2 * Note: this file was generated by the Gromacs avx_256_single kernel generator.
4 * This source code is part of
8 * Copyright (c) 2001-2012, The GROMACS Development Team
10 * Gromacs is a library for molecular simulation and trajectory analysis,
11 * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
12 * a full list of developers and information, check out http://www.gromacs.org
14 * This program is free software; you can redistribute it and/or modify it under
15 * the terms of the GNU Lesser General Public License as published by the Free
16 * Software Foundation; either version 2 of the License, or (at your option) any
19 * To help fund GROMACS development, we humbly ask that you cite
20 * the papers people have written on it - you can find them on the website.
28 #include "../nb_kernel.h"
29 #include "types/simple.h"
33 #include "gmx_math_x86_avx_256_single.h"
34 #include "kernelutil_x86_avx_256_single.h"
37 * Gromacs nonbonded kernel: nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_256_single
38 * Electrostatics interaction: None
39 * VdW interaction: CubicSplineTable
40 * Geometry: Particle-Particle
41 * Calculate force/pot: PotentialAndForce
44 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_VF_avx_256_single
45 (t_nblist * gmx_restrict nlist,
46 rvec * gmx_restrict xx,
47 rvec * gmx_restrict ff,
48 t_forcerec * gmx_restrict fr,
49 t_mdatoms * gmx_restrict mdatoms,
50 nb_kernel_data_t * gmx_restrict kernel_data,
51 t_nrnb * gmx_restrict nrnb)
53 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
54 * just 0 for non-waters.
55 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
56 * jnr indices corresponding to data put in the four positions in the SIMD register.
58 int i_shift_offset,i_coord_offset,outeriter,inneriter;
59 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
60 int jnrA,jnrB,jnrC,jnrD;
61 int jnrE,jnrF,jnrG,jnrH;
62 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
63 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
64 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
65 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
66 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
68 real *shiftvec,*fshift,*x,*f;
69 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
71 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
72 real * vdwioffsetptr0;
73 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
74 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
75 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
76 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
78 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
81 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
82 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
84 __m128i vfitab_lo,vfitab_hi;
85 __m128i ifour = _mm_set1_epi32(4);
86 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
88 __m256 dummy_mask,cutoff_mask;
89 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
90 __m256 one = _mm256_set1_ps(1.0);
91 __m256 two = _mm256_set1_ps(2.0);
97 jindex = nlist->jindex;
99 shiftidx = nlist->shift;
101 shiftvec = fr->shift_vec[0];
102 fshift = fr->fshift[0];
103 nvdwtype = fr->ntype;
105 vdwtype = mdatoms->typeA;
107 vftab = kernel_data->table_vdw->data;
108 vftabscale = _mm256_set1_ps(kernel_data->table_vdw->scale);
110 /* Avoid stupid compiler warnings */
111 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
124 for(iidx=0;iidx<4*DIM;iidx++)
129 /* Start outer loop over neighborlists */
130 for(iidx=0; iidx<nri; iidx++)
132 /* Load shift vector for this list */
133 i_shift_offset = DIM*shiftidx[iidx];
135 /* Load limits for loop over neighbors */
136 j_index_start = jindex[iidx];
137 j_index_end = jindex[iidx+1];
139 /* Get outer coordinate index */
141 i_coord_offset = DIM*inr;
143 /* Load i particle coords and add shift vector */
144 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
146 fix0 = _mm256_setzero_ps();
147 fiy0 = _mm256_setzero_ps();
148 fiz0 = _mm256_setzero_ps();
150 /* Load parameters for i particles */
151 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
153 /* Reset potential sums */
154 vvdwsum = _mm256_setzero_ps();
156 /* Start inner kernel loop */
157 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
160 /* Get j neighbor index, and coordinate index */
169 j_coord_offsetA = DIM*jnrA;
170 j_coord_offsetB = DIM*jnrB;
171 j_coord_offsetC = DIM*jnrC;
172 j_coord_offsetD = DIM*jnrD;
173 j_coord_offsetE = DIM*jnrE;
174 j_coord_offsetF = DIM*jnrF;
175 j_coord_offsetG = DIM*jnrG;
176 j_coord_offsetH = DIM*jnrH;
178 /* load j atom coordinates */
179 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
180 x+j_coord_offsetC,x+j_coord_offsetD,
181 x+j_coord_offsetE,x+j_coord_offsetF,
