12#if defined(__OFFLOAD) && !defined(__NO_OFFLOAD_NNP)
21#define NNP_SPLINE_CACHE_MAX 256
30int nnp_gpu_max_se_c(
void);
31int nnp_gpu_spline_cache_max_c(
void);
39struct nnp_gpu_cache_entry {
56typedef struct nnp_gpu_state {
57 offloadStream_t stream;
61 struct nnp_gpu_cache_entry spline_cache[NNP_SPLINE_CACHE_MAX];
62 int spline_cache_size;
67 double *image_table_dev;
68 size_t image_table_dev_cap;
72 double *coord_scaled_dev;
73 size_t coord_scaled_dev_cap;
74 int *ele_ind_global_dev;
75 size_t ele_ind_global_dev_cap;
76 int *cell_image_atom_dev;
77 size_t cell_image_atom_dev_cap;
78 int *cell_image_shift_dev;
79 size_t cell_image_shift_dev_cap;
81 size_t cell_head_dev_cap;
83 size_t cell_next_dev_cap;
89 size_t pair_n_rad_dev_cap;
91 size_t pair_n_ang1_dev_cap;
93 size_t pair_n_ang2_dev_cap;
94 double *pair_max_cutoff_dev;
95 size_t pair_max_cutoff_dev_cap;
96 int *pair_rad_groups_dev;
97 size_t pair_rad_groups_dev_cap;
98 int *pair_ang1_groups_dev;
99 size_t pair_ang1_groups_dev_cap;
100 int *pair_ang2_groups_dev;
101 size_t pair_ang2_groups_dev_cap;
102 double *rad_cutoff_dev;
103 size_t rad_cutoff_dev_cap;
104 double *ang_cutoff_dev;
105 size_t ang_cutoff_dev_cap;
106 int pair_map_uploaded;
108 int pair_map_max_pair_grps;
109 int pair_map_max_n_radgrp;
110 int pair_map_max_n_anggrp;
116 int *walk_atoms_global_dev;
117 size_t walk_atoms_global_dev_cap;
118 int *walk_n_atoms_in_e_dev;
119 size_t walk_n_atoms_in_e_dev_cap;
120 int *walk_rad_ind_dev;
121 size_t walk_rad_ind_dev_cap;
122 int *walk_rad_image_idx_dev;
123 size_t walk_rad_image_idx_dev_cap;
124 int *walk_ang1_ind_dev;
125 size_t walk_ang1_ind_dev_cap;
126 int *walk_ang1_image_idx_dev;
127 size_t walk_ang1_image_idx_dev_cap;
128 int *walk_ang2_ind_dev;
129 size_t walk_ang2_ind_dev_cap;
130 int *walk_ang2_image_idx_dev;
131 size_t walk_ang2_image_idx_dev_cap;
132 int *walk_n_rad_pa_dev;
133 size_t walk_n_rad_pa_dev_cap;
134 int *walk_n_ang1_pa_dev;
135 size_t walk_n_ang1_pa_dev_cap;
136 int *walk_n_ang2_pa_dev;
137 size_t walk_n_ang2_pa_dev_cap;
138 int *walk_overflow_dev;
139 size_t walk_overflow_cap;
147 double *G_dev_scaled;
148 size_t G_dev_scaled_cap;
154 size_t expol_flag_dev_cap;
161 size_t pbuf_symf_map_cap;
163 size_t pbuf_atom_map_cap;
167 int *pbuf_n_neigh_pa;
168 size_t pbuf_n_neigh_pa_cap;
170 size_t pbuf_n_ang1_pa_cap;
172 size_t pbuf_n_ang2_pa_cap;
174 size_t pbuf_izeta_cap;
179 size_t pbuf_rad_ind_cap;
180 int *pbuf_rad_image_idx;
181 size_t pbuf_rad_image_idx_cap;
182 int *pbuf_atoms_global_rad;
183 size_t pbuf_atoms_global_rad_cap;
185 size_t pbuf_ang1_ind_cap;
187 size_t pbuf_ang2_ind_cap;
188 int *pbuf_ang1_image_idx;
189 size_t pbuf_ang1_image_idx_cap;
190 int *pbuf_ang2_image_idx;
191 size_t pbuf_ang2_image_idx_cap;
192 int *pbuf_atoms_global_ang;
193 size_t pbuf_atoms_global_ang_cap;
194 int *scatter_neigh_ind;
195 size_t scatter_neigh_ind_cap;
199 float *pbuf_ang1_disp;
200 size_t pbuf_ang1_disp_cap;
201 float *pbuf_ang2_disp;
202 size_t pbuf_ang2_disp_cap;
207 double *scaling_loc_av_dev;
208 size_t scaling_loc_av_cap;
209 double *scaling_loc_min_dev;
210 size_t scaling_loc_min_cap;
211 double *scaling_loc_max_dev;
212 size_t scaling_loc_max_cap;
213 double *scaling_sigma_dev;
214 size_t scaling_sigma_cap;
215 int *scaling_n_input_nodes_dev;
216 size_t scaling_n_input_nodes_cap;
217 int scaling_uploaded;
223 int scaling_n_input_max;
227 int *se_static_rad_symf_packed_dev;
228 size_t se_static_rad_symf_packed_cap;
229 int *se_static_rad_symf_offsets_dev;
230 size_t se_static_rad_symf_offsets_cap;
231 int *se_static_ang_symf_packed_dev;
232 size_t se_static_ang_symf_packed_cap;
233 int *se_static_ang_symf_offsets_dev;
