Line data Source code
1 : /*----------------------------------------------------------------------------*/
2 : /* CP2K: A general program to perform molecular dynamics simulations */
3 : /* Copyright 2000-2026 CP2K developers group <https://cp2k.org> */
4 : /* */
5 : /* SPDX-License-Identifier: BSD-3-Clause */
6 : /*----------------------------------------------------------------------------*/
7 :
8 : #include <assert.h>
9 : #include <math.h>
10 : #include <omp.h>
11 : #include <stdint.h>
12 : #include <stdio.h>
13 : #include <stdlib.h>
14 : #include <string.h>
15 :
16 : #include "../common/grid_common.h"
17 : #include "grid_ref_collocate.h"
18 : #include "grid_ref_integrate.h"
19 : #include "grid_ref_task_list.h"
20 : #include "grid_ref_task_list_internal.h"
21 :
22 : /*******************************************************************************
23 : * \brief Comperator passed to qsort to compare two tasks.
24 : * \author Ole Schuett
25 : ******************************************************************************/
26 71637007 : static int compare_tasks(const void *a, const void *b) {
27 71637007 : const grid_ref_task *task_a = a, *task_b = b;
28 71637007 : if (task_a->level != task_b->level) {
29 5203194 : return task_a->level - task_b->level;
30 66433813 : } else if (task_a->block_num != task_b->block_num) {
31 37329420 : return task_a->block_num - task_b->block_num;
32 29104393 : } else if (task_a->iset != task_b->iset) {
33 3363475 : return task_a->iset - task_b->iset;
34 : } else {
35 25740918 : return task_a->jset - task_b->jset;
36 : }
37 : }
38 : /*******************************************************************************
39 : * \brief Allocates a task list for the reference backend.
40 : * See grid_task_list.h for details.
41 : * \author Ole Schuett
42 : ******************************************************************************/
43 17366 : void grid_ref_create_task_list(
44 : const bool orthorhombic, const int ntasks, const int nlevels,
45 : const int natoms, const int nkinds, const int nblocks,
46 : const int block_offsets[nblocks], const double atom_positions[natoms][3],
47 : const int atom_kinds[natoms], const grid_basis_set *basis_sets[nkinds],
48 : const int level_list[ntasks], const int iatom_list[ntasks],
49 : const int jatom_list[ntasks], const int iset_list[ntasks],
50 : const int jset_list[ntasks], const int ipgf_list[ntasks],
51 : const int jpgf_list[ntasks], const int border_mask_list[ntasks],
52 : const int block_num_list[ntasks], const double radius_list[ntasks],
53 : const double rab_list[ntasks][3], const int npts_global[nlevels][3],
54 : const int npts_local[nlevels][3], const int shift_local[nlevels][3],
55 : const int border_width[nlevels][3], const double dh[nlevels][3][3],
56 : const double dh_inv[nlevels][3][3], grid_ref_task_list **task_list_out) {
57 :
58 17366 : if (*task_list_out != NULL) {
59 : // This is actually an opportunity to reuse some buffers.
60 7145 : grid_ref_free_task_list(*task_list_out);
61 : }
62 :
63 17366 : grid_ref_task_list_internal *task_list =
64 17366 : malloc(sizeof(grid_ref_task_list_internal));
65 17366 : assert(task_list != NULL);
66 :
67 17366 : task_list->orthorhombic = orthorhombic;
68 17366 : task_list->ntasks = ntasks;
69 17366 : task_list->nlevels = nlevels;
70 17366 : task_list->natoms = natoms;
71 17366 : task_list->nkinds = nkinds;
72 17366 : task_list->nblocks = nblocks;
73 :
74 17366 : size_t size = nblocks * sizeof(int);
75 17366 : task_list->block_offsets = malloc(size);
76 17366 : assert(task_list->block_offsets != NULL || size == 0);
77 17366 : if (size != 0) {
78 17366 : memcpy(task_list->block_offsets, block_offsets, size);
79 : }
80 :
81 17366 : size = 3 * natoms * sizeof(double);
82 17366 : task_list->atom_positions = malloc(size);
83 17366 : assert(task_list->atom_positions != NULL || size == 0);
84 17366 : if (size != 0) {
85 17366 : memcpy(task_list->atom_positions, atom_positions, size);
86 : }
87 :
88 17366 : size = natoms * sizeof(int);