182 x+j_coord_offsetG,x+j_coord_offsetH,
185 /* Calculate displacement vector */
186 dx00 = _mm256_sub_ps(ix0,jx0);
187 dy00 = _mm256_sub_ps(iy0,jy0);
188 dz00 = _mm256_sub_ps(iz0,jz0);
190 /* Calculate squared distance and things based on it */
191 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
193 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
195 /* Load parameters for j particles */
196 vdwjidx0A = 2*vdwtype[jnrA+0];
197 vdwjidx0B = 2*vdwtype[jnrB+0];
198 vdwjidx0C = 2*vdwtype[jnrC+0];
199 vdwjidx0D = 2*vdwtype[jnrD+0];
200 vdwjidx0E = 2*vdwtype[jnrE+0];
201 vdwjidx0F = 2*vdwtype[jnrF+0];
202 vdwjidx0G = 2*vdwtype[jnrG+0];
203 vdwjidx0H = 2*vdwtype[jnrH+0];
205 /**************************
206 * CALCULATE INTERACTIONS *
207 **************************/
209 r00 = _mm256_mul_ps(rsq00,rinv00);
211 /* Compute parameters for interactions between i and j atoms */
212 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
213 vdwioffsetptr0+vdwjidx0B,
214 vdwioffsetptr0+vdwjidx0C,
215 vdwioffsetptr0+vdwjidx0D,
216 vdwioffsetptr0+vdwjidx0E,
217 vdwioffsetptr0+vdwjidx0F,
218 vdwioffsetptr0+vdwjidx0G,
219 vdwioffsetptr0+vdwjidx0H,
222 /* Calculate table index by multiplying r with table scale and truncate to integer */
223 rt = _mm256_mul_ps(r00,vftabscale);
224 vfitab = _mm256_cvttps_epi32(rt);
225 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
226 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
227 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
228 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
229 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
230 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
232 /* CUBIC SPLINE TABLE DISPERSION */
233 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
234 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
235 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
236 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
237 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
238 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
239 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
240 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
241 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
242 Heps = _mm256_mul_ps(vfeps,H);
243 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
244 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
245 vvdw6 = _mm256_mul_ps(c6_00,VV);
246 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
247 fvdw6 = _mm256_mul_ps(c6_00,FF);
249 /* CUBIC SPLINE TABLE REPULSION */
250 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
251 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
252 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
253 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
254 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
255 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
256 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
257 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
258 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
259 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
260 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
261 Heps = _mm256_mul_ps(vfeps,H);
262 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
263 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
264 vvdw12 = _mm256_mul_ps(c12_00,VV);
265 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
266 fvdw12 = _mm256_mul_ps(c12_00,FF);
267 vvdw = _mm256_add_ps(vvdw12,vvdw6);
268 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
270 /* Update potential sum for this i atom from the interaction with this j atom. */
271 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
275 /* Calculate temporary vectorial force */
276 tx = _mm256_mul_ps(fscal,dx00);
277 ty = _mm256_mul_ps(fscal,dy00);
278 tz = _mm256_mul_ps(fscal,dz00);
280 /* Update vectorial force */
281 fix0 = _mm256_add_ps(fix0,tx);
282 fiy0 = _mm256_add_ps(fiy0,ty);
283 fiz0 = _mm256_add_ps(fiz0,tz);
285 fjptrA = f+j_coord_offsetA;
286 fjptrB = f+j_coord_offsetB;
287 fjptrC = f+j_coord_offsetC;
288 fjptrD = f+j_coord_offsetD;
289 fjptrE = f+j_coord_offsetE;
290 fjptrF = f+j_coord_offsetF;
291 fjptrG = f+j_coord_offsetG;
292 fjptrH = f+j_coord_offsetH;
293 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
295 /* Inner loop uses 56 flops */
301 /* Get j neighbor index, and coordinate index */
302 jnrlistA = jjnr[jidx];
303 jnrlistB = jjnr[jidx+1];
304 jnrlistC = jjnr[jidx+2];
305 jnrlistD = jjnr[jidx+3];
306 jnrlistE = jjnr[jidx+4];
307 jnrlistF = jjnr[jidx+5];
308 jnrlistG = jjnr[jidx+6];
309 jnrlistH = jjnr[jidx+7];
310 /* Sign of each element will be negative for non-real atoms.
