2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, 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.
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.
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.
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.
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.
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.
36 * Note: this file was generated by the GROMACS avx_256_single kernel generator.
42 #include "../nb_kernel.h"
43 #include "types/simple.h"
44 #include "gromacs/math/vec.h"
47 #include "gromacs/simd/math_x86_avx_256_single.h"
48 #include "kernelutil_x86_avx_256_single.h"
51 * Gromacs nonbonded kernel: nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_256_single
52 * Electrostatics interaction: None
53 * VdW interaction: CubicSplineTable
54 * Geometry: Particle-Particle
55 * Calculate force/pot: PotentialAndForce
58 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_256_single
59 (t_nblist * gmx_restrict nlist,
60 rvec * gmx_restrict xx,
61 rvec * gmx_restrict ff,
62 t_forcerec * gmx_restrict fr,
63 t_mdatoms * gmx_restrict mdatoms,
64 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
65 t_nrnb * gmx_restrict nrnb)
67 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
68 * just 0 for non-waters.
69 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
70 * jnr indices corresponding to data put in the four positions in the SIMD register.
72 int i_shift_offset,i_coord_offset,outeriter,inneriter;
73 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
74 int jnrA,jnrB,jnrC,jnrD;
75 int jnrE,jnrF,jnrG,jnrH;
76 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
77 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
78 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
79 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
80 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
82 real *shiftvec,*fshift,*x,*f;
83 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
85 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
86 real * vdwioffsetptr0;
87 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
88 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
89 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
90 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
92 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
95 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
96 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
98 __m128i vfitab_lo,vfitab_hi;
99 __m128i ifour = _mm_set1_epi32(4);
100 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
102 __m256 dummy_mask,cutoff_mask;
103 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
104 __m256 one = _mm256_set1_ps(1.0);
105 __m256 two = _mm256_set1_ps(2.0);
111 jindex = nlist->jindex;
113 shiftidx = nlist->shift;
115 shiftvec = fr->shift_vec[0];
116 fshift = fr->fshift[0];
117 nvdwtype = fr->ntype;
119 vdwtype = mdatoms->typeA;
121 vftab = kernel_data->table_vdw->data;
122 vftabscale = _mm256_set1_ps(kernel_data->table_vdw->scale);
124 /* Avoid stupid compiler warnings */
125 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
138 for(iidx=0;iidx<4*DIM;iidx++)
143 /* Start outer loop over neighborlists */
144 for(iidx=0; iidx<nri; iidx++)
146 /* Load shift vector for this list */
147 i_shift_offset = DIM*shiftidx[iidx];
149 /* Load limits for loop over neighbors */
150 j_index_start = jindex[iidx];
151 j_index_end = jindex[iidx+1];
153 /* Get outer coordinate index */
155 i_coord_offset = DIM*inr;
157 /* Load i particle coords and add shift vector */
158 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
160 fix0 = _mm256_setzero_ps();
161 fiy0 = _mm256_setzero_ps();
162 fiz0 = _mm256_setzero_ps();
164 /* Load parameters for i particles */
165 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
167 /* Reset potential sums */
168 vvdwsum = _mm256_setzero_ps();
170 /* Start inner kernel loop */
171 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
174 /* Get j neighbor index, and coordinate index */
183 j_coord_offsetA = DIM*jnrA;
184 j_coord_offsetB = DIM*jnrB;
185 j_coord_offsetC = DIM*jnrC;
186 j_coord_offsetD = DIM*jnrD;
187 j_coord_offsetE = DIM*jnrE;
188 j_coord_offsetF = DIM*jnrF;
189 j_coord_offsetG = DIM*jnrG;
190 j_coord_offsetH = DIM*jnrH;
192 /* load j atom coordinates */
193 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
194 x+j_coord_offsetC,x+j_coord_offsetD,
195 x+j_coord_offsetE,x+j_coord_offsetF,
196 x+j_coord_offsetG,x+j_coord_offsetH,
199 /* Calculate displacement vector */
200 dx00 = _mm256_sub_ps(ix0,jx0);
