2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013, 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.
44 #include "../nb_kernel.h"
45 #include "types/simple.h"
49 #include "gromacs/simd/math_x86_avx_256_single.h"
50 #include "kernelutil_x86_avx_256_single.h"
53 * Gromacs nonbonded kernel: nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_256_single
54 * Electrostatics interaction: None
55 * VdW interaction: CubicSplineTable
56 * Geometry: Particle-Particle
57 * Calculate force/pot: PotentialAndForce
60 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_256_single
61 (t_nblist * gmx_restrict nlist,
62 rvec * gmx_restrict xx,
63 rvec * gmx_restrict ff,
64 t_forcerec * gmx_restrict fr,
65 t_mdatoms * gmx_restrict mdatoms,
66 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
67 t_nrnb * gmx_restrict nrnb)
69 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
70 * just 0 for non-waters.
71 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
72 * jnr indices corresponding to data put in the four positions in the SIMD register.
74 int i_shift_offset,i_coord_offset,outeriter,inneriter;
75 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
76 int jnrA,jnrB,jnrC,jnrD;
77 int jnrE,jnrF,jnrG,jnrH;
78 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
79 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
80 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
81 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
82 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
84 real *shiftvec,*fshift,*x,*f;
85 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
87 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
88 real * vdwioffsetptr0;
89 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
90 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
91 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
92 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
94 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
97 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
98 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
100 __m128i vfitab_lo,vfitab_hi;
101 __m128i ifour = _mm_set1_epi32(4);
102 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
104 __m256 dummy_mask,cutoff_mask;
105 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
106 __m256 one = _mm256_set1_ps(1.0);
107 __m256 two = _mm256_set1_ps(2.0);
113 jindex = nlist->jindex;
115 shiftidx = nlist->shift;
117 shiftvec = fr->shift_vec[0];
118 fshift = fr->fshift[0];
119 nvdwtype = fr->ntype;
121 vdwtype = mdatoms->typeA;
123 vftab = kernel_data->table_vdw->data;
124 vftabscale = _mm256_set1_ps(kernel_data->table_vdw->scale);
126 /* Avoid stupid compiler warnings */
127 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
140 for(iidx=0;iidx<4*DIM;iidx++)
145 /* Start outer loop over neighborlists */
146 for(iidx=0; iidx<nri; iidx++)
148 /* Load shift vector for this list */
149 i_shift_offset = DIM*shiftidx[iidx];
151 /* Load limits for loop over neighbors */
152 j_index_start = jindex[iidx];
153 j_index_end = jindex[iidx+1];
155 /* Get outer coordinate index */
157 i_coord_offset = DIM*inr;
159 /* Load i particle coords and add shift vector */
160 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
162 fix0 = _mm256_setzero_ps();
163 fiy0 = _mm256_setzero_ps();
164 fiz0 = _mm256_setzero_ps();
166 /* Load parameters for i particles */
167 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
169 /* Reset potential sums */
170 vvdwsum = _mm256_setzero_ps();
172 /* Start inner kernel loop */
173 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
176 /* Get j neighbor index, and coordinate index */
185 j_coord_offsetA = DIM*jnrA;
186 j_coord_offsetB = DIM*jnrB;
187 j_coord_offsetC = DIM*jnrC;
188 j_coord_offsetD = DIM*jnrD;
189 j_coord_offsetE = DIM*jnrE;
190 j_coord_offsetF = DIM*jnrF;
191 j_coord_offsetG = DIM*jnrG;
192 j_coord_offsetH = DIM*jnrH;
194 /* load j atom coordinates */
195 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
196 x+j_coord_offsetC,x+j_coord_offsetD,
197 x+j_coord_offsetE,x+j_coord_offsetF,
198 x+j_coord_offsetG,x+j_coord_offsetH,
201 /* Calculate displacement vector */
202 dx00 = _mm256_sub_ps(ix0,jx0);
203 dy00 = _mm256_sub_ps(iy0,jy0);
204 dz00 = _mm256_sub_ps(iz0,jz0);
206 /* Calculate squared distance and things based on it */
207 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
