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.
44 #include "../nb_kernel.h"
45 #include "gromacs/legacyheaders/types/simple.h"
46 #include "gromacs/math/vec.h"
47 #include "gromacs/legacyheaders/nrnb.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_ElecGB_VdwNone_GeomP1P1_VF_avx_256_single
54 * Electrostatics interaction: GeneralizedBorn
55 * VdW interaction: None
56 * Geometry: Particle-Particle
57 * Calculate force/pot: PotentialAndForce
60 nb_kernel_ElecGB_VdwNone_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;
93 __m256 velec,felec,velecsum,facel,crf,krf,krf2;
96 __m128i gbitab_lo,gbitab_hi;
97 __m256 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
98 __m256 minushalf = _mm256_set1_ps(-0.5);
99 real *invsqrta,*dvda,*gbtab;
101 __m128i vfitab_lo,vfitab_hi;
102 __m128i ifour = _mm_set1_epi32(4);
103 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
105 __m256 dummy_mask,cutoff_mask;
106 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
107 __m256 one = _mm256_set1_ps(1.0);
108 __m256 two = _mm256_set1_ps(2.0);
114 jindex = nlist->jindex;
116 shiftidx = nlist->shift;
118 shiftvec = fr->shift_vec[0];
119 fshift = fr->fshift[0];
120 facel = _mm256_set1_ps(fr->epsfac);
121 charge = mdatoms->chargeA;
123 invsqrta = fr->invsqrta;
125 gbtabscale = _mm256_set1_ps(fr->gbtab.scale);
126 gbtab = fr->gbtab.data;
127 gbinvepsdiff = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
129 /* Avoid stupid compiler warnings */
130 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
143 for(iidx=0;iidx<4*DIM;iidx++)
148 /* Start outer loop over neighborlists */
149 for(iidx=0; iidx<nri; iidx++)
151 /* Load shift vector for this list */
152 i_shift_offset = DIM*shiftidx[iidx];
154 /* Load limits for loop over neighbors */
155 j_index_start = jindex[iidx];
156 j_index_end = jindex[iidx+1];
158 /* Get outer coordinate index */
160 i_coord_offset = DIM*inr;
162 /* Load i particle coords and add shift vector */
163 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
165 fix0 = _mm256_setzero_ps();
166 fiy0 = _mm256_setzero_ps();
167 fiz0 = _mm256_setzero_ps();
169 /* Load parameters for i particles */
170 iq0 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
171 isai0 = _mm256_set1_ps(invsqrta[inr+0]);
173 /* Reset potential sums */
174 velecsum = _mm256_setzero_ps();
175 vgbsum = _mm256_setzero_ps();
176 dvdasum = _mm256_setzero_ps();
178 /* Start inner kernel loop */
179 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
182 /* Get j neighbor index, and coordinate index */
191 j_coord_offsetA = DIM*jnrA;
192 j_coord_offsetB = DIM*jnrB;
193 j_coord_offsetC = DIM*jnrC;
194 j_coord_offsetD = DIM*jnrD;
195 j_coord_offsetE = DIM*jnrE;
196 j_coord_offsetF = DIM*jnrF;
197 j_coord_offsetG = DIM*jnrG;
198 j_coord_offsetH = DIM*jnrH;
200 /* load j atom coordinates */
201 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
202 x+j_coord_offsetC,x+j_coord_offsetD,
203 x+j_coord_offsetE,x+j_coord_offsetF,
204 x+j_coord_offsetG,x+j_coord_offsetH,
207 /* Calculate displacement vector */
208 dx00 = _mm256_sub_ps(ix0,jx0);
209 dy00 = _mm256_sub_ps(iy0,jy0);
210 dz00 = _mm256_sub_ps(iz0,jz0);
212 /* Calculate squared distance and things based on it */
213 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
215 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
217 /* Load parameters for j particles */
218 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
219 charge+jnrC+0,charge+jnrD+0,
220 charge+jnrE+0,charge+jnrF+0,
221 charge+jnrG+0,charge+jnrH+0);
222 isaj0 = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
223 invsqrta+jnrC+0,invsqrta+jnrD+0,
224 invsqrta+jnrE+0,invsqrta+jnrF+0,
225 invsqrta+jnrG+0,invsqrta+jnrH+0);
227 /**************************
228 * CALCULATE INTERACTIONS *
229 **************************/
231 r00 = _mm256_mul_ps(rsq00,rinv00);
233 /* Compute parameters for interactions between i and j atoms */
234 qq00 = _mm256_mul_ps(iq0,jq0);
236 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
237 isaprod = _mm256_mul_ps(isai0,isaj0);
238 gbqqfactor = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
239 gbscale = _mm256_mul_ps(isaprod,gbtabscale);
241 /* Calculate generalized born table index - this is a separate table from the normal one,
242 * but we use the same procedure by multiplying r with scale and truncating to integer.
