx];
190 i_coord_offset = DIM*inr;
191
192 /* Load i particle coords and add shift vector */
193 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
194 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
195
196 fix0 = _mm_setzero_ps();
197 fiy0 = _mm_setzero_ps();
198 fiz0 = _mm_setzero_ps();
199 fix1 = _mm_setzero_ps();
200 fiy1 = _mm_setzero_ps();
201 fiz1 = _mm_setzero_ps();
202 fix2 = _mm_setzero_ps();
203 fiy2 = _mm_setzero_ps();
204 fiz2 = _mm_setzero_ps();
205 fix3 = _mm_setzero_ps();
206 fiy3 = _mm_setzero_ps();
207 fiz3 = _mm_setzero_ps();
208
209 /* Reset potential sums */
210 velecsum = _mm_setzero_ps();
211 vvdwsum = _mm_setzero_ps();
212
213 /* Start inner kernel loop */
214 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
215 {
216
217 /* Get j neighbor index, and coordinate index */
218 jnrA = jjnr[jidx];
219 jnrB = jjnr[jidx+1];
220 jnrC = jjnr[jidx+2];
221 jnrD = jjnr[jidx+3];
222 j_coord_offsetA = DIM*jnrA;
223 j_coord_offsetB = DIM*jnrB;
224 j_coord_offsetC = DIM*jnrC;
225 j_coord_offsetD = DIM*jnrD;
226
227 /* load j atom coordinates */
228 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
229 x+j_coord_offsetC,x+j_coord_offsetD,
230 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
231 &jy2,&jz2,&jx3,&jy3,&jz3);
232
233 /* Calculate displacement vector */
234 dx00 = _mm_sub_ps(ix0,jx0);
235 dy00 = _mm_sub_ps(iy0,jy0);
236 dz00 = _mm_sub_ps(iz0,jz0);
237 dx11 = _mm_sub_ps(ix1,jx1);
238 dy11 = _mm_sub_ps(iy1,jy1);
239 dz11 = _mm_sub_ps(iz1,jz1);
240 dx12 = _mm_sub_ps(ix1,jx2);
241 dy12 = _mm_sub_ps(iy1,jy2);
242 dz12 = _mm_sub_ps(iz1,jz2);
243 dx13 = _mm_sub_ps(ix1,jx3);
244 dy13 = _mm_sub_ps(iy1,jy3);
245 dz13 = _mm_sub_ps(iz1,jz3);
246 dx21 = _mm_sub_ps(ix2,jx1);
247 dy21 = _mm_sub_ps(iy2,jy1);
248 dz21 = _mm_sub_ps(iz2,jz1);
249 dx22 = _mm_sub_ps(ix2,jx2);
250 dy22 = _mm_sub_ps(iy2,jy2);
251 dz22 = _mm_sub_ps(iz2,jz2);
252 dx23 = _mm_sub_ps(ix2,jx3);
253 dy23 = _mm_sub_ps(iy2,jy3);
254 dz23 = _mm_sub_ps(iz2,jz3);
255 dx31 = _mm_sub_ps(ix3,jx1);
256 dy31 = _mm_sub_ps(iy3,jy1);
257 dz31 = _mm_sub_ps(iz3,jz1);
258 dx32 = _mm_sub_ps(ix3,jx2);
259 dy32 = _mm_sub_ps(iy3,jy2);
260 dz32 = _mm_sub_ps(iz3,jz2);
261 dx33 = _mm_sub_ps(ix3,jx3);
262 dy33 = _mm_sub_ps(iy3,jy3);
263 dz33 = _mm_sub_ps(iz3,jz3);
264
265 /* Calculate squared distance and things based on it */
266 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
267 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
268 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
269 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
270 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
271 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
272 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
273 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
274 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
275 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
276
277 rinv11 = avx128fma_invsqrt_f(rsq11);
278 rinv12 = avx128fma_invsqrt_f(rsq12);
279 rinv13 = avx128fma_invsqrt_f(rsq13);
280 rinv21 = avx128fma_invsqrt_f(rsq21);
281 rinv22 = avx128fma_invsqrt_f(rsq22);
282 rinv23 = avx128fma_invsqrt_f(rsq23);
283 rinv31 = avx128fma_invsqrt_f(rsq31);
284 rinv32 = avx128fma_invsqrt_f(rsq32);
285 rinv33 = avx128fma_invsqrt_f(rsq33);
286
287 rinvsq00 = avx128fma_inv_f(rsq00);
288 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
289 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
290 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
291 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
292 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
293 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
294 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
295 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
296 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
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 fjx3 = _mm_setzero_ps();
308 fjy3 = _mm_setzero_ps();
309 fjz3 = _mm_setzero_ps();
310
311 /**************************
312 * CALCULATE INTERACTIONS *
313 **************************/
314
315 /* LENNARD-JONES DISPERSION/REPULSION */
316
317 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
318 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
319 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
320 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
321 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
322
323 /* Update potential sum for this i atom from the interaction with this j atom. */
324 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
325
326 fscal = fvdw;
327
328 /* Update vectorial force */
329 fix0 = _mm_macc_ps(dx00,fscal,fix0);
330 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
331 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
332
333 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
334 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
335 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
336
337 /**************************
338 * CALCULATE INTERACTIONS *
339 **************************/
340
341 /* REACTION-FIELD ELECTROSTATICS */
342 velec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
343 felec = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
344
345 /* Update potential sum for this i atom from the interaction with this j atom. */
346 velecsum = _mm_add_ps(velecsum,velec);
347
348 fscal = felec;
349
350 /* Update vectorial force */
351 fix1 = _mm_macc_ps(dx11,fscal,fix1);
352 fiy1 = _mm_macc_ps(dy11,fscal,fiy1);
353 fiz1 = _mm_macc_ps(dz11,fscal,fiz1);
354
355 fjx1 = _mm_macc_ps(dx11,fscal,fjx1);
356 fjy1 = _mm_macc_ps(dy11,fscal,fjy1);
357 fjz1 = _mm_macc_ps(dz11,fscal,fjz1);
358
359 /**************************
360 * CALCULATE INTERACTIONS *
361 **************************/
362
363 /* REACTION-FIELD ELECTROSTATICS */
364 velec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
365 felec = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
