Line data Source code
1 : /*----------------------------------------------------------------------------*/
2 : /* CP2K: A general program to perform molecular dynamics simulations */
3 : /* Copyright 2000-2025 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_cpu_collocate.h"
18 : #include "grid_cpu_integrate.h"
19 : #include "grid_cpu_task_list.h"
20 :
21 : /*******************************************************************************
22 : * \brief Comperator passed to qsort to compare two tasks.
23 : * \author Ole Schuett
24 : ******************************************************************************/
25 48647944 : static int compare_tasks(const void *a, const void *b) {
26 48647944 : const grid_cpu_task *task_a = a, *task_b = b;
27 48647944 : if (task_a->level != task_b->level) {
28 3686441 : return task_a->level - task_b->level;
29 44961503 : } else if (task_a->block_num != task_b->block_num) {
30 20968069 : return task_a->block_num - task_b->block_num;
31 23993434 : } else if (task_a->iset != task_b->iset) {
32 2868103 : return task_a->iset - task_b->iset;
33 : } else {
34 21125331 : return task_a->jset - task_b->jset;
35 : }
36 : }
37 : /*******************************************************************************
38 : * \brief Allocates a task list for the cpu backend.
39 : * See grid_task_list.h for details.
40 : * \author Ole Schuett
41 : ******************************************************************************/
42 14138 : void grid_cpu_create_task_list(
43 : const bool orthorhombic, const int ntasks, const int nlevels,
44 : const int natoms, const int nkinds, const int nblocks,
45 : const int block_offsets[nblocks], const double atom_positions[natoms][3],
46 : const int atom_kinds[natoms], const grid_basis_set *basis_sets[nkinds],
47 : const int level_list[ntasks], const int iatom_list[ntasks],
48 : const int jatom_list[ntasks], const int iset_list[ntasks],
49 : const int jset_list[ntasks], const int ipgf_list[ntasks],
50 : const int jpgf_list[ntasks], const int border_mask_list[ntasks],
51 : const int block_num_list[ntasks], const double radius_list[ntasks],
52 : const double rab_list[ntasks][3], const int npts_global[nlevels][3],
53 : const int npts_local[nlevels][3], const int shift_local[nlevels][3],
54 : const int border_width[nlevels][3], const double dh[nlevels][3][3],
55 : const double dh_inv[nlevels][3][3], grid_cpu_task_list **task_list_out) {
56 :
57 14138 : if (*task_list_out != NULL) {
58 : // This is actually an opportunity to reuse some buffers.
59 5666 : grid_cpu_free_task_list(*task_list_out);
60 : }
61 :
62 14138 : grid_cpu_task_list *task_list = malloc(sizeof(grid_cpu_task_list));
63 14138 : assert(task_list != NULL);
64 :
65 14138 : task_list->orthorhombic = orthorhombic;
66 14138 : task_list->ntasks = ntasks;
67 14138 : task_list->nlevels = nlevels;
68 14138 : task_list->natoms = natoms;
69 14138 : task_list->nkinds = nkinds;
70 14138 : task_list->nblocks = nblocks;
71 :
72 14138 : size_t size = nblocks * sizeof(int);
73 14138 : task_list->block_offsets = malloc(size);
74 14138 : assert(task_list->block_offsets != NULL || size == 0);
75 14138 : if (size != 0) {
76 13908 : memcpy(task_list->block_offsets, block_offsets, size);
77 : }
78 :
79 14138 : size = 3 * natoms * sizeof(double);
80 14138 : task_list->atom_positions = malloc(size);
81 14138 : assert(task_list->atom_positions != NULL || size == 0);
82 14138 : if (size != 0) {
83 14138 : memcpy(task_list->atom_positions, atom_positions, size);
84 : }
85 :
86 14138 : size = natoms * sizeof(int);
87 14138 : task_list->atom_kinds = malloc(size);
88 14138 : assert(task_list->atom_kinds != NULL || size == 0);
89 14138 : if (size != 0) {
90 14138 : memcpy(task_list->atom_kinds, atom_kinds, size);
91 : }
92 :
93 14138 : size = nkinds * sizeof(grid_basis_set *);
94 14138 : task_list->basis_sets = malloc(size);
95 14138 : assert(task_list->basis_sets != NULL || size == 0);
96 14138 : if (size != 0) {
97 14138 : memcpy(task_list->basis_sets, basis_sets, size);
98 : }
99 :
100 14138 : size = ntasks * sizeof(grid_cpu_task);
101 14138 : task_list->tasks = malloc(size);
102 14138 : assert(task_list->tasks != NULL || size == 0);
103 7820555 : for (int i = 0; i < ntasks; i++) {
104 7806417 : task_list->tasks[i].level = level_list[i];
105 7806417 : task_list->tasks[i].iatom = iatom_list[i];
106 7806417 : task_list->tasks[i].jatom = jatom_list[i];
107 7806417 : task_list->tasks[i].iset = iset_list[i];
108 7806417 : task_list->tasks[i].jset = jset_list[i];
109 7806417 : task_list->tasks[i].ipgf = ipgf_list[i];
110 7806417 : task_list->tasks[i].jpgf = jpgf_list[i];
111 7806417 : task_list->tasks[i].border_mask = border_mask_list[i];
112 7806417 : task_list->tasks[i].block_num = block_num_list[i];
113 7806417 : task_list->tasks[i].radius = radius_list[i];
114 7806417 : task_list->tasks[i].rab[0] = rab_list[i][0];
115 7806417 : task_list->tasks[i].rab[1] = rab_list[i][1];
116 7806417 : task_list->tasks[i].rab[2] = rab_list[i][2];
117 : }
118 :
119 : // Store grid layouts.