244 rt = _mm256_mul_ps(r00,gbscale);
245 gbitab = _mm256_cvttps_epi32(rt);
246 gbeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
247 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
248 gbitab_lo = _mm256_extractf128_si256(gbitab,0x0);
249 gbitab_hi = _mm256_extractf128_si256(gbitab,0x1);
250 gbitab_lo = _mm_slli_epi32(gbitab_lo,2);
251 gbitab_hi = _mm_slli_epi32(gbitab_hi,2);
252 Y = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
253 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
254 F = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
255 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
256 G = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
257 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
258 H = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
259 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
260 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
261 Heps = _mm256_mul_ps(gbeps,H);
262 Fp = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
263 VV = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
264 vgb = _mm256_mul_ps(gbqqfactor,VV);
266 FF = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
267 fgb = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
268 dvdatmp = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
269 dvdasum = _mm256_add_ps(dvdasum,dvdatmp);
278 gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
279 _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
280 velec = _mm256_mul_ps(qq00,rinv00);
281 felec = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
283 /* Update potential sum for this i atom from the interaction with this j atom. */
284 velecsum = _mm256_add_ps(velecsum,velec);
285 vgbsum = _mm256_add_ps(vgbsum,vgb);
289 /* Calculate temporary vectorial force */
290 tx = _mm256_mul_ps(fscal,dx00);
291 ty = _mm256_mul_ps(fscal,dy00);
292 tz = _mm256_mul_ps(fscal,dz00);
294 /* Update vectorial force */
295 fix0 = _mm256_add_ps(fix0,tx);
296 fiy0 = _mm256_add_ps(fiy0,ty);
297 fiz0 = _mm256_add_ps(fiz0,tz);
299 fjptrA = f+j_coord_offsetA;
300 fjptrB = f+j_coord_offsetB;
301 fjptrC = f+j_coord_offsetC;
302 fjptrD = f+j_coord_offsetD;
303 fjptrE = f+j_coord_offsetE;
304 fjptrF = f+j_coord_offsetF;
305 fjptrG = f+j_coord_offsetG;
306 fjptrH = f+j_coord_offsetH;
307 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
309 /* Inner loop uses 57 flops */
315 /* Get j neighbor index, and coordinate index */
316 jnrlistA = jjnr[jidx];
317 jnrlistB = jjnr[jidx+1];
318 jnrlistC = jjnr[jidx+2];
319 jnrlistD = jjnr[jidx+3];
320 jnrlistE = jjnr[jidx+4];
321 jnrlistF = jjnr[jidx+5];
322 jnrlistG = jjnr[jidx+6];
323 jnrlistH = jjnr[jidx+7];
324 /* Sign of each element will be negative for non-real atoms.
