2 #error This file must be processed with the Gromacs pre-preprocessor
4 /* #if INCLUDE_HEADER */
11 #include "../nb_kernel.h"
12 #include "types/simple.h"
16 #include "gmx_math_x86_avx_256_double.h"
17 #include "kernelutil_x86_avx_256_double.h"
20 /* ## List of variables set by the generating script: */
22 /* ## Setttings that apply to the entire kernel: */
23 /* ## KERNEL_ELEC: String, choice for electrostatic interactions */
24 /* ## KERNEL_VDW: String, choice for van der Waals interactions */
25 /* ## KERNEL_NAME: String, name of this kernel */
26 /* ## KERNEL_VF: String telling if we calculate potential, force, or both */
27 /* ## GEOMETRY_I/GEOMETRY_J: String, name of each geometry, e.g. 'Water3' or '1Particle' */
29 /* ## Setttings that apply to particles in the outer (I) or inner (J) loops: */
30 /* ## PARTICLES_I[]/ Arrays with lists of i/j particles to use in kernel. It is */
31 /* ## PARTICLES_J[]: just [0] for particle geometry, but can be longer for water */
32 /* ## PARTICLES_ELEC_I[]/ Arrays with lists of i/j particle that have electrostatics */
33 /* ## PARTICLES_ELEC_J[]: interactions that should be calculated in this kernel. */
34 /* ## PARTICLES_VDW_I[]/ Arrays with the list of i/j particle that have VdW */
35 /* ## PARTICLES_VDW_J[]: interactions that should be calculated in this kernel. */
37 /* ## Setttings for pairs of interactions (e.g. 2nd i particle against 1st j particle) */
38 /* ## PAIRS_IJ[]: Array with (i,j) tuples of pairs for which interactions */
39 /* ## should be calculated in this kernel. Zero-charge particles */
40 /* ## do not have interactions with particles without vdw, and */
41 /* ## Vdw-only interactions are not evaluated in a no-vdw-kernel. */
42 /* ## INTERACTION_FLAGS[][]: 2D matrix, dimension e.g. 3*3 for water-water interactions. */
43 /* ## For each i-j pair, the element [I][J] is a list of strings */
44 /* ## defining properties/flags of this interaction. Examples */
45 /* ## include 'electrostatics'/'vdw' if that type of interaction */
46 /* ## should be evaluated, 'rsq'/'rinv'/'rinvsq' if those values */
47 /* ## are needed, and 'exactcutoff' or 'shift','switch' to */
48 /* ## decide if the force/potential should be modified. This way */
49 /* ## we only calculate values absolutely needed for each case. */
51 /* ## Calculate the size and offset for (merged/interleaved) table data */
54 * Gromacs nonbonded kernel: {KERNEL_NAME}
55 * Electrostatics interaction: {KERNEL_ELEC}
56 * VdW interaction: {KERNEL_VDW}
57 * Geometry: {GEOMETRY_I}-{GEOMETRY_J}
58 * Calculate force/pot: {KERNEL_VF}
62 (t_nblist * gmx_restrict nlist,
63 rvec * gmx_restrict xx,
64 rvec * gmx_restrict ff,
65 t_forcerec * gmx_restrict fr,
66 t_mdatoms * gmx_restrict mdatoms,
67 nb_kernel_data_t * gmx_restrict kernel_data,
68 t_nrnb * gmx_restrict nrnb)
70 /* ## Not all variables are used for all kernels, but any optimizing compiler fixes that, */
71 /* ## so there is no point in going to extremes to exclude variables that are not needed. */
72 /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
73 * just 0 for non-waters.
74 * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
75 * jnr indices corresponding to data put in the four positions in the SIMD register.
77 int i_shift_offset,i_coord_offset,outeriter,inneriter;
78 int j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
79 int jnrA,jnrB,jnrC,jnrD;
80 int jnrlistA,jnrlistB,jnrlistC,jnrlistD;
81 int jnrlistE,jnrlistF,jnrlistG,jnrlistH;
82 int j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
83 int *iinr,*jindex,*jjnr,*shiftidx,*gid;
85 real *shiftvec,*fshift,*x,*f;
86 real *fjptrA,*fjptrB,*fjptrC,*fjptrD;
88 __m256d tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
89 /* #for I in PARTICLES_I */
90 real * vdwioffsetptr{I};
91 __m256d ix{I},iy{I},iz{I},fix{I},fiy{I},fiz{I},iq{I},isai{I};
93 /* #for J in PARTICLES_J */
94 int vdwjidx{J}A,vdwjidx{J}B,vdwjidx{J}C,vdwjidx{J}D;
95 __m256d jx{J},jy{J},jz{J},fjx{J},fjy{J},fjz{J},jq{J},isaj{J};
97 /* #for I,J in PAIRS_IJ */
98 __m256d dx{I}{J},dy{I}{J},dz{I}{J},rsq{I}{J},rinv{I}{J},rinvsq{I}{J},r{I}{J},qq{I}{J},c6_{I}{J},c12_{I}{J};
100 /* #if KERNEL_ELEC != 'None' */
101 __m256d velec,felec,velecsum,facel,crf,krf,krf2;
104 /* #if 'GeneralizedBorn' in KERNEL_ELEC */
106 __m256d vgb,fgb,vgbsum,dvdasum,gbscale,gbtabscale,isaprod,gbqqfactor,gbinvepsdiff,gbeps,dvdatmp;
107 __m256d minushalf = _mm256_set1_pd(-0.5);
108 real *invsqrta,*dvda,*gbtab;
110 /* #if KERNEL_VDW != 'None' */
112 __m256d rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
115 __m256d one_sixth = _mm256_set1_pd(1.0/6.0);
116 __m256d one_twelfth = _mm256_set1_pd(1.0/12.0);
118 /* #if 'Table' in KERNEL_ELEC or 'GeneralizedBorn' in KERNEL_ELEC or 'Table' in KERNEL_VDW */