89 17366 : task_list->atom_kinds = malloc(size);
90 17366 : assert(task_list->atom_kinds != NULL || size == 0);
91 17366 : if (size != 0) {
92 17366 : memcpy(task_list->atom_kinds, atom_kinds, size);
93 : }
94 :
95 17366 : size = nkinds * sizeof(grid_basis_set *);
96 17366 : task_list->basis_sets = malloc(size);
97 17366 : assert(task_list->basis_sets != NULL || size == 0);
98 17366 : if (size != 0) {
99 17366 : memcpy(task_list->basis_sets, basis_sets, size);
100 : }
101 :
102 17366 : size = ntasks * sizeof(grid_ref_task);
103 17366 : task_list->tasks = malloc(size);
104 17366 : assert(task_list->tasks != NULL || size == 0);
105 17366 : #pragma omp parallel for schedule(static) if (ntasks > GRID_OMP_MIN_ITERATIONS)
106 : for (int i = 0; i < ntasks; i++) {
107 : task_list->tasks[i].level = level_list[i];
108 : task_list->tasks[i].iatom = iatom_list[i];
109 : task_list->tasks[i].jatom = jatom_list[i];
110 : task_list->tasks[i].iset = iset_list[i];
111 : task_list->tasks[i].jset = jset_list[i];
112 : task_list->tasks[i].ipgf = ipgf_list[i];
113 : task_list->tasks[i].jpgf = jpgf_list[i];
114 : task_list->tasks[i].border_mask = border_mask_list[i];
115 : task_list->tasks[i].block_num = block_num_list[i];
116 : task_list->tasks[i].radius = radius_list[i];
117 : task_list->tasks[i].rab[0] = rab_list[i][0];
118 : task_list->tasks[i].rab[1] = rab_list[i][1];
119 : task_list->tasks[i].rab[2] = rab_list[i][2];
120 : }
121 :
122 : // Store grid layouts.
123 17366 : size = nlevels * sizeof(grid_ref_layout);
124 17366 : task_list->layouts = malloc(size);
125 17366 : assert(task_list->layouts != NULL || size == 0);
126 86044 : for (int level = 0; level < nlevels; level++) {
127 274712 : for (int i = 0; i < 3; i++) {
128 206034 : task_list->layouts[level].npts_global[i] = npts_global[level][i];
129 206034 : task_list->layouts[level].npts_local[i] = npts_local[level][i];
130 206034 : task_list->layouts[level].shift_local[i] = shift_local[level][i];
131 206034 : task_list->layouts[level].border_width[i] = border_width[level][i];
132 824136 : for (int j = 0; j < 3; j++) {
133 618102 : task_list->layouts[level].dh[i][j] = dh[level][i][j];
134 618102 : task_list->layouts[level].dh_inv[i][j] = dh_inv[level][i][j];
135 : }
136 : }
137 : }
138 :
139 : // Sort tasks by level, block_num, iset, and jset.
140 17366 : qsort(task_list->tasks, ntasks, sizeof(grid_ref_task), &compare_tasks);
141 :
142 : // Find first and last task for each level and block.
143 17366 : size = nlevels * nblocks * sizeof(int);
144 17366 : task_list->first_level_block_task = malloc(size);
145 17366 : assert(task_list->first_level_block_task != NULL || size == 0);
146 17366 : task_list->last_level_block_task = malloc(size);
147 17366 : assert(task_list->last_level_block_task != NULL || size == 0);
148 17366 : const int nlevel_blocks = nlevels * nblocks;
149 17366 : #pragma omp parallel for schedule(static) if (nlevel_blocks > \
150 : GRID_OMP_MIN_ITERATIONS)
151 : for (int i = 0; i < nlevel_blocks; i++) {
152 : task_list->first_level_block_task[i] = 0;
153 : task_list->last_level_block_task[i] = -1; // last < first means no tasks
154 : }
155 10686806 : for (int itask = 0; itask < ntasks; itask++) {
156 10669440 : const int level = task_list->tasks[itask].level - 1;
157 10669440 : const int block_num = task_list->tasks[itask].block_num - 1;
158 10669440 : if (itask == 0 || task_list->tasks[itask - 1].level - 1 != level ||
159 10621224 : task_list->tasks[itask - 1].block_num - 1 != block_num) {
160 1314389 : task_list->first_level_block_task[level * nblocks + block_num] = itask;
161 : }
162 10669440 : task_list->last_level_block_task[level * nblocks + block_num] = itask;
163 : }
164 :
165 : // Find largest Cartesian subblock size.
166 17366 : task_list->maxco = 0;
167 48032 : for (int i = 0; i < nkinds; i++) {
168 30666 : task_list->maxco = imax(task_list->maxco, task_list->basis_sets[i]->maxco);
169 : }
170 :
171 : // Initialize thread-local storage.