325 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
326 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
328 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
329 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
331 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
332 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
333 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
334 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
335 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
336 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
337 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
338 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
339 j_coord_offsetA = DIM*jnrA;
340 j_coord_offsetB = DIM*jnrB;
341 j_coord_offsetC = DIM*jnrC;
342 j_coord_offsetD = DIM*jnrD;
343 j_coord_offsetE = DIM*jnrE;
344 j_coord_offsetF = DIM*jnrF;
345 j_coord_offsetG = DIM*jnrG;
346 j_coord_offsetH = DIM*jnrH;
348 /* load j atom coordinates */
349 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
350 x+j_coord_offsetC,x+j_coord_offsetD,
351 x+j_coord_offsetE,x+j_coord_offsetF,
352 x+j_coord_offsetG,x+j_coord_offsetH,
355 /* Calculate displacement vector */
356 dx00 = _mm256_sub_ps(ix0,jx0);
357 dy00 = _mm256_sub_ps(iy0,jy0);
358 dz00 = _mm256_sub_ps(iz0,jz0);
360 /* Calculate squared distance and things based on it */
361 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
363 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
365 /* Load parameters for j particles */
366 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
367 charge+jnrC+0,charge+jnrD+0,
368 charge+jnrE+0,charge+jnrF+0,
369 charge+jnrG+0,charge+jnrH+0);
370 isaj0 = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
371 invsqrta+jnrC+0,invsqrta+jnrD+0,
372 invsqrta+jnrE+0,invsqrta+jnrF+0,
373 invsqrta+jnrG+0,invsqrta+jnrH+0);
375 /**************************
376 * CALCULATE INTERACTIONS *
377 **************************/
379 r00 = _mm256_mul_ps(rsq00,rinv00);
380 r00 = _mm256_andnot_ps(dummy_mask,r00);
382 /* Compute parameters for interactions between i and j atoms */
383 qq00 = _mm256_mul_ps(iq0,jq0);
385 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
386 isaprod = _mm256_mul_ps(isai0,isaj0);
387 gbqqfactor = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
388 gbscale = _mm256_mul_ps(isaprod,gbtabscale);
390 /* Calculate generalized born table index - this is a separate table from the normal one,
391 * but we use the same procedure by multiplying r with scale and truncating to integer.
393 rt = _mm256_mul_ps(r00,gbscale);
394 gbitab = _mm256_cvttps_epi32(rt);
395 gbeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
396 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
397 gbitab_lo = _mm256_extractf128_si256(gbitab,0x0);
398 gbitab_hi = _mm256_extractf128_si256(gbitab,0x1);
399 gbitab_lo = _mm_slli_epi32(gbitab_lo,2);
400 gbitab_hi = _mm_slli_epi32(gbitab_hi,2);
401 Y = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
402 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
403 F = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
404 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
405 G = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
406 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
407 H = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
408 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
409 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
410 Heps = _mm256_mul_ps(gbeps,H);
411 Fp = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
412 VV = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
413 vgb = _mm256_mul_ps(gbqqfactor,VV);
415 FF = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
416 fgb = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
417 dvdatmp = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
418 dvdatmp = _mm256_andnot_ps(dummy_mask,dvdatmp);
419 dvdasum = _mm256_add_ps(dvdasum,dvdatmp);
420 /* 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. */
421 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
422 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
423 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
424 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
425 fjptrE = (jnrlistE>=0) ? dvda+jnrE : scratch;
426 fjptrF = (jnrlistF>=0) ? dvda+jnrF : scratch;
427 fjptrG = (jnrlistG>=0) ? dvda+jnrG : scratch;
428 fjptrH = (jnrlistH>=0) ? dvda+jnrH : scratch;
429 gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
430 _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
431 velec = _mm256_mul_ps(qq00,rinv00);
432 felec = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
434 /* Update potential sum for this i atom from the interaction with this j atom. */
435 velec = _mm256_andnot_ps(dummy_mask,velec);
436 velecsum = _mm256_add_ps(velecsum,velec);