366
367 /* Update potential sum for this i atom from the interaction with this j atom. */
368 velecsum = _mm_add_ps(velecsum,velec);
369
370 fscal = felec;
371
372 /* Update vectorial force */
373 fix1 = _mm_macc_ps(dx12,fscal,fix1);
374 fiy1 = _mm_macc_ps(dy12,fscal,fiy1);
375 fiz1 = _mm_macc_ps(dz12,fscal,fiz1);
376
377 fjx2 = _mm_macc_ps(dx12,fscal,fjx2);
378 fjy2 = _mm_macc_ps(dy12,fscal,fjy2);
379 fjz2 = _mm_macc_ps(dz12,fscal,fjz2);
380
381 /**************************
382 * CALCULATE INTERACTIONS *
383 **************************/
384
385 /* REACTION-FIELD ELECTROSTATICS */
386 velec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
387 felec = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
388
389 /* Update potential sum for this i atom from the interaction with this j atom. */
390 velecsum = _mm_add_ps(velecsum,velec);
391
392 fscal = felec;
393
394 /* Update vectorial force */
395 fix1 = _mm_macc_ps(dx13,fscal,fix1);
396 fiy1 = _mm_macc_ps(dy13,fscal,fiy1);
397 fiz1 = _mm_macc_ps(dz13,fscal,fiz1);
398
399 fjx3 = _mm_macc_ps(dx13,fscal,fjx3);
400 fjy3 = _mm_macc_ps(dy13,fscal,fjy3);
401 fjz3 = _mm_macc_ps(dz13,fscal,fjz3);
402
403 /**************************
404 * CALCULATE INTERACTIONS *
405 **************************/
406
407 /* REACTION-FIELD ELECTROSTATICS */
408 velec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
409 felec = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
410
411 /* Update potential sum for this i atom from the interaction with this j atom. */
412 velecsum = _mm_add_ps(velecsum,velec);
413
414 fscal = felec;
415
416 /* Update vectorial force */
417 fix2 = _mm_macc_ps(dx21,fscal,fix2);
418 fiy2 = _mm_macc_ps(dy21,fscal,fiy2);
419 fiz2 = _mm_macc_ps(dz21,fscal,fiz2);
420
421 fjx1 = _mm_macc_ps(dx21,fscal,fjx1);
422 fjy1 = _mm_macc_ps(dy21,fscal,fjy1);
423 fjz1 = _mm_macc_ps(dz21,fscal,fjz1);
424
425 /**************************
426 * CALCULATE INTERACTIONS *
427 **************************/
428
429 /* REACTION-FIELD ELECTROSTATICS */
430 velec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
431 felec = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
432
433 /* Update potential sum for this i atom from the interaction with this j atom. */
434 velecsum = _mm_add_ps(velecsum,velec);
435
436 fscal = felec;
437
438 /* Update vectorial force */
439 fix2 = _mm_macc_ps(dx22,fscal,fix2);
440 fiy2 = _mm_macc_ps(dy22,fscal,fiy2);
441 fiz2 = _mm_macc_ps(dz22,fscal,fiz2);
442
443 fjx2 = _mm_macc_ps(dx22,fscal,fjx2);
444 fjy2 = _mm_macc_ps(dy22,fscal,fjy2);
445 fjz2 = _mm_macc_ps(dz22,fscal,fjz2);
446
447 /**************************
448 * CALCULATE INTERACTIONS *
449 **************************/
450
451 /* REACTION-FIELD ELECTROSTATICS */
452 velec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
453 felec = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
454
455 /* Update potential sum for this i atom from the interaction with this j atom. */
456 velecsum = _mm_add_ps(velecsum,velec);
457
458 fscal = felec;
459
460 /* Update vectorial force */
461 fix2 = _mm_macc_ps(dx23,fscal,fix2);
462 fiy2 = _mm_macc_ps(dy23,fscal,fiy2);
463 fiz2 = _mm_macc_ps(dz23,fscal,fiz2);
464
465 fjx3 = _mm_macc_ps(dx23,fscal,fjx3);
466 fjy3 = _mm_macc_ps(dy23,fscal,fjy3);
467 fjz3 = _mm_macc_ps(dz23,fscal,fjz3);
468
469 /**************************
470 * CALCULATE INTERACTIONS *
471 **************************/
472
473 /* REACTION-FIELD ELECTROSTATICS */
474 velec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
475 felec = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
476
477 /* Update potential sum for this i atom from the interaction with this j atom. */
478 velecsum = _mm_add_ps(velecsum,velec);
479
480 fscal = felec;
481
482 /* Update vectorial force */
483 fix3 = _mm_macc_ps(dx31,fscal,fix3);
484 fiy3 = _mm_macc_ps(dy31,fscal,fiy3);
485 fiz3 = _mm_macc_ps(dz31,fscal,fiz3);
486
487 fjx1 = _mm_macc_ps(dx31,fscal,fjx1);
488 fjy1 = _mm_macc_ps(dy31,fscal,fjy1);
489 fjz1 = _mm_macc_ps(dz31,fscal,fjz1);
490
491 /**************************
492 * CALCULATE INTERACTIONS *
493 **************************/
494
495 /* REACTION-FIELD ELECTROSTATICS */
496 velec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
497 felec = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
498
499 /* Update potential sum for this i atom from the interaction with this j atom. */
500 velecsum = _mm_add_ps(velecsum,velec);
501
502 fscal = felec;
503
504 /* Update vectorial force */
505 fix3 = _mm_macc_ps(dx32,fscal,fix3);
506 fiy3 = _mm_macc_ps(dy32,fscal,fiy3);
507 fiz3 = _mm_macc_ps(dz32,fscal,fiz3);
508
509 fjx2 = _mm_macc_ps(dx32,fscal,fjx2);
510 fjy2 = _mm_macc_ps(dy32,fscal,fjy2);
511 fjz2 = _mm_macc_ps(dz32,fscal,fjz2);
512
513 /**************************
514 * CALCULATE INTERACTIONS *
515 **************************/
516
517 /* REACTION-FIELD ELECTROSTATICS */
518 velec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
519 felec = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,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 /* Update vectorial force */
527 fix3 = _mm_macc_ps(dx33,fscal,fix3);
528 fiy3 = _mm_macc_ps(dy33,fscal,fiy3);
529 fiz3 = _mm_macc_ps(dz33,fscal,fiz3);
530
531 fjx3 = _mm_macc_ps(dx33,fscal,fjx3);
532 fjy3 = _mm_macc_ps(dy33,fscal,fjy3);
533 fjz3 = _mm_macc_ps(dz33,fscal,fjz3);
534
535 fjptrA = f+j_coord_offsetA;
536 fjptrB = f+j_coord_offsetB;
537 fjptrC = f+j_coord_offsetC;
538 fjptrD = f+j_coord_offsetD;
539
540 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
541 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
542 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
543
544 /* Inner loop uses 353 flops */
545 }
546
547 if(jidx<j_index_end)
548 {
549
550 /* Get j neighbor index, and coordinate index */
551 jnrlistA = jjnr[jidx];
552 jnrlistB = jjnr[jidx+1];
553 jnrlistC = jjnr[jidx+2];
554 jnrlistD = jjnr[jidx+3];
555 /* Sign of each element will be negative for non-real atoms.