201 dy00 = _mm256_sub_ps(iy0,jy0);
202 dz00 = _mm256_sub_ps(iz0,jz0);
204 /* Calculate squared distance and things based on it */
205 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
207 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
209 /* Load parameters for j particles */
210 vdwjidx0A = 2*vdwtype[jnrA+0];
211 vdwjidx0B = 2*vdwtype[jnrB+0];
212 vdwjidx0C = 2*vdwtype[jnrC+0];
213 vdwjidx0D = 2*vdwtype[jnrD+0];
214 vdwjidx0E = 2*vdwtype[jnrE+0];
215 vdwjidx0F = 2*vdwtype[jnrF+0];
216 vdwjidx0G = 2*vdwtype[jnrG+0];
217 vdwjidx0H = 2*vdwtype[jnrH+0];
219 /**************************
220 * CALCULATE INTERACTIONS *
221 **************************/
223 r00 = _mm256_mul_ps(rsq00,rinv00);
225 /* Compute parameters for interactions between i and j atoms */
226 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
227 vdwioffsetptr0+vdwjidx0B,
228 vdwioffsetptr0+vdwjidx0C,
229 vdwioffsetptr0+vdwjidx0D,
230 vdwioffsetptr0+vdwjidx0E,
231 vdwioffsetptr0+vdwjidx0F,
232 vdwioffsetptr0+vdwjidx0G,
233 vdwioffsetptr0+vdwjidx0H,
236 /* Calculate table index by multiplying r with table scale and truncate to integer */
237 rt = _mm256_mul_ps(r00,vftabscale);
238 vfitab = _mm256_cvttps_epi32(rt);
239 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
240 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
241 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
242 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
243 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
244 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
246 /* CUBIC SPLINE TABLE DISPERSION */
247 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
248 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
249 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
250 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
251 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
252 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
253 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
254 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
255 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
256 Heps = _mm256_mul_ps(vfeps,H);
257 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
258 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
259 vvdw6 = _mm256_mul_ps(c6_00,VV);
260 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
261 fvdw6 = _mm256_mul_ps(c6_00,FF);
263 /* CUBIC SPLINE TABLE REPULSION */
264 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
265 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
266 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
267 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
268 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
269 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
270 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
271 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
272 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
273 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
274 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
275 Heps = _mm256_mul_ps(vfeps,H);
276 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
277 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
278 vvdw12 = _mm256_mul_ps(c12_00,VV);
279 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
280 fvdw12 = _mm256_mul_ps(c12_00,FF);
281 vvdw = _mm256_add_ps(vvdw12,vvdw6);
282 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
284 /* Update potential sum for this i atom from the interaction with this j atom. */
285 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
289 /* Calculate temporary vectorial force */
290 tx = _mm256_mul_ps(fscal,dx00);
291 ty = _mm256_mul_ps(fscal,dy00);
292 tz = _mm256_mul_ps(fscal,dz00);
294 /* Update vectorial force */
295 fix0 = _mm256_add_ps(fix0,tx);
296 fiy0 = _mm256_add_ps(fiy0,ty);
297 fiz0 = _mm256_add_ps(fiz0,tz);
299 fjptrA = f+j_coord_offsetA;
300 fjptrB = f+j_coord_offsetB;
301 fjptrC = f+j_coord_offsetC;
302 fjptrD = f+j_coord_offsetD;
303 fjptrE = f+j_coord_offsetE;
304 fjptrF = f+j_coord_offsetF;
305 fjptrG = f+j_coord_offsetG;
306 fjptrH = f+j_coord_offsetH;
307 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
309 /* Inner loop uses 56 flops */
315 /* Get j neighbor index, and coordinate index */
316 jnrlistA = jjnr[jidx];
317 jnrlistB = jjnr[jidx+1];
318 jnrlistC = jjnr[jidx+2];
319 jnrlistD = jjnr[jidx+3];
320 jnrlistE = jjnr[jidx+4];
321 jnrlistF = jjnr[jidx+5];
322 jnrlistG = jjnr[jidx+6];
323 jnrlistH = jjnr[jidx+7];
324 /* Sign of each element will be negative for non-real atoms.