120 __m128i ifour = _mm_set1_epi32(4);
121 __m256d rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
124 /* #if 'Ewald' in KERNEL_ELEC */
126 __m256d ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
127 __m256d beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
130 /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
131 __m256d rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
132 real rswitch_scalar,d_scalar;
134 __m256d dummy_mask,cutoff_mask;
135 __m128 tmpmask0,tmpmask1;
136 __m256d signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
137 __m256d one = _mm256_set1_pd(1.0);
138 __m256d two = _mm256_set1_pd(2.0);
144 jindex = nlist->jindex;
146 shiftidx = nlist->shift;
148 shiftvec = fr->shift_vec[0];
149 fshift = fr->fshift[0];
150 /* #if KERNEL_ELEC != 'None' */
151 facel = _mm256_set1_pd(fr->epsfac);
152 charge = mdatoms->chargeA;
153 /* #if 'ReactionField' in KERNEL_ELEC */
154 krf = _mm256_set1_pd(fr->ic->k_rf);
155 krf2 = _mm256_set1_pd(fr->ic->k_rf*2.0);
156 crf = _mm256_set1_pd(fr->ic->c_rf);
159 /* #if KERNEL_VDW != 'None' */
160 nvdwtype = fr->ntype;
162 vdwtype = mdatoms->typeA;
165 /* #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW */
166 vftab = kernel_data->table_elec_vdw->data;
167 vftabscale = _mm256_set1_pd(kernel_data->table_elec_vdw->scale);
168 /* #elif 'Table' in KERNEL_ELEC */
169 vftab = kernel_data->table_elec->data;
170 vftabscale = _mm256_set1_pd(kernel_data->table_elec->scale);
171 /* #elif 'Table' in KERNEL_VDW */
172 vftab = kernel_data->table_vdw->data;
173 vftabscale = _mm256_set1_pd(kernel_data->table_vdw->scale);
176 /* #if 'Ewald' in KERNEL_ELEC */
177 sh_ewald = _mm256_set1_pd(fr->ic->sh_ewald);
178 beta = _mm256_set1_pd(fr->ic->ewaldcoeff);
179 beta2 = _mm256_mul_pd(beta,beta);
180 beta3 = _mm256_mul_pd(beta,beta2);
182 /* #if KERNEL_VF=='Force' and KERNEL_MOD_ELEC!='PotentialSwitch' */
183 ewtab = fr->ic->tabq_coul_F;
184 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
185 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
187 ewtab = fr->ic->tabq_coul_FDV0;
188 ewtabscale = _mm256_set1_pd(fr->ic->tabq_scale);
189 ewtabhalfspace = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
193 /* #if KERNEL_ELEC=='GeneralizedBorn' */
194 invsqrta = fr->invsqrta;
196 gbtabscale = _mm256_set1_pd(fr->gbtab.scale);
197 gbtab = fr->gbtab.data;
198 gbinvepsdiff = _mm256_set1_pd((1.0/fr->epsilon_r) - (1.0/fr->gb_epsilon_solvent));
201 /* #if 'Water' in GEOMETRY_I */
202 /* Setup water-specific parameters */
203 inr = nlist->iinr[0];
204 /* #for I in PARTICLES_ELEC_I */
205 iq{I} = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+{I}]));
207 /* #for I in PARTICLES_VDW_I */
208 vdwioffsetptr{I} = vdwparam+2*nvdwtype*vdwtype[inr+{I}];
212 /* #if 'Water' in GEOMETRY_J */
213 /* #for J in PARTICLES_ELEC_J */
214 jq{J} = _mm256_set1_pd(charge[inr+{J}]);
216 /* #for J in PARTICLES_VDW_J */
217 vdwjidx{J}A = 2*vdwtype[inr+{J}];
219 /* #for I,J in PAIRS_IJ */
220 /* #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
221 qq{I}{J} = _mm256_mul_pd(iq{I},jq{J});
223 /* #if 'vdw' in INTERACTION_FLAGS[I][J] */
224 c6_{I}{J} = _mm256_set1_pd(vdwioffsetptr{I}[vdwjidx{J}A]);
225 c12_{I}{J} = _mm256_set1_pd(vdwioffsetptr{I}[vdwjidx{J}A+1]);
230 /* #if KERNEL_MOD_ELEC!='None' or KERNEL_MOD_VDW!='None' */
231 /* #if KERNEL_ELEC!='None' */
232 /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
233 rcutoff_scalar = fr->rcoulomb;
235 rcutoff_scalar = fr->rvdw;
237 rcutoff = _mm256_set1_pd(rcutoff_scalar);
238 rcutoff2 = _mm256_mul_pd(rcutoff,rcutoff);
241 /* #if KERNEL_MOD_VDW=='PotentialShift' */
242 sh_vdw_invrcut6 = _mm256_set1_pd(fr->ic->sh_invrc6);
243 rvdw = _mm256_set1_pd(fr->rvdw);
246 /* #if 'PotentialSwitch' in [KERNEL_MOD_ELEC,KERNEL_MOD_VDW] */
247 /* #if KERNEL_MOD_ELEC=='PotentialSwitch' */
248 rswitch_scalar = fr->rcoulomb_switch;
249 rswitch = _mm256_set1_pd(rswitch_scalar);
251 rswitch_scalar = fr->rvdw_switch;
252 rswitch = _mm256_set1_pd(rswitch_scalar);
254 /* Setup switch parameters */
255 d_scalar = rcutoff_scalar-rswitch_scalar;
256 d = _mm256_set1_pd(d_scalar);
257 swV3 = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
258 swV4 = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
259 swV5 = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
260 /* #if 'Force' in KERNEL_VF */
261 swF2 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
262 swF3 = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
263 swF4 = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
267 /* Avoid stupid compiler warnings */
268 jnrA = jnrB = jnrC = jnrD = 0;
274 /* ## Keep track of the floating point operations we issue for reporting! */
275 /* #define OUTERFLOPS 0 */