234 size_t se_static_ang_symf_offsets_cap;
235 int se_static_uploaded;
240 int se_static_max_n_radgrp;
241 int se_static_max_n_anggrp;
242 int se_static_total_rad_symf;
243 int se_static_total_ang_symf;
248 double *per_se_rad_self[NNP_MAX_SE];
249 size_t per_se_rad_self_cap[NNP_MAX_SE];
250 double *per_se_rad_force[NNP_MAX_SE];
251 size_t per_se_rad_force_cap[NNP_MAX_SE];
252 double *per_se_ang_self[NNP_MAX_SE];
253 size_t per_se_ang_self_cap[NNP_MAX_SE];
254 double *per_se_ang_jj[NNP_MAX_SE];
255 size_t per_se_ang_jj_cap[NNP_MAX_SE];
256 double *per_se_ang_kk[NNP_MAX_SE];
257 size_t per_se_ang_kk_cap[NNP_MAX_SE];
262 int *ang_izeta_dev[NNP_MAX_SE];
263 size_t ang_izeta_cap[NNP_MAX_SE];
264 char *ang_uiz_dev[NNP_MAX_SE];
265 size_t ang_uiz_cap[NNP_MAX_SE];
266 int ang_zeta_n[NNP_MAX_SE];
276 int *step_ele_ind_dev;
277 size_t step_ele_ind_cap;
278 int *step_ele_ind_host;
279 size_t step_ele_ind_host_cap;
286 size_t nn_n_nodes_dev_cap;
287 int *nn_act_fnct_dev;
288 size_t nn_act_fnct_dev_cap;
289 double *nn_atom_energies_dev;
290 size_t nn_atom_energies_dev_cap;
291 double *nn_W_packed_dev;
292 size_t nn_W_packed_dev_cap;
293 double *nn_b_packed_dev;
294 size_t nn_b_packed_dev_cap;
295 int nn_packed_uploaded;
296 int nn_packed_n_committee;
297 int nn_packed_n_layer;
298 int nn_packed_max_n_nodes;
303 double *pbuf_atomE_out;
304 size_t pbuf_atomE_out_cap;
305 double *pbuf_dEdG_out;
306 size_t pbuf_dEdG_out_cap;
326 size_t virial_dev_cap;
334nnp_gpu_state_t *nnp_gpu_state_get(
void);
341void nnp_gpu_state_release(
void);
350void nnp_gpu_reset_caches(
void);
361int nnp_gpu_register_env(
void);
367void nnp_gpu_unregister_env(
void);
375void nnp_gpu_release_env(
void);
386int nnp_gpu_pbuf_ensure_named(
void **pbuf,
size_t *pcap,
size_t n_bytes,
387 const char *bufname);
388#define nnp_gpu_pbuf_ensure(pbuf, pcap, n_bytes) \
389 nnp_gpu_pbuf_ensure_named((pbuf), (pcap), (n_bytes), #pbuf)
392unsigned long nnp_gpu_pbuf_grow_events(
void);
395size_t nnp_gpu_pbuf_total_bytes(
void);
412double *nnp_gpu_get_or_upload_spline(
const double *h_ptr,
size_t n_doubles);
419int nnp_gpu_set_precision(
int use_fp32);
422int nnp_gpu_precision_fp32(
void);
430int nnp_gpu_upload_positions_step_c(
int num_atoms,
const double *h_coord);
432int nnp_gpu_upload_image_table_c(
int n_images,
const double *h_image_table);
434int nnp_gpu_upload_cell_list_c(
int n_images,
int n_cells,
435 const int *h_image_atom,
436 const int *h_image_shift,
const int *h_head,
439int nnp_gpu_upload_pair_map_c(
int n_ele,
int max_pair_grps,
int max_n_radgrp,
440 int max_n_anggrp,
const int *h_n_rad,
441 const int *h_n_ang1,
const int *h_n_ang2,
442 const double *h_pair_max_cutoff,
443 const int *h_rad_groups,
const int *h_ang1_groups,
444 const int *h_ang2_groups,
445 const double *h_rad_cutoff,
446 const double *h_ang_cutoff);
448int nnp_gpu_cell_list_walk_c(
449 int num_atoms_global,
int n_ele,
int n_images,
int n_cells,
450 int n_atoms_in_e_max,
int max_n_neigh_rad,
int max_n_neigh_ang1,
451 int max_n_neigh_ang2,
int max_n_radgrp,
int max_n_anggrp,
const int *h_nbin,
452 const double *h_bin_lower,
const double *h_bin_width,
const int *h_bin_span,
453 const int *h_perd,
const int *h_exact_pbc_copies,
454 const double *h_coord_scaled,
const int *h_ele_ind_global,
455 const int *h_atoms_global_per_e,
const int *h_n_atoms_in_e);
457int nnp_gpu_walk_counts_d2h_c(
int n_atoms_in_e_max,
int max_n_radgrp,
458 int max_n_anggrp,
int n_ele,
int *h_n_rad_pa,
459 int *h_n_ang1_pa,
int *h_n_ang2_pa);
468int nnp_gpu_begin_descriptor_step(
int n_atoms,
int n_input_max);
470int nnp_gpu_upload_scaling_tables_c(
int n_ele,
int n_input_max,
471 const double *h_loc_av_per_ele,