311 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
312 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
314 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
315 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
317 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
318 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
319 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
320 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
321 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
322 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
323 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
324 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
325 j_coord_offsetA = DIM*jnrA;
326 j_coord_offsetB = DIM*jnrB;
327 j_coord_offsetC = DIM*jnrC;
328 j_coord_offsetD = DIM*jnrD;
329 j_coord_offsetE = DIM*jnrE;
330 j_coord_offsetF = DIM*jnrF;
331 j_coord_offsetG = DIM*jnrG;
332 j_coord_offsetH = DIM*jnrH;
334 /* load j atom coordinates */
335 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
336 x+j_coord_offsetC,x+j_coord_offsetD,
337 x+j_coord_offsetE,x+j_coord_offsetF,
338 x+j_coord_offsetG,x+j_coord_offsetH,
341 /* Calculate displacement vector */
342 dx00 = _mm256_sub_ps(ix0,jx0);
343 dy00 = _mm256_sub_ps(iy0,jy0);
344 dz00 = _mm256_sub_ps(iz0,jz0);
346 /* Calculate squared distance and things based on it */
347 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
349 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
351 /* Load parameters for j particles */
352 vdwjidx0A = 2*vdwtype[jnrA+0];
353 vdwjidx0B = 2*vdwtype[jnrB+0];
354 vdwjidx0C = 2*vdwtype[jnrC+0];
355 vdwjidx0D = 2*vdwtype[jnrD+0];
356 vdwjidx0E = 2*vdwtype[jnrE+0];
357 vdwjidx0F = 2*vdwtype[jnrF+0];
358 vdwjidx0G = 2*vdwtype[jnrG+0];
359 vdwjidx0H = 2*vdwtype[jnrH+0];
361 /**************************
362 * CALCULATE INTERACTIONS *
363 **************************/
365 r00 = _mm256_mul_ps(rsq00,rinv00);
366 r00 = _mm256_andnot_ps(dummy_mask,r00);
368 /* Compute parameters for interactions between i and j atoms */
369 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
370 vdwioffsetptr0+vdwjidx0B,
371 vdwioffsetptr0+vdwjidx0C,
372 vdwioffsetptr0+vdwjidx0D,
373 vdwioffsetptr0+vdwjidx0E,
374 vdwioffsetptr0+vdwjidx0F,
375 vdwioffsetptr0+vdwjidx0G,
376 vdwioffsetptr0+vdwjidx0H,
379 /* Calculate table index by multiplying r with table scale and truncate to integer */
380 rt = _mm256_mul_ps(r00,vftabscale);
381 vfitab = _mm256_cvttps_epi32(rt);
382 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
383 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
384 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
385 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
386 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
387 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
389 /* CUBIC SPLINE TABLE DISPERSION */
390 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
391 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
392 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
393 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
394 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
395 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
396 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
397 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
398 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
399 Heps = _mm256_mul_ps(vfeps,H);
400 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
401 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
402 vvdw6 = _mm256_mul_ps(c6_00,VV);
403 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
404 fvdw6 = _mm256_mul_ps(c6_00,FF);
406 /* CUBIC SPLINE TABLE REPULSION */
407 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
408 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
409 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
410 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
411 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
412 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
413 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
414 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
415 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
416 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
417 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
418 Heps = _mm256_mul_ps(vfeps,H);
419 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
420 VV = _mm256_add_ps(Y,_mm256_mul_ps(vfeps,Fp));
421 vvdw12 = _mm256_mul_ps(c12_00,VV);
422 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
423 fvdw12 = _mm256_mul_ps(c12_00,FF);
424 vvdw = _mm256_add_ps(vvdw12,vvdw6);
425 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
427 /* Update potential sum for this i atom from the interaction with this j atom. */
428 vvdw = _mm256_andnot_ps(dummy_mask,vvdw);
429 vvdwsum = _mm256_add_ps(vvdwsum,vvdw);
433 fscal = _mm256_andnot_ps(dummy_mask,fscal);
435 /* Calculate temporary vectorial force */
436 tx = _mm256_mul_ps(fscal,dx00);
437 ty = _mm256_mul_ps(fscal,dy00);
438 tz = _mm256_mul_ps(fscal,dz00);
440 /* Update vectorial force */
441 fix0 = _mm256_add_ps(fix0,tx);
442 fiy0 = _mm256_add_ps(fiy0,ty);
443 fiz0 = _mm256_add_ps(fiz0,tz);
445 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
446 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
447 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
448 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
449 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
450 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
451 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
452 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
453 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
455 /* Inner loop uses 57 flops */
458 /* End of innermost loop */
460 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
461 f+i_coord_offset,fshift+i_shift_offset);
464 /* Update potential energies */
465 gmx_mm256_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
467 /* Increment number of inner iterations */
468 inneriter += j_index_end - j_index_start;
470 /* Outer loop uses 7 flops */
473 /* Increment number of outer iterations */
476 /* Update outer/inner flops */
478 inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*57);
481 * Gromacs nonbonded kernel: nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_single
482 * Electrostatics interaction: None
483 * VdW interaction: CubicSplineTable
484 * Geometry: Particle-Particle
485 * Calculate force/pot: Force
488 nb_kernel_ElecNone_VdwCSTab_GeomP1P1_F_avx_256_single
489 (t_nblist * gmx_restrict nlist,
490 rvec * gmx_restrict xx,
491 rvec * gmx_restrict ff,
492 t_forcerec * gmx_restrict fr,
493 t_mdatoms * gmx_restrict mdatoms,
494 nb_kernel_data_t * gmx_restrict kernel_data,
495 t_nrnb * gmx_restrict nrnb)
497 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
498 * just 0 for non-waters.