279 for(iidx=0;iidx<4*DIM;iidx++)
284 /* Start outer loop over neighborlists */
285 for(iidx=0; iidx<nri; iidx++)
287 /* Load shift vector for this list */
288 i_shift_offset = DIM*shiftidx[iidx];
290 /* Load limits for loop over neighbors */
291 j_index_start = jindex[iidx];
292 j_index_end = jindex[iidx+1];
294 /* Get outer coordinate index */
296 i_coord_offset = DIM*inr;
298 /* Load i particle coords and add shift vector */
299 /* #if GEOMETRY_I == 'Particle' */
300 gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
301 /* #elif GEOMETRY_I == 'Water3' */
302 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
303 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
304 /* #elif GEOMETRY_I == 'Water4' */
305 /* #if 0 in PARTICLES_I */
306 gmx_mm256_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
307 &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
309 gmx_mm256_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
310 &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
314 /* #if 'Force' in KERNEL_VF */
315 /* #for I in PARTICLES_I */
316 fix{I} = _mm256_setzero_pd();
317 fiy{I} = _mm256_setzero_pd();
318 fiz{I} = _mm256_setzero_pd();
322 /* ## For water we already preloaded parameters at the start of the kernel */
323 /* #if not 'Water' in GEOMETRY_I */
324 /* Load parameters for i particles */
325 /* #for I in PARTICLES_ELEC_I */
326 iq{I} = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+{I}]));
327 /* #define OUTERFLOPS OUTERFLOPS+1 */
328 /* #if KERNEL_ELEC=='GeneralizedBorn' */
329 isai{I} = _mm256_set1_pd(invsqrta[inr+{I}]);
332 /* #for I in PARTICLES_VDW_I */
333 vdwioffsetptr{I} = vdwparam+2*nvdwtype*vdwtype[inr+{I}];
337 /* #if 'Potential' in KERNEL_VF */
338 /* Reset potential sums */
339 /* #if KERNEL_ELEC != 'None' */
340 velecsum = _mm256_setzero_pd();
342 /* #if 'GeneralizedBorn' in KERNEL_ELEC */
343 vgbsum = _mm256_setzero_pd();
345 /* #if KERNEL_VDW != 'None' */
346 vvdwsum = _mm256_setzero_pd();
349 /* #if 'GeneralizedBorn' in KERNEL_ELEC and 'Force' in KERNEL_VF */
350 dvdasum = _mm256_setzero_pd();
353 /* #for ROUND in ['Loop','Epilogue'] */
355 /* #if ROUND =='Loop' */
356 /* Start inner kernel loop */
357 for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
359 /* ## First round is normal loop (next statement resets indentation) */
366 /* ## Second round is epilogue */
368 /* #define INNERFLOPS 0 */
370 /* Get j neighbor index, and coordinate index */
371 /* #if ROUND =='Loop' */
377 jnrlistA = jjnr[jidx];
378 jnrlistB = jjnr[jidx+1];
379 jnrlistC = jjnr[jidx+2];
380 jnrlistD = jjnr[jidx+3];
381 /* Sign of each element will be negative for non-real atoms.
382 * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
383 * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
385 tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
387 tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
388 tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
389 dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
391 jnrA = (jnrlistA>=0) ? jnrlistA : 0;
392 jnrB = (jnrlistB>=0) ? jnrlistB : 0;
393 jnrC = (jnrlistC>=0) ? jnrlistC : 0;
394 jnrD = (jnrlistD>=0) ? jnrlistD : 0;
396 j_coord_offsetA = DIM*jnrA;
397 j_coord_offsetB = DIM*jnrB;
398 j_coord_offsetC = DIM*jnrC;
399 j_coord_offsetD = DIM*jnrD;
401 /* load j atom coordinates */
402 /* #if GEOMETRY_J == 'Particle' */
403 gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
404 x+j_coord_offsetC,x+j_coord_offsetD,
406 /* #elif GEOMETRY_J == 'Water3' */
407 gmx_mm256_load_3rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
408 x+j_coord_offsetC,x+j_coord_offsetD,
409 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
410 /* #elif GEOMETRY_J == 'Water4' */
411 /* #if 0 in PARTICLES_J */
412 gmx_mm256_load_4rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
413 x+j_coord_offsetC,x+j_coord_offsetD,
414 &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
415 &jy2,&jz2,&jx3,&jy3,&jz3);
417 gmx_mm256_load_3rvec_4ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
418 x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
419 &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
423 /* Calculate displacement vector */
424 /* #for I,J in PAIRS_IJ */
425 dx{I}{J} = _mm256_sub_pd(ix{I},jx{J});
426 dy{I}{J} = _mm256_sub_pd(iy{I},jy{J});
427 dz{I}{J} = _mm256_sub_pd(iz{I},jz{J});
428 /* #define INNERFLOPS INNERFLOPS+3 */
431 /* Calculate squared distance and things based on it */
432 /* #for I,J in PAIRS_IJ */
433 rsq{I}{J} = gmx_mm256_calc_rsq_pd(dx{I}{J},dy{I}{J},dz{I}{J});
434 /* #define INNERFLOPS INNERFLOPS+5 */
437 /* #for I,J in PAIRS_IJ */
438 /* #if 'rinv' in INTERACTION_FLAGS[I][J] */
439 rinv{I}{J} = gmx_mm256_invsqrt_pd(rsq{I}{J});