120 14138 : size = nlevels * sizeof(grid_cpu_layout);
121 14138 : task_list->layouts = malloc(size);
122 14138 : assert(task_list->layouts != NULL || size == 0);
123 70074 : for (int level = 0; level < nlevels; level++) {
124 223744 : for (int i = 0; i < 3; i++) {
125 167808 : task_list->layouts[level].npts_global[i] = npts_global[level][i];
126 167808 : task_list->layouts[level].npts_local[i] = npts_local[level][i];
127 167808 : task_list->layouts[level].shift_local[i] = shift_local[level][i];
128 167808 : task_list->layouts[level].border_width[i] = border_width[level][i];
129 671232 : for (int j = 0; j < 3; j++) {
130 503424 : task_list->layouts[level].dh[i][j] = dh[level][i][j];
131 503424 : task_list->layouts[level].dh_inv[i][j] = dh_inv[level][i][j];
132 : }
133 : }
134 : }
135 :
136 : // Sort tasks by level, block_num, iset, and jset.
137 14138 : qsort(task_list->tasks, ntasks, sizeof(grid_cpu_task), &compare_tasks);
138 :
139 : // Find first and last task for each level and block.
140 14138 : size = nlevels * nblocks * sizeof(int);
141 14138 : task_list->first_level_block_task = malloc(size);
142 14138 : assert(task_list->first_level_block_task != NULL || size == 0);
143 14138 : task_list->last_level_block_task = malloc(size);
144 14138 : assert(task_list->last_level_block_task != NULL || size == 0);
145 1105352 : for (int i = 0; i < nlevels * nblocks; i++) {
146 1091214 : task_list->first_level_block_task[i] = 0;
147 1091214 : task_list->last_level_block_task[i] = -1; // last < first means no tasks
148 : }
149 7820555 : for (int itask = 0; itask < ntasks; itask++) {
150 7806417 : const int level = task_list->tasks[itask].level - 1;
151 7806417 : const int block_num = task_list->tasks[itask].block_num - 1;
152 7806417 : if (itask == 0 || task_list->tasks[itask - 1].level - 1 != level ||
153 7768081 : task_list->tasks[itask - 1].block_num - 1 != block_num) {
154 551641 : task_list->first_level_block_task[level * nblocks + block_num] = itask;
155 : }
156 7806417 : task_list->last_level_block_task[level * nblocks + block_num] = itask;
157 : }
158 :
159 : // Find largest Cartesian subblock size.
160 14138 : task_list->maxco = 0;
161 39367 : for (int i = 0; i < nkinds; i++) {
162 25229 : task_list->maxco = imax(task_list->maxco, task_list->basis_sets[i]->maxco);
163 : }
164 :
165 : // Initialize thread-local storage.
166 14138 : size = omp_get_max_threads() * sizeof(double *);
167 14138 : task_list->threadlocals = malloc(size);
168 14138 : assert(task_list->threadlocals != NULL);
169 14138 : memset(task_list->threadlocals, 0, size);
170 14138 : size = omp_get_max_threads() * sizeof(size_t);
171 14138 : task_list->threadlocal_sizes = malloc(size);
172 14138 : assert(task_list->threadlocal_sizes != NULL);
173 14138 : memset(task_list->threadlocal_sizes, 0, size);
174 :
175 14138 : *task_list_out = task_list;
176 14138 : }
177 :
178 : /*******************************************************************************
179 : * \brief Deallocates given task list, basis_sets have to be freed separately.