556 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
557 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
558 */
559 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
560 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
561 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
562 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
563 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
564 j_coord_offsetA = DIM*jnrA;
565 j_coord_offsetB = DIM*jnrB;
566 j_coord_offsetC = DIM*jnrC;
567 j_coord_offsetD = DIM*jnrD;
568
569 /* load j atom coordinates */
570 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
571 x+j_coord_offsetC,x+j_coord_offsetD,
572 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
573 &jy2,&jz2,&jx3,&jy3,&jz3);
574
575 /* Calculate displacement vector */
576 dx00 = _mm_sub_ps(ix0,jx0);
577 dy00 = _mm_sub_ps(iy0,jy0);
578 dz00 = _mm_sub_ps(iz0,jz0);
579 dx11 = _mm_sub_ps(ix1,jx1);
580 dy11 = _mm_sub_ps(iy1,jy1);
581 dz11 = _mm_sub_ps(iz1,jz1);
582 dx12 = _mm_sub_ps(ix1,jx2);
583 dy12 = _mm_sub_ps(iy1,jy2);
584 dz12 = _mm_sub_ps(iz1,jz2);
585 dx13 = _mm_sub_ps(ix1,jx3);
586 dy13 = _mm_sub_ps(iy1,jy3);
587 dz13 = _mm_sub_ps(iz1,jz3);
588 dx21 = _mm_sub_ps(ix2,jx1);
589 dy21 = _mm_sub_ps(iy2,jy1);
590 dz21 = _mm_sub_ps(iz2,jz1);
591 dx22 = _mm_sub_ps(ix2,jx2);
592 dy22 = _mm_sub_ps(iy2,jy2);
593 dz22 = _mm_sub_ps(iz2,jz2);
594 dx23 = _mm_sub_ps(ix2,jx3);
595 dy23 = _mm_sub_ps(iy2,jy3);
596 dz23 = _mm_sub_ps(iz2,jz3);
597 dx31 = _mm_sub_ps(ix3,jx1);
598 dy31 = _mm_sub_ps(iy3,jy1);
599 dz31 = _mm_sub_ps(iz3,jz1);
600 dx32 = _mm_sub_ps(ix3,jx2);
601 dy32 = _mm_sub_ps(iy3,jy2);
602 dz32 = _mm_sub_ps(iz3,jz2);
603 dx33 = _mm_sub_ps(ix3,jx3);
604 dy33 = _mm_sub_ps(iy3,jy3);
605 dz33 = _mm_sub_ps(iz3,jz3);
606
607 /* Calculate squared distance and things based on it */
608 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
609 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
610 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
611 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
612 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
613 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
614 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
615 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
616 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
617 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
618
619 rinv11 = avx128fma_invsqrt_f(rsq11);
620 rinv12 = avx128fma_invsqrt_f(rsq12);
621 rinv13 = avx128fma_invsqrt_f(rsq13);
622 rinv21 = avx128fma_invsqrt_f(rsq21);
623 rinv22 = avx128fma_invsqrt_f(rsq22);
624 rinv23 = avx128fma_invsqrt_f(rsq23);
625 rinv31 = avx128fma_invsqrt_f(rsq31);
626 rinv32 = avx128fma_invsqrt_f(rsq32);
627 rinv33 = avx128fma_invsqrt_f(rsq33);
628
629 rinvsq00 = avx128fma_inv_f(rsq00);
630 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
631 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
632 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
633 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
634 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
635 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
636 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
637 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
638 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
639
640 fjx0 = _mm_setzero_ps();
641 fjy0 = _mm_setzero_ps();
642 fjz0 = _mm_setzero_ps();
643 fjx1 = _mm_setzero_ps();
644 fjy1 = _mm_setzero_ps();
645 fjz1 = _mm_setzero_ps();
646 fjx2 = _mm_setzero_ps();
647 fjy2 = _mm_setzero_ps();
648 fjz2 = _mm_setzero_ps();
649 fjx3 = _mm_setzero_ps();
650 fjy3 = _mm_setzero_ps();
651 fjz3 = _mm_setzero_ps();
652
653 /**************************
654 * CALCULATE INTERACTIONS *
655 **************************/
656
657 /* LENNARD-JONES DISPERSION/REPULSION */
658
659 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
660 vvdw6 = _mm_mul_ps(c6_00,rinvsix);
661 vvdw12 = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
662 vvdw = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
663 fvdw = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
664
665 /* Update potential sum for this i atom from the interaction with this j atom. */
666 vvdw = _mm_andnot_ps(dummy_mask,vvdw);
667 vvdwsum = _mm_add_ps(vvdwsum,vvdw);
668
669 fscal = fvdw;
670
671 fscal = _mm_andnot_ps(dummy_mask,fscal);
672
673 /* Update vectorial force */
674 fix0 = _mm_macc_ps(dx00,fscal,fix0);
675 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
676 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
677
678 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
679 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
680 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
681
682 /**************************
683 * CALCULATE INTERACTIONS *
684 **************************/
685
686 /* REACTION-FIELD ELECTROSTATICS */
687 velec = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
688 felec = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
689
690 /* Update potential sum for this i atom from the interaction with this j atom. */
691 velec = _mm_andnot_ps(dummy_mask,velec);
692 velecsum = _mm_add_ps(velecsum,velec);
693
694 fscal = felec;
695
696 fscal = _mm_andnot_ps(dummy_mask,fscal);
697
698 /* Update vectorial force */
699 fix1 = _mm_macc_ps(dx11,fscal,fix1);
700 fiy1 = _mm_macc_ps(dy11,fscal,fiy1);
701 fiz1 = _mm_macc_ps(dz11,fscal,fiz1);
702
703 fjx1 = _mm_macc_ps(dx11,fscal,fjx1);
704 fjy1 = _mm_macc_ps(dy11,fscal,fjy1);
705 fjz1 = _mm_macc_ps(dz11,fscal,fjz1);
706
707 /**************************
708 * CALCULATE INTERACTIONS *
709 **************************/
710
711 /* REACTION-FIELD ELECTROSTATICS */
712 velec = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
713 felec = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
714
715 /* Update potential sum for this i atom from the interaction with this j atom. */
716 velec = _mm_andnot_ps(dummy_mask,velec);
717 velecsum = _mm_add_ps(velecsum,velec);
718
719 fscal = felec;
720
721 fscal = _mm_andnot_ps(dummy_mask,fscal);
722
723 /* Update vectorial force */
724 fix1 = _mm_macc_ps(dx12,fscal,fix1);
725 fiy1 = _mm_macc_ps(dy12,fscal,fiy1);
726 fiz1 = _mm_macc_ps(dz12,fscal,fiz1);
727
728 fjx2 = _mm_macc_ps(dx12,fscal,fjx2);
729 fjy2 = _mm_macc_ps(dy12,fscal,fjy2);
730 fjz2 = _mm_macc_ps(dz12,fscal,fjz2);
731
732 /**************************
733 * CALCULATE INTERACTIONS *
734 **************************/
735
736 /* REACTION-FIELD ELECTROSTATICS */
737 velec = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
738 felec = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
739
740 /* Update potential sum for this i atom from the interaction with this j atom. */
741 velec = _mm_andnot_ps(dummy_mask,velec);
742 velecsum = _mm_add_ps(velecsum,velec);
743
744 fscal = felec;
745
746 fscal = _mm_andnot_ps(dummy_mask,fscal);
747
748 /* Update vectorial force */
749 fix1 = _mm_macc_ps(dx13,fscal,fix1);
750 fiy1 = _mm_macc_ps(dy13,fscal,fiy1);