440 /* #define INNERFLOPS INNERFLOPS+5 */
444 /* #for I,J in PAIRS_IJ */
445 /* #if 'rinvsq' in INTERACTION_FLAGS[I][J] */
446 /* # if 'rinv' not in INTERACTION_FLAGS[I][J] */
447 rinvsq{I}{J} = gmx_mm256_inv_pd(rsq{I}{J});
448 /* #define INNERFLOPS INNERFLOPS+4 */
450 rinvsq{I}{J} = _mm256_mul_pd(rinv{I}{J},rinv{I}{J});
451 /* #define INNERFLOPS INNERFLOPS+1 */
456 /* #if not 'Water' in GEOMETRY_J */
457 /* Load parameters for j particles */
458 /* #for J in PARTICLES_ELEC_J */
459 jq{J} = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+{J},charge+jnrB+{J},
460 charge+jnrC+{J},charge+jnrD+{J});
461 /* #if KERNEL_ELEC=='GeneralizedBorn' */
462 isaj{J} = gmx_mm256_load_4real_swizzle_pd(invsqrta+jnrA+{J},invsqrta+jnrB+{J},
463 invsqrta+jnrC+{J},invsqrta+jnrD+{J});
466 /* #for J in PARTICLES_VDW_J */
467 vdwjidx{J}A = 2*vdwtype[jnrA+{J}];
468 vdwjidx{J}B = 2*vdwtype[jnrB+{J}];
469 vdwjidx{J}C = 2*vdwtype[jnrC+{J}];
470 vdwjidx{J}D = 2*vdwtype[jnrD+{J}];
474 /* #if 'Force' in KERNEL_VF and not 'Particle' in GEOMETRY_I */
475 /* #for J in PARTICLES_J */
476 fjx{J} = _mm256_setzero_pd();
477 fjy{J} = _mm256_setzero_pd();
478 fjz{J} = _mm256_setzero_pd();
482 /* #for I,J in PAIRS_IJ */
484 /**************************
485 * CALCULATE INTERACTIONS *
486 **************************/
488 /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
489 /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
490 /* ## We always calculate rinv/rinvsq above to enable pipelineing in compilers (performance tested on x86) */
491 if (gmx_mm256_any_lt(rsq{I}{J},rcutoff2))
493 /* #if 0 ## this and the next two lines is a hack to maintain auto-indentation in template file */
496 /* #define INNERFLOPS INNERFLOPS+1 */
499 /* #if 'r' in INTERACTION_FLAGS[I][J] */
500 r{I}{J} = _mm256_mul_pd(rsq{I}{J},rinv{I}{J});
501 /* #if ROUND == 'Epilogue' */
502 r{I}{J} = _mm256_andnot_pd(dummy_mask,r{I}{J});
503 /* #define INNERFLOPS INNERFLOPS+1 */
505 /* #define INNERFLOPS INNERFLOPS+1 */
508 /* ## For water geometries we already loaded parameters at the start of the kernel */
509 /* #if not 'Water' in GEOMETRY_J */
510 /* Compute parameters for interactions between i and j atoms */
511 /* #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
512 qq{I}{J} = _mm256_mul_pd(iq{I},jq{J});
513 /* #define INNERFLOPS INNERFLOPS+1 */
515 /* #if 'vdw' in INTERACTION_FLAGS[I][J] */
516 gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr{I}+vdwjidx{J}A,
517 vdwioffsetptr{I}+vdwjidx{J}B,
518 vdwioffsetptr{I}+vdwjidx{J}C,
519 vdwioffsetptr{I}+vdwjidx{J}D,
520 &c6_{I}{J},&c12_{I}{J});
524 /* #if 'table' in INTERACTION_FLAGS[I][J] */
525 /* Calculate table index by multiplying r with table scale and truncate to integer */
526 rt = _mm256_mul_pd(r{I}{J},vftabscale);
527 vfitab = _mm256_cvttpd_epi32(rt);
528 vfeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
529 /* #define INNERFLOPS INNERFLOPS+4 */
530 /* #if 'Table' in KERNEL_ELEC and 'Table' in KERNEL_VDW */
531 /* ## 3 tables, 4 bytes per point: multiply index by 12 */
532 vfitab = _mm_slli_epi32(_mm_add_epi32(vfitab,_mm_slli_epi32(vfitab,1)),2);
533 /* #elif 'Table' in KERNEL_ELEC */
534 /* ## 1 table, 4 bytes per point: multiply index by 4 */
535 vfitab = _mm_slli_epi32(vfitab,2);
536 /* #elif 'Table' in KERNEL_VDW */
537 /* ## 2 tables, 4 bytes per point: multiply index by 8 */
538 vfitab = _mm_slli_epi32(vfitab,3);
542 /* ## ELECTROSTATIC INTERACTIONS */
543 /* #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
545 /* #if KERNEL_ELEC=='Coulomb' */
547 /* COULOMB ELECTROSTATICS */
548 velec = _mm256_mul_pd(qq{I}{J},rinv{I}{J});
549 /* #define INNERFLOPS INNERFLOPS+1 */
550 /* #if 'Force' in KERNEL_VF */
551 felec = _mm256_mul_pd(velec,rinvsq{I}{J});
552 /* #define INNERFLOPS INNERFLOPS+1 */
555 /* #elif KERNEL_ELEC=='ReactionField' */
557 /* REACTION-FIELD ELECTROSTATICS */
558 /* #if 'Potential' in KERNEL_VF */
559 velec = _mm256_mul_pd(qq{I}{J},_mm256_sub_pd(_mm256_add_pd(rinv{I}{J},_mm256_mul_pd(krf,rsq{I}{J})),crf));
560 /* #define INNERFLOPS INNERFLOPS+4 */
562 /* #if 'Force' in KERNEL_VF */
563 felec = _mm256_mul_pd(qq{I}{J},_mm256_sub_pd(_mm256_mul_pd(rinv{I}{J},rinvsq{I}{J}),krf2));
564 /* #define INNERFLOPS INNERFLOPS+3 */
567 /* #elif KERNEL_ELEC=='GeneralizedBorn' */
569 /* GENERALIZED BORN AND COULOMB ELECTROSTATICS */
570 isaprod = _mm256_mul_pd(isai{I},isaj{J});
571 gbqqfactor = _mm256_xor_pd(signbit,_mm256_mul_pd(qq{I}{J},_mm256_mul_pd(isaprod,gbinvepsdiff)));
572 gbscale = _mm256_mul_pd(isaprod,gbtabscale);
573 /* #define INNERFLOPS INNERFLOPS+5 */
575 /* Calculate generalized born table index - this is a separate table from the normal one,
576 * but we use the same procedure by multiplying r with scale and truncating to integer.