437 vgb = _mm256_andnot_ps(dummy_mask,vgb);
438 vgbsum = _mm256_add_ps(vgbsum,vgb);
442 fscal = _mm256_andnot_ps(dummy_mask,fscal);
444 /* Calculate temporary vectorial force */
445 tx = _mm256_mul_ps(fscal,dx00);
446 ty = _mm256_mul_ps(fscal,dy00);
447 tz = _mm256_mul_ps(fscal,dz00);
449 /* Update vectorial force */
450 fix0 = _mm256_add_ps(fix0,tx);
451 fiy0 = _mm256_add_ps(fiy0,ty);
452 fiz0 = _mm256_add_ps(fiz0,tz);
454 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
455 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
456 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
457 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
458 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
459 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
460 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
461 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
462 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
464 /* Inner loop uses 58 flops */
467 /* End of innermost loop */
469 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
470 f+i_coord_offset,fshift+i_shift_offset);
473 /* Update potential energies */
474 gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
475 gmx_mm256_update_1pot_ps(vgbsum,kernel_data->energygrp_polarization+ggid);
476 dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
477 gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
479 /* Increment number of inner iterations */
480 inneriter += j_index_end - j_index_start;
482 /* Outer loop uses 9 flops */
485 /* Increment number of outer iterations */
488 /* Update outer/inner flops */
490 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*9 + inneriter*58);
493 * Gromacs nonbonded kernel: nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_single
494 * Electrostatics interaction: GeneralizedBorn
495 * VdW interaction: None
496 * Geometry: Particle-Particle
497 * Calculate force/pot: Force
500 nb_kernel_ElecGB_VdwNone_GeomP1P1_F_avx_256_single
501 (t_nblist * gmx_restrict nlist,
502 rvec * gmx_restrict xx,
503 rvec * gmx_restrict ff,
504 t_forcerec * gmx_restrict fr,
505 t_mdatoms * gmx_restrict mdatoms,
506 nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
507 t_nrnb * gmx_restrict nrnb)
509 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
510 * just 0 for non-waters.
511 * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
512 * jnr indices corresponding to data put in the four positions in the SIMD register.
514 int i_shift_offset,i_coord_offset,outeriter,inneriter;
515 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
516 int jnrA,jnrB,jnrC,jnrD;
517 int jnrE,jnrF,jnrG,jnrH;
518 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
519 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
520 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
521 int j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
522 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
524 real *shiftvec,*fshift,*x,*f;
525 real *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
527 __m256 tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
528 real * vdwioffsetptr0;
529 __m256 ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
530 int vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
531 __m256 jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
532 __m256 dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
533 __m256 velec,felec,velecsum,facel,crf,krf,krf2;
536 __m128i gbitab_lo,gbitab_hi;
537 __m256 vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
538 __m256 minushalf = _mm256_set1_ps(-0.5);
539 real *invsqrta,*dvda,*gbtab;
541 __m128i vfitab_lo,vfitab_hi;
542 __m128i ifour = _mm_set1_epi32(4);
543 __m256 rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
545 __m256 dummy_mask,cutoff_mask;
546 __m256 signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
547 __m256 one = _mm256_set1_ps(1.0);
548 __m256 two = _mm256_set1_ps(2.0);
554 jindex = nlist->jindex;
556 shiftidx = nlist->shift;
558 shiftvec = fr->shift_vec[0];
559 fshift = fr->fshift[0];
560 facel = _mm256_set1_ps(fr->epsfac);
561 charge = mdatoms->chargeA;
563 invsqrta = fr->invsqrta;
565 gbtabscale = _mm256_set1_ps(fr->gbtab.scale);
566 gbtab = fr->gbtab.data;
567 gbinvepsdiff = _mm256_set1_ps((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
569 /* Avoid stupid compiler warnings */
570 jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
583 for(iidx=0;iidx<4*DIM;iidx++)
588 /* Start outer loop over neighborlists */
589 for(iidx=0; iidx<nri; iidx++)
591 /* Load shift vector for this list */
592 i_shift_offset = DIM*shiftidx[iidx];
594 /* Load limits for loop over neighbors */
595 j_index_start = jindex[iidx];
596 j_index_end = jindex[iidx+1];
598 /* Get outer coordinate index */
600 i_coord_offset = DIM*inr;
602 /* Load i particle coords and add shift vector */