180 : * \author Ole Schuett
181 : ******************************************************************************/
182 14138 : void grid_cpu_free_task_list(grid_cpu_task_list *task_list) {
183 14138 : free(task_list->block_offsets);
184 14138 : free(task_list->atom_positions);
185 14138 : free(task_list->atom_kinds);
186 14138 : free(task_list->basis_sets);
187 14138 : free(task_list->tasks);
188 14138 : free(task_list->layouts);
189 14138 : free(task_list->first_level_block_task);
190 14138 : free(task_list->last_level_block_task);
191 28276 : for (int i = 0; i < omp_get_max_threads(); i++) {
192 14138 : if (task_list->threadlocals[i] != NULL) {
193 12480 : free(task_list->threadlocals[i]);
194 : }
195 : }
196 14138 : free(task_list->threadlocals);
197 14138 : free(task_list->threadlocal_sizes);
198 14138 : free(task_list);
199 14138 : }
200 :
201 : /*******************************************************************************
202 : * \brief Prototype for BLAS dgemm.
203 : * \author Ole Schuett
204 : ******************************************************************************/
205 : void dgemm_(const char *transa, const char *transb, const int *m, const int *n,
206 : const int *k, const double *alpha, const double *a, const int *lda,
207 : const double *b, const int *ldb, const double *beta, double *c,
208 : const int *ldc);
209 :
210 : /*******************************************************************************
211 : * \brief Convenient wrapper to hide Fortran nature of dgemm_, swapping a and b.
212 : * \author Ole Schuett
213 : ******************************************************************************/
214 39856102 : static void dgemm(const char transa, const char transb, const int m,
215 : const int n, const int k, const double alpha, const double *a,
216 : const int lda, const double *b, const int ldb,
217 : const double beta, double *c, const int ldc) {
218 39856102 : dgemm_(&transb, &transa, &n, &m, &k, &alpha, b, &ldb, a, &lda, &beta, c,
219 : &ldc);
220 39856102 : }
221 :
222 : /*******************************************************************************
223 : * \brief Transforms pab from contracted spherical to prim. cartesian basis.
224 : * \author Ole Schuett
225 : ******************************************************************************/
226 10624968 : static void load_pab(const grid_basis_set *ibasis, const grid_basis_set *jbasis,
227 : const int iset, const int jset, const bool transpose,
228 10624968 : const double *block, double *pab) {
229 :
230 : // Define some more convenient aliases.
231 10624968 : const int ncoseta = ncoset(ibasis->lmax[iset]);
232 10624968 : const int ncosetb = ncoset(jbasis->lmax[jset]);
233 10624968 : const int ncoa = ibasis->npgf[iset] * ncoseta; // size of carthesian set
234 10624968 : const int ncob = jbasis->npgf[jset] * ncosetb;
235 :
236 10624968 : const int nsgf_seta = ibasis->nsgf_set[iset]; // size of spherical set
237 10624968 : const int nsgf_setb = jbasis->nsgf_set[jset];
238 10624968 : const int nsgfa = ibasis->nsgf; // size of entire spherical basis
239 10624968 : const int nsgfb = jbasis->nsgf;
240 10624968 : const int sgfa = ibasis->first_sgf[iset] - 1; // start of spherical set
241 10624968 : const int sgfb = jbasis->first_sgf[jset] - 1;
242 10624968 : const int maxcoa = ibasis->maxco;
243 10624968 : const int maxcob = jbasis->maxco;
244 :
245 10624968 : double work[nsgf_setb * ncoa];
246 10624968 : if (transpose) {
247 : // work[nsgf_setb][ncoa] = MATMUL(subblock, ibasis->sphi)
248 6463882 : dgemm('N', 'N', nsgf_setb, ncoa, nsgf_seta, 1.0,
249 6463882 : &block[sgfb * nsgfa + sgfa], nsgfa, &ibasis->sphi[sgfa * maxcoa],
250 : maxcoa, 0.0, work, ncoa);
251 : } else {
252 : // work[nsgf_setb][ncoa] = MATMUL(TRANSPOSE(subblock), ibasis->sphi)
253 4161086 : dgemm('T', 'N', nsgf_setb, ncoa, nsgf_seta, 1.0,
254 4161086 : &block[sgfa * nsgfb + sgfb], nsgfb, &ibasis->sphi[sgfa * maxcoa],
255 : maxcoa, 0.0, work, ncoa);
256 : }
257 : // pab[ncob][ncoa] = MATMUL(TRANSPOSE(jbasis->sphi), work)
258 10624968 : dgemm('T', 'N', ncob, ncoa, nsgf_setb, 1.0, &jbasis->sphi[sgfb * maxcob],
259 : maxcob, work, ncoa, 0.0, pab, ncoa);
260 10624968 : }
261 :
262 : /*******************************************************************************
263 : * \brief Collocate a range of tasks which are destined for the same grid level.