578 rt = _mm256_mul_pd(r{I}{J},gbscale);
579 gbitab = _mm256_cvttpd_epi32(rt);
580 gbeps = _mm256_sub_pd(rt,_mm256_round_pd(rt, _MM_FROUND_FLOOR));
581 gbitab = _mm_slli_epi32(gbitab,2);
582 Y = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,0) );
583 F = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,1) );
584 G = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,2) );
585 H = _mm256_load_pd( gbtab + _mm_extract_epi32(gbitab,3) );
586 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
587 Heps = _mm256_mul_pd(gbeps,H);
588 Fp = _mm256_add_pd(F,_mm256_mul_pd(gbeps,_mm256_add_pd(G,Heps)));
589 VV = _mm256_add_pd(Y,_mm256_mul_pd(gbeps,Fp));
590 vgb = _mm256_mul_pd(gbqqfactor,VV);
591 /* #define INNERFLOPS INNERFLOPS+10 */
593 /* #if 'Force' in KERNEL_VF */
594 FF = _mm256_add_pd(Fp,_mm256_mul_pd(gbeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
595 fgb = _mm256_mul_pd(gbqqfactor,_mm256_mul_pd(FF,gbscale));
596 dvdatmp = _mm256_mul_pd(minushalf,_mm256_add_pd(vgb,_mm256_mul_pd(fgb,r{I}{J})));
597 dvdasum = _mm256_add_pd(dvdasum,dvdatmp);
598 /* #if ROUND == 'Loop' */
604 /* 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. */
605 fjptrA = (jnrlistA>=0) ? dvda+jnrA : scratch;
606 fjptrB = (jnrlistB>=0) ? dvda+jnrB : scratch;
607 fjptrC = (jnrlistC>=0) ? dvda+jnrC : scratch;
608 fjptrD = (jnrlistD>=0) ? dvda+jnrD : scratch;
610 gmx_mm256_increment_4real_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
611 _mm256_mul_pd(dvdatmp,_mm256_mul_pd(isaj{J},isaj{J})));
612 /* #define INNERFLOPS INNERFLOPS+12 */
614 velec = _mm256_mul_pd(qq{I}{J},rinv{I}{J});
615 /* #define INNERFLOPS INNERFLOPS+1 */
616 /* #if 'Force' in KERNEL_VF */
617 felec = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(velec,rinv{I}{J}),fgb),rinv{I}{J});
618 /* #define INNERFLOPS INNERFLOPS+3 */
621 /* #elif KERNEL_ELEC=='Ewald' */
622 /* EWALD ELECTROSTATICS */
624 /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
625 ewrt = _mm256_mul_pd(r{I}{J},ewtabscale);
626 ewitab = _mm256_cvttpd_epi32(ewrt);
627 eweps = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
628 /* #define INNERFLOPS INNERFLOPS+4 */
629 /* #if 'Potential' in KERNEL_VF or KERNEL_MOD_ELEC=='PotentialSwitch' */
630 ewitab = _mm_slli_epi32(ewitab,2);
631 ewtabF = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
632 ewtabD = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
633 ewtabV = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
634 ewtabFn = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
635 GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
636 felec = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
637 /* #define INNERFLOPS INNERFLOPS+2 */
638 /* #if KERNEL_MOD_ELEC=='PotentialShift' */
639 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
640 velec = _mm256_mul_pd(qq{I}{J},_mm256_sub_pd(_mm256_sub_pd(rinv{I}{J},sh_ewald),velec));
641 /* #define INNERFLOPS INNERFLOPS+7 */
643 velec = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
644 velec = _mm256_mul_pd(qq{I}{J},_mm256_sub_pd(rinv{I}{J},velec));
645 /* #define INNERFLOPS INNERFLOPS+6 */
647 /* #if 'Force' in KERNEL_VF */
648 felec = _mm256_mul_pd(_mm256_mul_pd(qq{I}{J},rinv{I}{J}),_mm256_sub_pd(rinvsq{I}{J},felec));
649 /* #define INNERFLOPS INNERFLOPS+3 */
651 /* #elif KERNEL_VF=='Force' */
652 gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
653 ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
655 felec = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
656 felec = _mm256_mul_pd(_mm256_mul_pd(qq{I}{J},rinv{I}{J}),_mm256_sub_pd(rinvsq{I}{J},felec));
657 /* #define INNERFLOPS INNERFLOPS+7 */
660 /* #elif KERNEL_ELEC=='CubicSplineTable' */
662 /* CUBIC SPLINE TABLE ELECTROSTATICS */
663 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
664 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
665 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
666 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
667 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
668 Heps = _mm256_mul_pd(vfeps,H);
669 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
670 /* #define INNERFLOPS INNERFLOPS+4 */
671 /* #if 'Potential' in KERNEL_VF */
672 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
673 velec = _mm256_mul_pd(qq{I}{J},VV);
674 /* #define INNERFLOPS INNERFLOPS+3 */
676 /* #if 'Force' in KERNEL_VF */
677 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