325 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
326 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
328 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
329 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
331 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
332 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
333 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
334 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
335 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
336 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
337 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
338 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
339 j_coord_offsetA = DIM*jnrA;
340 j_coord_offsetB = DIM*jnrB;
341 j_coord_offsetC = DIM*jnrC;
342 j_coord_offsetD = DIM*jnrD;
343 j_coord_offsetE = DIM*jnrE;
344 j_coord_offsetF = DIM*jnrF;
345 j_coord_offsetG = DIM*jnrG;
346 j_coord_offsetH = DIM*jnrH;
348 /* load j atom coordinates */
349 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
350 x+j_coord_offsetC,x+j_coord_offsetD,
351 x+j_coord_offsetE,x+j_coord_offsetF,
352 x+j_coord_offsetG,x+j_coord_offsetH,
355 /* Calculate displacement vector */
356 dx00 = _mm256_sub_ps(ix0,jx0);
357 dy00 = _mm256_sub_ps(iy0,jy0);
358 dz00 = _mm256_sub_ps(iz0,jz0);
360 /* Calculate squared distance and things based on it */
361 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
363 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
365 /* Load parameters for j particles */
366 vdwjidx0A = 2*vdwtype[jnrA+0];
367 vdwjidx0B = 2*vdwtype[jnrB+0];
368 vdwjidx0C = 2*vdwtype[jnrC+0];
369 vdwjidx0D = 2*vdwtype[jnrD+0];
370 vdwjidx0E = 2*vdwtype[jnrE+0];
371 vdwjidx0F = 2*vdwtype[jnrF+0];
372 vdwjidx0G = 2*vdwtype[jnrG+0];
373 vdwjidx0H = 2*vdwtype[jnrH+0];
375 /**************************
376 * CALCULATE INTERACTIONS *
377 **************************/
379 r00 = _mm256_mul_ps(rsq00,rinv00);
380 r00 = _mm256_andnot_ps(dummy_mask,r00);
382 /* Compute parameters for interactions between i and j atoms */
383 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
384 vdwioffsetptr0+vdwjidx0B,
385 vdwioffsetptr0+vdwjidx0C,
386 vdwioffsetptr0+vdwjidx0D,
387 vdwioffsetptr0+vdwjidx0E,
388 vdwioffsetptr0+vdwjidx0F,
389 vdwioffsetptr0+vdwjidx0G,
390 vdwioffsetptr0+vdwjidx0H,
393 /* Calculate table index by multiplying r with table scale and truncate to integer */
394 rt = _mm256_mul_ps(r00,vftabscale);
395 vfitab = _mm256_cvttps_epi32(rt);
396 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
397 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
398 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
399 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
400 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
401 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
403 /* CUBIC SPLINE TABLE DISPERSION */
404 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
405 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
406 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
407 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
408 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
409 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
410 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
411 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
412 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
413 Heps = _mm256_mul_ps(vfeps,H);
414 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
415 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
416 vvdw6 = _mm256_mul_ps(c6_00,VV);
417 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
418 fvdw6 = _mm256_mul_ps(c6_00,FF);
420 /* CUBIC SPLINE TABLE REPULSION */
421 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
422 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
423 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
424 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
425 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
426 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
427 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
428 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
429 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
430 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
431 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
432 Heps = _mm256_mul_ps(vfeps,H);
433 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
434 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
435 vvdw12 = _mm256_mul_ps(c12_00,VV);
436 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
437 fvdw12 = _mm256_mul_ps(c12_00,FF);
438 vvdw = _mm256_add_ps(vvdw12,vvdw6);
439 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
441 /* Update potential sum for this i atom from the interaction with this j atom. */
442 vvdw = _mm256_andnot_ps(dummy_mask,vvdw);
443 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
447 fscal = _mm256_andnot_ps(dummy_mask,fscal);
449 /* Calculate temporary vectorial force */
450 tx = _mm256_mul_ps(fscal,dx00);
451 ty = _mm256_mul_ps(fscal,dy00);
452 tz = _mm256_mul_ps(fscal,dz00);
454 /* Update vectorial force */
455 fix0 = _mm256_add_ps(fix0,tx);
456 fiy0 = _mm256_add_ps(fiy0,ty);
457 fiz0 = _mm256_add_ps(fiz0,tz);
459 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
460 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
461 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
462 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
463 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
464 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
465 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
466 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
467 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
469 /* Inner loop uses 57 flops */
472 /* End of innermost loop */
474 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
475 f+i_coord_offset,fshift+i_shift_offset);
478 /* Update potential energies */
479 gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
481 /* Increment number of inner iterations */
482 inneriter += j_index_end - j_index_start;
484 /* Outer loop uses 7 flops */
487 /* Increment number of outer iterations */
490 /* Update outer/inner flops */
492 inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*57);
495 * Gromacs nonbonded kernel: nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_single
496 * Electrostatics interaction: None
497 * VdW interaction: CubicSplineTable
498 * Geometry: Particle-Particle
499 * Calculate force/pot: Force
502 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_single
503 (t_nblist * gmx_restrict nlist,
504 rvec * gmx_restrict xx,
505 rvec * gmx_restrict ff,
506 t_forcerec * gmx_restrict fr,
507 t_mdatoms * gmx_restrict mdatoms,
508 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
509 t_nrnb * gmx_restrict nrnb)
511 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
512 * just 0 for non-waters.