209 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
211 /* Load parameters for j particles */
212 vdwjidx0A = 2*vdwtype[jnrA+0];
213 vdwjidx0B = 2*vdwtype[jnrB+0];
214 vdwjidx0C = 2*vdwtype[jnrC+0];
215 vdwjidx0D = 2*vdwtype[jnrD+0];
216 vdwjidx0E = 2*vdwtype[jnrE+0];
217 vdwjidx0F = 2*vdwtype[jnrF+0];
218 vdwjidx0G = 2*vdwtype[jnrG+0];
219 vdwjidx0H = 2*vdwtype[jnrH+0];
221 /**************************
222 * CALCULATE INTERACTIONS *
223 **************************/
225 r00 = _mm256_mul_ps(rsq00,rinv00);
227 /* Compute parameters for interactions between i and j atoms */
228 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
229 vdwioffsetptr0+vdwjidx0B,
230 vdwioffsetptr0+vdwjidx0C,
231 vdwioffsetptr0+vdwjidx0D,
232 vdwioffsetptr0+vdwjidx0E,
233 vdwioffsetptr0+vdwjidx0F,
234 vdwioffsetptr0+vdwjidx0G,
235 vdwioffsetptr0+vdwjidx0H,
238 /* Calculate table index by multiplying r with table scale and truncate to integer */
239 rt = _mm256_mul_ps(r00,vftabscale);
240 vfitab = _mm256_cvttps_epi32(rt);
241 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
242 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
243 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
244 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
245 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
246 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
248 /* CUBIC SPLINE TABLE DISPERSION */
249 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
250 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
251 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
252 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
253 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
254 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
255 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
256 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
257 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
258 Heps = _mm256_mul_ps(vfeps,H);
259 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
260 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
261 vvdw6 = _mm256_mul_ps(c6_00,VV);
262 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
263 fvdw6 = _mm256_mul_ps(c6_00,FF);
265 /* CUBIC SPLINE TABLE REPULSION */
266 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
267 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
268 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
269 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
270 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
271 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
272 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
273 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
274 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
275 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
276 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
277 Heps = _mm256_mul_ps(vfeps,H);
278 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
279 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
280 vvdw12 = _mm256_mul_ps(c12_00,VV);
281 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
282 fvdw12 = _mm256_mul_ps(c12_00,FF);
283 vvdw = _mm256_add_ps(vvdw12,vvdw6);
284 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
286 /* Update potential sum for this i atom from the interaction with this j atom. */
287 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
291 /* Calculate temporary vectorial force */
292 tx = _mm256_mul_ps(fscal,dx00);
293 ty = _mm256_mul_ps(fscal,dy00);
294 tz = _mm256_mul_ps(fscal,dz00);
296 /* Update vectorial force */
297 fix0 = _mm256_add_ps(fix0,tx);
298 fiy0 = _mm256_add_ps(fiy0,ty);
299 fiz0 = _mm256_add_ps(fiz0,tz);
301 fjptrA = f+j_coord_offsetA;
302 fjptrB = f+j_coord_offsetB;
303 fjptrC = f+j_coord_offsetC;
304 fjptrD = f+j_coord_offsetD;
305 fjptrE = f+j_coord_offsetE;
306 fjptrF = f+j_coord_offsetF;
307 fjptrG = f+j_coord_offsetG;
308 fjptrH = f+j_coord_offsetH;
309 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
311 /* Inner loop uses 56 flops */
317 /* Get j neighbor index, and coordinate index */
318 jnrlistA = jjnr[jidx];
319 jnrlistB = jjnr[jidx+1];
320 jnrlistC = jjnr[jidx+2];
321 jnrlistD = jjnr[jidx+3];
322 jnrlistE = jjnr[jidx+4];
323 jnrlistF = jjnr[jidx+5];
324 jnrlistG = jjnr[jidx+6];
325 jnrlistH = jjnr[jidx+7];
326 /* Sign of each element will be negative for non-real atoms.