264 : * \author Ole Schuett
265 : ******************************************************************************/
266 823652 : static void collocate_one_grid_level(
267 : const grid_cpu_task_list *task_list, const int *first_block_task,
268 : const int *last_block_task, const enum grid_func func,
269 : const int npts_global[3], const int npts_local[3], const int shift_local[3],
270 : const int border_width[3], const double dh[3][3], const double dh_inv[3][3],
271 : const double *pab_blocks, offload_buffer *grid) {
272 :
273 : // Using default(shared) because with GCC 9 the behavior around const changed:
274 : // https://www.gnu.org/software/gcc/gcc-9/porting_to.html
275 823652 : #pragma omp parallel default(shared)
276 : {
277 : const int ithread = omp_get_thread_num();
278 : const int nthreads = omp_get_num_threads();
279 :
280 : // Initialize variables to detect when a new subblock has to be fetched.
281 : int old_offset = -1, old_iset = -1, old_jset = -1;
282 :
283 : // Matrix pab is re-used across tasks.
284 : double pab[imax(task_list->maxco * task_list->maxco, 1)];
285 :
286 : // Ensure that grid can fit into thread-local storage, reallocate if needed.
287 : const int npts_local_total = npts_local[0] * npts_local[1] * npts_local[2];
288 : const size_t grid_size = npts_local_total * sizeof(double);
289 : if (task_list->threadlocal_sizes[ithread] < grid_size) {
290 : if (task_list->threadlocals[ithread] != NULL) {
291 : free(task_list->threadlocals[ithread]);
292 : }
293 : task_list->threadlocals[ithread] = malloc(grid_size);
294 : assert(task_list->threadlocals[ithread] != NULL);
295 : task_list->threadlocal_sizes[ithread] = grid_size;
296 : }
297 :
298 : // Zero thread-local copy of the grid.
299 : double *const my_grid = task_list->threadlocals[ithread];
300 : memset(my_grid, 0, grid_size);
301 :
302 : // Parallelize over blocks to avoid unnecessary calls to load_pab.
303 : const int chunk_size = imax(1, task_list->nblocks / (nthreads * 50));
304 : #pragma omp for schedule(dynamic, chunk_size)
305 : for (int block_num = 0; block_num < task_list->nblocks; block_num++) {
306 : const int first_task = first_block_task[block_num];
307 : const int last_task = last_block_task[block_num];
308 :
309 : for (int itask = first_task; itask <= last_task; itask++) {
310 : // Define some convenient aliases.
311 : const grid_cpu_task *task = &task_list->tasks[itask];
312 : const int iatom = task->iatom - 1;
313 : const int jatom = task->jatom - 1;
314 : const int iset = task->iset - 1;
315 : const int jset = task->jset - 1;
316 : const int ipgf = task->ipgf - 1;
317 : const int jpgf = task->jpgf - 1;
318 : const int ikind = task_list->atom_kinds[iatom] - 1;
319 : const int jkind = task_list->atom_kinds[jatom] - 1;
320 : const grid_basis_set *ibasis = task_list->basis_sets[ikind];
321 : const grid_basis_set *jbasis = task_list->basis_sets[jkind];
322 : const double zeta = ibasis->zet[iset * ibasis->maxpgf + ipgf];
323 : const double zetb = jbasis->zet[jset * jbasis->maxpgf + jpgf];
324 : const int ncoseta = ncoset(ibasis->lmax[iset]);
325 : const int ncosetb = ncoset(jbasis->lmax[jset]);
326 : const int ncoa = ibasis->npgf[iset] * ncoseta; // size of carthesian set
327 : const int ncob = jbasis->npgf[jset] * ncosetb;
328 : const int block_num = task->block_num - 1;
329 : const int block_offset = task_list->block_offsets[block_num];
330 : const double *block = &pab_blocks[block_offset];
331 : const bool transpose = (iatom <= jatom);
332 :
333 : // Load subblock from buffer and decontract into Cartesian sublock pab.