751 fiz1 = _mm_macc_ps(dz13,fscal,fiz1);
752
753 fjx3 = _mm_macc_ps(dx13,fscal,fjx3);
754 fjy3 = _mm_macc_ps(dy13,fscal,fjy3);
755 fjz3 = _mm_macc_ps(dz13,fscal,fjz3);
756
757 /**************************
758 * CALCULATE INTERACTIONS *
759 **************************/
760
761 /* REACTION-FIELD ELECTROSTATICS */
762 velec = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
763 felec = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
764
765 /* Update potential sum for this i atom from the interaction with this j atom. */
766 velec = _mm_andnot_ps(dummy_mask,velec);
767 velecsum = _mm_add_ps(velecsum,velec);
768
769 fscal = felec;
770
771 fscal = _mm_andnot_ps(dummy_mask,fscal);
772
773 /* Update vectorial force */
774 fix2 = _mm_macc_ps(dx21,fscal,fix2);
775 fiy2 = _mm_macc_ps(dy21,fscal,fiy2);
776 fiz2 = _mm_macc_ps(dz21,fscal,fiz2);
777
778 fjx1 = _mm_macc_ps(dx21,fscal,fjx1);
779 fjy1 = _mm_macc_ps(dy21,fscal,fjy1);
780 fjz1 = _mm_macc_ps(dz21,fscal,fjz1);
781
782 /**************************
783 * CALCULATE INTERACTIONS *
784 **************************/
785
786 /* REACTION-FIELD ELECTROSTATICS */
787 velec = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
788 felec = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
789
790 /* Update potential sum for this i atom from the interaction with this j atom. */
791 velec = _mm_andnot_ps(dummy_mask,velec);
792 velecsum = _mm_add_ps(velecsum,velec);
793
794 fscal = felec;
795
796 fscal = _mm_andnot_ps(dummy_mask,fscal);
797
798 /* Update vectorial force */
799 fix2 = _mm_macc_ps(dx22,fscal,fix2);
800 fiy2 = _mm_macc_ps(dy22,fscal,fiy2);
801 fiz2 = _mm_macc_ps(dz22,fscal,fiz2);
802
803 fjx2 = _mm_macc_ps(dx22,fscal,fjx2);
804 fjy2 = _mm_macc_ps(dy22,fscal,fjy2);
805 fjz2 = _mm_macc_ps(dz22,fscal,fjz2);
806
807 /**************************
808 * CALCULATE INTERACTIONS *
809 **************************/
810
811 /* REACTION-FIELD ELECTROSTATICS */
812 velec = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
813 felec = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
814
815 /* Update potential sum for this i atom from the interaction with this j atom. */
816 velec = _mm_andnot_ps(dummy_mask,velec);
817 velecsum = _mm_add_ps(velecsum,velec);
818
819 fscal = felec;
820
821 fscal = _mm_andnot_ps(dummy_mask,fscal);
822
823 /* Update vectorial force */
824 fix2 = _mm_macc_ps(dx23,fscal,fix2);
825 fiy2 = _mm_macc_ps(dy23,fscal,fiy2);
826 fiz2 = _mm_macc_ps(dz23,fscal,fiz2);
827
828 fjx3 = _mm_macc_ps(dx23,fscal,fjx3);
829 fjy3 = _mm_macc_ps(dy23,fscal,fjy3);
830 fjz3 = _mm_macc_ps(dz23,fscal,fjz3);
831
832 /**************************
833 * CALCULATE INTERACTIONS *
834 **************************/
835
836 /* REACTION-FIELD ELECTROSTATICS */
837 velec = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
838 felec = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
839
840 /* Update potential sum for this i atom from the interaction with this j atom. */
841 velec = _mm_andnot_ps(dummy_mask,velec);
842 velecsum = _mm_add_ps(velecsum,velec);
843
844 fscal = felec;
845
846 fscal = _mm_andnot_ps(dummy_mask,fscal);
847
848 /* Update vectorial force */
849 fix3 = _mm_macc_ps(dx31,fscal,fix3);
850 fiy3 = _mm_macc_ps(dy31,fscal,fiy3);
851 fiz3 = _mm_macc_ps(dz31,fscal,fiz3);
852
853 fjx1 = _mm_macc_ps(dx31,fscal,fjx1);
854 fjy1 = _mm_macc_ps(dy31,fscal,fjy1);
855 fjz1 = _mm_macc_ps(dz31,fscal,fjz1);
856
857 /**************************
858 * CALCULATE INTERACTIONS *
859 **************************/
860
861 /* REACTION-FIELD ELECTROSTATICS */
862 velec = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
863 felec = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
864
865 /* Update potential sum for this i atom from the interaction with this j atom. */
866 velec = _mm_andnot_ps(dummy_mask,velec);
867 velecsum = _mm_add_ps(velecsum,velec);
868
869 fscal = felec;
870
871 fscal = _mm_andnot_ps(dummy_mask,fscal);
872
873 /* Update vectorial force */
874 fix3 = _mm_macc_ps(dx32,fscal,fix3);
875 fiy3 = _mm_macc_ps(dy32,fscal,fiy3);
876 fiz3 = _mm_macc_ps(dz32,fscal,fiz3);
877
878 fjx2 = _mm_macc_ps(dx32,fscal,fjx2);
879 fjy2 = _mm_macc_ps(dy32,fscal,fjy2);
880 fjz2 = _mm_macc_ps(dz32,fscal,fjz2);
881
882 /**************************
883 * CALCULATE INTERACTIONS *
884 **************************/
885
886 /* REACTION-FIELD ELECTROSTATICS */
887 velec = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
888 felec = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
889
890 /* Update potential sum for this i atom from the interaction with this j atom. */
891 velec = _mm_andnot_ps(dummy_mask,velec);
892 velecsum = _mm_add_ps(velecsum,velec);
893
894 fscal = felec;
895
896 fscal = _mm_andnot_ps(dummy_mask,fscal);
897
898 /* Update vectorial force */
899 fix3 = _mm_macc_ps(dx33,fscal,fix3);
900 fiy3 = _mm_macc_ps(dy33,fscal,fiy3);
901 fiz3 = _mm_macc_ps(dz33,fscal,fiz3);
902
903 fjx3 = _mm_macc_ps(dx33,fscal,fjx3);
904 fjy3 = _mm_macc_ps(dy33,fscal,fjy3);
905 fjz3 = _mm_macc_ps(dz33,fscal,fjz3);
906
907 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
908 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
909 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
910 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
911
912 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
913 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
914 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
915
916 /* Inner loop uses 353 flops */
917 }
918
919 /* End of innermost loop */
920
921 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
922 f+i_coord_offset,fshift+i_shift_offset);
923
924 ggid = gid[iidx];
925 /* Update potential energies */
926 gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
927 gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
928
929 /* Increment number of inner iterations */
930 inneriter += j_index_end - j_index_start;
931
932 /* Outer loop uses 26 flops */
933 }
934
935 /* Increment number of outer iterations */
936 outeriter += nri;
937
938 /* Update outer/inner flops */
939
940 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*353);
941 }
942 /*
943 * Gromacs nonbonded kernel: nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_avx_128_fma_single
944 * Electrostatics interaction: ReactionField
945 * VdW interaction: LennardJones
946 * Geometry: Water4-Water4
947 * Calculate force/pot: Force
948 */
949 void
950 nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_avx_128_fma_single
951 (t_nblist * gmx_restrict nlist,
952 rvec * gmx_restrict xx,
953 rvec * gmx_restrict ff,
954 struct t_forcerec * gmx_restrict fr,
955 t_mdatoms * gmx_restrict mdatoms,
956 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
957 t_nrnb * gmx_restrict nrnb)
958 {
959 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
960 * just 0 for non-waters.
961 * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
962 * jnr indices corresponding to data put in the four positions in the SIMD register.