327 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
328 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
330 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
331 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
333 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
334 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
335 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
336 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
337 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
338 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
339 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
340 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
341 j_coord_offsetA = DIM*jnrA;
342 j_coord_offsetB = DIM*jnrB;
343 j_coord_offsetC = DIM*jnrC;
344 j_coord_offsetD = DIM*jnrD;
345 j_coord_offsetE = DIM*jnrE;
346 j_coord_offsetF = DIM*jnrF;
347 j_coord_offsetG = DIM*jnrG;
348 j_coord_offsetH = DIM*jnrH;
350 /* load j atom coordinates */
351 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
352 x+j_coord_offsetC,x+j_coord_offsetD,
353 x+j_coord_offsetE,x+j_coord_offsetF,
354 x+j_coord_offsetG,x+j_coord_offsetH,
357 /* Calculate displacement vector */
358 dx00 = _mm256_sub_ps(ix0,jx0);
359 dy00 = _mm256_sub_ps(iy0,jy0);
360 dz00 = _mm256_sub_ps(iz0,jz0);
362 /* Calculate squared distance and things based on it */
363 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
365 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
367 /* Load parameters for j particles */
368 vdwjidx0A = 2*vdwtype[jnrA+0];
369 vdwjidx0B = 2*vdwtype[jnrB+0];
370 vdwjidx0C = 2*vdwtype[jnrC+0];
371 vdwjidx0D = 2*vdwtype[jnrD+0];
372 vdwjidx0E = 2*vdwtype[jnrE+0];
373 vdwjidx0F = 2*vdwtype[jnrF+0];
374 vdwjidx0G = 2*vdwtype[jnrG+0];
375 vdwjidx0H = 2*vdwtype[jnrH+0];
377 /**************************
378 * CALCULATE INTERACTIONS *
379 **************************/
381 r00 = _mm256_mul_ps(rsq00,rinv00);
382 r00 = _mm256_andnot_ps(dummy_mask,r00);
384 /* Compute parameters for interactions between i and j atoms */
385 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
386 vdwioffsetptr0+vdwjidx0B,
387 vdwioffsetptr0+vdwjidx0C,
388 vdwioffsetptr0+vdwjidx0D,
389 vdwioffsetptr0+vdwjidx0E,
390 vdwioffsetptr0+vdwjidx0F,
391 vdwioffsetptr0+vdwjidx0G,
392 vdwioffsetptr0+vdwjidx0H,
395 /* Calculate table index by multiplying r with table scale and truncate to integer */
396 rt = _mm256_mul_ps(r00,vftabscale);
397 vfitab = _mm256_cvttps_epi32(rt);
398 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
399 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
400 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
401 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
402 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
403 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
405 /* CUBIC SPLINE TABLE DISPERSION */
406 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
407 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
408 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
409 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
410 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
411 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
412 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
413 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
414 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
415 Heps = _mm256_mul_ps(vfeps,H);
416 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
417 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
418 vvdw6 = _mm256_mul_ps(c6_00,VV);
419 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
420 fvdw6 = _mm256_mul_ps(c6_00,FF);
422 /* CUBIC SPLINE TABLE REPULSION */
423 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
424 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
425 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
426 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
427 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
428 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
429 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
430 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
431 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
432 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
433 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
434 Heps = _mm256_mul_ps(vfeps,H);
435 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
436 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
437 vvdw12 = _mm256_mul_ps(c12_00,VV);
438 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
439 fvdw12 = _mm256_mul_ps(c12_00,FF);
440 vvdw = _mm256_add_ps(vvdw12,vvdw6);
441 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
443 /* Update potential sum for this i atom from the interaction with this j atom. */
444 vvdw = _mm256_andnot_ps(dummy_mask,vvdw);
445 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
449 fscal = _mm256_andnot_ps(dummy_mask,fscal);
451 /* Calculate temporary vectorial force */
452 tx = _mm256_mul_ps(fscal,dx00);
453 ty = _mm256_mul_ps(fscal,dy00);
454 tz = _mm256_mul_ps(fscal,dz00);
456 /* Update vectorial force */
457 fix0 = _mm256_add_ps(fix0,tx);
458 fiy0 = _mm256_add_ps(fiy0,ty);
459 fiz0 = _mm256_add_ps(fiz0,tz);
461 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
462 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
463 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
464 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
465 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
466 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
467 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
468 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
469 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
471 /* Inner loop uses 57 flops */
474 /* End of innermost loop */
476 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
477 f+i_coord_offset,fshift+i_shift_offset);
480 /* Update potential energies */
481 gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
483 /* Increment number of inner iterations */
484 inneriter += j_index_end - j_index_start;
486 /* Outer loop uses 7 flops */
489 /* Increment number of outer iterations */
492 /* Update outer/inner flops */
494 inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*57);
497 * Gromacs nonbonded kernel: nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_single
498 * Electrostatics interaction: None
499 * VdW interaction: CubicSplineTable
500 * Geometry: Particle-Particle
501 * Calculate force/pot: Force
504 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_single
505 (t_nblist * gmx_restrict nlist,
506 rvec * gmx_restrict xx,
507 rvec * gmx_restrict ff,
508 t_forcerec * gmx_restrict fr,
509 t_mdatoms * gmx_restrict mdatoms,
510 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
511 t_nrnb * gmx_restrict nrnb)
513 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
514 * just 0 for non-waters.