334 : // The previous pab can be reused when only ipgf or jpgf has changed.
335 : if (block_offset != old_offset || iset != old_iset ||
336 : jset != old_jset) {
337 : old_offset = block_offset;
338 : old_iset = iset;
339 : old_jset = jset;
340 : load_pab(ibasis, jbasis, iset, jset, transpose, block, pab);
341 : }
342 :
343 : grid_cpu_collocate_pgf_product(
344 : /*orthorhombic=*/task_list->orthorhombic,
345 : /*border_mask=*/task->border_mask,
346 : /*func=*/func,
347 : /*la_max=*/ibasis->lmax[iset],
348 : /*la_min=*/ibasis->lmin[iset],
349 : /*lb_max=*/jbasis->lmax[jset],
350 : /*lb_min=*/jbasis->lmin[jset],
351 : /*zeta=*/zeta,
352 : /*zetb=*/zetb,
353 : /*rscale=*/(iatom == jatom) ? 1 : 2,
354 : /*dh=*/dh,
355 : /*dh_inv=*/dh_inv,
356 : /*ra=*/&task_list->atom_positions[3 * iatom],
357 : /*rab=*/task->rab,
358 : /*npts_global=*/npts_global,
359 : /*npts_local=*/npts_local,
360 : /*shift_local=*/shift_local,
361 : /*border_width=*/border_width,
362 : /*radius=*/task->radius,
363 : /*o1=*/ipgf * ncoseta,
364 : /*o2=*/jpgf * ncosetb,
365 : /*n1=*/ncoa,
366 : /*n2=*/ncob,
367 : /*pab=*/(const double(*)[ncoa])pab,
368 : /*grid=*/my_grid);
369 :
370 : } // end of task loop
371 : } // end of block loop
372 :
373 : // While there should be an implicit barrier at the end of the block loop, this
374 : // explicit barrier eliminates occasional seg faults with icc compiled binaries.
375 : #pragma omp barrier
376 :
377 : // Merge thread-local grids via an efficient tree reduction.
378 : const int nreduction_cycles = ceil(log(nthreads) / log(2)); // tree depth
379 : for (int icycle = 1; icycle <= nreduction_cycles; icycle++) {
380 : // Threads are divided into groups, whose size doubles with each cycle.
381 : // After a cycle the reduced data is stored at first thread of each group.
382 : const int group_size = 1 << icycle; // 2**icycle
383 : const int igroup = ithread / group_size;
384 : const int dest_thread = igroup * group_size;
385 : const int src_thread = dest_thread + group_size / 2;
386 : // The last group might actually be a bit smaller.
387 : const int actual_group_size = imin(group_size, nthreads - dest_thread);
388 : // Parallelize summation by dividing grid points across group members.
389 : const int rank = modulo(ithread, group_size); // position within the group
390 : const int64_t lb = ((int64_t)npts_local_total * rank) / actual_group_size;
391 : const int64_t ub =
392 : ((int64_t)npts_local_total * (rank + 1)) / actual_group_size;
393 : if (src_thread < nthreads) {
394 : for (int i = (int)lb; i < (int)ub; i++) {
395 : task_list->threadlocals[dest_thread][i] +=
396 : task_list->threadlocals[src_thread][i];
397 : }
398 : }
399 : #pragma omp barrier
400 : }
401 :
402 : // Copy final result from first thread into shared grid.
403 : const int64_t lb = ((int64_t)npts_local_total * ithread) / nthreads;
404 : const int64_t ub = ((int64_t)npts_local_total * (ithread + 1)) / nthreads;
405 : for (int i = (int)lb; i < (int)ub; i++) {
406 : grid->host_buffer[i] = task_list->threadlocals[0][i];
407 : }
408 :
409 : } // end of omp parallel region
410 823652 : }
411 :
412 : /*******************************************************************************
413 : * \brief Collocate all tasks of in given list onto given grids.
414 : * See grid_task_list.h for details.