499 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
500 * jnr indices corresponding to data put in the four positions in the SIMD register.
502 int i_shift_offset,i_coord_offset,outeriter,inneriter;
503 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
504 int jnrA,jnrB,jnrC,jnrD;
505 int jnrE,jnrF,jnrG,jnrH;
506 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
507 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
508 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
509 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
510 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
512 real *shiftvec,*fshift,*x,*f;
513 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
515 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
516 real * vdwioffsetptr0;
517 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
518 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
519 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
520 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
522 __m256 rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
525 __m256 one_sixth = _mm256_set1_ps(1.0/6.0);
526 __m256 one_twelfth = _mm256_set1_ps(1.0/12.0);
528 __m128i vfitab_lo,vfitab_hi;
529 __m128i ifour = _mm_set1_epi32(4);
530 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
532 __m256 dummy_mask,cutoff_mask;
533 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
534 __m256 one = _mm256_set1_ps(1.0);
535 __m256 two = _mm256_set1_ps(2.0);
541 jindex = nlist->jindex;
543 shiftidx = nlist->shift;
545 shiftvec = fr->shift_vec[0];
546 fshift = fr->fshift[0];
547 nvdwtype = fr->ntype;
549 vdwtype = mdatoms->typeA;
551 vftab = kernel_data->table_vdw->data;
552 vftabscale = _mm256_set1_ps(kernel_data->table_vdw->scale);
554 /* Avoid stupid compiler warnings */
555 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
568 for(iidx=0;iidx<4*DIM;iidx++)
573 /* Start outer loop over neighborlists */
574 for(iidx=0; iidx<nri; iidx++)
576 /* Load shift vector for this list */
577 i_shift_offset = DIM*shiftidx[iidx];
579 /* Load limits for loop over neighbors */
580 j_index_start = jindex[iidx];
581 j_index_end = jindex[iidx+1];
583 /* Get outer coordinate index */
585 i_coord_offset = DIM*inr;
587 /* Load i particle coords and add shift vector */
588 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
590 fix0 = _mm256_setzero_ps();
591 fiy0 = _mm256_setzero_ps();
592 fiz0 = _mm256_setzero_ps();
594 /* Load parameters for i particles */
595 vdwioffsetptr0 = vdwparam+2*nvdwtype*vdwtype[inr+0];
597 /* Start inner kernel loop */
598 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
601 /* Get j neighbor index, and coordinate index */
610 j_coord_offsetA = DIM*jnrA;
611 j_coord_offsetB = DIM*jnrB;
612 j_coord_offsetC = DIM*jnrC;
613 j_coord_offsetD = DIM*jnrD;
614 j_coord_offsetE = DIM*jnrE;
615 j_coord_offsetF = DIM*jnrF;
616 j_coord_offsetG = DIM*jnrG;
617 j_coord_offsetH = DIM*jnrH;
619 /* load j atom coordinates */
620 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
621 x+j_coord_offsetC,x+j_coord_offsetD,
622 x+j_coord_offsetE,x+j_coord_offsetF,
623 x+j_coord_offsetG,x+j_coord_offsetH,
626 /* Calculate displacement vector */
627 dx00 = _mm256_sub_ps(ix0,jx0);
628 dy00 = _mm256_sub_ps(iy0,jy0);
629 dz00 = _mm256_sub_ps(iz0,jz0);
631 /* Calculate squared distance and things based on it */
632 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
634 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
636 /* Load parameters for j particles */
637 vdwjidx0A = 2*vdwtype[jnrA+0];
638 vdwjidx0B = 2*vdwtype[jnrB+0];
639 vdwjidx0C = 2*vdwtype[jnrC+0];
640 vdwjidx0D = 2*vdwtype[jnrD+0];
641 vdwjidx0E = 2*vdwtype[jnrE+0];
642 vdwjidx0F = 2*vdwtype[jnrF+0];