678 felec = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_mul_pd(qq{I}{J},FF),_mm256_mul_pd(vftabscale,rinv{I}{J})));
679 /* #define INNERFLOPS INNERFLOPS+7 */
682 /* ## End of check for electrostatics interaction forms */
684 /* ## END OF ELECTROSTATIC INTERACTION CHECK FOR PAIR I-J */
686 /* #if 'vdw' in INTERACTION_FLAGS[I][J] */
688 /* #if KERNEL_VDW=='LennardJones' */
690 /* LENNARD-JONES DISPERSION/REPULSION */
692 rinvsix = _mm256_mul_pd(_mm256_mul_pd(rinvsq{I}{J},rinvsq{I}{J}),rinvsq{I}{J});
693 /* #define INNERFLOPS INNERFLOPS+2 */
694 /* #if 'Potential' in KERNEL_VF or KERNEL_MOD_VDW=='PotentialSwitch' */
695 vvdw6 = _mm256_mul_pd(c6_{I}{J},rinvsix);
696 vvdw12 = _mm256_mul_pd(c12_{I}{J},_mm256_mul_pd(rinvsix,rinvsix));
697 /* #define INNERFLOPS INNERFLOPS+3 */
698 /* #if KERNEL_MOD_VDW=='PotentialShift' */
699 vvdw = _mm256_sub_pd(_mm256_mul_pd( _mm256_sub_pd(vvdw12 , _mm256_mul_pd(c12_{I}{J},_mm256_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
700 _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_{I}{J},sh_vdw_invrcut6)),one_sixth));
701 /* #define INNERFLOPS INNERFLOPS+8 */
703 vvdw = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
704 /* #define INNERFLOPS INNERFLOPS+3 */
706 /* ## Check for force inside potential check, i.e. this means we already did the potential part */
707 /* #if 'Force' in KERNEL_VF */
708 fvdw = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq{I}{J});
709 /* #define INNERFLOPS INNERFLOPS+2 */
711 /* #elif KERNEL_VF=='Force' */
712 /* ## Force-only LennardJones makes it possible to save 1 flop (they do add up...) */
713 fvdw = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_{I}{J},rinvsix),c6_{I}{J}),_mm256_mul_pd(rinvsix,rinvsq{I}{J}));
714 /* #define INNERFLOPS INNERFLOPS+4 */
717 /* #elif KERNEL_VDW=='CubicSplineTable' */
719 /* CUBIC SPLINE TABLE DISPERSION */
720 /* #if 'Table' in KERNEL_ELEC */
721 vfitab = _mm_add_epi32(vfitab,ifour);
723 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
724 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
725 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
726 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
727 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
728 Heps = _mm256_mul_pd(vfeps,H);
729 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
730 /* #define INNERFLOPS INNERFLOPS+4 */
731 /* #if 'Potential' in KERNEL_VF */
732 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
733 vvdw6 = _mm256_mul_pd(c6_{I}{J},VV);
734 /* #define INNERFLOPS INNERFLOPS+3 */
736 /* #if 'Force' in KERNEL_VF */
737 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
738 fvdw6 = _mm256_mul_pd(c6_{I}{J},FF);
739 /* #define INNERFLOPS INNERFLOPS+4 */
742 /* CUBIC SPLINE TABLE REPULSION */
743 vfitab = _mm_add_epi32(vfitab,ifour);
744 Y = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
745 F = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
746 G = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,2) );
747 H = _mm256_load_pd( vftab + _mm_extract_epi32(vfitab,3) );
748 GMX_MM256_FULLTRANSPOSE4_PD(Y,F,G,H);
749 Heps = _mm256_mul_pd(vfeps,H);
750 Fp = _mm256_add_pd(F,_mm256_mul_pd(vfeps,_mm256_add_pd(G,Heps)));
751 /* #define INNERFLOPS INNERFLOPS+4 */
752 /* #if 'Potential' in KERNEL_VF */
753 VV = _mm256_add_pd(Y,_mm256_mul_pd(vfeps,Fp));
754 vvdw12 = _mm256_mul_pd(c12_{I}{J},VV);
755 /* #define INNERFLOPS INNERFLOPS+3 */
757 /* #if 'Force' in KERNEL_VF */
758 FF = _mm256_add_pd(Fp,_mm256_mul_pd(vfeps,_mm256_add_pd(G,_mm256_add_pd(Heps,Heps))));
759 fvdw12 = _mm256_mul_pd(c12_{I}{J},FF);
760 /* #define INNERFLOPS INNERFLOPS+5 */
762 /* #if 'Potential' in KERNEL_VF */
763 vvdw = _mm256_add_pd(vvdw12,vvdw6);
764 /* #define INNERFLOPS INNERFLOPS+1 */
766 /* #if 'Force' in KERNEL_VF */
767 fvdw = _mm256_xor_pd(signbit,_mm256_mul_pd(_mm256_add_pd(fvdw6,fvdw12),_mm256_mul_pd(vftabscale,rinv{I}{J})));
768 /* #define INNERFLOPS INNERFLOPS+4 */
771 /* ## End of check for vdw interaction forms */
773 /* ## END OF VDW INTERACTION CHECK FOR PAIR I-J */
775 /* #if 'switch' in INTERACTION_FLAGS[I][J] */
776 d = _mm256_sub_pd(r{I}{J},rswitch);
777 d = _mm256_max_pd(d,_mm256_setzero_pd());
778 d2 = _mm256_mul_pd(d,d);
779 sw = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
780 /* #define INNERFLOPS INNERFLOPS+10 */
782 /* #if 'Force' in KERNEL_VF */
783 dsw = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
784 /* #define INNERFLOPS INNERFLOPS+5 */
787 /* Evaluate switch function */
