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 : #include "dbm_multiply.h"
8 : #include "../offload/offload_mempool.h"
9 : #include "../offload/offload_runtime.h"
10 : #include "dbm_hyperparams.h"
11 : #include "dbm_internal.h"
12 : #include "dbm_library.h"
13 : #include "dbm_multiply_comm.h"
14 : #include "dbm_multiply_cpu.h"
15 : #include "dbm_multiply_gpu.h"
16 :
17 : #include <assert.h>
18 : #include <limits.h>
19 : #include <math.h>
20 : #include <omp.h>
21 : #include <stdio.h>
22 : #include <stdlib.h>
23 : #include <string.h>
24 :
25 : /*******************************************************************************
26 : * \brief Private routine for computing the max filter threshold for each row.
27 : * \author Ole Schuett
28 : ******************************************************************************/
29 242058 : static float *compute_rows_max_eps(const bool trans, const dbm_matrix_t *matrix,
30 : const double filter_eps) {
31 242058 : const int nrows = (trans) ? matrix->ncols : matrix->nrows;
32 242058 : int *nblocks_per_row = calloc(nrows, sizeof(int));
33 242058 : float *row_max_eps = malloc(nrows * sizeof(float));
34 242058 : assert((nblocks_per_row != NULL && row_max_eps != NULL) || nrows == 0);
35 :
36 242058 : #pragma omp parallel
37 : {
38 : #pragma omp for
39 : for (int ishard = 0; ishard < dbm_get_num_shards(matrix); ishard++) {
40 : dbm_shard_t *shard = &matrix->shards[ishard];
41 : for (int iblock = 0; iblock < shard->nblocks; iblock++) {
42 : const dbm_block_t *blk = &shard->blocks[iblock];
43 : const int row = (trans) ? blk->col : blk->row;
44 : #pragma omp atomic
45 : ++nblocks_per_row[row];
46 : }
47 : }
48 : #pragma omp master
49 : cp_mpi_sum_int(nblocks_per_row, nrows, matrix->dist->comm);
50 : #pragma omp barrier
51 : #pragma omp for
52 : for (int i = 0; i < nrows; i++) {
53 : const float f =
54 : ((float)filter_eps) / ((float)imax(1, nblocks_per_row[i]));
55 : row_max_eps[i] = f * f;
56 : }
57 : } // end of omp parallel region
58 :
59 242058 : free(nblocks_per_row);
60 242058 : return row_max_eps; // Ownership of row_max_eps transfers to caller.
61 : }
62 :
63 : /*******************************************************************************
64 : * \brief Private struct for storing the context of the multiplication backend.
65 : * \author Ole Schuett
66 : ******************************************************************************/
67 : typedef struct {
68 : #if defined(__OFFLOAD) && !defined(__NO_OFFLOAD_DBM)
69 : dbm_multiply_gpu_context_t gpu;
70 : #endif
71 : int cpu_options; // Binary or'ed dbm_multiply_cpu_options (enum).
72 : } backend_context_t;
73 :
74 : /*******************************************************************************
75 : * \brief Private routine for initializing the multiplication backend.
76 : * \author Ole Schuett
77 : ******************************************************************************/
78 242058 : static backend_context_t *backend_start(const dbm_matrix_t *matrix_c) {
79 242058 : backend_context_t *const ctx = calloc(1, sizeof(backend_context_t));
80 : // BLAS and LIBXS benefit in general from DBM_MULTIPLY_TASK_REORDER.
81 242058 : ctx->cpu_options = DBM_MULTIPLY_TASK_REORDER;
82 :
83 : #if defined(__OFFLOAD) && !defined(__NO_OFFLOAD_DBM)
84 : dbm_multiply_gpu_start(DBM_MAX_BATCH_SIZE, dbm_get_num_shards(matrix_c),
85 : matrix_c->shards, &ctx->gpu);
86 : #else
87 242058 : (void)matrix_c; // mark as used
88 : #endif
89 :
90 242058 : return ctx;
91 : }
92 :
93 : /*******************************************************************************
94 : * \brief Private routine for handing newly arrived packs to the backend.