515 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
516 * jnr indices corresponding to data put in the four positions in the SIMD register.
518 int i_shift_offset,i_coord_offset,outeriter,inneriter;
519 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
520 int jnrA,jnrB,jnrC,jnrD;
521 int jnrE,jnrF,jnrG,jnrH;
522 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
523 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
524 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
525 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
526 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
528 real *shiftvec,*fshift,*x,*f;
529 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
531 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
532 real * vdwioffsetptr0;
533 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
534 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
535 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
536 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
538 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
541 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
542 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
544 __m128i vfitab_lo,vfitab_hi;
545 __m128i ifour = _mm_set1_epi32(4);
546 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
548 __m256 dummy_mask,cutoff_mask;
549 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
550 __m256 one = _mm256_set1_ps(1.0);
551 __m256 two = _mm256_set1_ps(2.0);
557 jindex = nlist->jindex;
559 shiftidx = nlist->shift;
561 shiftvec = fr->shift_vec[0];
562 fshift = fr->fshift[0];
563 nvdwtype = fr->ntype;
565 vdwtype = mdatoms->typeA;
567 vftab = kernel_data->table_vdw->data;
568 vftabscale = _mm256_set1_ps(kernel_data->table_vdw->scale);
570 /* Avoid stupid compiler warnings */
571 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
584 for(iidx=0;iidx<4*DIM;iidx++)
589 /* Start outer loop over neighborlists */
590 for(iidx=0; iidx<nri; iidx++)
592 /* Load shift vector for this list */
593 i_shift_offset = DIM*shiftidx[iidx];
595 /* Load limits for loop over neighbors */
596 j_index_start = jindex[iidx];
597 j_index_end = jindex[iidx+1];
599 /* Get outer coordinate index */
601 i_coord_offset = DIM*inr;
603 /* Load i particle coords and add shift vector */
604 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
606 fix0 = _mm256_setzero_ps();
607 fiy0 = _mm256_setzero_ps();
608 fiz0 = _mm256_setzero_ps();
610 /* Load parameters for i particles */
611 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
613 /* Start inner kernel loop */
614 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
617 /* Get j neighbor index, and coordinate index */
626 j_coord_offsetA = DIM*jnrA;
627 j_coord_offsetB = DIM*jnrB;
628 j_coord_offsetC = DIM*jnrC;
629 j_coord_offsetD = DIM*jnrD;
630 j_coord_offsetE = DIM*jnrE;
631 j_coord_offsetF = DIM*jnrF;
632 j_coord_offsetG = DIM*jnrG;
633 j_coord_offsetH = DIM*jnrH;
635 /* load j atom coordinates */
636 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
637 x+j_coord_offsetC,x+j_coord_offsetD,
638 x+j_coord_offsetE,x+j_coord_offsetF,
639 x+j_coord_offsetG,x+j_coord_offsetH,
642 /* Calculate displacement vector */
643 dx00 = _mm256_sub_ps(ix0,jx0);
644 dy00 = _mm256_sub_ps(iy0,jy0);
645 dz00 = _mm256_sub_ps(iz0,jz0);
647 /* Calculate squared distance and things based on it */
648 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
650 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
652 /* Load parameters for j particles */
653 vdwjidx0A = 2*vdwtype[jnrA+0];
654 vdwjidx0B = 2*vdwtype[jnrB+0];
655 vdwjidx0C = 2*vdwtype[jnrC+0];
656 vdwjidx0D = 2*vdwtype[jnrD+0];
657 vdwjidx0E = 2*vdwtype[jnrE+0];
658 vdwjidx0F = 2*vdwtype[jnrF+0];
