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 "gromacs/legacyheaders/types/simple.h"
44 #include "gromacs/math/vec.h"
45 #include "gromacs/legacyheaders/nrnb.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_ElecGB_VdwNone_GeomP1P1_VF_avx_256_single
52 * Electrostatics interaction: GeneralizedBorn
53 * VdW interaction: None
54 * Geometry: Particle-Particle
55 * Calculate force/pot: PotentialAndForce
58 nb_kernel_ElecGB_VdwNone_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;
91 __m256 velec,felec,velecsum,facel,crf,krf,krf2;
94 __m128i gbitab_lo,gbitab_hi;
95 __m256 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
96 __m256 minushalf = _mm256_set1_ps(-0.5);
97 real *invsqrta,*dvda,*gbtab;
99 __m128i vfitab_lo,vfitab_hi;
100 __m128i ifour = _mm_set1_epi32(4);
101 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
103 __m256 dummy_mask,cutoff_mask;
104 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
105 __m256 one = _mm256_set1_ps(1.0);
106 __m256 two = _mm256_set1_ps(2.0);
112 jindex = nlist->jindex;
114 shiftidx = nlist->shift;
116 shiftvec = fr->shift_vec[0];
117 fshift = fr->fshift[0];
118 facel = _mm256_set1_ps(fr->epsfac);
119 charge = mdatoms->chargeA;
121 invsqrta = fr->invsqrta;
123 gbtabscale = _mm256_set1_ps(fr->gbtab.scale);
124 gbtab = fr->gbtab.data;
125 gbinvepsdiff = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
127 /* Avoid stupid compiler warnings */
128 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
141 for(iidx=0;iidx<4*DIM;iidx++)
146 /* Start outer loop over neighborlists */
147 for(iidx=0; iidx<nri; iidx++)
149 /* Load shift vector for this list */
150 i_shift_offset = DIM*shiftidx[iidx];
152 /* Load limits for loop over neighbors */
153 j_index_start = jindex[iidx];
154 j_index_end = jindex[iidx+1];
156 /* Get outer coordinate index */
158 i_coord_offset = DIM*inr;
160 /* Load i particle coords and add shift vector */
161 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
163 fix0 = _mm256_setzero_ps();
164 fiy0 = _mm256_setzero_ps();
165 fiz0 = _mm256_setzero_ps();
167 /* Load parameters for i particles */
168 iq0 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
169 isai0 = _mm256_set1_ps(invsqrta[inr+0]);
171 /* Reset potential sums */
172 velecsum = _mm256_setzero_ps();
173 vgbsum = _mm256_setzero_ps();
174 dvdasum = _mm256_setzero_ps();
176 /* Start inner kernel loop */
177 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
180 /* Get j neighbor index, and coordinate index */
189 j_coord_offsetA = DIM*jnrA;
190 j_coord_offsetB = DIM*jnrB;
191 j_coord_offsetC = DIM*jnrC;
192 j_coord_offsetD = DIM*jnrD;
193 j_coord_offsetE = DIM*jnrE;
194 j_coord_offsetF = DIM*jnrF;
195 j_coord_offsetG = DIM*jnrG;
196 j_coord_offsetH = DIM*jnrH;
198 /* load j atom coordinates */
199 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
200 x+j_coord_offsetC,x+j_coord_offsetD,
201 x+j_coord_offsetE,x+j_coord_offsetF,
202 x+j_coord_offsetG,x+j_coord_offsetH,
205 /* Calculate displacement vector */
206 dx00 = _mm256_sub_ps(ix0,jx0);
207 dy00 = _mm256_sub_ps(iy0,jy0);
208 dz00 = _mm256_sub_ps(iz0,jz0);
210 /* Calculate squared distance and things based on it */
211 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
213 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
215 /* Load parameters for j particles */
216 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
217 charge+jnrC+0,charge+jnrD+0,
218 charge+jnrE+0,charge+jnrF+0,
219 charge+jnrG+0,charge+jnrH+0);
220 isaj0 = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
221 invsqrta+jnrC+0,invsqrta+jnrD+0,
222 invsqrta+jnrE+0,invsqrta+jnrF+0,
223 invsqrta+jnrG+0,invsqrta+jnrH+0);
225 /**************************
226 * CALCULATE INTERACTIONS *
227 **************************/
229 r00 = _mm256_mul_ps(rsq00,rinv00);
231 /* Compute parameters for interactions between i and j atoms */
232 qq00 = _mm256_mul_ps(iq0,jq0);
234 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
235 isaprod = _mm256_mul_ps(isai0,isaj0);
236 gbqqfactor = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
237 gbscale = _mm256_mul_ps(isaprod,gbtabscale);
239 /* Calculate generalized born table index - this is a separate table from the normal one,
240 * but we use the same procedure by multiplying r with scale and truncating to integer.