95 : * \author Ole Schuett
96 : ******************************************************************************/
97 0 : static bool backend_upload_packs(const dbm_pack_t *pack_a,
98 : const dbm_pack_t *pack_b,
99 : backend_context_t *ctx) {
100 : #if defined(__OFFLOAD) && !defined(__NO_OFFLOAD_DBM)
101 : return dbm_multiply_gpu_upload_packs(pack_a, pack_b, &ctx->gpu);
102 : #else
103 0 : (void)pack_a; // mark as used
104 0 : (void)pack_b;
105 0 : (void)ctx;
106 0 : return false;
107 : #endif
108 : }
109 :
110 : /*******************************************************************************
111 : * \brief Private routine for sending a batch to the multiplication backend.
112 : * \author Ole Schuett
113 : ******************************************************************************/
114 263542 : static void backend_process_batch(const int ntasks,
115 : const dbm_task_t batch[ntasks],
116 : const double alpha, const dbm_pack_t *pack_a,
117 : const dbm_pack_t *pack_b, const int kshard,
118 : dbm_shard_t *shard_c, const bool finish,
119 : const bool force_cpu,
120 : backend_context_t *ctx) {
121 263542 : if (NULL != ctx) {
122 : #if defined(__OFFLOAD) && !defined(__NO_OFFLOAD_DBM)
123 : if (!force_cpu) {
124 : dbm_multiply_gpu_process_batch(ntasks, batch, alpha, shard_c, kshard,
125 : finish, &ctx->gpu);
126 : } else
127 : #endif
128 : {
129 263542 : (void)kshard;
130 263542 : (void)finish;
131 263542 : (void)force_cpu;
132 263542 : dbm_multiply_cpu_process_batch(ntasks, batch, alpha, pack_a, pack_b,
133 : shard_c, ctx->cpu_options);
134 : }
135 : } else { // Validate against host (aka CPU).
136 0 : dbm_multiply_cpu_process_batch(ntasks, batch, alpha, pack_a, pack_b,
137 : shard_c, DBM_MULTIPLY_BLAS_LIBRARY);
138 : }
139 263542 : }
140 :
141 : /*******************************************************************************
142 : * \brief Private routine for shutting down the multiplication backend.
143 : * \author Ole Schuett
144 : ******************************************************************************/
145 242058 : static void backend_stop(backend_context_t *ctx) {
146 : #if defined(__OFFLOAD) && !defined(__NO_OFFLOAD_DBM)
147 : dbm_multiply_gpu_stop(&ctx->gpu);
148 : #endif
149 242058 : free(ctx);
150 242058 : }
151 :
152 : /*******************************************************************************
153 : * \brief Private routine for multipling two packs.
154 : * \author Ole Schuett
155 : ******************************************************************************/
156 263464 : static void multiply_packs(const bool transa, const bool transb,
157 : const double alpha, const dbm_pack_t *pack_a,
158 : const dbm_pack_t *pack_b,
159 : const dbm_matrix_t *matrix_a,
160 : const dbm_matrix_t *matrix_b, dbm_matrix_t *matrix_c,
161 : const float *rows_max_eps,
162 : const bool retain_sparsity, const bool force_cpu,
163 : int64_t *flop, backend_context_t *ctx) {
164 : // For validation, FLOPS do not count, and relying on ctx is not necessary.
165 263464 : backend_context_t *const context = (NULL != flop ? ctx : NULL);
166 263464 : const float alpha2 = (float)(alpha * alpha);
167 263464 : int64_t flop_sum = 0;
168 :
169 263464 : const int nshard_rows = matrix_c->dist->rows.nshards;
170 263464 : const int nshard_cols = matrix_c->dist->cols.nshards;
171 263464 : int *shard_row_start = calloc(nshard_rows, sizeof(int));
172 263464 : int *shard_col_start = calloc(nshard_cols, sizeof(int));
173 263464 : assert(NULL != shard_row_start && NULL != shard_col_start);
174 :
175 263464 : const int *sum_index_sizes_a =
176 : (transa) ? matrix_a->row_sizes : matrix_a->col_sizes;
177 263464 : const int *sum_index_sizes_b =
178 : (transb) ? matrix_b->col_sizes : matrix_b->row_sizes;
179 263464 : const int *free_index_sizes_a =
180 : (transa) ? matrix_a->col_sizes : matrix_a->row_sizes;
181 263464 : const int *free_index_sizes_b =
182 : (transb) ? matrix_b->row_sizes : matrix_b->col_sizes;
183 :
184 263464 : #pragma omp parallel reduction(+ : flop_sum)
185 : {
186 : // Thread-private array covering given work in piece-wise fashion.