172 17366 : size = omp_get_max_threads() * sizeof(double *);
173 17366 : task_list->threadlocals = malloc(size);
174 17366 : assert(task_list->threadlocals != NULL);
175 17366 : memset(task_list->threadlocals, 0, size);
176 17366 : size = omp_get_max_threads() * sizeof(size_t);
177 17366 : task_list->threadlocal_sizes = malloc(size);
178 17366 : assert(task_list->threadlocal_sizes != NULL);
179 17366 : memset(task_list->threadlocal_sizes, 0, size);
180 :
181 17366 : *task_list_out = task_list;
182 17366 : }
183 :
184 : /*******************************************************************************
185 : * \brief Deallocates given task list, basis_sets have to be freed separately.
186 : * \author Ole Schuett
187 : ******************************************************************************/
188 17365 : void grid_ref_free_task_list(grid_ref_task_list *ptr) {
189 17365 : if (ptr == NULL)
190 : return;
191 :
192 17365 : grid_ref_task_list_internal *task_list = (grid_ref_task_list_internal *)ptr;
193 :
194 17365 : free(task_list->block_offsets);
195 17365 : free(task_list->atom_positions);
196 17365 : free(task_list->atom_kinds);
197 17365 : free(task_list->basis_sets);
198 17365 : free(task_list->tasks);
199 17365 : free(task_list->layouts);
200 17365 : free(task_list->first_level_block_task);
201 17365 : free(task_list->last_level_block_task);
202 34730 : for (int i = 0; i < omp_get_max_threads(); i++) {
203 17365 : if (task_list->threadlocals[i] != NULL) {
204 6 : free(task_list->threadlocals[i]);
205 : }
206 : }
207 17365 : free(task_list->threadlocals);
208 17365 : free(task_list->threadlocal_sizes);
209 17365 : free(task_list);
210 : }
211 :
212 : /*******************************************************************************
213 : * \brief Prototype for BLAS dgemm.
214 : * \author Ole Schuett
215 : ******************************************************************************/
216 : void dgemm_(const char *transa, const char *transb, const int *m, const int *n,
217 : const int *k, const double *alpha, const double *a, const int *lda,
218 : const double *b, const int *ldb, const double *beta, double *c,
219 : const int *ldc);
220 :
221 : /*******************************************************************************
222 : * \brief Convenient wrapper to hide Fortran nature of dgemm_, swapping a and b.
223 : * \author Ole Schuett
224 : ******************************************************************************/
225 1928 : static void dgemm(const char transa, const char transb, const int m,
226 : const int n, const int k, const double alpha, const double *a,
227 : const int lda, const double *b, const int ldb,
228 : const double beta, double *c, const int ldc) {
229 1928 : dgemm_(&transb, &transa, &n, &m, &k, &alpha, b, &ldb, a, &lda, &beta, c,
230 : &ldc);
231 1928 : }
232 :
233 : /*******************************************************************************
234 : * \brief Transforms pab from contracted spherical to prim. cartesian basis.
235 : * \author Ole Schuett
236 : ******************************************************************************/
237 580 : static void load_pab(const grid_basis_set *ibasis, const grid_basis_set *jbasis,
238 : const int iset, const int jset, const bool transpose,
239 580 : const double *block, double *pab) {
240 :
241 : // Define some more convenient aliases.
242 580 : const int ncoseta = ncoset(ibasis->lmax[iset]);
243 580 : const int ncosetb = ncoset(jbasis->lmax[jset]);
244 580 : const int ncoa = ibasis->npgf[iset] * ncoseta; // size of carthesian set
245 580 : const int ncob = jbasis->npgf[jset] * ncosetb;
246 :
247 580 : const int nsgf_seta = ibasis->nsgf_set[iset]; // size of spherical set
248 580 : const int nsgf_setb = jbasis->nsgf_set[jset];
249 580 : const int nsgfa = ibasis->nsgf; // size of entire spherical basis
250 580 : const int nsgfb = jbasis->nsgf;
251 580 : const int sgfa = ibasis->first_sgf[iset] - 1; // start of spherical set
252 580 : const int sgfb = jbasis->first_sgf[jset] - 1;
253 580 : const int maxcoa = ibasis->maxco;
254 580 : const int maxcob = jbasis->maxco;
255 :
256 580 : double work[nsgf_setb * ncoa];
257 580 : if (transpose) {
258 : // work[nsgf_setb][ncoa] = MATMUL(subblock, ibasis->sphi)
259 412 : dgemm('N', 'N', nsgf_setb, ncoa, nsgf_seta, 1.0,
260 412 : &block[sgfb * nsgfa + sgfa], nsgfa, &ibasis->sphi[sgfa * maxcoa],
261 : maxcoa, 0.0, work, ncoa);
262 : } else {
263 : // work[nsgf_setb][ncoa] = MATMUL(TRANSPOSE(subblock), ibasis->sphi)
264 168 : dgemm('T', 'N', nsgf_setb, ncoa, nsgf_seta, 1.0,
265 168 : &block[sgfa * nsgfb + sgfb], nsgfb, &ibasis->sphi[sgfa * maxcoa],
266 : maxcoa, 0.0, work, ncoa);
267 : }
268 : // pab[ncob][ncoa] = MATMUL(TRANSPOSE(jbasis->sphi), work)
269 580 : dgemm('T', 'N', ncob, ncoa, nsgf_setb, 1.0, &jbasis->sphi[sgfb * maxcob],
270 : maxcob, work, ncoa, 0.0, pab, ncoa);
271 580 : }
272 :
273 : /*******************************************************************************
274 : * \brief Collocate a range of tasks which are destined for the same grid level.