242 rt = _mm256_mul_ps(r00,gbscale);
243 gbitab = _mm256_cvttps_epi32(rt);
244 gbeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
245 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
246 gbitab_lo = _mm256_extractf128_si256(gbitab,0x0);
247 gbitab_hi = _mm256_extractf128_si256(gbitab,0x1);
248 gbitab_lo = _mm_slli_epi32(gbitab_lo,2);
249 gbitab_hi = _mm_slli_epi32(gbitab_hi,2);
250 Y = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
251 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
252 F = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
253 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
254 G = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
255 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
256 H = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
257 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
258 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
259 Heps = _mm256_mul_ps(gbeps,H);
260 Fp = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
261 VV = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
262 vgb = _mm256_mul_ps(gbqqfactor,VV);
264 FF = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
265 fgb = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
266 dvdatmp = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
267 dvdasum = _mm256_add_ps(dvdasum,dvdatmp);
276 gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
277 _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
278 velec = _mm256_mul_ps(qq00,rinv00);
279 felec = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
281 /* Update potential sum for this i atom from the interaction with this j atom. */
282 velecsum = _mm256_add_ps(velecsum,velec);
283 vgbsum = _mm256_add_ps(vgbsum,vgb);
287 /* Calculate temporary vectorial force */
288 tx = _mm256_mul_ps(fscal,dx00);
289 ty = _mm256_mul_ps(fscal,dy00);
290 tz = _mm256_mul_ps(fscal,dz00);
292 /* Update vectorial force */
293 fix0 = _mm256_add_ps(fix0,tx);
294 fiy0 = _mm256_add_ps(fiy0,ty);
295 fiz0 = _mm256_add_ps(fiz0,tz);
297 fjptrA = f+j_coord_offsetA;
298 fjptrB = f+j_coord_offsetB;
299 fjptrC = f+j_coord_offsetC;
300 fjptrD = f+j_coord_offsetD;
301 fjptrE = f+j_coord_offsetE;
302 fjptrF = f+j_coord_offsetF;
303 fjptrG = f+j_coord_offsetG;
304 fjptrH = f+j_coord_offsetH;
305 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
307 /* Inner loop uses 57 flops */
313 /* Get j neighbor index, and coordinate index */
314 jnrlistA = jjnr[jidx];
315 jnrlistB = jjnr[jidx+1];
316 jnrlistC = jjnr[jidx+2];
317 jnrlistD = jjnr[jidx+3];
318 jnrlistE = jjnr[jidx+4];
319 jnrlistF = jjnr[jidx+5];
320 jnrlistG = jjnr[jidx+6];
321 jnrlistH = jjnr[jidx+7];
322 /* Sign of each element will be negative for non-real atoms.