513 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
514 * jnr indices corresponding to data put in the four positions in the SIMD register.
516 int i_shift_offset,i_coord_offset,outeriter,inneriter;
517 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
518 int jnrA,jnrB,jnrC,jnrD;
519 int jnrE,jnrF,jnrG,jnrH;
520 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
521 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
522 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
523 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
524 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
526 real *shiftvec,*fshift,*x,*f;
527 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
529 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
530 real * vdwioffsetptr0;
531 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
532 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
533 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
534 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
536 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
539 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
540 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
542 __m128i vfitab_lo,vfitab_hi;
543 __m128i ifour = _mm_set1_epi32(4);
544 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
546 __m256 dummy_mask,cutoff_mask;
547 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
548 __m256 one = _mm256_set1_ps(1.0);
549 __m256 two = _mm256_set1_ps(2.0);
555 jindex = nlist->jindex;
557 shiftidx = nlist->shift;
559 shiftvec = fr->shift_vec[0];
560 fshift = fr->fshift[0];
561 nvdwtype = fr->ntype;
563 vdwtype = mdatoms->typeA;
565 vftab = kernel_data->table_vdw->data;
566 vftabscale = _mm256_set1_ps(kernel_data->table_vdw->scale);
568 /* Avoid stupid compiler warnings */
569 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
582 for(iidx=0;iidx<4*DIM;iidx++)
587 /* Start outer loop over neighborlists */
588 for(iidx=0; iidx<nri; iidx++)
590 /* Load shift vector for this list */
591 i_shift_offset = DIM*shiftidx[iidx];
593 /* Load limits for loop over neighbors */
594 j_index_start = jindex[iidx];
595 j_index_end = jindex[iidx+1];
597 /* Get outer coordinate index */
599 i_coord_offset = DIM*inr;
601 /* Load i particle coords and add shift vector */
602 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
604 fix0 = _mm256_setzero_ps();
605 fiy0 = _mm256_setzero_ps();
606 fiz0 = _mm256_setzero_ps();
608 /* Load parameters for i particles */
609 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
611 /* Start inner kernel loop */
612 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
615 /* Get j neighbor index, and coordinate index */
624 j_coord_offsetA = DIM*jnrA;
625 j_coord_offsetB = DIM*jnrB;
626 j_coord_offsetC = DIM*jnrC;
627 j_coord_offsetD = DIM*jnrD;
628 j_coord_offsetE = DIM*jnrE;
629 j_coord_offsetF = DIM*jnrF;
630 j_coord_offsetG = DIM*jnrG;
631 j_coord_offsetH = DIM*jnrH;
633 /* load j atom coordinates */
634 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
635 x+j_coord_offsetC,x+j_coord_offsetD,
636 x+j_coord_offsetE,x+j_coord_offsetF,
637 x+j_coord_offsetG,x+j_coord_offsetH,
640 /* Calculate displacement vector */
641 dx00 = _mm256_sub_ps(ix0,jx0);
642 dy00 = _mm256_sub_ps(iy0,jy0);
643 dz00 = _mm256_sub_ps(iz0,jz0);
645 /* Calculate squared distance and things based on it */
646 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
648 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
650 /* Load parameters for j particles */
651 vdwjidx0A = 2*vdwtype[jnrA+0];
652 vdwjidx0B = 2*vdwtype[jnrB+0];
653 vdwjidx0C = 2*vdwtype[jnrC+0];
654 vdwjidx0D = 2*vdwtype[jnrD+0];
655 vdwjidx0E = 2*vdwtype[jnrE+0];
656 vdwjidx0F = 2*vdwtype[jnrF+0];