275 : * \author Ole Schuett
276 : ******************************************************************************/
277 168 : static void collocate_one_grid_level(
278 : const grid_ref_task_list *ptr, const int *first_block_task,
279 : const int *last_block_task, const enum grid_func func,
280 : const int npts_global[3], const int npts_local[3], const int shift_local[3],
281 : const int border_width[3], const double dh[3][3], const double dh_inv[3][3],
282 : const double *pab_blocks, offload_buffer *grid) {
283 :
284 168 : if (ptr == NULL)
285 : return;
286 :
287 168 : grid_ref_task_list_internal *task_list = (grid_ref_task_list_internal *)ptr;
288 :
289 : // Using default(shared) because with GCC 9 the behavior around const changed:
290 : // https://www.gnu.org/software/gcc/gcc-9/porting_to.html
291 168 : #pragma omp parallel default(shared)
292 : {
293 : const int ithread = omp_get_thread_num();
294 : const int nthreads = omp_get_num_threads();
295 :
296 : // Initialize variables to detect when a new subblock has to be fetched.
297 : int old_offset = -1, old_iset = -1, old_jset = -1;
298 :
299 : // Matrix pab is re-used across tasks.
300 : double pab[imax(task_list->maxco * task_list->maxco, 1)];
301 :
302 : // Ensure that grid can fit into thread-local storage, reallocate if needed.
303 : const int npts_local_total = npts_local[0] * npts_local[1] * npts_local[2];
304 : const size_t grid_size = npts_local_total * sizeof(double);
305 : if (task_list->threadlocal_sizes[ithread] < grid_size) {
306 : if (task_list->threadlocals[ithread] != NULL) {
307 : free(task_list->threadlocals[ithread]);
308 : }
309 : task_list->threadlocals[ithread] = malloc(grid_size);
310 : assert(task_list->threadlocals[ithread] != NULL);
311 : task_list->threadlocal_sizes[ithread] = grid_size;
312 : }
313 :
314 : // Zero thread-local copy of the grid.
315 : double *const my_grid = task_list->threadlocals[ithread];
316 : memset(my_grid, 0, grid_size);
317 :
318 : // Parallelize over blocks to avoid unnecessary calls to load_pab.
319 : const int chunk_size = imax(1, task_list->nblocks / (nthreads * 50));
320 : #pragma omp for schedule(dynamic, chunk_size)
321 : for (int block_num = 0; block_num < task_list->nblocks; block_num++) {
322 : const int first_task = first_block_task[block_num];
323 : const int last_task = last_block_task[block_num];
324 :
325 : for (int itask = first_task; itask <= last_task; itask++) {
326 : // Define some convenient aliases.
327 : const grid_ref_task *task = &task_list->tasks[itask];
328 : const int iatom = task->iatom - 1;
329 : const int jatom = task->jatom - 1;
330 : const int iset = task->iset - 1;
331 : const int jset = task->jset - 1;
332 : const int ipgf = task->ipgf - 1;
333 : const int jpgf = task->jpgf - 1;
334 : const int ikind = task_list->atom_kinds[iatom] - 1;
335 : const int jkind = task_list->atom_kinds[jatom] - 1;
336 : const grid_basis_set *ibasis = task_list->basis_sets[ikind];
337 : const grid_basis_set *jbasis = task_list->basis_sets[jkind];
338 : const double zeta = ibasis->zet[iset * ibasis->maxpgf + ipgf];
339 : const double zetb = jbasis->zet[jset * jbasis->maxpgf + jpgf];
340 : const int ncoseta = ncoset(ibasis->lmax[iset]);
341 : const int ncosetb = ncoset(jbasis->lmax[jset]);
342 : const int ncoa = ibasis->npgf[iset] * ncoseta; // size of carthesian set
343 : const int ncob = jbasis->npgf[jset] * ncosetb;
344 : const int block_num = task->block_num - 1;
345 : const int block_offset = task_list->block_offsets[block_num];
346 : const double *block = &pab_blocks[block_offset];
347 : const bool transpose = (iatom <= jatom);
348 :
349 : // Load subblock from buffer and decontract into Cartesian sublock pab.