323 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
324 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
326 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
327 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
329 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
330 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
331 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
332 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
333 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
334 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
335 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
336 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
337 j_coord_offsetA = DIM*jnrA;
338 j_coord_offsetB = DIM*jnrB;
339 j_coord_offsetC = DIM*jnrC;
340 j_coord_offsetD = DIM*jnrD;
341 j_coord_offsetE = DIM*jnrE;
342 j_coord_offsetF = DIM*jnrF;
343 j_coord_offsetG = DIM*jnrG;
344 j_coord_offsetH = DIM*jnrH;
346 /* load j atom coordinates */
347 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
348 x+j_coord_offsetC,x+j_coord_offsetD,
349 x+j_coord_offsetE,x+j_coord_offsetF,
350 x+j_coord_offsetG,x+j_coord_offsetH,
353 /* Calculate displacement vector */
354 dx00 = _mm256_sub_ps(ix0,jx0);
355 dy00 = _mm256_sub_ps(iy0,jy0);
356 dz00 = _mm256_sub_ps(iz0,jz0);
358 /* Calculate squared distance and things based on it */
359 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
361 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
363 /* Load parameters for j particles */
364 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
365 charge+jnrC+0,charge+jnrD+0,
366 charge+jnrE+0,charge+jnrF+0,
367 charge+jnrG+0,charge+jnrH+0);
368 isaj0 = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
369 invsqrta+jnrC+0,invsqrta+jnrD+0,
370 invsqrta+jnrE+0,invsqrta+jnrF+0,
371 invsqrta+jnrG+0,invsqrta+jnrH+0);
373 /**************************
374 * CALCULATE INTERACTIONS *
375 **************************/
377 r00 = _mm256_mul_ps(rsq00,rinv00);
378 r00 = _mm256_andnot_ps(dummy_mask,r00);
380 /* Compute parameters for interactions between i and j atoms */
381 qq00 = _mm256_mul_ps(iq0,jq0);
383 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
384 isaprod = _mm256_mul_ps(isai0,isaj0);
385 gbqqfactor = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
386 gbscale = _mm256_mul_ps(isaprod,gbtabscale);
388 /* Calculate generalized born table index - this is a separate table from the normal one,
389 * but we use the same procedure by multiplying r with scale and truncating to integer.
391 rt = _mm256_mul_ps(r00,gbscale);
392 gbitab = _mm256_cvttps_epi32(rt);
393 gbeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
394 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
395 gbitab_lo = _mm256_extractf128_si256(gbitab,0x0);
396 gbitab_hi = _mm256_extractf128_si256(gbitab,0x1);
397 gbitab_lo = _mm_slli_epi32(gbitab_lo,2);
398 gbitab_hi = _mm_slli_epi32(gbitab_hi,2);
399 Y = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
400 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
401 F = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
402 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
403 G = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
404 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
405 H = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
406 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
407 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
408 Heps = _mm256_mul_ps(gbeps,H);
409 Fp = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
410 VV = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
411 vgb = _mm256_mul_ps(gbqqfactor,VV);
413 FF = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
414 fgb = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
415 dvdatmp = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
416 dvdatmp = _mm256_andnot_ps(dummy_mask,dvdatmp);
417 dvdasum = _mm256_add_ps(dvdasum,dvdatmp);
418 /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
419 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
420 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
421 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
422 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
423 fjptrE = (jnrlistE>=0) ? dvda+jnrE : scratch;
424 fjptrF = (jnrlistF>=0) ? dvda+jnrF : scratch;
425 fjptrG = (jnrlistG>=0) ? dvda+jnrG : scratch;
426 fjptrH = (jnrlistH>=0) ? dvda+jnrH : scratch;
427 gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
428 _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
429 velec = _mm256_mul_ps(qq00,rinv00);
430 felec = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
432 /* Update potential sum for this i atom from the interaction with this j atom. */
433 velec = _mm256_andnot_ps(dummy_mask,velec);
434 velecsum = _mm256_add_ps(velecsum,velec);
435 vgb = _mm256_andnot_ps(dummy_mask,vgb);
436 vgbsum = _mm256_add_ps(vgbsum,vgb);
440 fscal = _mm256_andnot_ps(dummy_mask,fscal);
442 /* Calculate temporary vectorial force */
443 tx = _mm256_mul_ps(fscal,dx00);
444 ty = _mm256_mul_ps(fscal,dy00);
445 tz = _mm256_mul_ps(fscal,dz00);
447 /* Update vectorial force */
448 fix0 = _mm256_add_ps(fix0,tx);
449 fiy0 = _mm256_add_ps(fiy0,ty);
450 fiz0 = _mm256_add_ps(fiz0,tz);
452 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
453 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