187 : dbm_task_t *batch =
188 : offload_mempool_host_malloc(sizeof(dbm_task_t) * DBM_MAX_BATCH_SIZE);
189 :
190 : // Blocks are ordered first by shard. Creating lookup tables of boundaries.
191 : #pragma omp for nowait
192 : for (int iblock = 1; iblock < pack_a->nblocks; iblock++) {
193 : const int shard_row = pack_a->blocks[iblock].free_index % nshard_rows;
194 : const int prev_shard_row =
195 : pack_a->blocks[iblock - 1].free_index % nshard_rows;
196 : if (prev_shard_row != shard_row) {
197 : shard_row_start[shard_row] = iblock;
198 : }
199 : }
200 : #pragma omp for
201 : for (int jblock = 1; jblock < pack_b->nblocks; jblock++) {
202 : const int shard_col = pack_b->blocks[jblock].free_index % nshard_cols;
203 : const int prev_shard_col =
204 : pack_b->blocks[jblock - 1].free_index % nshard_cols;
205 : if (prev_shard_col != shard_col) {
206 : shard_col_start[shard_col] = jblock;
207 : }
208 : }
209 :
210 : #pragma omp for collapse(2) DBM_OMP_SCHEDULE
211 : for (int shard_row = 0; shard_row < nshard_rows; shard_row++) {
212 : for (int shard_col = 0; shard_col < nshard_cols; shard_col++) {
213 : const int ishard = shard_row * nshard_cols + shard_col;
214 : dbm_shard_t *const shard_c = &matrix_c->shards[ishard];
215 : int ntasks = 0;
216 :
217 : // Determine contiguous block ranges for this shard in A and B.
218 : // Use a merge-join to find pairs of blocks with matching sum indices.
219 : // This utilizes that blocks within a shard are ordered by sum_index.
220 : const int iblock_start = shard_row_start[shard_row];
221 : int iblock_end = pack_a->nblocks;
222 : for (int t = iblock_start; t < pack_a->nblocks; ++t) {
223 : if (pack_a->blocks[t].free_index % nshard_rows != shard_row) {
224 : iblock_end = t;
225 : break;
226 : }
227 : }
228 : const int jblock_start = shard_col_start[shard_col];
229 : int jblock_end = pack_b->nblocks;
230 : for (int t = jblock_start; t < pack_b->nblocks; ++t) {
231 : if (pack_b->blocks[t].free_index % nshard_cols != shard_col) {
232 : jblock_end = t;
233 : break;
234 : }
235 : }
236 : if (iblock_start >= iblock_end || jblock_start >= jblock_end) {
237 : backend_process_batch(ntasks, batch, alpha, pack_a, pack_b, ishard,
238 : shard_c, true, force_cpu, context);
239 : continue;
240 : }
241 :
242 : // Merge over sum_index (both ranges sorted by sum_index).
243 : int i = iblock_start, j = jblock_start, last_sum_index = -1;
244 : int b_range_start = -1, b_range_end = -1;
245 :
246 : while (i < iblock_end) {
247 : const dbm_pack_block_t *blk_a = &pack_a->blocks[i];
248 : const int sum_a = blk_a->sum_index;
249 :
250 : // Advance j until sum_b >= sum_a.
251 : while (j < jblock_end && pack_b->blocks[j].sum_index < sum_a) {
252 : ++j;
253 : }
254 : if (j >= jblock_end) {
255 : break; // No more matches possible.
256 : }
257 :
258 : const int sum_b = pack_b->blocks[j].sum_index;
259 : if (sum_b > sum_a) {
260 : ++i;
261 : continue; // Need next A block with higher sum_index.
262 : }
263 :
264 : // sum_a == sum_b: establish (or reuse) B range with this sum_index.
265 : if (sum_a != last_sum_index) {
266 : b_range_start = j;
267 : int t = j + 1;
268 : while (t < jblock_end && pack_b->blocks[t].sum_index == sum_a) {
269 : ++t;
270 : }
271 : b_range_end = t;
272 : last_sum_index = sum_a;
273 : }
274 :
275 : // Iterate over B blocks in current sum_index range.