643 vdwjidx0G = 2*vdwtype[jnrG+0];
644 vdwjidx0H = 2*vdwtype[jnrH+0];
646 /**************************
647 * CALCULATE INTERACTIONS *
648 **************************/
650 r00 = _mm256_mul_ps(rsq00,rinv00);
652 /* Compute parameters for interactions between i and j atoms */
653 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
654 vdwioffsetptr0+vdwjidx0B,
655 vdwioffsetptr0+vdwjidx0C,
656 vdwioffsetptr0+vdwjidx0D,
657 vdwioffsetptr0+vdwjidx0E,
658 vdwioffsetptr0+vdwjidx0F,
659 vdwioffsetptr0+vdwjidx0G,
660 vdwioffsetptr0+vdwjidx0H,
663 /* Calculate table index by multiplying r with table scale and truncate to integer */
664 rt = _mm256_mul_ps(r00,vftabscale);
665 vfitab = _mm256_cvttps_epi32(rt);
666 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
667 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
668 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
669 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
670 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
671 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
673 /* CUBIC SPLINE TABLE DISPERSION */
674 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
675 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
676 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
677 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
678 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
679 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
680 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
681 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
682 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
683 Heps = _mm256_mul_ps(vfeps,H);
684 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
685 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
686 fvdw6 = _mm256_mul_ps(c6_00,FF);
688 /* CUBIC SPLINE TABLE REPULSION */
689 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
690 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
691 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
692 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
693 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
694 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
695 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
696 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
697 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
698 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
699 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
700 Heps = _mm256_mul_ps(vfeps,H);
701 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
702 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
703 fvdw12 = _mm256_mul_ps(c12_00,FF);
704 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
708 /* Calculate temporary vectorial force */
709 tx = _mm256_mul_ps(fscal,dx00);
710 ty = _mm256_mul_ps(fscal,dy00);
711 tz = _mm256_mul_ps(fscal,dz00);
713 /* Update vectorial force */
714 fix0 = _mm256_add_ps(fix0,tx);
715 fiy0 = _mm256_add_ps(fiy0,ty);
716 fiz0 = _mm256_add_ps(fiz0,tz);
718 fjptrA = f+j_coord_offsetA;
719 fjptrB = f+j_coord_offsetB;
720 fjptrC = f+j_coord_offsetC;
721 fjptrD = f+j_coord_offsetD;
722 fjptrE = f+j_coord_offsetE;
723 fjptrF = f+j_coord_offsetF;
724 fjptrG = f+j_coord_offsetG;
725 fjptrH = f+j_coord_offsetH;
726 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
728 /* Inner loop uses 48 flops */
734 /* Get j neighbor index, and coordinate index */
735 jnrlistA = jjnr[jidx];
736 jnrlistB = jjnr[jidx+1];
737 jnrlistC = jjnr[jidx+2];
738 jnrlistD = jjnr[jidx+3];
739 jnrlistE = jjnr[jidx+4];
740 jnrlistF = jjnr[jidx+5];
741 jnrlistG = jjnr[jidx+6];
742 jnrlistH = jjnr[jidx+7];
743 /* Sign of each element will be negative for non-real atoms.