963 */
964 int i_shift_offset,i_coord_offset,outeriter,inneriter;
965 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
966 int jnrA,jnrB,jnrC,jnrD;
967 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
968 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
969 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
970 real rcutoff_scalar;
971 real *shiftvec,*fshift,*x,*f;
972 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
973 real scratch[4*DIM];
974 __m128 fscal,rcutoff,rcutoff2,jidxall;
975 int vdwioffset0;
976 __m128 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
977 int vdwioffset1;
978 __m128 ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
979 int vdwioffset2;
980 __m128 ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
981 int vdwioffset3;
982 __m128 ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
983 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
984 __m128 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
985 int vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
986 __m128 jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
987 int vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
988 __m128 jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
989 int vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
990 __m128 jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
991 __m128 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
992 __m128 dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
993 __m128 dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
994 __m128 dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
995 __m128 dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
996 __m128 dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
997 __m128 dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
998 __m128 dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
999 __m128 dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1000 __m128 dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1001 __m128 velec,felec,velecsum,facel,crf,krf,krf2;
1002 real *charge;
1003 int nvdwtype;
1004 __m128 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1005 int *vdwtype;
1006 real *vdwparam;
1007 __m128 one_sixth = _mm_set1_ps(1.0/6.0);
1008 __m128 one_twelfth = _mm_set1_ps(1.0/12.0);
1009 __m128 dummy_mask,cutoff_mask;
1010 __m128 signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1011 __m128 one = _mm_set1_ps(1.0);
1012 __m128 two = _mm_set1_ps(2.0);
1013 x = xx[0];
1014 f = ff[0];
1015
1016 nri = nlist->nri;
1017 iinr = nlist->iinr;
1018 jindex = nlist->jindex;
1019 jjnr = nlist->jjnr;
1020 shiftidx = nlist->shift;
1021 gid = nlist->gid;
1022 shiftvec = fr->shift_vec[0];
1023 fshift = fr->fshift[0];
1024 facel = _mm_set1_ps(fr->ic->epsfac);
1025 charge = mdatoms->chargeA;
1026 krf = _mm_set1_ps(fr->ic->k_rf);
1027 krf2 = _mm_set1_ps(fr->ic->k_rf*2.0);
1028 crf = _mm_set1_ps(fr->ic->c_rf);
1029 nvdwtype = fr->ntype;
1030 vdwparam = fr->nbfp;
1031 vdwtype = mdatoms->typeA;
1032
1033 /* Setup water-specific parameters */
1034 inr = nlist->iinr[0];
1035 iq1 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1036 iq2 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1037 iq3 = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1038 vdwioffset0 = 2*nvdwtype*vdwtype[inr+0];
1039
1040 jq1 = _mm_set1_ps(charge[inr+1]);
1041 jq2 = _mm_set1_ps(charge[inr+2]);
1042 jq3 = _mm_set1_ps(charge[inr+3]);
1043 vdwjidx0A = 2*vdwtype[inr+0];
1044 c6_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1045 c12_00 = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1046 qq11 = _mm_mul_ps(iq1,jq1);
1047 qq12 = _mm_mul_ps(iq1,jq2);
1048 qq13 = _mm_mul_ps(iq1,jq3);
1049 qq21 = _mm_mul_ps(iq2,jq1);
1050 qq22 = _mm_mul_ps(iq2,jq2);
1051 qq23 = _mm_mul_ps(iq2,jq3);
1052 qq31 = _mm_mul_ps(iq3,jq1);
1053 qq32 = _mm_mul_ps(iq3,jq2);
1054 qq33 = _mm_mul_ps(iq3,jq3);
1055
1056 /* Avoid stupid compiler warnings */
1057 jnrA = jnrB = jnrC = jnrD = 0;
1058 j_coord_offsetA = 0;
1059 j_coord_offsetB = 0;
1060 j_coord_offsetC = 0;
1061 j_coord_offsetD = 0;
1062
1063 outeriter = 0;
1064 inneriter = 0;
1065
1066 for(iidx=0;iidx<4*DIM;iidx++)
1067 {
1068 scratch[iidx] = 0.0;
1069 }
1070
1071 /* Start outer loop over neighborlists */
1072 for(iidx=0; iidx<nri; iidx++)
1073 {
1074 /* Load shift vector for this list */
1075 i_shift_offset = DIM*shiftidx[iidx];
1076
1077 /* Load limits for loop over neighbors */
1078 j_index_start = jindex[iidx];
1079 j_index_end = jindex[iidx+1];
1080
1081 /* Get outer coordinate index */
1082 inr = iinr[iidx];
1083 i_coord_offset = DIM*inr;
1084
1085 /* Load i particle coords and add shift vector */
1086 gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1087 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1088
1089 fix0 = _mm_setzero_ps();
1090 fiy0 = _mm_setzero_ps();
1091 fiz0 = _mm_setzero_ps();
1092 fix1 = _mm_setzero_ps();
1093 fiy1 = _mm_setzero_ps();
1094 fiz1 = _mm_setzero_ps();
1095 fix2 = _mm_setzero_ps();
1096 fiy2 = _mm_setzero_ps();
1097 fiz2 = _mm_setzero_ps();
1098 fix3 = _mm_setzero_ps();
1099 fiy3 = _mm_setzero_ps();
1100 fiz3 = _mm_setzero_ps();
1101
1102 /* Start inner kernel loop */
1103 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1104 {
1105
1106 /* Get j neighbor index, and coordinate index */
1107 jnrA = jjnr[jidx];
1108 jnrB = jjnr[jidx+1];
1109 jnrC = jjnr[jidx+2];
1110 jnrD = jjnr[jidx+3];
1111 j_coord_offsetA = DIM*jnrA;
1112 j_coord_offsetB = DIM*jnrB;
1113 j_coord_offsetC = DIM*jnrC;
1114 j_coord_offsetD = DIM*jnrD;
1115
1116 /* load j atom coordinates */
1117 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1118 x+j_coord_offsetC,x+j_coord_offsetD,
1119 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1120 &jy2,&jz2,&jx3,&jy3,&jz3);
1121
1122 /* Calculate displacement vector */
1123 dx00 = _mm_sub_ps(ix0,jx0);
1124 dy00 = _mm_sub_ps(iy0,jy0);
1125 dz00 = _mm_sub_ps(iz0,jz0);
1126 dx11 = _mm_sub_ps(ix1,jx1);
1127 dy11 = _mm_sub_ps(iy1,jy1);
1128 dz11 = _mm_sub_ps(iz1,jz1);
1129 dx12 = _mm_sub_ps(ix1,jx2);
1130 dy12 = _mm_sub_ps(iy1,jy2);
1131 dz12 = _mm_sub_ps(iz1,jz2);
1132 dx13 = _mm_sub_ps(ix1,jx3);
1133 dy13 = _mm_sub_ps(iy1,jy3);
1134 dz13 = _mm_sub_ps(iz1,jz3);
1135 dx21 = _mm_sub_ps(ix2,jx1);
1136 dy21 = _mm_sub_ps(iy2,jy1);
1137 dz21 = _mm_sub_ps(iz2,jz1);
1138 dx22 = _mm_sub_ps(ix2,jx2);
1139 dy22 = _mm_sub_ps(iy2,jy2);
1140 dz22 = _mm_sub_ps(iz2,jz2);
1141 dx23 = _mm_sub_ps(ix2,jx3);
1142 dy23 = _mm_sub_ps(iy2,jy3);
1143 dz23 = _mm_sub_ps(iz2,jz3);
1144 dx31 = _mm_sub_ps(ix3,jx1);
1145 dy31 = _mm_sub_ps(iy3,jy1);
1146 dz31 = _mm_sub_ps(iz3,jz1);
1147 dx32 = _mm_sub_ps(ix3,jx2);
1148 dy32 = _mm_sub_ps(iy3,jy2);
1149 dz32 = _mm_sub_ps(iz3,jz2);
1150 dx33 = _mm_sub_ps(ix3,jx3);
1151 dy33 = _mm_sub_ps(iy3,jy3);
1152 dz33 = _mm_sub_ps(iz3,jz3);
1153
1154 /* Calculate squared distance and things based on it */
1155 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1156 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1157 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1158 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1159 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1160 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1161 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1162 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1163 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1164 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1165