415 : * \author Ole Schuett
416 : ******************************************************************************/
417 208020 : void grid_cpu_collocate_task_list(const grid_cpu_task_list *task_list,
418 : const enum grid_func func, const int nlevels,
419 : const offload_buffer *pab_blocks,
420 : offload_buffer *grids[nlevels]) {
421 :
422 208020 : assert(task_list->nlevels == nlevels);
423 :
424 1031672 : for (int level = 0; level < task_list->nlevels; level++) {
425 823652 : const int idx = level * task_list->nblocks;
426 823652 : const int *first_block_task = &task_list->first_level_block_task[idx];
427 823652 : const int *last_block_task = &task_list->last_level_block_task[idx];
428 823652 : const grid_cpu_layout *layout = &task_list->layouts[level];
429 823652 : collocate_one_grid_level(
430 823652 : task_list, first_block_task, last_block_task, func, layout->npts_global,
431 823652 : layout->npts_local, layout->shift_local, layout->border_width,
432 823652 : layout->dh, layout->dh_inv, pab_blocks->host_buffer, grids[level]);
433 : }
434 208020 : }
435 :
436 : /*******************************************************************************
437 : * \brief Transforms hab from prim. cartesian to contracted spherical basis.
438 : * \author Ole Schuett
439 : ******************************************************************************/
440 9303083 : static inline void store_hab(const grid_basis_set *ibasis,
441 : const grid_basis_set *jbasis, const int iset,
442 : const int jset, const bool transpose,
443 9303083 : const double *hab, double *block) {
444 :
445 : // Define some more convenient aliases.
446 9303083 : const int ncoseta = ncoset(ibasis->lmax[iset]);
447 9303083 : const int ncosetb = ncoset(jbasis->lmax[jset]);
448 9303083 : const int ncoa = ibasis->npgf[iset] * ncoseta; // size of carthesian set
449 9303083 : const int ncob = jbasis->npgf[jset] * ncosetb;
450 :
451 9303083 : const int nsgf_seta = ibasis->nsgf_set[iset]; // size of spherical set
452 9303083 : const int nsgf_setb = jbasis->nsgf_set[jset];
453 9303083 : const int nsgfa = ibasis->nsgf; // size of entire spherical basis
454 9303083 : const int nsgfb = jbasis->nsgf;
455 9303083 : const int sgfa = ibasis->first_sgf[iset] - 1; // start of spherical set
456 9303083 : const int sgfb = jbasis->first_sgf[jset] - 1;
457 9303083 : const int maxcoa = ibasis->maxco;
458 9303083 : const int maxcob = jbasis->maxco;
459 :
460 9303083 : double work[nsgf_setb * ncoa];
461 :
462 : // work[nsgf_setb][ncoa] = MATMUL(jbasis->sphi, hab)
463 9303083 : dgemm('N', 'N', nsgf_setb, ncoa, ncob, 1.0, &jbasis->sphi[sgfb * maxcob],
464 : maxcob, hab, ncoa, 0.0, work, ncoa);
465 :
466 9303083 : if (transpose) {
467 : // subblock[nsgf_setb][nsgf_seta] += MATMUL(work, TRANSPOSE(ibasis->sphi))
468 5649853 : dgemm('N', 'T', nsgf_setb, nsgf_seta, ncoa, 1.0, work, ncoa,
469 5649853 : &ibasis->sphi[sgfa * maxcoa], maxcoa, 1.0,
470 5649853 : &block[sgfb * nsgfa + sgfa], nsgfa);
471 : } else {
472 : // subblock[nsgf_seta][nsgf_setb] += MATMUL(ibasis->sphi, TRANSPOSE(work))
473 3653230 : dgemm('N', 'T', nsgf_seta, nsgf_setb, ncoa, 1.0,
474 3653230 : &ibasis->sphi[sgfa * maxcoa], maxcoa, work, ncoa, 1.0,
475 3653230 : &block[sgfa * nsgfb + sgfb], nsgfb);
476 : }
477 9303083 : }
478 :
479 : /*******************************************************************************
480 : * \brief Integrate a range of tasks that belong to the same grid level.
481 : * \author Ole Schuett
482 : ******************************************************************************/
483 754478 : static void integrate_one_grid_level(
484 : const grid_cpu_task_list *task_list, const int *first_block_task,
485 : const int *last_block_task, const bool compute_tau, const int natoms,
486 : const int npts_global[3], const int npts_local[3], const int shift_local[3],
487 : const int border_width[3], const double dh[3][3], const double dh_inv[3][3],
488 : const offload_buffer *pab_blocks, const offload_buffer *grid,
489 : offload_buffer *hab_blocks, double forces[natoms][3], double virial[3][3]) {
490 :
491 : // Using default(shared) because with GCC 9 the behavior around const changed:
492 : // https://www.gnu.org/software/gcc/gcc-9/porting_to.html
493 754478 : #pragma omp parallel default(shared)
494 : {
495 : // Initialize variables to detect when a new subblock has to be fetched.