744 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
745 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
747 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
748 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
750 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
751 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
752 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
753 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
754 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
755 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
756 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
757 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
758 j_coord_offsetA = DIM*jnrA;
759 j_coord_offsetB = DIM*jnrB;
760 j_coord_offsetC = DIM*jnrC;
761 j_coord_offsetD = DIM*jnrD;
762 j_coord_offsetE = DIM*jnrE;
763 j_coord_offsetF = DIM*jnrF;
764 j_coord_offsetG = DIM*jnrG;
765 j_coord_offsetH = DIM*jnrH;
767 /* load j atom coordinates */
768 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
769 x+j_coord_offsetC,x+j_coord_offsetD,
770 x+j_coord_offsetE,x+j_coord_offsetF,
771 x+j_coord_offsetG,x+j_coord_offsetH,
774 /* Calculate displacement vector */
775 dx00 = _mm256_sub_ps(ix0,jx0);
776 dy00 = _mm256_sub_ps(iy0,jy0);
777 dz00 = _mm256_sub_ps(iz0,jz0);
779 /* Calculate squared distance and things based on it */
780 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
782 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
784 /* Load parameters for j particles */
785 vdwjidx0A = 2*vdwtype[jnrA+0];
786 vdwjidx0B = 2*vdwtype[jnrB+0];
787 vdwjidx0C = 2*vdwtype[jnrC+0];
788 vdwjidx0D = 2*vdwtype[jnrD+0];
789 vdwjidx0E = 2*vdwtype[jnrE+0];
790 vdwjidx0F = 2*vdwtype[jnrF+0];
791 vdwjidx0G = 2*vdwtype[jnrG+0];
792 vdwjidx0H = 2*vdwtype[jnrH+0];
794 /**************************
795 * CALCULATE INTERACTIONS *
796 **************************/
798 r00 = _mm256_mul_ps(rsq00,rinv00);
799 r00 = _mm256_andnot_ps(dummy_mask,r00);
801 /* Compute parameters for interactions between i and j atoms */
802 gmx_mm256_load_8pair_swizzle_ps(vdwioffsetptr0+vdwjidx0A,
803 vdwioffsetptr0+vdwjidx0B,
804 vdwioffsetptr0+vdwjidx0C,
805 vdwioffsetptr0+vdwjidx0D,
806 vdwioffsetptr0+vdwjidx0E,
807 vdwioffsetptr0+vdwjidx0F,
808 vdwioffsetptr0+vdwjidx0G,
809 vdwioffsetptr0+vdwjidx0H,
812 /* Calculate table index by multiplying r with table scale and truncate to integer */
813 rt = _mm256_mul_ps(r00,vftabscale);
814 vfitab = _mm256_cvttps_epi32(rt);
815 vfeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
816 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
817 vfitab_lo = _mm256_extractf128_si256(vfitab,0x0);
818 vfitab_hi = _mm256_extractf128_si256(vfitab,0x1);
819 vfitab_lo = _mm_slli_epi32(vfitab_lo,3);
820 vfitab_hi = _mm_slli_epi32(vfitab_hi,3);
822 /* CUBIC SPLINE TABLE DISPERSION */
823 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
824 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
825 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
826 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
827 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
828 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
829 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
830 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
831 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
832 Heps = _mm256_mul_ps(vfeps,H);
833 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
834 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
835 fvdw6 = _mm256_mul_ps(c6_00,FF);
837 /* CUBIC SPLINE TABLE REPULSION */
838 vfitab_lo = _mm_add_epi32(vfitab_lo,ifour);
839 vfitab_hi = _mm_add_epi32(vfitab_hi,ifour);
840 Y = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,0)),
841 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,0)));
842 F = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,1)),
843 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,1)));
844 G = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,2)),
845 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,2)));
846 H = gmx_mm256_set_m128(_mm_load_ps(vftab + _mm_extract_epi32(vfitab_hi,3)),
847 _mm_load_ps(vftab + _mm_extract_epi32(vfitab_lo,3)));
848 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
849 Heps = _mm256_mul_ps(vfeps,H);
850 Fp = _mm256_add_ps(F,_mm256_mul_ps(vfeps,_mm256_add_ps(G,Heps)));
851 FF = _mm256_add_ps(Fp,_mm256_mul_ps(vfeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
852 fvdw12 = _mm256_mul_ps(c12_00,FF);
853 fvdw = _mm256_xor_ps(signbit,_mm256_mul_ps(_mm256_add_ps(fvdw6,fvdw12),_mm256_mul_ps(vftabscale,rinv00)));
857 fscal = _mm256_andnot_ps(dummy_mask,fscal);
859 /* Calculate temporary vectorial force */
860 tx = _mm256_mul_ps(fscal,dx00);
861 ty = _mm256_mul_ps(fscal,dy00);
862 tz = _mm256_mul_ps(fscal,dz00);
864 /* Update vectorial force */
865 fix0 = _mm256_add_ps(fix0,tx);
866 fiy0 = _mm256_add_ps(fiy0,ty);
867 fiz0 = _mm256_add_ps(fiz0,tz);
869 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
870 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
871 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
872 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
873 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
874 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
875 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
876 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
877 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
879 /* Inner loop uses 49 flops */
882 /* End of innermost loop */
884 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
885 f+i_coord_offset,fshift+i_shift_offset);
887 /* Increment number of inner iterations */
888 inneriter += j_index_end - j_index_start;
890 /* Outer loop uses 6 flops */
893 /* Increment number of outer iterations */
896 /* Update outer/inner flops */
898 inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*49);