657 vdwjidx0G = 2*vdwtype[jnrG+0];
658 vdwjidx0H = 2*vdwtype[jnrH+0];
660 /**************************
661 * CALCULATE INTERACTIONS *
662 **************************/
664 r00 = _mm256_mul_ps(rsq00,rinv00);
666 /* Compute parameters for interactions between i and j atoms */
667 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
668 vdwioffsetptr0+vdwjidx0B,
669 vdwioffsetptr0+vdwjidx0C,
670 vdwioffsetptr0+vdwjidx0D,
671 vdwioffsetptr0+vdwjidx0E,
672 vdwioffsetptr0+vdwjidx0F,
673 vdwioffsetptr0+vdwjidx0G,
674 vdwioffsetptr0+vdwjidx0H,
677 /* Calculate table index by multiplying r with table scale and truncate to integer */
678 rt = _mm256_mul_ps(r00,vftabscale);
679 vfitab = _mm256_cvttps_epi32(rt);
680 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
681 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
682 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
683 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
684 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
685 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
687 /* CUBIC SPLINE TABLE DISPERSION */
688 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
689 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
690 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
691 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
692 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
693 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
694 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
695 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
696 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
697 Heps = _mm256_mul_ps(vfeps,H);
698 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
699 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
700 fvdw6 = _mm256_mul_ps(c6_00,FF);
702 /* CUBIC SPLINE TABLE REPULSION */
703 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
704 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
705 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
706 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
707 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
708 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
709 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
710 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
711 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
712 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
713 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
714 Heps = _mm256_mul_ps(vfeps,H);
715 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
716 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
717 fvdw12 = _mm256_mul_ps(c12_00,FF);
718 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
722 /* Calculate temporary vectorial force */
723 tx = _mm256_mul_ps(fscal,dx00);
724 ty = _mm256_mul_ps(fscal,dy00);
725 tz = _mm256_mul_ps(fscal,dz00);
727 /* Update vectorial force */
728 fix0 = _mm256_add_ps(fix0,tx);
729 fiy0 = _mm256_add_ps(fiy0,ty);
730 fiz0 = _mm256_add_ps(fiz0,tz);
732 fjptrA = f+j_coord_offsetA;
733 fjptrB = f+j_coord_offsetB;
734 fjptrC = f+j_coord_offsetC;
735 fjptrD = f+j_coord_offsetD;
736 fjptrE = f+j_coord_offsetE;
737 fjptrF = f+j_coord_offsetF;
738 fjptrG = f+j_coord_offsetG;
739 fjptrH = f+j_coord_offsetH;
740 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
742 /* Inner loop uses 48 flops */
748 /* Get j neighbor index, and coordinate index */
749 jnrlistA = jjnr[jidx];
750 jnrlistB = jjnr[jidx+1];
751 jnrlistC = jjnr[jidx+2];
752 jnrlistD = jjnr[jidx+3];
753 jnrlistE = jjnr[jidx+4];
754 jnrlistF = jjnr[jidx+5];
755 jnrlistG = jjnr[jidx+6];
756 jnrlistH = jjnr[jidx+7];
757 /* Sign of each element will be negative for non-real atoms.