454 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
455 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
456 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
457 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
458 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
459 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
460 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
462 /* Inner loop uses 58 flops */
465 /* End of innermost loop */
467 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
468 f+i_coord_offset,fshift+i_shift_offset);
471 /* Update potential energies */
472 gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
473 gmx_mm256_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
474 dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
475 gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
477 /* Increment number of inner iterations */
478 inneriter += j_index_end - j_index_start;
480 /* Outer loop uses 9 flops */
483 /* Increment number of outer iterations */
486 /* Update outer/inner flops */
488 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*58);
491 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_single
492 * Electrostatics interaction: GeneralizedBorn
493 * VdW interaction: None
494 * Geometry: Particle-Particle
495 * Calculate force/pot: Force
498 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_single
499 (t_nblist * gmx_restrict nlist,
500 rvec * gmx_restrict xx,
501 rvec * gmx_restrict ff,
502 t_forcerec * gmx_restrict fr,
503 t_mdatoms * gmx_restrict mdatoms,
504 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
505 t_nrnb * gmx_restrict nrnb)
507 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
508 * just 0 for non-waters.
509 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
510 * jnr indices corresponding to data put in the four positions in the SIMD register.
512 int i_shift_offset,i_coord_offset,outeriter,inneriter;
513 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
514 int jnrA,jnrB,jnrC,jnrD;
515 int jnrE,jnrF,jnrG,jnrH;
516 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
517 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
518 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
519 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
520 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
522 real *shiftvec,*fshift,*x,*f;
523 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
525 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
526 real * vdwioffsetptr0;
527 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
528 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
529 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
530 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
531 __m256 velec,felec,velecsum,facel,crf,krf,krf2;
534 __m128i gbitab_lo,gbitab_hi;
535 __m256 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
536 __m256 minushalf = _mm256_set1_ps(-0.5);
537 real *invsqrta,*dvda,*gbtab;
539 __m128i vfitab_lo,vfitab_hi;
540 __m128i ifour = _mm_set1_epi32(4);
541 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
543 __m256 dummy_mask,cutoff_mask;
544 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
545 __m256 one = _mm256_set1_ps(1.0);
546 __m256 two = _mm256_set1_ps(2.0);
552 jindex = nlist->jindex;
554 shiftidx = nlist->shift;
556 shiftvec = fr->shift_vec[0];
557 fshift = fr->fshift[0];
558 facel = _mm256_set1_ps(fr->epsfac);
559 charge = mdatoms->chargeA;
561 invsqrta = fr->invsqrta;
563 gbtabscale = _mm256_set1_ps(fr->gbtab.scale);
564 gbtab = fr->gbtab.data;
565 gbinvepsdiff = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
567 /* Avoid stupid compiler warnings */
568 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
581 for(iidx=0;iidx<4*DIM;iidx++)
586 /* Start outer loop over neighborlists */
587 for(iidx=0; iidx<nri; iidx++)
589 /* Load shift vector for this list */
590 i_shift_offset = DIM*shiftidx[iidx];
592 /* Load limits for loop over neighbors */
593 j_index_start = jindex[iidx];
594 j_index_end = jindex[iidx+1];
596 /* Get outer coordinate index */
598 i_coord_offset = DIM*inr;
600 /* Load i particle coords and add shift vector */
601 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
603 fix0 = _mm256_setzero_ps();
604 fiy0 = _mm256_setzero_ps();
605 fiz0 = _mm256_setzero_ps();
607 /* Load parameters for i particles */
608 iq0 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
609 isai0 = _mm256_set1_ps(invsqrta[inr+0]);
611 dvdasum = _mm256_setzero_ps();
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 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
654 charge+jnrC+0,charge+jnrD+0,
655 charge+jnrE+0,charge+jnrF+0,
656 charge+jnrG+0,charge+jnrH+0);
657 isaj0 = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
658 invsqrta+jnrC+0,invsqrta+jnrD+0,
659 invsqrta+jnrE+0,invsqrta+jnrF+0,
660 invsqrta+jnrG+0,invsqrta+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 qq00 = _mm256_mul_ps(iq0,jq0);
671 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
672 isaprod = _mm256_mul_ps(isai0,isaj0);
673 gbqqfactor = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
674 gbscale = _mm256_mul_ps(isaprod,gbtabscale);
676 /* Calculate generalized born table index - this is a separate table from the normal one,
677 * but we use the same procedure by multiplying r with scale and truncating to integer.