1166 rinv11 = avx128fma_invsqrt_f(rsq11);
1167 rinv12 = avx128fma_invsqrt_f(rsq12);
1168 rinv13 = avx128fma_invsqrt_f(rsq13);
1169 rinv21 = avx128fma_invsqrt_f(rsq21);
1170 rinv22 = avx128fma_invsqrt_f(rsq22);
1171 rinv23 = avx128fma_invsqrt_f(rsq23);
1172 rinv31 = avx128fma_invsqrt_f(rsq31);
1173 rinv32 = avx128fma_invsqrt_f(rsq32);
1174 rinv33 = avx128fma_invsqrt_f(rsq33);
1175
1176 rinvsq00 = avx128fma_inv_f(rsq00);
1177 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1178 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1179 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1180 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1181 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1182 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1183 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1184 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1185 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1186
1187 fjx0 = _mm_setzero_ps();
1188 fjy0 = _mm_setzero_ps();
1189 fjz0 = _mm_setzero_ps();
1190 fjx1 = _mm_setzero_ps();
1191 fjy1 = _mm_setzero_ps();
1192 fjz1 = _mm_setzero_ps();
1193 fjx2 = _mm_setzero_ps();
1194 fjy2 = _mm_setzero_ps();
1195 fjz2 = _mm_setzero_ps();
1196 fjx3 = _mm_setzero_ps();
1197 fjy3 = _mm_setzero_ps();
1198 fjz3 = _mm_setzero_ps();
1199
1200 /**************************
1201 * CALCULATE INTERACTIONS *
1202 **************************/
1203
1204 /* LENNARD-JONES DISPERSION/REPULSION */
1205
1206 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1207 fvdw = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1208
1209 fscal = fvdw;
1210
1211 /* Update vectorial force */
1212 fix0 = _mm_macc_ps(dx00,fscal,fix0);
1213 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
1214 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
1215
1216 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
1217 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
1218 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
1219
1220 /**************************
1221 * CALCULATE INTERACTIONS *
1222 **************************/
1223
1224 /* REACTION-FIELD ELECTROSTATICS */
1225 felec = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1226
1227 fscal = felec;
1228
1229 /* Update vectorial force */
1230 fix1 = _mm_macc_ps(dx11,fscal,fix1);
1231 fiy1 = _mm_macc_ps(dy11,fscal,fiy1);
1232 fiz1 = _mm_macc_ps(dz11,fscal,fiz1);
1233
1234 fjx1 = _mm_macc_ps(dx11,fscal,fjx1);
1235 fjy1 = _mm_macc_ps(dy11,fscal,fjy1);
1236 fjz1 = _mm_macc_ps(dz11,fscal,fjz1);
1237
1238 /**************************
1239 * CALCULATE INTERACTIONS *
1240 **************************/
1241
1242 /* REACTION-FIELD ELECTROSTATICS */
1243 felec = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1244
1245 fscal = felec;
1246
1247 /* Update vectorial force */
1248 fix1 = _mm_macc_ps(dx12,fscal,fix1);
1249 fiy1 = _mm_macc_ps(dy12,fscal,fiy1);
1250 fiz1 = _mm_macc_ps(dz12,fscal,fiz1);
1251
1252 fjx2 = _mm_macc_ps(dx12,fscal,fjx2);
1253 fjy2 = _mm_macc_ps(dy12,fscal,fjy2);
1254 fjz2 = _mm_macc_ps(dz12,fscal,fjz2);
1255
1256 /**************************
1257 * CALCULATE INTERACTIONS *
1258 **************************/
1259
1260 /* REACTION-FIELD ELECTROSTATICS */
1261 felec = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1262
1263 fscal = felec;
1264
1265 /* Update vectorial force */
1266 fix1 = _mm_macc_ps(dx13,fscal,fix1);
1267 fiy1 = _mm_macc_ps(dy13,fscal,fiy1);
1268 fiz1 = _mm_macc_ps(dz13,fscal,fiz1);
1269
1270 fjx3 = _mm_macc_ps(dx13,fscal,fjx3);
1271 fjy3 = _mm_macc_ps(dy13,fscal,fjy3);
1272 fjz3 = _mm_macc_ps(dz13,fscal,fjz3);
1273
1274 /**************************
1275 * CALCULATE INTERACTIONS *
1276 **************************/
1277
1278 /* REACTION-FIELD ELECTROSTATICS */
1279 felec = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1280
1281 fscal = felec;
1282
1283 /* Update vectorial force */
1284 fix2 = _mm_macc_ps(dx21,fscal,fix2);
1285 fiy2 = _mm_macc_ps(dy21,fscal,fiy2);
1286 fiz2 = _mm_macc_ps(dz21,fscal,fiz2);
1287
1288 fjx1 = _mm_macc_ps(dx21,fscal,fjx1);
1289 fjy1 = _mm_macc_ps(dy21,fscal,fjy1);
1290 fjz1 = _mm_macc_ps(dz21,fscal,fjz1);
1291
1292 /**************************
1293 * CALCULATE INTERACTIONS *
1294 **************************/
1295
1296 /* REACTION-FIELD ELECTROSTATICS */
1297 felec = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1298
1299 fscal = felec;
1300
1301 /* Update vectorial force */
1302 fix2 = _mm_macc_ps(dx22,fscal,fix2);
1303 fiy2 = _mm_macc_ps(dy22,fscal,fiy2);
1304 fiz2 = _mm_macc_ps(dz22,fscal,fiz2);
1305
1306 fjx2 = _mm_macc_ps(dx22,fscal,fjx2);
1307 fjy2 = _mm_macc_ps(dy22,fscal,fjy2);
1308 fjz2 = _mm_macc_ps(dz22,fscal,fjz2);
1309
1310 /**************************
1311 * CALCULATE INTERACTIONS *
1312 **************************/
1313
1314 /* REACTION-FIELD ELECTROSTATICS */
1315 felec = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1316
1317 fscal = felec;
1318
1319 /* Update vectorial force */
1320 fix2 = _mm_macc_ps(dx23,fscal,fix2);
1321 fiy2 = _mm_macc_ps(dy23,fscal,fiy2);
1322 fiz2 = _mm_macc_ps(dz23,fscal,fiz2);
1323
1324 fjx3 = _mm_macc_ps(dx23,fscal,fjx3);
1325 fjy3 = _mm_macc_ps(dy23,fscal,fjy3);
1326 fjz3 = _mm_macc_ps(dz23,fscal,fjz3);
1327
1328 /**************************
1329 * CALCULATE INTERACTIONS *
1330 **************************/
1331
1332 /* REACTION-FIELD ELECTROSTATICS */
1333 felec = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1334
1335 fscal = felec;
1336
1337 /* Update vectorial force */
1338 fix3 = _mm_macc_ps(dx31,fscal,fix3);
1339 fiy3 = _mm_macc_ps(dy31,fscal,fiy3);
1340 fiz3 = _mm_macc_ps(dz31,fscal,fiz3);
1341
1342 fjx1 = _mm_macc_ps(dx31,fscal,fjx1);
1343 fjy1 = _mm_macc_ps(dy31,fscal,fjy1);
1344 fjz1 = _mm_macc_ps(dz31,fscal,fjz1);
1345
1346 /**************************
1347 * CALCULATE INTERACTIONS *
1348 **************************/
1349
1350 /* REACTION-FIELD ELECTROSTATICS */
1351 felec = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1352
1353 fscal = felec;
1354
1355 /* Update vectorial force */
1356 fix3 = _mm_macc_ps(dx32,fscal,fix3);
1357 fiy3 = _mm_macc_ps(dy32,fscal,fiy3);
1358 fiz3 = _mm_macc_ps(dz32,fscal,fiz3);
1359
1360 fjx2 = _mm_macc_ps(dx32,fscal,fjx2);
1361 fjy2 = _mm_macc_ps(dy32,fscal,fjy2);
1362 fjz2 = _mm_macc_ps(dz32,fscal,fjz2);
1363
1364 /**************************
1365 * CALCULATE INTERACTIONS *
1366 **************************/
1367
1368 /* REACTION-FIELD ELECTROSTATICS */
1369 felec = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1370
1371 fscal = felec;
1372
1373 /* Update vectorial force */
1374 fix3 = _mm_macc_ps(dx33,fscal,fix3);
1375 fiy3 = _mm_macc_ps(dy33,fscal,fiy3);
1376 fiz3 = _mm_macc_ps(dz33,fscal,fiz3);
1377
1378 fjx3 = _mm_macc_ps(dx33,fscal,fjx3);
1379 fjy3 = _mm_macc_ps(dy33,fscal,fjy3);
1380 fjz3 = _mm_macc_ps(dz33,fscal,fjz3);
1381
1382 fjptrA = f+j_coord_offsetA;
1383 fjptrB = f+j_coord_offsetB;
1384 fjptrC = f+j_coord_offsetC;
1385 fjptrD = f+j_coord_offsetD;
1386
1387 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1388 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1389 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1390
1391 /* Inner loop uses 303 flops */
1392 }
1393
1394 if(jidx<j_index_end)
1395 {
1396
1397 /* Get j neighbor index, and coordinate index */
1398 jnrlistA = jjnr[jidx];
1399 jnrlistB = jjnr[jidx+1];
1400 jnrlistC = jjnr[jidx+2];
1401 jnrlistD = jjnr[jidx+3];
1402 /* Sign of each element will be negative for non-real atoms.