350 : // The previous pab can be reused when only ipgf or jpgf has changed.
351 : if (block_offset != old_offset || iset != old_iset ||
352 : jset != old_jset) {
353 : old_offset = block_offset;
354 : old_iset = iset;
355 : old_jset = jset;
356 : load_pab(ibasis, jbasis, iset, jset, transpose, block, pab);
357 : }
358 :
359 : grid_ref_collocate_pgf_product(
360 : /*orthorhombic=*/task_list->orthorhombic,
361 : /*border_mask=*/task->border_mask,
362 : /*func=*/func,
363 : /*la_max=*/ibasis->lmax[iset],
364 : /*la_min=*/ibasis->lmin[iset],
365 : /*lb_max=*/jbasis->lmax[jset],
366 : /*lb_min=*/jbasis->lmin[jset],
367 : /*zeta=*/zeta,
368 : /*zetb=*/zetb,
369 : /*rscale=*/(iatom == jatom) ? 1 : 2,
370 : /*dh=*/dh,
371 : /*dh_inv=*/dh_inv,
372 : /*ra=*/&task_list->atom_positions[3 * iatom],
373 : /*rab=*/task->rab,
374 : /*npts_global=*/npts_global,
375 : /*npts_local=*/npts_local,
376 : /*shift_local=*/shift_local,
377 : /*border_width=*/border_width,
378 : /*radius=*/task->radius,
379 : /*o1=*/ipgf * ncoseta,
380 : /*o2=*/jpgf * ncosetb,
381 : /*n1=*/ncoa,
382 : /*n2=*/ncob,
383 : /*pab=*/(const double (*)[ncoa])pab,
384 : /*grid=*/my_grid);
385 :
386 : } // end of task loop
387 : } // end of block loop
388 :
389 : // While there should be an implicit barrier at the end of the block loop, this
390 : // explicit barrier eliminates occasional seg faults with icc compiled binaries.
391 : #pragma omp barrier
392 :
393 : // Merge thread-local grids via an efficient tree reduction.
394 : const int nreduction_cycles = ceil(log(nthreads) / log(2)); // tree depth
395 : for (int icycle = 1; icycle <= nreduction_cycles; icycle++) {
396 : // Threads are divided into groups, whose size doubles with each cycle.
397 : // After a cycle the reduced data is stored at first thread of each group.
398 : const int group_size = 1 << icycle; // 2**icycle
399 : const int igroup = ithread / group_size;
400 : const int dest_thread = igroup * group_size;
401 : const int src_thread = dest_thread + group_size / 2;
402 : // The last group might actually be a bit smaller.
403 : const int actual_group_size = imin(group_size, nthreads - dest_thread);
404 : // Parallelize summation by dividing grid points across group members.
405 : const int rank = modulo(ithread, group_size); // position within the group
406 : const int64_t lb = ((int64_t)npts_local_total * rank) / actual_group_size;
407 : const int64_t ub =
408 : ((int64_t)npts_local_total * (rank + 1)) / actual_group_size;
409 : if (src_thread < nthreads) {
410 : GRID_PRAGMA_SIMD_LOOP
411 : for (int i = (int)lb; i < (int)ub; i++) {
412 : task_list->threadlocals[dest_thread][i] +=
413 : task_list->threadlocals[src_thread][i];
414 : }
415 : }
416 : #pragma omp barrier
417 : }
418 :
419 : // Copy final result from first thread into shared grid.
420 : const int64_t lb = ((int64_t)npts_local_total * ithread) / nthreads;
421 : const int64_t ub = ((int64_t)npts_local_total * (ithread + 1)) / nthreads;
422 : GRID_PRAGMA_SIMD_LOOP
423 : for (int i = (int)lb; i < (int)ub; i++) {
424 : grid->host_buffer[i] = task_list->threadlocals[0][i];
425 : }
426 :
427 : } // end of omp parallel region
428 : }
429 :
430 : /*******************************************************************************
431 : * \brief Collocate all tasks of in given list onto given grids.
432 : * See grid_task_list.h for details.