679 rt = _mm256_mul_ps(r00,gbscale);
680 gbitab = _mm256_cvttps_epi32(rt);
681 gbeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
682 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
683 gbitab_lo = _mm256_extractf128_si256(gbitab,0x0);
684 gbitab_hi = _mm256_extractf128_si256(gbitab,0x1);
685 gbitab_lo = _mm_slli_epi32(gbitab_lo,2);
686 gbitab_hi = _mm_slli_epi32(gbitab_hi,2);
687 Y = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
688 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
689 F = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
690 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
691 G = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
692 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
693 H = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
694 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
695 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
696 Heps = _mm256_mul_ps(gbeps,H);
697 Fp = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
698 VV = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
699 vgb = _mm256_mul_ps(gbqqfactor,VV);
701 FF = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
702 fgb = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
703 dvdatmp = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
704 dvdasum = _mm256_add_ps(dvdasum,dvdatmp);
713 gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
714 _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
715 velec = _mm256_mul_ps(qq00,rinv00);
716 felec = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
720 /* Calculate temporary vectorial force */
721 tx = _mm256_mul_ps(fscal,dx00);
722 ty = _mm256_mul_ps(fscal,dy00);
723 tz = _mm256_mul_ps(fscal,dz00);
725 /* Update vectorial force */
726 fix0 = _mm256_add_ps(fix0,tx);
727 fiy0 = _mm256_add_ps(fiy0,ty);
728 fiz0 = _mm256_add_ps(fiz0,tz);
730 fjptrA = f+j_coord_offsetA;
731 fjptrB = f+j_coord_offsetB;
732 fjptrC = f+j_coord_offsetC;
733 fjptrD = f+j_coord_offsetD;
734 fjptrE = f+j_coord_offsetE;
735 fjptrF = f+j_coord_offsetF;
736 fjptrG = f+j_coord_offsetG;
737 fjptrH = f+j_coord_offsetH;
738 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
740 /* Inner loop uses 55 flops */
746 /* Get j neighbor index, and coordinate index */
747 jnrlistA = jjnr[jidx];
748 jnrlistB = jjnr[jidx+1];
749 jnrlistC = jjnr[jidx+2];
750 jnrlistD = jjnr[jidx+3];
751 jnrlistE = jjnr[jidx+4];
752 jnrlistF = jjnr[jidx+5];
753 jnrlistG = jjnr[jidx+6];
754 jnrlistH = jjnr[jidx+7];
755 /* Sign of each element will be negative for non-real atoms.