1403 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1404 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1405 */
1406 dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1407 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
1408 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
1409 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
1410 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
1411 j_coord_offsetA = DIM*jnrA;
1412 j_coord_offsetB = DIM*jnrB;
1413 j_coord_offsetC = DIM*jnrC;
1414 j_coord_offsetD = DIM*jnrD;
1415
1416 /* load j atom coordinates */
1417 gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1418 x+j_coord_offsetC,x+j_coord_offsetD,
1419 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1420 &jy2,&jz2,&jx3,&jy3,&jz3);
1421
1422 /* Calculate displacement vector */
1423 dx00 = _mm_sub_ps(ix0,jx0);
1424 dy00 = _mm_sub_ps(iy0,jy0);
1425 dz00 = _mm_sub_ps(iz0,jz0);
1426 dx11 = _mm_sub_ps(ix1,jx1);
1427 dy11 = _mm_sub_ps(iy1,jy1);
1428 dz11 = _mm_sub_ps(iz1,jz1);
1429 dx12 = _mm_sub_ps(ix1,jx2);
1430 dy12 = _mm_sub_ps(iy1,jy2);
1431 dz12 = _mm_sub_ps(iz1,jz2);
1432 dx13 = _mm_sub_ps(ix1,jx3);
1433 dy13 = _mm_sub_ps(iy1,jy3);
1434 dz13 = _mm_sub_ps(iz1,jz3);
1435 dx21 = _mm_sub_ps(ix2,jx1);
1436 dy21 = _mm_sub_ps(iy2,jy1);
1437 dz21 = _mm_sub_ps(iz2,jz1);
1438 dx22 = _mm_sub_ps(ix2,jx2);
1439 dy22 = _mm_sub_ps(iy2,jy2);
1440 dz22 = _mm_sub_ps(iz2,jz2);
1441 dx23 = _mm_sub_ps(ix2,jx3);
1442 dy23 = _mm_sub_ps(iy2,jy3);
1443 dz23 = _mm_sub_ps(iz2,jz3);
1444 dx31 = _mm_sub_ps(ix3,jx1);
1445 dy31 = _mm_sub_ps(iy3,jy1);
1446 dz31 = _mm_sub_ps(iz3,jz1);
1447 dx32 = _mm_sub_ps(ix3,jx2);
1448 dy32 = _mm_sub_ps(iy3,jy2);
1449 dz32 = _mm_sub_ps(iz3,jz2);
1450 dx33 = _mm_sub_ps(ix3,jx3);
1451 dy33 = _mm_sub_ps(iy3,jy3);
1452 dz33 = _mm_sub_ps(iz3,jz3);
1453
1454 /* Calculate squared distance and things based on it */
1455 rsq00 = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1456 rsq11 = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1457 rsq12 = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1458 rsq13 = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1459 rsq21 = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1460 rsq22 = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1461 rsq23 = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1462 rsq31 = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1463 rsq32 = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1464 rsq33 = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1465
1466 rinv11 = avx128fma_invsqrt_f(rsq11);
1467 rinv12 = avx128fma_invsqrt_f(rsq12);
1468 rinv13 = avx128fma_invsqrt_f(rsq13);
1469 rinv21 = avx128fma_invsqrt_f(rsq21);
1470 rinv22 = avx128fma_invsqrt_f(rsq22);
1471 rinv23 = avx128fma_invsqrt_f(rsq23);
1472 rinv31 = avx128fma_invsqrt_f(rsq31);
1473 rinv32 = avx128fma_invsqrt_f(rsq32);
1474 rinv33 = avx128fma_invsqrt_f(rsq33);
1475
1476 rinvsq00 = avx128fma_inv_f(rsq00);
1477 rinvsq11 = _mm_mul_ps(rinv11,rinv11);
1478 rinvsq12 = _mm_mul_ps(rinv12,rinv12);
1479 rinvsq13 = _mm_mul_ps(rinv13,rinv13);
1480 rinvsq21 = _mm_mul_ps(rinv21,rinv21);
1481 rinvsq22 = _mm_mul_ps(rinv22,rinv22);
1482 rinvsq23 = _mm_mul_ps(rinv23,rinv23);
1483 rinvsq31 = _mm_mul_ps(rinv31,rinv31);
1484 rinvsq32 = _mm_mul_ps(rinv32,rinv32);
1485 rinvsq33 = _mm_mul_ps(rinv33,rinv33);
1486
1487 fjx0 = _mm_setzero_ps();
1488 fjy0 = _mm_setzero_ps();
1489 fjz0 = _mm_setzero_ps();
1490 fjx1 = _mm_setzero_ps();
1491 fjy1 = _mm_setzero_ps();
1492 fjz1 = _mm_setzero_ps();
1493 fjx2 = _mm_setzero_ps();
1494 fjy2 = _mm_setzero_ps();
1495 fjz2 = _mm_setzero_ps();
1496 fjx3 = _mm_setzero_ps();
1497 fjy3 = _mm_setzero_ps();
1498 fjz3 = _mm_setzero_ps();
1499
1500 /**************************
1501 * CALCULATE INTERACTIONS *
1502 **************************/
1503
1504 /* LENNARD-JONES DISPERSION/REPULSION */
1505
1506 rinvsix = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1507 fvdw = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1508
1509 fscal = fvdw;
1510
1511 fscal = _mm_andnot_ps(dummy_mask,fscal);
1512
1513 /* Update vectorial force */
1514 fix0 = _mm_macc_ps(dx00,fscal,fix0);
1515 fiy0 = _mm_macc_ps(dy00,fscal,fiy0);
1516 fiz0 = _mm_macc_ps(dz00,fscal,fiz0);
1517
1518 fjx0 = _mm_macc_ps(dx00,fscal,fjx0);
1519 fjy0 = _mm_macc_ps(dy00,fscal,fjy0);
1520 fjz0 = _mm_macc_ps(dz00,fscal,fjz0);
1521
1522 /**************************
1523 * CALCULATE INTERACTIONS *
1524 **************************/
1525
1526 /* REACTION-FIELD ELECTROSTATICS */
1527 felec = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1528
1529 fscal = felec;
1530
1531 fscal = _mm_andnot_ps(dummy_mask,fscal);
1532
1533 /* Update vectorial force */
1534 fix1 = _mm_macc_ps(dx11,fscal,fix1);
1535 fiy1 = _mm_macc_ps(dy11,fscal,fiy1);
1536 fiz1 = _mm_macc_ps(dz11,fscal,fiz1);
1537
1538 fjx1 = _mm_macc_ps(dx11,fscal,fjx1);
1539 fjy1 = _mm_macc_ps(dy11,fscal,fjy1);
1540 fjz1 = _mm_macc_ps(dz11,fscal,fjz1);
1541
1542 /**************************
1543 * CALCULATE INTERACTIONS *
1544 **************************/
1545
1546 /* REACTION-FIELD ELECTROSTATICS */
1547 felec = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1548
1549 fscal = felec;
1550
1551 fscal = _mm_andnot_ps(dummy_mask,fscal);
1552
1553 /* Update vectorial force */
1554 fix1 = _mm_macc_ps(dx12,fscal,fix1);