788 /* #if 'Force' in KERNEL_VF */
789 /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
790 /* #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
791 felec = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv{I}{J},_mm256_mul_pd(velec,dsw)) );
792 /* #define INNERFLOPS INNERFLOPS+4 */
794 /* #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
795 fvdw = _mm256_sub_pd( _mm256_mul_pd(fvdw,sw) , _mm256_mul_pd(rinv{I}{J},_mm256_mul_pd(vvdw,dsw)) );
796 /* #define INNERFLOPS INNERFLOPS+4 */
799 /* #if 'Potential' in KERNEL_VF */
800 /* #if 'electrostatics' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_ELEC=='PotentialSwitch' */
801 velec = _mm256_mul_pd(velec,sw);
802 /* #define INNERFLOPS INNERFLOPS+1 */
804 /* #if 'vdw' in INTERACTION_FLAGS[I][J] and KERNEL_MOD_VDW=='PotentialSwitch' */
805 vvdw = _mm256_mul_pd(vvdw,sw);
806 /* #define INNERFLOPS INNERFLOPS+1 */
810 /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
811 /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
812 cutoff_mask = _mm256_cmp_pd(rsq{I}{J},rcutoff2,_CMP_LT_OQ);
813 /* #define INNERFLOPS INNERFLOPS+1 */
816 /* #if 'Potential' in KERNEL_VF */
817 /* Update potential sum for this i atom from the interaction with this j atom. */
818 /* #if 'electrostatics' in INTERACTION_FLAGS[I][J] */
819 /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
820 velec = _mm256_and_pd(velec,cutoff_mask);
821 /* #define INNERFLOPS INNERFLOPS+1 */
823 /* #if ROUND == 'Epilogue' */
824 velec = _mm256_andnot_pd(dummy_mask,velec);
826 velecsum = _mm256_add_pd(velecsum,velec);
827 /* #define INNERFLOPS INNERFLOPS+1 */
828 /* #if KERNEL_ELEC=='GeneralizedBorn' */
829 /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] */
830 vgb = _mm256_and_pd(vgb,cutoff_mask);
831 /* #define INNERFLOPS INNERFLOPS+1 */
833 /* #if ROUND == 'Epilogue' */
834 vgb = _mm256_andnot_pd(dummy_mask,vgb);
836 vgbsum = _mm256_add_pd(vgbsum,vgb);
837 /* #define INNERFLOPS INNERFLOPS+1 */
840 /* #if 'vdw' in INTERACTION_FLAGS[I][J] */
841 /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
842 /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
843 vvdw = _mm256_and_pd(vvdw,cutoff_mask);
844 /* #define INNERFLOPS INNERFLOPS+1 */
846 /* #if ROUND == 'Epilogue' */
847 vvdw = _mm256_andnot_pd(dummy_mask,vvdw);
849 vvdwsum = _mm256_add_pd(vvdwsum,vvdw);
850 /* #define INNERFLOPS INNERFLOPS+1 */
854 /* #if 'Force' in KERNEL_VF */
856 /* #if 'electrostatics' in INTERACTION_FLAGS[I][J] and 'vdw' in INTERACTION_FLAGS[I][J] */
857 fscal = _mm256_add_pd(felec,fvdw);
858 /* #define INNERFLOPS INNERFLOPS+1 */
859 /* #elif 'electrostatics' in INTERACTION_FLAGS[I][J] */
861 /* #elif 'vdw' in INTERACTION_FLAGS[I][J] */
865 /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
866 /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
867 fscal = _mm256_and_pd(fscal,cutoff_mask);
868 /* #define INNERFLOPS INNERFLOPS+1 */
871 /* #if ROUND == 'Epilogue' */
872 fscal = _mm256_andnot_pd(dummy_mask,fscal);
875 /* Calculate temporary vectorial force */
876 tx = _mm256_mul_pd(fscal,dx{I}{J});
877 ty = _mm256_mul_pd(fscal,dy{I}{J});
878 tz = _mm256_mul_pd(fscal,dz{I}{J});
880 /* Update vectorial force */
881 fix{I} = _mm256_add_pd(fix{I},tx);
882 fiy{I} = _mm256_add_pd(fiy{I},ty);
883 fiz{I} = _mm256_add_pd(fiz{I},tz);
884 /* #define INNERFLOPS INNERFLOPS+6 */
886 /* #if GEOMETRY_I == 'Particle' */
887 /* #if ROUND == 'Loop' */
888 fjptrA = f+j_coord_offsetA;
889 fjptrB = f+j_coord_offsetB;
890 fjptrC = f+j_coord_offsetC;
891 fjptrD = f+j_coord_offsetD;
893 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
894 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
895 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
896 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
898 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
899 /* #define INNERFLOPS INNERFLOPS+3 */
901 fjx{J} = _mm256_add_pd(fjx{J},tx);
902 fjy{J} = _mm256_add_pd(fjy{J},ty);
903 fjz{J} = _mm256_add_pd(fjz{J},tz);
904 /* #define INNERFLOPS INNERFLOPS+3 */
909 /* ## Note special check for TIP4P-TIP4P. Since we are cutting of all hydrogen interactions we also cut the LJ-only O-O interaction */
910 /* #if 'exactcutoff' in INTERACTION_FLAGS[I][J] or (GEOMETRY_I=='Water4' and GEOMETRY_J=='Water4' and 'exactcutoff' in INTERACTION_FLAGS[1][1]) */
911 /* #if 0 ## This and next two lines is a hack to maintain indentation in template file */
916 /* ## End of check for the interaction being outside the cutoff */