433 : * \author Ole Schuett
434 : ******************************************************************************/
435 42 : void grid_ref_collocate_task_list(const grid_ref_task_list *ptr,
436 : const enum grid_func func, const int nlevels,
437 : const offload_buffer *pab_blocks,
438 : offload_buffer *grids[nlevels]) {
439 :
440 42 : if (ptr == NULL)
441 : return;
442 :
443 42 : grid_ref_task_list_internal *task_list = (grid_ref_task_list_internal *)ptr;
444 :
445 42 : assert(task_list->nlevels == nlevels);
446 :
447 210 : for (int level = 0; level < task_list->nlevels; level++) {
448 168 : const int idx = level * task_list->nblocks;
449 168 : const int *first_block_task = &task_list->first_level_block_task[idx];
450 168 : const int *last_block_task = &task_list->last_level_block_task[idx];
451 168 : const grid_ref_layout *layout = &task_list->layouts[level];
452 168 : collocate_one_grid_level(
453 168 : task_list, first_block_task, last_block_task, func, layout->npts_global,
454 168 : layout->npts_local, layout->shift_local, layout->border_width,
455 168 : layout->dh, layout->dh_inv, pab_blocks->host_buffer, grids[level]);
456 : }
457 : }
458 :
459 : /*******************************************************************************
460 : * \brief Transforms hab from prim. cartesian to contracted spherical basis.
461 : * \author Ole Schuett
462 : ******************************************************************************/
463 384 : static inline void store_hab(const grid_basis_set *ibasis,
464 : const grid_basis_set *jbasis, const int iset,
465 : const int jset, const bool transpose,
466 384 : const double *hab, double *block) {
467 :
468 : // Define some more convenient aliases.
469 384 : const int ncoseta = ncoset(ibasis->lmax[iset]);
470 384 : const int ncosetb = ncoset(jbasis->lmax[jset]);
471 384 : const int ncoa = ibasis->npgf[iset] * ncoseta; // size of carthesian set
472 384 : const int ncob = jbasis->npgf[jset] * ncosetb;
473 :
474 384 : const int nsgf_seta = ibasis->nsgf_set[iset]; // size of spherical set
475 384 : const int nsgf_setb = jbasis->nsgf_set[jset];
476 384 : const int nsgfa = ibasis->nsgf; // size of entire spherical basis
477 384 : const int nsgfb = jbasis->nsgf;
478 384 : const int sgfa = ibasis->first_sgf[iset] - 1; // start of spherical set
479 384 : const int sgfb = jbasis->first_sgf[jset] - 1;
480 384 : const int maxcoa = ibasis->maxco;
481 384 : const int maxcob = jbasis->maxco;
482 :
483 384 : double work[nsgf_setb * ncoa];
484 :
485 : // work[nsgf_setb][ncoa] = MATMUL(jbasis->sphi, hab)
486 384 : dgemm('N', 'N', nsgf_setb, ncoa, ncob, 1.0, &jbasis->sphi[sgfb * maxcob],
487 : maxcob, hab, ncoa, 0.0, work, ncoa);
488 :
489 384 : if (transpose) {
490 : // subblock[nsgf_setb][nsgf_seta] += MATMUL(work, TRANSPOSE(ibasis->sphi))
491 288 : dgemm('N', 'T', nsgf_setb, nsgf_seta, ncoa, 1.0, work, ncoa,
492 288 : &ibasis->sphi[sgfa * maxcoa], maxcoa, 1.0,
493 288 : &block[sgfb * nsgfa + sgfa], nsgfa);
494 : } else {
495 : // subblock[nsgf_seta][nsgf_setb] += MATMUL(ibasis->sphi, TRANSPOSE(work))
496 96 : dgemm('N', 'T', nsgf_seta, nsgf_setb, ncoa, 1.0,
497 96 : &ibasis->sphi[sgfa * maxcoa], maxcoa, work, ncoa, 1.0,
498 96 : &block[sgfa * nsgfb + sgfb], nsgfb);
499 : }
500 384 : }
501 :
502 : /*******************************************************************************
503 : * \brief Integrate a range of tasks that belong to the same grid level.