1555 fiy1 = _mm_macc_ps(dy12,fscal,fiy1);
1556 fiz1 = _mm_macc_ps(dz12,fscal,fiz1);
1557
1558 fjx2 = _mm_macc_ps(dx12,fscal,fjx2);
1559 fjy2 = _mm_macc_ps(dy12,fscal,fjy2);
1560 fjz2 = _mm_macc_ps(dz12,fscal,fjz2);
1561
1562 /**************************
1563 * CALCULATE INTERACTIONS *
1564 **************************/
1565
1566 /* REACTION-FIELD ELECTROSTATICS */
1567 felec = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1568
1569 fscal = felec;
1570
1571 fscal = _mm_andnot_ps(dummy_mask,fscal);
1572
1573 /* Update vectorial force */
1574 fix1 = _mm_macc_ps(dx13,fscal,fix1);
1575 fiy1 = _mm_macc_ps(dy13,fscal,fiy1);
1576 fiz1 = _mm_macc_ps(dz13,fscal,fiz1);
1577
1578 fjx3 = _mm_macc_ps(dx13,fscal,fjx3);
1579 fjy3 = _mm_macc_ps(dy13,fscal,fjy3);
1580 fjz3 = _mm_macc_ps(dz13,fscal,fjz3);
1581
1582 /**************************
1583 * CALCULATE INTERACTIONS *
1584 **************************/
1585
1586 /* REACTION-FIELD ELECTROSTATICS */
1587 felec = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1588
1589 fscal = felec;
1590
1591 fscal = _mm_andnot_ps(dummy_mask,fscal);
1592
1593 /* Update vectorial force */
1594 fix2 = _mm_macc_ps(dx21,fscal,fix2);
1595 fiy2 = _mm_macc_ps(dy21,fscal,fiy2);
1596 fiz2 = _mm_macc_ps(dz21,fscal,fiz2);
1597
1598 fjx1 = _mm_macc_ps(dx21,fscal,fjx1);
1599 fjy1 = _mm_macc_ps(dy21,fscal,fjy1);
1600 fjz1 = _mm_macc_ps(dz21,fscal,fjz1);
1601
1602 /**************************
1603 * CALCULATE INTERACTIONS *
1604 **************************/
1605
1606 /* REACTION-FIELD ELECTROSTATICS */
1607 felec = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1608
1609 fscal = felec;
1610
1611 fscal = _mm_andnot_ps(dummy_mask,fscal);
1612
1613 /* Update vectorial force */
1614 fix2 = _mm_macc_ps(dx22,fscal,fix2);
1615 fiy2 = _mm_macc_ps(dy22,fscal,fiy2);
1616 fiz2 = _mm_macc_ps(dz22,fscal,fiz2);
1617
1618 fjx2 = _mm_macc_ps(dx22,fscal,fjx2);
1619 fjy2 = _mm_macc_ps(dy22,fscal,fjy2);
1620 fjz2 = _mm_macc_ps(dz22,fscal,fjz2);
1621
1622 /**************************
1623 * CALCULATE INTERACTIONS *
1624 **************************/
1625
1626 /* REACTION-FIELD ELECTROSTATICS */
1627 felec = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1628
1629 fscal = felec;
1630
1631 fscal = _mm_andnot_ps(dummy_mask,fscal);
1632
1633 /* Update vectorial force */
1634 fix2 = _mm_macc_ps(dx23,fscal,fix2);
1635 fiy2 = _mm_macc_ps(dy23,fscal,fiy2);
1636 fiz2 = _mm_macc_ps(dz23,fscal,fiz2);
1637
1638 fjx3 = _mm_macc_ps(dx23,fscal,fjx3);
1639 fjy3 = _mm_macc_ps(dy23,fscal,fjy3);
1640 fjz3 = _mm_macc_ps(dz23,fscal,fjz3);
1641
1642 /**************************
1643 * CALCULATE INTERACTIONS *
1644 **************************/
1645
1646 /* REACTION-FIELD ELECTROSTATICS */
1647 felec = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1648
1649 fscal = felec;
1650
1651 fscal = _mm_andnot_ps(dummy_mask,fscal);
1652
1653 /* Update vectorial force */
1654 fix3 = _mm_macc_ps(dx31,fscal,fix3);
1655 fiy3 = _mm_macc_ps(dy31,fscal,fiy3);
1656 fiz3 = _mm_macc_ps(dz31,fscal,fiz3);
1657
1658 fjx1 = _mm_macc_ps(dx31,fscal,fjx1);
1659 fjy1 = _mm_macc_ps(dy31,fscal,fjy1);
1660 fjz1 = _mm_macc_ps(dz31,fscal,fjz1);
1661
1662 /**************************
1663 * CALCULATE INTERACTIONS *
1664 **************************/
1665
1666 /* REACTION-FIELD ELECTROSTATICS */
1667 felec = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1668
1669 fscal = felec;
1670
1671 fscal = _mm_andnot_ps(dummy_mask,fscal);
1672
1673 /* Update vectorial force */
1674 fix3 = _mm_macc_ps(dx32,fscal,fix3);
1675 fiy3 = _mm_macc_ps(dy32,fscal,fiy3);
1676 fiz3 = _mm_macc_ps(dz32,fscal,fiz3);
1677
1678 fjx2 = _mm_macc_ps(dx32,fscal,fjx2);
1679 fjy2 = _mm_macc_ps(dy32,fscal,fjy2);
1680 fjz2 = _mm_macc_ps(dz32,fscal,fjz2);
1681
1682 /**************************
1683 * CALCULATE INTERACTIONS *
1684 **************************/
1685
1686 /* REACTION-FIELD ELECTROSTATICS */
1687 felec = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1688
1689 fscal = felec;
1690
1691 fscal = _mm_andnot_ps(dummy_mask,fscal);
1692
1693 /* Update vectorial force */
1694 fix3 = _mm_macc_ps(dx33,fscal,fix3);
1695 fiy3 = _mm_macc_ps(dy33,fscal,fiy3);
1696 fiz3 = _mm_macc_ps(dz33,fscal,fiz3);
1697
1698 fjx3 = _mm_macc_ps(dx33,fscal,fjx3);
1699 fjy3 = _mm_macc_ps(dy33,fscal,fjy3);
1700 fjz3 = _mm_macc_ps(dz33,fscal,fjz3);
1701
1702 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1703 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1704 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1705 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1706
1707 gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1708 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1709 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1710
1711 /* Inner loop uses 303 flops */
1712 }
1713
1714 /* End of innermost loop */
1715
1716 gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1717 f+i_coord_offset,fshift+i_shift_offset);
1718
1719 /* Increment number of inner iterations */
1720 inneriter += j_index_end - j_index_start;
1721
1722 /* Outer loop uses 24 flops */
1723 }
1724
1725 /* Increment number of outer iterations */
1726 outeriter += nri;
1727
1728 /* Update outer/inner flops */
1729
1730 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
1731 }
The ultimate molecular dynamics simulation package