919 /* ## End of loop over i-j interaction pairs */
921 /* #if GEOMETRY_I != 'Particle' */
922 /* #if ROUND == 'Loop' */
923 fjptrA = f+j_coord_offsetA;
924 fjptrB = f+j_coord_offsetB;
925 fjptrC = f+j_coord_offsetC;
926 fjptrD = f+j_coord_offsetD;
928 fjptrA = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
929 fjptrB = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
930 fjptrC = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
931 fjptrD = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
935 /* #if 'Water' in GEOMETRY_I and GEOMETRY_J == 'Particle' */
936 gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,fjx0,fjy0,fjz0);
937 /* #define INNERFLOPS INNERFLOPS+3 */
938 /* #elif GEOMETRY_J == 'Water3' */
939 gmx_mm256_decrement_3rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
940 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
941 /* #define INNERFLOPS INNERFLOPS+9 */
942 /* #elif GEOMETRY_J == 'Water4' */
943 /* #if 0 in PARTICLES_J */
944 gmx_mm256_decrement_4rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,
945 fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
946 fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
947 /* #define INNERFLOPS INNERFLOPS+12 */
949 gmx_mm256_decrement_3rvec_4ptr_swizzle_pd(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
950 fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
951 /* #define INNERFLOPS INNERFLOPS+9 */
955 /* Inner loop uses {INNERFLOPS} flops */
960 /* End of innermost loop */
962 /* #if 'Force' in KERNEL_VF */
963 /* #if GEOMETRY_I == 'Particle' */
964 gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
965 f+i_coord_offset,fshift+i_shift_offset);
966 /* #define OUTERFLOPS OUTERFLOPS+6 */
967 /* #elif GEOMETRY_I == 'Water3' */
968 gmx_mm256_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
969 f+i_coord_offset,fshift+i_shift_offset);
970 /* #define OUTERFLOPS OUTERFLOPS+18 */
971 /* #elif GEOMETRY_I == 'Water4' */
972 /* #if 0 in PARTICLES_I */
973 gmx_mm256_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
974 f+i_coord_offset,fshift+i_shift_offset);
975 /* #define OUTERFLOPS OUTERFLOPS+24 */
977 gmx_mm256_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
978 f+i_coord_offset+DIM,fshift+i_shift_offset);
979 /* #define OUTERFLOPS OUTERFLOPS+18 */
984 /* #if 'Potential' in KERNEL_VF */
986 /* Update potential energies */
987 /* #if KERNEL_ELEC != 'None' */
988 gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
989 /* #define OUTERFLOPS OUTERFLOPS+1 */
991 /* #if 'GeneralizedBorn' in KERNEL_ELEC */
992 gmx_mm256_update_1pot_pd(vgbsum,kernel_data->energygrp_polarization+ggid);
993 /* #define OUTERFLOPS OUTERFLOPS+1 */
995 /* #if KERNEL_VDW != 'None' */
996 gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
997 /* #define OUTERFLOPS OUTERFLOPS+1 */
1000 /* #if 'GeneralizedBorn' in KERNEL_ELEC and 'Force' in KERNEL_VF */
1001 dvdasum = _mm256_mul_pd(dvdasum, _mm256_mul_pd(isai{I},isai{I}));
1002 gmx_mm256_update_1pot_pd(dvdasum,dvda+inr);
1005 /* Increment number of inner iterations */
1006 inneriter += j_index_end - j_index_start;
1008 /* Outer loop uses {OUTERFLOPS} flops */
1011 /* Increment number of outer iterations */
1014 /* Update outer/inner flops */
1015 /* ## NB: This is not important, it just affects the flopcount. However, since our preprocessor is */
1016 /* ## primitive and replaces aggressively even in strings inside these directives, we need to */
1017 /* ## assemble the main part of the name (containing KERNEL/ELEC/VDW) directly in the source. */
1018 /* #if GEOMETRY_I == 'Water3' */
1019 /* #define ISUFFIX '_W3' */
1020 /* #elif GEOMETRY_I == 'Water4' */
1021 /* #define ISUFFIX '_W4' */
1023 /* #define ISUFFIX '' */
1025 /* #if GEOMETRY_J == 'Water3' */
1026 /* #define JSUFFIX 'W3' */
1027 /* #elif GEOMETRY_J == 'Water4' */
1028 /* #define JSUFFIX 'W4' */
1030 /* #define JSUFFIX '' */
1032 /* #if 'PotentialAndForce' in KERNEL_VF */
1033 /* #define VFSUFFIX '_VF' */
1034 /* #elif 'Potential' in KERNEL_VF */
1035 /* #define VFSUFFIX '_V' */
1037 /* #define VFSUFFIX '_F' */
1040 /* #if KERNEL_ELEC != 'None' and KERNEL_VDW != 'None' */
1041 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1042 /* #elif KERNEL_ELEC != 'None' */
1043 inc_nrnb(nrnb,eNR_NBKERNEL_ELEC{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});
1045 inc_nrnb(nrnb,eNR_NBKERNEL_VDW{ISUFFIX}{JSUFFIX}{VFSUFFIX},outeriter*{OUTERFLOPS} + inneriter*{INNERFLOPS});