603 gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
605 fix0 = _mm256_setzero_ps();
606 fiy0 = _mm256_setzero_ps();
607 fiz0 = _mm256_setzero_ps();
609 /* Load parameters for i particles */
610 iq0 = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
611 isai0 = _mm256_set1_ps(invsqrta[inr+0]);
613 dvdasum = _mm256_setzero_ps();
615 /* Start inner kernel loop */
616 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
619 /* Get j neighbor index, and coordinate index */
628 j_coord_offsetA = DIM*jnrA;
629 j_coord_offsetB = DIM*jnrB;
630 j_coord_offsetC = DIM*jnrC;
631 j_coord_offsetD = DIM*jnrD;
632 j_coord_offsetE = DIM*jnrE;
633 j_coord_offsetF = DIM*jnrF;
634 j_coord_offsetG = DIM*jnrG;
635 j_coord_offsetH = DIM*jnrH;
637 /* load j atom coordinates */
638 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
639 x+j_coord_offsetC,x+j_coord_offsetD,
640 x+j_coord_offsetE,x+j_coord_offsetF,
641 x+j_coord_offsetG,x+j_coord_offsetH,
644 /* Calculate displacement vector */
645 dx00 = _mm256_sub_ps(ix0,jx0);
646 dy00 = _mm256_sub_ps(iy0,jy0);
647 dz00 = _mm256_sub_ps(iz0,jz0);
649 /* Calculate squared distance and things based on it */
650 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
652 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
654 /* Load parameters for j particles */
655 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
656 charge+jnrC+0,charge+jnrD+0,
657 charge+jnrE+0,charge+jnrF+0,
658 charge+jnrG+0,charge+jnrH+0);
659 isaj0 = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
660 invsqrta+jnrC+0,invsqrta+jnrD+0,
661 invsqrta+jnrE+0,invsqrta+jnrF+0,
662 invsqrta+jnrG+0,invsqrta+jnrH+0);
664 /**************************
665 * CALCULATE INTERACTIONS *
666 **************************/
668 r00 = _mm256_mul_ps(rsq00,rinv00);
670 /* Compute parameters for interactions between i and j atoms */
671 qq00 = _mm256_mul_ps(iq0,jq0);
673 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
674 isaprod = _mm256_mul_ps(isai0,isaj0);
675 gbqqfactor = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
676 gbscale = _mm256_mul_ps(isaprod,gbtabscale);
678 /* Calculate generalized born table index - this is a separate table from the normal one,
679 * but we use the same procedure by multiplying r with scale and truncating to integer.
681 rt = _mm256_mul_ps(r00,gbscale);
682 gbitab = _mm256_cvttps_epi32(rt);
683 gbeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
684 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
685 gbitab_lo = _mm256_extractf128_si256(gbitab,0x0);
686 gbitab_hi = _mm256_extractf128_si256(gbitab,0x1);
687 gbitab_lo = _mm_slli_epi32(gbitab_lo,2);
688 gbitab_hi = _mm_slli_epi32(gbitab_hi,2);
689 Y = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
690 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
691 F = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
692 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
693 G = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
694 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
695 H = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
696 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
697 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
698 Heps = _mm256_mul_ps(gbeps,H);
699 Fp = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
700 VV = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
701 vgb = _mm256_mul_ps(gbqqfactor,VV);
703 FF = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
704 fgb = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
705 dvdatmp = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
706 dvdasum = _mm256_add_ps(dvdasum,dvdatmp);
715 gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
716 _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
717 velec = _mm256_mul_ps(qq00,rinv00);
718 felec = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
722 /* Calculate temporary vectorial force */
723 tx = _mm256_mul_ps(fscal,dx00);
724 ty = _mm256_mul_ps(fscal,dy00);
725 tz = _mm256_mul_ps(fscal,dz00);
727 /* Update vectorial force */
728 fix0 = _mm256_add_ps(fix0,tx);
729 fiy0 = _mm256_add_ps(fiy0,ty);
730 fiz0 = _mm256_add_ps(fiz0,tz);
732 fjptrA = f+j_coord_offsetA;
733 fjptrB = f+j_coord_offsetB;
734 fjptrC = f+j_coord_offsetC;
735 fjptrD = f+j_coord_offsetD;
736 fjptrE = f+j_coord_offsetE;
737 fjptrF = f+j_coord_offsetF;
738 fjptrG = f+j_coord_offsetG;
739 fjptrH = f+j_coord_offsetH;
740 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
742 /* Inner loop uses 55 flops */
748 /* Get j neighbor index, and coordinate index */
749 jnrlistA = jjnr[jidx];
750 jnrlistB = jjnr[jidx+1];
751 jnrlistC = jjnr[jidx+2];
752 jnrlistD = jjnr[jidx+3];
753 jnrlistE = jjnr[jidx+4];
754 jnrlistF = jjnr[jidx+5];
755 jnrlistG = jjnr[jidx+6];
756 jnrlistH = jjnr[jidx+7];
757 /* Sign of each element will be negative for non-real atoms.