276 : for (int jb = b_range_start; jb < b_range_end; ++jb) {
277 : const dbm_pack_block_t *const blk_b = &pack_b->blocks[jb];
278 :
279 : // Norm filter first (early reject).
280 : const float result_norm = alpha2 * blk_a->norm * blk_b->norm;
281 : if (result_norm < rows_max_eps[blk_a->free_index]) {
282 : continue;
283 : }
284 :
285 : // Check block sizes.
286 : const int m = free_index_sizes_a[blk_a->free_index];
287 : const int n = free_index_sizes_b[blk_b->free_index];
288 : const int k = sum_index_sizes_a[sum_a];
289 : assert(m == matrix_c->row_sizes[blk_a->free_index]);
290 : assert(n == matrix_c->col_sizes[blk_b->free_index]);
291 : assert(k == sum_index_sizes_b[blk_b->sum_index]);
292 :
293 : if (m == 0 || n == 0 || k == 0) {
294 : continue;
295 : }
296 :
297 : // Get C block.
298 : const int row = blk_a->free_index, col = blk_b->free_index;
299 : dbm_block_t *blk_c = dbm_shard_lookup(shard_c, row, col);
300 : if (blk_c == NULL) {
301 : if (retain_sparsity) {
302 : continue;
303 : }
304 : assert(dbm_get_shard_index(matrix_c, row, col) == ishard);
305 : assert(dbm_get_stored_coordinates(matrix_c, row, col) ==
306 : matrix_c->dist->my_rank);
307 : blk_c = dbm_shard_promise_new_block(shard_c, row, col, m * n);
308 : }
309 :
310 : // Count flops.
311 : const int64_t task_flops = 2LL * m * n * k;
312 : flop_sum += task_flops;
313 : dbm_library_counter_increment(m, n, k);
314 :
315 : // Add block multiplication to batch.
316 : dbm_task_t *const tptr = &batch[ntasks];
317 : tptr->offset_a = blk_a->offset;
318 : tptr->offset_b = blk_b->offset;
319 : tptr->offset_c = blk_c->offset;
320 : tptr->m = m;
321 : tptr->n = n;
322 : tptr->k = k;
323 : ++ntasks;
324 :
325 : if (ntasks == DBM_MAX_BATCH_SIZE) {
326 : backend_process_batch(ntasks, batch, alpha, pack_a, pack_b,
327 : ishard, shard_c, false, force_cpu, context);
328 : ntasks = 0;
329 : }
330 : }
331 :
332 : // Advance i; if next A block has same sum_index, B range is reused.
333 : ++i;
334 : }
335 : backend_process_batch(ntasks, batch, alpha, pack_a, pack_b, ishard,
336 : shard_c, true, force_cpu, context);
337 : }
338 : }
339 :
340 : offload_mempool_host_free(batch);
341 : }
342 :
343 263464 : free(shard_row_start);
344 263464 : free(shard_col_start);
345 :
346 263464 : if (NULL != flop) {
347 263464 : *flop += flop_sum;
348 : }
349 263464 : }
350 :
351 : /*******************************************************************************
352 : * \brief Performs a multiplication of two dbm_matrix_t matrices.
353 : * See dbm_matrix.h for details.
354 : * \author Ole Schuett
355 : ******************************************************************************/
356 242058 : void dbm_multiply(const bool transa, const bool transb, const double alpha,
357 : const dbm_matrix_t *matrix_a, const dbm_matrix_t *matrix_b,
358 : const double beta, dbm_matrix_t *matrix_c,
359 : const bool retain_sparsity, const double filter_eps,
360 : int64_t *flop) {
361 242058 : assert(omp_get_num_threads() == 1);
362 242058 : assert(matrix_a != NULL && matrix_b != NULL && matrix_c != NULL);
363 :
364 : // Throughout the matrix multiplication code the "sum_index" and "free_index"
365 : // denote the summation (aka dummy) and free index from the Einstein notation.
366 242058 : const int num_sum_index_a = (transa) ? matrix_a->nrows : matrix_a->ncols;
367 242058 : const int num_sum_index_b = (transb) ? matrix_b->ncols : matrix_b->nrows;
368 242058 : const int num_free_index_a = (transa) ? matrix_a->ncols : matrix_a->nrows;
369 242058 : const int num_free_index_b = (transb) ? matrix_b->nrows : matrix_b->ncols;
370 :
371 : // Sanity check matrix dimensions.