472 const double *h_loc_min_per_ele,
473 const double *h_loc_max_per_ele,
474 const double *h_sigma_per_ele,
475 const int *h_n_input_nodes_per_ele);
477int nnp_gpu_upload_se_static_tables_c(
int n_ele,
int max_n_radgrp,
478 int max_n_anggrp,
int total_rad_symf,
480 const int *h_rad_symf_packed,
481 const int *h_rad_symf_offsets,
482 const int *h_ang_symf_packed,
483 const int *h_ang_symf_offsets);
485int nnp_gpu_scale_G_dev_c(
int n_atoms,
int n_input_max,
int n_ele,
486 int scaling_mode,
double scmin,
double scmax,
487 const int *h_ele_ind_for_atoms);
492int nnp_gpu_extrapolation_flag_c(
void);
494int nnp_gpu_radial_sf_to_G_dev_c(
495 const int n_atoms_in_e,
const int max_n_neigh_per_se,
const int n_symf,
496 const int spline_n,
const double spline_dx,
const double spline_dx_inv,
497 const double spline_x_max,
const double *h_spline_y,
498 const double *h_spline_dy,
const int s_0based,
const int e_0based,
499 const int max_n_radgrp_walk,
const int n_atoms_in_e_max_walk,
500 const int max_n_neigh_walk,
const int istart_1based,
const int n_input_max,
503int nnp_gpu_angular_sf_to_G_dev_c(
504 const int homo,
const int cut_type,
const int max_n_ang1_per_se,
505 const int max_n_ang2_per_se,
const int max_n_kk_per_se,
const int n_symf,
506 const int n_atoms_in_e,
const double cutoff_s,
const double cutoff_sqr,
507 const double *h_pack_lam,
const double *h_pack_zeta,
508 const double *h_pack_eta,
const double *h_pack_prefzeta,
509 const int *h_pack_izeta,
const char *h_pack_use_int_zeta,
510 const int s_0based,
const int e_0based,
const int max_n_anggrp_walk,
511 const int n_atoms_in_e_max_walk,
const int max_n_neigh_ang1_walk,
512 const int max_n_neigh_ang2_walk,
const int istart_1based,
513 const int n_rad_for_e,
const int n_input_max,
const int se_idx);
522int nnp_gpu_upload_nn_packed_c(
int n_committee,
int n_layer,
int max_n_nodes,
523 int n_ele,
const int *h_n_nodes,
524 const int *h_act_fnct,
525 const double *h_atom_energies,
526 const double *h_W_packed,
527 const double *h_b_packed);
534int nnp_gpu_step_ele_ind_ensure(nnp_gpu_state_t *state,
const int *h_ele_ind,
539int nnp_gpu_nn_atomwise_with_G_dev_c(
const int n_atoms,
const int n_committee,
540 const int n_layer,
const int max_n_nodes,
541 const int n_input_max,
const int normnodes,
542 const int n_ele,
const int *h_ele_ind,
543 double *h_atomic_energy,
double *h_dE_dG);
552int nnp_gpu_force_scatter_begin_c(
int num_atoms_global,
int n_committee,
553 const double *h_force);
555int nnp_gpu_force_scatter_end_c(
int num_atoms_global,
int n_committee,
563int nnp_gpu_force_scatter_virial_d2h_c(
int n_committee,
double *h_virial);
565int nnp_gpu_force_scatter_radial_c(
566 const int n_atoms_in_e,
const int n_symf,
const int max_n_neigh_per_se,
567 const int n_committee,
const int n_input_max,
const int num_atoms_global,
568 const int ele_idx,
const int scaling_mode,
const double scmin,
569 const double scmax,
const int s_0based,
const int max_n_radgrp_walk,
570 const int n_atoms_in_e_max_walk,
const int max_n_neigh_walk,
571 const int istart_1based,
const int se_idx,
const int accumulate_virial);
573int nnp_gpu_force_scatter_angular_c(
574 const int n_atoms_in_e,
const int n_symf,
const int max_n_ang1_per_se,
575 const int max_n_kk_per_se,
const int n_committee,
const int n_input_max,
576 const int num_atoms_global,
const int ele_idx,
const int n_rad_offset,
577 const int scaling_mode,
const int homo,
const double scmin,
578 const double scmax,
const int s_0based,
const int max_n_anggrp_walk,
579 const int n_atoms_in_e_max_walk,
const int max_n_neigh_ang1_walk,
580 const int max_n_neigh_ang2_walk,
const int istart_1based,
const int se_idx,
581 const int accumulate_virial);