758 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
759 * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
761 dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
762 gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
764 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
765 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
766 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
767 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
768 jnrE = (jnrlistE>=0) ? jnrlistE : 0;
769 jnrF = (jnrlistF>=0) ? jnrlistF : 0;
770 jnrG = (jnrlistG>=0) ? jnrlistG : 0;
771 jnrH = (jnrlistH>=0) ? jnrlistH : 0;
772 j_coord_offsetA = DIM*jnrA;
773 j_coord_offsetB = DIM*jnrB;
774 j_coord_offsetC = DIM*jnrC;
775 j_coord_offsetD = DIM*jnrD;
776 j_coord_offsetE = DIM*jnrE;
777 j_coord_offsetF = DIM*jnrF;
778 j_coord_offsetG = DIM*jnrG;
779 j_coord_offsetH = DIM*jnrH;
781 /* load j atom coordinates */
782 gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
783 x+j_coord_offsetC,x+j_coord_offsetD,
784 x+j_coord_offsetE,x+j_coord_offsetF,
785 x+j_coord_offsetG,x+j_coord_offsetH,
788 /* Calculate displacement vector */
789 dx00 = _mm256_sub_ps(ix0,jx0);
790 dy00 = _mm256_sub_ps(iy0,jy0);
791 dz00 = _mm256_sub_ps(iz0,jz0);
793 /* Calculate squared distance and things based on it */
794 rsq00 = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
796 rinv00 = gmx_mm256_invsqrt_ps(rsq00);
798 /* Load parameters for j particles */
799 jq0 = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
800 charge+jnrC+0,charge+jnrD+0,
801 charge+jnrE+0,charge+jnrF+0,
802 charge+jnrG+0,charge+jnrH+0);
803 isaj0 = gmx_mm256_load_8real_swizzle_ps(invsqrta+jnrA+0,invsqrta+jnrB+0,
804 invsqrta+jnrC+0,invsqrta+jnrD+0,
805 invsqrta+jnrE+0,invsqrta+jnrF+0,
806 invsqrta+jnrG+0,invsqrta+jnrH+0);
808 /**************************
809 * CALCULATE INTERACTIONS *
810 **************************/
812 r00 = _mm256_mul_ps(rsq00,rinv00);
813 r00 = _mm256_andnot_ps(dummy_mask,r00);
815 /* Compute parameters for interactions between i and j atoms */
816 qq00 = _mm256_mul_ps(iq0,jq0);
818 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
819 isaprod = _mm256_mul_ps(isai0,isaj0);
820 gbqqfactor = _mm256_xor_ps(signbit,_mm256_mul_ps(qq00,_mm256_mul_ps(isaprod,gbinvepsdiff)));
821 gbscale = _mm256_mul_ps(isaprod,gbtabscale);
823 /* Calculate generalized born table index - this is a separate table from the normal one,
824 * but we use the same procedure by multiplying r with scale and truncating to integer.