496 : int old_offset = -1, old_iset = -1, old_jset = -1;
497 : grid_basis_set *old_ibasis = NULL, *old_jbasis = NULL;
498 : bool old_transpose = false;
499 :
500 : // Matrix pab and hab are re-used across tasks.
501 : double pab[imax(task_list->maxco * task_list->maxco, 1)];
502 : double hab[imax(task_list->maxco * task_list->maxco, 1)];
503 :
504 : // Parallelize over blocks to avoid concurred access to hab_blocks.
505 : const int nthreads = omp_get_num_threads();
506 : const int chunk_size = imax(1, task_list->nblocks / (nthreads * 50));
507 : #pragma omp for schedule(dynamic, chunk_size)
508 : for (int block_num = 0; block_num < task_list->nblocks; block_num++) {
509 : const int first_task = first_block_task[block_num];
510 : const int last_task = last_block_task[block_num];
511 :
512 : // Accumulate forces per block as it corresponds to a pair of atoms.
513 : const int iatom = task_list->tasks[first_task].iatom - 1;
514 : const int jatom = task_list->tasks[first_task].jatom - 1;
515 : double my_forces[2][3] = {0};
516 : double my_virials[2][3][3] = {0};
517 :
518 : for (int itask = first_task; itask <= last_task; itask++) {
519 : // Define some convenient aliases.
520 : const grid_cpu_task *task = &task_list->tasks[itask];
521 : assert(task->block_num - 1 == block_num);
522 : assert(task->iatom - 1 == iatom && task->jatom - 1 == jatom);
523 : const int ikind = task_list->atom_kinds[iatom] - 1;
524 : const int jkind = task_list->atom_kinds[jatom] - 1;
525 : grid_basis_set *ibasis = task_list->basis_sets[ikind];
526 : grid_basis_set *jbasis = task_list->basis_sets[jkind];
527 : const int iset = task->iset - 1;
528 : const int jset = task->jset - 1;
529 : const int ipgf = task->ipgf - 1;
530 : const int jpgf = task->jpgf - 1;
531 : const double zeta = ibasis->zet[iset * ibasis->maxpgf + ipgf];
532 : const double zetb = jbasis->zet[jset * jbasis->maxpgf + jpgf];
533 : const int ncoseta = ncoset(ibasis->lmax[iset]);
534 : const int ncosetb = ncoset(jbasis->lmax[jset]);
535 : const int ncoa = ibasis->npgf[iset] * ncoseta; // size of carthesian set
536 : const int ncob = jbasis->npgf[jset] * ncosetb;
537 : const int block_offset = task_list->block_offsets[block_num];
538 : const bool transpose = (iatom <= jatom);
539 : const bool pab_required = (forces != NULL || virial != NULL);
540 :
541 : // Load pab and store hab subblocks when needed.
542 : // Previous hab and pab can be reused when only ipgf or jpgf changed.
543 : if (block_offset != old_offset || iset != old_iset ||
544 : jset != old_jset) {
545 : if (pab_required) {
546 : load_pab(ibasis, jbasis, iset, jset, transpose,
547 : &pab_blocks->host_buffer[block_offset], pab);
548 : }
549 : if (old_offset >= 0) { // skip first iteration
550 : store_hab(old_ibasis, old_jbasis, old_iset, old_jset, old_transpose,
551 : hab, &hab_blocks->host_buffer[old_offset]);
552 : }
553 : memset(hab, 0, ncoa * ncob * sizeof(double));
554 : old_offset = block_offset;
555 : old_iset = iset;
556 : old_jset = jset;
557 : old_ibasis = ibasis;
558 : old_jbasis = jbasis;
559 : old_transpose = transpose;
560 : }
561 :
562 : grid_cpu_integrate_pgf_product(
563 : /*orthorhombic=*/task_list->orthorhombic,
564 : /*compute_tau=*/compute_tau,
565 : /*border_mask=*/task->border_mask,
566 : /*la_max=*/ibasis->lmax[iset],
567 : /*la_min=*/ibasis->lmin[iset],
568 : /*lb_max=*/jbasis->lmax[jset],
569 : /*lb_min=*/jbasis->lmin[jset],
570 : /*zeta=*/zeta,
571 : /*zetb=*/zetb,
572 : /*dh=*/dh,
573 : /*dh_inv=*/dh_inv,
574 : /*ra=*/&task_list->atom_positions[3 * iatom],
575 : /*rab=*/task->rab,
576 : /*npts_global=*/npts_global,
577 : /*npts_local=*/npts_local,
578 : /*shift_local=*/shift_local,
579 : /*border_width=*/border_width,
580 : /*radius=*/task->radius,
581 : /*o1=*/ipgf * ncoseta,
582 : /*o2=*/jpgf * ncosetb,
583 : /*n1=*/ncoa,
584 : /*n2=*/ncob,
585 : /*grid=*/grid->host_buffer,
586 : /*hab=*/(double(*)[ncoa])hab,
587 : /*pab=*/(pab_required) ? (const double(*)[ncoa])pab : NULL,
588 : /*forces=*/(forces != NULL) ? my_forces : NULL,
589 : /*virials=*/(virial != NULL) ? my_virials : NULL,
590 : /*hdab=*/NULL,
591 : /*hadb=*/NULL,
592 : /*a_hdab=*/NULL);
593 :
594 : } // end of task loop
595 :
596 : // Merge thread-local forces and virial into shared ones.