659 vdwjidx0G = 2*vdwtype[jnrG+0];
660 vdwjidx0H = 2*vdwtype[jnrH+0];
662 /**************************
663 * CALCULATE INTERACTIONS *
664 **************************/
666 r00 = _mm256_mul_ps(rsq00,rinv00);
668 /* Compute parameters for interactions between i and j atoms */
669 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
670 vdwioffsetptr0+vdwjidx0B,
671 vdwioffsetptr0+vdwjidx0C,
672 vdwioffsetptr0+vdwjidx0D,
673 vdwioffsetptr0+vdwjidx0E,
674 vdwioffsetptr0+vdwjidx0F,
675 vdwioffsetptr0+vdwjidx0G,
676 vdwioffsetptr0+vdwjidx0H,
679 /* Calculate table index by multiplying r with table scale and truncate to integer */
680 rt = _mm256_mul_ps(r00,vftabscale);
681 vfitab = _mm256_cvttps_epi32(rt);
682 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
683 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
684 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
685 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
686 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
687 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
689 /* CUBIC SPLINE TABLE DISPERSION */
690 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
691 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
692 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
693 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
694 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
695 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
696 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
697 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
698 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
699 Heps = _mm256_mul_ps(vfeps,H);
700 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
701 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
702 fvdw6 = _mm256_mul_ps(c6_00,FF);
704 /* CUBIC SPLINE TABLE REPULSION */
705 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
706 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
707 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
708 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
709 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
710 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
711 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
712 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
713 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
714 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
715 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
716 Heps = _mm256_mul_ps(vfeps,H);
717 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
718 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
719 fvdw12 = _mm256_mul_ps(c12_00,FF);
720 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
724 /* Calculate temporary vectorial force */
725 tx = _mm256_mul_ps(fscal,dx00);
726 ty = _mm256_mul_ps(fscal,dy00);
727 tz = _mm256_mul_ps(fscal,dz00);
729 /* Update vectorial force */
730 fix0 = _mm256_add_ps(fix0,tx);
731 fiy0 = _mm256_add_ps(fiy0,ty);
732 fiz0 = _mm256_add_ps(fiz0,tz);
734 fjptrA = f+j_coord_offsetA;
735 fjptrB = f+j_coord_offsetB;
736 fjptrC = f+j_coord_offsetC;
737 fjptrD = f+j_coord_offsetD;
738 fjptrE = f+j_coord_offsetE;
739 fjptrF = f+j_coord_offsetF;
740 fjptrG = f+j_coord_offsetG;
741 fjptrH = f+j_coord_offsetH;
742 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
744 /* Inner loop uses 48 flops */
750 /* Get j neighbor index, and coordinate index */
751 jnrlistA = jjnr[jidx];
752 jnrlistB = jjnr[jidx+1];
753 jnrlistC = jjnr[jidx+2];
754 jnrlistD = jjnr[jidx+3];
755 jnrlistE = jjnr[jidx+4];
756 jnrlistF = jjnr[jidx+5];
757 jnrlistG = jjnr[jidx+6];
758 jnrlistH = jjnr[jidx+7];
759 /* Sign of each element will be negative for non-real atoms.