826 rt = _mm256_mul_ps(r00,gbscale);
827 gbitab = _mm256_cvttps_epi32(rt);
828 gbeps = _mm256_sub_ps(rt,_mm256_round_ps(rt, _MM_FROUND_FLOOR));
829 /* AVX1 does not support 256-bit integer operations, so now we go to 128-bit mode... */
830 gbitab_lo = _mm256_extractf128_si256(gbitab,0x0);
831 gbitab_hi = _mm256_extractf128_si256(gbitab,0x1);
832 gbitab_lo = _mm_slli_epi32(gbitab_lo,2);
833 gbitab_hi = _mm_slli_epi32(gbitab_hi,2);
834 Y = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,0)),
835 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,0)));
836 F = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,1)),
837 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,1)));
838 G = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,2)),
839 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,2)));
840 H = gmx_mm256_set_m128(_mm_load_ps(gbtab + _mm_extract_epi32(gbitab_hi,3)),
841 _mm_load_ps(gbtab + _mm_extract_epi32(gbitab_lo,3)));
842 GMX_MM256_HALFTRANSPOSE4_PS(Y,F,G,H);
843 Heps = _mm256_mul_ps(gbeps,H);
844 Fp = _mm256_add_ps(F,_mm256_mul_ps(gbeps,_mm256_add_ps(G,Heps)));
845 VV = _mm256_add_ps(Y,_mm256_mul_ps(gbeps,Fp));
846 vgb = _mm256_mul_ps(gbqqfactor,VV);
848 FF = _mm256_add_ps(Fp,_mm256_mul_ps(gbeps,_mm256_add_ps(G,_mm256_add_ps(Heps,Heps))));
849 fgb = _mm256_mul_ps(gbqqfactor,_mm256_mul_ps(FF,gbscale));
850 dvdatmp = _mm256_mul_ps(minushalf,_mm256_add_ps(vgb,_mm256_mul_ps(fgb,r00)));
851 dvdatmp = _mm256_andnot_ps(dummy_mask,dvdatmp);
852 dvdasum = _mm256_add_ps(dvdasum,dvdatmp);
853 /* 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. */
854 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
855 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
856 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
857 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
858 fjptrE = (jnrlistE>=0) ? dvda+jnrE : scratch;
859 fjptrF = (jnrlistF>=0) ? dvda+jnrF : scratch;
860 fjptrG = (jnrlistG>=0) ? dvda+jnrG : scratch;
861 fjptrH = (jnrlistH>=0) ? dvda+jnrH : scratch;
862 gmx_mm256_increment_8real_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,
863 _mm256_mul_ps(dvdatmp,_mm256_mul_ps(isaj0,isaj0)));
864 velec = _mm256_mul_ps(qq00,rinv00);
865 felec = _mm256_mul_ps(_mm256_sub_ps(_mm256_mul_ps(velec,rinv00),fgb),rinv00);
869 fscal = _mm256_andnot_ps(dummy_mask,fscal);
871 /* Calculate temporary vectorial force */
872 tx = _mm256_mul_ps(fscal,dx00);
873 ty = _mm256_mul_ps(fscal,dy00);
874 tz = _mm256_mul_ps(fscal,dz00);
876 /* Update vectorial force */
877 fix0 = _mm256_add_ps(fix0,tx);
878 fiy0 = _mm256_add_ps(fiy0,ty);
879 fiz0 = _mm256_add_ps(fiz0,tz);
881 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
882 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
883 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
884 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
885 fjptrE = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
886 fjptrF = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
887 fjptrG = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
888 fjptrH = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
889 gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
891 /* Inner loop uses 56 flops */
894 /* End of innermost loop */
896 gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
897 f+i_coord_offset,fshift+i_shift_offset);
899 dvdasum = _mm256_mul_ps(dvdasum, _mm256_mul_ps(isai0,isai0));
900 gmx_mm256_update_1pot_ps(dvdasum,dvda+inr);
902 /* Increment number of inner iterations */
903 inneriter += j_index_end - j_index_start;
905 /* Outer loop uses 7 flops */
908 /* Increment number of outer iterations */
911 /* Update outer/inner flops */
913 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*56);