597 : // It does not seem worth the trouble to accumulate them thread-locally.
598 : const double scalef = (iatom == jatom) ? 1.0 : 2.0;
599 : if (forces != NULL) {
600 : #pragma omp critical(forces)
601 : for (int i = 0; i < 3; i++) {
602 : forces[iatom][i] += scalef * my_forces[0][i];
603 : forces[jatom][i] += scalef * my_forces[1][i];
604 : }
605 : }
606 : if (virial != NULL) {
607 : #pragma omp critical(virial)
608 : for (int i = 0; i < 3; i++) {
609 : for (int j = 0; j < 3; j++) {
610 : virial[i][j] += scalef * my_virials[0][i][j];
611 : virial[i][j] += scalef * my_virials[1][i][j];
612 : }
613 : }
614 : }
615 :
616 : } // end of block loop
617 :
618 : // store final hab
619 : if (old_offset >= 0) {
620 : store_hab(old_ibasis, old_jbasis, old_iset, old_jset, old_transpose, hab,
621 : &hab_blocks->host_buffer[old_offset]);
622 : }
623 :
624 : } // end of omp parallel region
625 754478 : }
626 :
627 : /*******************************************************************************
628 : * \brief Integrate all tasks of in given list from given grids.
629 : * See grid_task_list.h for details.
630 : * \author Ole Schuett
631 : ******************************************************************************/
632 190660 : void grid_cpu_integrate_task_list(
633 : const grid_cpu_task_list *task_list, const bool compute_tau,
634 : const int natoms, const int nlevels, const offload_buffer *pab_blocks,
635 : const offload_buffer *grids[nlevels], offload_buffer *hab_blocks,
636 : double forces[natoms][3], double virial[3][3]) {
637 :
638 190660 : assert(task_list->nlevels == nlevels);
639 190660 : assert(task_list->natoms == natoms);
640 190660 : assert(hab_blocks != NULL);
641 :
642 : // Zero result arrays.
643 190660 : if (hab_blocks->size != 0) {
644 188859 : memset(hab_blocks->host_buffer, 0, hab_blocks->size);
645 : }
646 190660 : if (forces != NULL) {
647 23559 : memset(forces, 0, natoms * 3 * sizeof(double));
648 : }
649 190660 : if (virial != NULL) {
650 3619 : memset(virial, 0, 9 * sizeof(double));
651 : }
652 :
653 945138 : for (int level = 0; level < task_list->nlevels; level++) {
654 754478 : const int idx = level * task_list->nblocks;
655 754478 : const int *first_block_task = &task_list->first_level_block_task[idx];
656 754478 : const int *last_block_task = &task_list->last_level_block_task[idx];
657 754478 : const grid_cpu_layout *layout = &task_list->layouts[level];
658 754478 : integrate_one_grid_level(
659 : task_list, first_block_task, last_block_task, compute_tau, natoms,
660 754478 : layout->npts_global, layout->npts_local, layout->shift_local,
661 754478 : layout->border_width, layout->dh, layout->dh_inv, pab_blocks,
662 754478 : grids[level], hab_blocks, forces, virial);
663 : }
664 190660 : }
665 :
666 : // EOF
|