760 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
761 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
763 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
764 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
766 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
767 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
768 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
769 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
770 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
771 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
772 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
773 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
774 j_coord_offsetA = DIM*jnrA;
775 j_coord_offsetB = DIM*jnrB;
776 j_coord_offsetC = DIM*jnrC;
777 j_coord_offsetD = DIM*jnrD;
778 j_coord_offsetE = DIM*jnrE;
779 j_coord_offsetF = DIM*jnrF;
780 j_coord_offsetG = DIM*jnrG;
781 j_coord_offsetH = DIM*jnrH;
783 /* load j atom coordinates */
784 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
785 x+j_coord_offsetC,x+j_coord_offsetD,
786 x+j_coord_offsetE,x+j_coord_offsetF,
787 x+j_coord_offsetG,x+j_coord_offsetH,
790 /* Calculate displacement vector */
791 dx00 = _mm256_sub_ps(ix0,jx0);
792 dy00 = _mm256_sub_ps(iy0,jy0);
793 dz00 = _mm256_sub_ps(iz0,jz0);
795 /* Calculate squared distance and things based on it */
796 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
798 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
800 /* Load parameters for j particles */
801 vdwjidx0A = 2*vdwtype[jnrA+0];
802 vdwjidx0B = 2*vdwtype[jnrB+0];
803 vdwjidx0C = 2*vdwtype[jnrC+0];
804 vdwjidx0D = 2*vdwtype[jnrD+0];
805 vdwjidx0E = 2*vdwtype[jnrE+0];
806 vdwjidx0F = 2*vdwtype[jnrF+0];
807 vdwjidx0G = 2*vdwtype[jnrG+0];
808 vdwjidx0H = 2*vdwtype[jnrH+0];
810 /**************************
811 * CALCULATE INTERACTIONS *
812 **************************/
814 r00 = _mm256_mul_ps(rsq00,rinv00);
815 r00 = _mm256_andnot_ps(dummy_mask,r00);
817 /* Compute parameters for interactions between i and j atoms */
818 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
819 vdwioffsetptr0+vdwjidx0B,
820 vdwioffsetptr0+vdwjidx0C,
821 vdwioffsetptr0+vdwjidx0D,
822 vdwioffsetptr0+vdwjidx0E,
823 vdwioffsetptr0+vdwjidx0F,
824 vdwioffsetptr0+vdwjidx0G,
825 vdwioffsetptr0+vdwjidx0H,
828 /* Calculate table index by multiplying r with table scale and truncate to integer */
829 rt = _mm256_mul_ps(r00,vftabscale);
830 vfitab = _mm256_cvttps_epi32(rt);
831 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
832 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
833 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
834 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
835 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
836 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
838 /* CUBIC SPLINE TABLE DISPERSION */
839 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
840 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
841 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
842 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
843 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
844 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
845 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
846 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
847 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
848 Heps = _mm256_mul_ps(vfeps,H);
849 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
850 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
851 fvdw6 = _mm256_mul_ps(c6_00,FF);
853 /* CUBIC SPLINE TABLE REPULSION */
854 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
855 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
856 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
857 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
858 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
859 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
860 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
861 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
862 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
863 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
864 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
865 Heps = _mm256_mul_ps(vfeps,H);
866 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
867 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
868 fvdw12 = _mm256_mul_ps(c12_00,FF);
869 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
873 fscal = _mm256_andnot_ps(dummy_mask,fscal);
875 /* Calculate temporary vectorial force */
876 tx = _mm256_mul_ps(fscal,dx00);
877 ty = _mm256_mul_ps(fscal,dy00);
878 tz = _mm256_mul_ps(fscal,dz00);
880 /* Update vectorial force */
881 fix0 = _mm256_add_ps(fix0,tx);
882 fiy0 = _mm256_add_ps(fiy0,ty);
883 fiz0 = _mm256_add_ps(fiz0,tz);
885 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
886 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
887 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
888 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
889 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
890 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
891 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
892 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
893 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
895 /* Inner loop uses 49 flops */
898 /* End of innermost loop */
900 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
901 f+i_coord_offset,fshift+i_shift_offset);
903 /* Increment number of inner iterations */
904 inneriter += j_index_end - j_index_start;
906 /* Outer loop uses 6 flops */
909 /* Increment number of outer iterations */
912 /* Update outer/inner flops */
914 inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*49);