756 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
757 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
759 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
760 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
762 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
763 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
764 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
765 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
766 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
767 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
768 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
769 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
770 j_coord_offsetA = DIM*jnrA;
771 j_coord_offsetB = DIM*jnrB;
772 j_coord_offsetC = DIM*jnrC;
773 j_coord_offsetD = DIM*jnrD;
774 j_coord_offsetE = DIM*jnrE;
775 j_coord_offsetF = DIM*jnrF;
776 j_coord_offsetG = DIM*jnrG;
777 j_coord_offsetH = DIM*jnrH;
779 /* load j atom coordinates */
780 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
781 x+j_coord_offsetC,x+j_coord_offsetD,
782 x+j_coord_offsetE,x+j_coord_offsetF,
783 x+j_coord_offsetG,x+j_coord_offsetH,
786 /* Calculate displacement vector */
787 dx00 = _mm256_sub_ps(ix0,jx0);
788 dy00 = _mm256_sub_ps(iy0,jy0);
789 dz00 = _mm256_sub_ps(iz0,jz0);
791 /* Calculate squared distance and things based on it */
792 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
794 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
796 /* Load parameters for j particles */
797 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
798 charge+jnrC+0,charge+jnrD+0,
799 charge+jnrE+0,charge+jnrF+0,
800 charge+jnrG+0,charge+jnrH+0);
801 isaj0 = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
802 invsqrta+jnrC+0,invsqrta+jnrD+0,
803 invsqrta+jnrE+0,invsqrta+jnrF+0,
804 invsqrta+jnrG+0,invsqrta+jnrH+0);
806 /**************************
807 * CALCULATE INTERACTIONS *
808 **************************/
810 r00 = _mm256_mul_ps(rsq00,rinv00);
811 r00 = _mm256_andnot_ps(dummy_mask,r00);
813 /* Compute parameters for interactions between i and j atoms */
814 qq00 = _mm256_mul_ps(iq0,jq0);
816 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
817 isaprod = _mm256_mul_ps(isai0,isaj0);
818 gbqqfactor = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
819 gbscale = _mm256_mul_ps(isaprod,gbtabscale);
821 /* Calculate generalized born table index - this is a separate table from the normal one,
822 * but we use the same procedure by multiplying r with scale and truncating to integer.
824 rt = _mm256_mul_ps(r00,gbscale);
825 gbitab = _mm256_cvttps_epi32(rt);
826 gbeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
827 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
828 gbitab_lo = _mm256_extractf128_si256(gbitab,0x0);
829 gbitab_hi = _mm256_extractf128_si256(gbitab,0x1);
830 gbitab_lo = _mm_slli_epi32(gbitab_lo,2);
831 gbitab_hi = _mm_slli_epi32(gbitab_hi,2);
832 Y = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
833 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
834 F = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
835 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
836 G = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
837 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
838 H = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
839 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
840 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
841 Heps = _mm256_mul_ps(gbeps,H);
842 Fp = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
843 VV = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
844 vgb = _mm256_mul_ps(gbqqfactor,VV);
846 FF = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
847 fgb = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
848 dvdatmp = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
849 dvdatmp = _mm256_andnot_ps(dummy_mask,dvdatmp);
850 dvdasum = _mm256_add_ps(dvdasum,dvdatmp);
851 /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
852 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
853 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
854 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
855 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
856 fjptrE = (jnrlistE>=0) ? dvda+jnrE : scratch;
857 fjptrF = (jnrlistF>=0) ? dvda+jnrF : scratch;
858 fjptrG = (jnrlistG>=0) ? dvda+jnrG : scratch;
859 fjptrH = (jnrlistH>=0) ? dvda+jnrH : scratch;
860 gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
861 _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
862 velec = _mm256_mul_ps(qq00,rinv00);
863 felec = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
867 fscal = _mm256_andnot_ps(dummy_mask,fscal);
869 /* Calculate temporary vectorial force */
870 tx = _mm256_mul_ps(fscal,dx00);
871 ty = _mm256_mul_ps(fscal,dy00);
872 tz = _mm256_mul_ps(fscal,dz00);
874 /* Update vectorial force */
875 fix0 = _mm256_add_ps(fix0,tx);
876 fiy0 = _mm256_add_ps(fiy0,ty);
877 fiz0 = _mm256_add_ps(fiz0,tz);
879 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
880 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
881 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
882 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
883 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
884 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
885 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
886 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
887 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
889 /* Inner loop uses 56 flops */
892 /* End of innermost loop */
894 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
895 f+i_coord_offset,fshift+i_shift_offset);
897 dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
898 gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
900 /* Increment number of inner iterations */
901 inneriter += j_index_end - j_index_start;
903 /* Outer loop uses 7 flops */
906 /* Increment number of outer iterations */
909 /* Update outer/inner flops */
911 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*56);