504 : * \author Ole Schuett
505 : ******************************************************************************/
506 136 : static void integrate_one_grid_level(
507 : const grid_ref_task_list *ptr, const int *first_block_task,
508 : const int *last_block_task, const bool compute_tau, const int natoms,
509 : const int npts_global[3], const int npts_local[3], const int shift_local[3],
510 : const int border_width[3], const double dh[3][3], const double dh_inv[3][3],
511 : const offload_buffer *pab_blocks, const offload_buffer *grid,
512 : offload_buffer *hab_blocks, double forces[natoms][3], double virial[3][3]) {
513 :
514 136 : if (ptr == NULL)
515 : return;
516 :
517 136 : grid_ref_task_list_internal *task_list = (grid_ref_task_list_internal *)ptr;
518 :
519 : // Using default(shared) because with GCC 9 the behavior around const changed:
520 : // https://www.gnu.org/software/gcc/gcc-9/porting_to.html
521 136 : #pragma omp parallel default(shared)
522 : {
523 : // Initialize variables to detect when a new subblock has to be fetched.
524 : int old_offset = -1, old_iset = -1, old_jset = -1;
525 : grid_basis_set *old_ibasis = NULL, *old_jbasis = NULL;
526 : bool old_transpose = false;
527 :
528 : // Matrix pab and hab are re-used across tasks.
529 : double pab[imax(task_list->maxco * task_list->maxco, 1)];
530 : double hab[imax(task_list->maxco * task_list->maxco, 1)];
531 :
532 : // Parallelize over blocks to avoid concurred access to hab_blocks.
533 : const int nthreads = omp_get_num_threads();
534 : const int chunk_size = imax(1, task_list->nblocks / (nthreads * 50));
535 : #pragma omp for schedule(dynamic, chunk_size)
536 : for (int block_num = 0; block_num < task_list->nblocks; block_num++) {
537 : const int first_task = first_block_task[block_num];
538 : const int last_task = last_block_task[block_num];
539 :
540 : // Accumulate forces per block as it corresponds to a pair of atoms.
541 : const int iatom = task_list->tasks[first_task].iatom - 1;
542 : const int jatom = task_list->tasks[first_task].jatom - 1;
543 : double my_forces[2][3] = {0};
544 : double my_virials[2][3][3] = {0};
545 :
546 : for (int itask = first_task; itask <= last_task; itask++) {
547 : // Define some convenient aliases.
548 : const grid_ref_task *task = &task_list->tasks[itask];
549 : assert(task->block_num - 1 == block_num);
550 : assert(task->iatom - 1 == iatom && task->jatom - 1 == jatom);
551 : const int ikind = task_list->atom_kinds[iatom] - 1;
552 : const int jkind = task_list->atom_kinds[jatom] - 1;
553 : grid_basis_set *ibasis = task_list->basis_sets[ikind];
554 : grid_basis_set *jbasis = task_list->basis_sets[jkind];
555 : const int iset = task->iset - 1;
556 : const int jset = task->jset - 1;
557 : const int ipgf = task->ipgf - 1;
558 : const int jpgf = task->jpgf - 1;
559 : const double zeta = ibasis->zet[iset * ibasis->maxpgf + ipgf];
560 : const double zetb = jbasis->zet[jset * jbasis->maxpgf + jpgf];
561 : const int ncoseta = ncoset(ibasis->lmax[iset]);
562 : const int ncosetb = ncoset(jbasis->lmax[jset]);
563 : const int ncoa = ibasis->npgf[iset] * ncoseta; // size of carthesian set
564 : const int ncob = jbasis->npgf[jset] * ncosetb;
565 : const int block_offset = task_list->block_offsets[block_num];
566 : const bool transpose = (iatom <= jatom);
567 : const bool pab_required = (forces != NULL || virial != NULL);
568 :
569 : // Load pab and store hab subblocks when needed.
570 : // Previous hab and pab can be reused when only ipgf or jpgf changed.