758 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
759 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
761 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
762 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
764 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
765 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
766 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
767 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
768 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
769 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
770 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
771 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
772 j_coord_offsetA = DIM*jnrA;
773 j_coord_offsetB = DIM*jnrB;
774 j_coord_offsetC = DIM*jnrC;
775 j_coord_offsetD = DIM*jnrD;
776 j_coord_offsetE = DIM*jnrE;
777 j_coord_offsetF = DIM*jnrF;
778 j_coord_offsetG = DIM*jnrG;
779 j_coord_offsetH = DIM*jnrH;
781 /* load j atom coordinates */
782 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
783 x+j_coord_offsetC,x+j_coord_offsetD,
784 x+j_coord_offsetE,x+j_coord_offsetF,
785 x+j_coord_offsetG,x+j_coord_offsetH,
788 /* Calculate displacement vector */
789 dx00 = _mm256_sub_ps(ix0,jx0);
790 dy00 = _mm256_sub_ps(iy0,jy0);
791 dz00 = _mm256_sub_ps(iz0,jz0);
793 /* Calculate squared distance and things based on it */
794 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
796 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
798 /* Load parameters for j particles */
799 vdwjidx0A = 2*vdwtype[jnrA+0];
800 vdwjidx0B = 2*vdwtype[jnrB+0];
801 vdwjidx0C = 2*vdwtype[jnrC+0];
802 vdwjidx0D = 2*vdwtype[jnrD+0];
803 vdwjidx0E = 2*vdwtype[jnrE+0];
804 vdwjidx0F = 2*vdwtype[jnrF+0];
805 vdwjidx0G = 2*vdwtype[jnrG+0];
806 vdwjidx0H = 2*vdwtype[jnrH+0];
808 /**************************
809 * CALCULATE INTERACTIONS *
810 **************************/
812 r00 = _mm256_mul_ps(rsq00,rinv00);
813 r00 = _mm256_andnot_ps(dummy_mask,r00);
815 /* Compute parameters for interactions between i and j atoms */
816 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
817 vdwioffsetptr0+vdwjidx0B,
818 vdwioffsetptr0+vdwjidx0C,
819 vdwioffsetptr0+vdwjidx0D,
820 vdwioffsetptr0+vdwjidx0E,
821 vdwioffsetptr0+vdwjidx0F,
822 vdwioffsetptr0+vdwjidx0G,
823 vdwioffsetptr0+vdwjidx0H,
826 /* Calculate table index by multiplying r with table scale and truncate to integer */
827 rt = _mm256_mul_ps(r00,vftabscale);
828 vfitab = _mm256_cvttps_epi32(rt);
829 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
830 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
831 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
832 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
833 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
834 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
836 /* CUBIC SPLINE TABLE DISPERSION */
837 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
838 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
839 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
840 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
841 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
842 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
843 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
844 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
845 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
846 Heps = _mm256_mul_ps(vfeps,H);
847 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
848 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
849 fvdw6 = _mm256_mul_ps(c6_00,FF);
851 /* CUBIC SPLINE TABLE REPULSION */
852 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
853 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
854 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
855 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
856 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
857 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
858 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
859 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
860 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
861 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
862 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
863 Heps = _mm256_mul_ps(vfeps,H);
864 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
865 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
866 fvdw12 = _mm256_mul_ps(c12_00,FF);
867 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
871 fscal = _mm256_andnot_ps(dummy_mask,fscal);
873 /* Calculate temporary vectorial force */
874 tx = _mm256_mul_ps(fscal,dx00);
875 ty = _mm256_mul_ps(fscal,dy00);
876 tz = _mm256_mul_ps(fscal,dz00);
878 /* Update vectorial force */
879 fix0 = _mm256_add_ps(fix0,tx);
880 fiy0 = _mm256_add_ps(fiy0,ty);
881 fiz0 = _mm256_add_ps(fiz0,tz);
883 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
884 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
885 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
886 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
887 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
888 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
889 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
890 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
891 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
893 /* Inner loop uses 49 flops */
896 /* End of innermost loop */
898 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
899 f+i_coord_offset,fshift+i_shift_offset);
901 /* Increment number of inner iterations */
902 inneriter += j_index_end - j_index_start;
904 /* Outer loop uses 6 flops */
907 /* Increment number of outer iterations */
910 /* Update outer/inner flops */
912 inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*49);