372 242058 : assert(num_sum_index_a == num_sum_index_b);
373 242058 : assert(num_free_index_a == matrix_c->nrows);
374 242058 : assert(num_free_index_b == matrix_c->ncols);
375 :
376 : // Prepare matrix_c (host).
377 242058 : dbm_scale(matrix_c, beta);
378 :
379 : // Determine if validation shall be performed.
380 242058 : const char *const maxeps_env = getenv("DBM_MULTIPLY_MAXEPS");
381 242058 : const char *const verify_env = getenv("DBM_MULTIPLY_VERIFY");
382 242058 : const double maxeps = (NULL == maxeps_env ? 1E-1 : fabs(atof(maxeps_env)));
383 484116 : const int verify =
384 242058 : (NULL == verify_env ? (NULL == maxeps_env ? 0 : 1) : atoi(verify_env));
385 242058 : dbm_matrix_t *matrix_d = NULL;
386 242058 : if (0 != verify) {
387 0 : dbm_distribution_t *const dist_shared = matrix_c->dist;
388 0 : dbm_create(&matrix_d, dist_shared, matrix_c->name, matrix_c->nrows,
389 0 : matrix_c->ncols, matrix_c->row_sizes, matrix_c->col_sizes);
390 0 : dbm_copy(matrix_d, matrix_c);
391 : }
392 :
393 : // Compute filter thresholds for each row.
394 242058 : float *rows_max_eps = compute_rows_max_eps(transa, matrix_a, filter_eps);
395 :
396 : // Start uploading matrix_c to the GPU.
397 242058 : backend_context_t *ctx = backend_start(matrix_c);
398 :
399 : // Redistribute matrix_a and matrix_b across MPI ranks.
400 242058 : dbm_comm_iterator_t *iter =
401 242058 : dbm_comm_iterator_start(transa, transb, matrix_a, matrix_b, matrix_c);
402 :
403 : // Count flops if requested.
404 242058 : if (NULL != flop) {
405 242058 : *flop = 0;
406 : }
407 :
408 : // Main loop.
409 : dbm_pack_t *pack_a, *pack_b;
410 505522 : while (dbm_comm_iterator_next(iter, &pack_a, &pack_b)) {
411 263464 : const bool uploaded = backend_upload_packs(pack_a, pack_b, ctx);
412 263464 : (void)uploaded; // mark used
413 263464 : multiply_packs(transa, transb, alpha, pack_a, pack_b, matrix_a, matrix_b,
414 : matrix_c, rows_max_eps, retain_sparsity, false /*!uploaded*/,
415 : flop, ctx);
416 : }
417 :
418 : // Wait for all other MPI ranks to complete, then release ressources.
419 242058 : dbm_comm_iterator_stop(iter);
420 242058 : backend_stop(ctx);
421 :
422 242058 : if (NULL != matrix_d) {
423 0 : ctx = backend_start(matrix_d);
424 0 : iter =
425 0 : dbm_comm_iterator_start(transa, transb, matrix_a, matrix_b, matrix_d);
426 0 : while (dbm_comm_iterator_next(iter, &pack_a, &pack_b)) {
427 0 : multiply_packs(transa, transb, alpha, pack_a, pack_b, matrix_a, matrix_b,
428 : matrix_d, rows_max_eps, retain_sparsity, true, NULL, ctx);
429 : }
430 0 : dbm_comm_iterator_stop(iter);
431 0 : backend_stop(ctx);
432 0 : const double epsilon = dbm_maxeps(matrix_d, matrix_c);
433 0 : if (maxeps < epsilon) {
434 0 : if (1 == verify) {
435 0 : fprintf(stderr, "WARN ACC/LIBDBM: diff=%g\n", epsilon);
436 : } else {
437 0 : fprintf(stderr, "ERROR ACC/LIBDBM: diff=%g\n", epsilon);
438 0 : exit(EXIT_FAILURE);
439 : }
440 : }
441 0 : dbm_release(matrix_d);
442 : }
443 :
444 : // Release filter thresholds.
445 242058 : free(rows_max_eps);
446 :
447 : // Final filter pass.
448 242058 : dbm_filter(matrix_c, filter_eps);
449 242058 : }
450 :
451 : // EOF
|