571 : if (block_offset != old_offset || iset != old_iset ||
572 : jset != old_jset) {
573 : if (pab_required) {
574 : load_pab(ibasis, jbasis, iset, jset, transpose,
575 : &pab_blocks->host_buffer[block_offset], pab);
576 : }
577 : if (old_offset >= 0) { // skip first iteration
578 : store_hab(old_ibasis, old_jbasis, old_iset, old_jset, old_transpose,
579 : hab, &hab_blocks->host_buffer[old_offset]);
580 : }
581 : memset(hab, 0, ncoa * ncob * sizeof(double));
582 : old_offset = block_offset;
583 : old_iset = iset;
584 : old_jset = jset;
585 : old_ibasis = ibasis;
586 : old_jbasis = jbasis;
587 : old_transpose = transpose;
588 : }
589 :
590 : grid_ref_integrate_pgf_product(
591 : /*orthorhombic=*/task_list->orthorhombic,
592 : /*compute_tau=*/compute_tau,
593 : /*border_mask=*/task->border_mask,
594 : /*la_max=*/ibasis->lmax[iset],
595 : /*la_min=*/ibasis->lmin[iset],
596 : /*lb_max=*/jbasis->lmax[jset],
597 : /*lb_min=*/jbasis->lmin[jset],
598 : /*zeta=*/zeta,
599 : /*zetb=*/zetb,
600 : /*dh=*/dh,
601 : /*dh_inv=*/dh_inv,
602 : /*ra=*/&task_list->atom_positions[3 * iatom],
603 : /*rab=*/task->rab,
604 : /*npts_global=*/npts_global,
605 : /*npts_local=*/npts_local,
606 : /*shift_local=*/shift_local,
607 : /*border_width=*/border_width,
608 : /*radius=*/task->radius,
609 : /*o1=*/ipgf * ncoseta,
610 : /*o2=*/jpgf * ncosetb,
611 : /*n1=*/ncoa,
612 : /*n2=*/ncob,
613 : /*grid=*/grid->host_buffer,
614 : /*hab=*/(double (*)[ncoa])hab,
615 : /*pab=*/(pab_required) ? (const double (*)[ncoa])pab : NULL,
616 : /*forces=*/(forces != NULL) ? my_forces : NULL,
617 : /*virials=*/(virial != NULL) ? my_virials : NULL,
618 : /*hdab=*/NULL,
619 : /*hadb=*/NULL,
620 : /*a_hdab=*/NULL);
621 :
622 : } // end of task loop
623 :
624 : // Merge thread-local forces and virial into shared ones.
625 : // It does not seem worth the trouble to accumulate them thread-locally.
626 : const double scalef = (iatom == jatom) ? 1.0 : 2.0;
627 : if (forces != NULL) {
628 : #pragma omp critical(forces)
629 : for (int i = 0; i < 3; i++) {
630 : forces[iatom][i] += scalef * my_forces[0][i];
631 : forces[jatom][i] += scalef * my_forces[1][i];
632 : }
633 : }
634 : if (virial != NULL) {
635 : #pragma omp critical(virial)
636 : for (int i = 0; i < 3; i++) {
637 : for (int j = 0; j < 3; j++) {
638 : virial[i][j] += scalef * my_virials[0][i][j];
639 : virial[i][j] += scalef * my_virials[1][i][j];
640 : }
641 : }
642 : }
643 :
644 : } // end of block loop
645 :
646 : // store final hab
647 : if (old_offset >= 0) {
648 : store_hab(old_ibasis, old_jbasis, old_iset, old_jset, old_transpose, hab,
649 : &hab_blocks->host_buffer[old_offset]);
650 : }
651 :
652 : } // end of omp parallel region
653 : }
654 :
655 : /*******************************************************************************
656 : * \brief Integrate all tasks of in given list from given grids.
657 : * See grid_task_list.h for details.
658 : * \author Ole Schuett
659 : ******************************************************************************/
660 34 : void grid_ref_integrate_task_list(
661 : const grid_ref_task_list *ptr, const bool compute_tau, const int natoms,
662 : const int nlevels, const offload_buffer *pab_blocks,
663 : const offload_buffer *grids[nlevels], offload_buffer *hab_blocks,
664 : double forces[natoms][3], double virial[3][3]) {
665 34 : if (ptr == NULL)
666 : return;
667 :
668 34 : grid_ref_task_list_internal *task_list = (grid_ref_task_list_internal *)ptr;
669 34 : assert(task_list->nlevels == nlevels);
670 34 : assert(task_list->natoms == natoms);
671 :
672 : // Zero result arrays.
673 34 : memset(hab_blocks->host_buffer, 0, hab_blocks->size);
674 34 : if (forces != NULL) {
675 6 : memset(forces, 0, natoms * 3 * sizeof(double));
676 : }
677 34 : if (virial != NULL) {
678 0 : memset(virial, 0, 9 * sizeof(double));
679 : }
680 :
681 170 : for (int level = 0; level < task_list->nlevels; level++) {
682 136 : const int idx = level * task_list->nblocks;
683 136 : const int *first_block_task = &task_list->first_level_block_task[idx];
684 136 : const int *last_block_task = &task_list->last_level_block_task[idx];
685 136 : const grid_ref_layout *layout = &task_list->layouts[level];
686 136 : integrate_one_grid_level(
687 : task_list, first_block_task, last_block_task, compute_tau, natoms,
688 136 : layout->npts_global, layout->npts_local, layout->shift_local,
689 136 : layout->border_width, layout->dh, layout->dh_inv, pab_blocks,
690 136 : grids[level], hab_blocks, forces, virial);
691 : }
692 : }
693 :
694 : // EOF
|
