1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
|
//------------------------------------------------------------------------------
// GB_split_sparse: split a sparse/hypersparse matrix into tiles
//------------------------------------------------------------------------------
// SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2025, All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0
//------------------------------------------------------------------------------
// Each output tile is first created in sparse/hyper form, matching the input
// matrix, and then conformed to its desired sparsity format.
#define GB_FREE_WORKSPACE \
GB_WERK_POP (C_ek_slicing, int64_t) ; \
GB_FREE_MEMORY (&Wp, Wp_size) ;
#define GB_FREE_ALL \
GB_FREE_WORKSPACE ; \
GB_Matrix_free (&C) ;
#include "split/GB_split.h"
#include "jitifyer/GB_stringify.h"
#include "apply/GB_apply.h"
GrB_Info GB_split_sparse // split a sparse matrix
(
GrB_Matrix *Tiles, // 2D row-major array of size m-by-n
const int64_t m,
const int64_t n,
const int64_t *restrict Tile_rows, // size m+1
const int64_t *restrict Tile_cols, // size n+1
const GrB_Matrix A, // input matrix
GB_Werk Werk
)
{
//--------------------------------------------------------------------------
// get inputs
//--------------------------------------------------------------------------
GrB_Info info ;
int A_sparsity = GB_sparsity (A) ;
bool A_is_hyper = (A_sparsity == GxB_HYPERSPARSE) ;
ASSERT (A_is_hyper || A_sparsity == GxB_SPARSE) ;
GrB_Matrix C = NULL ;
GB_WERK_DECLARE (C_ek_slicing, int64_t) ;
ASSERT_MATRIX_OK (A, "A sparse for split", GB0) ;
ASSERT (!GB_JUMBLED (A)) ;
ASSERT (!GB_ZOMBIES (A)) ;
ASSERT (!GB_PENDING (A)) ;
int sparsity_control = A->sparsity_control ;
float hyper_switch = A->hyper_switch ;
bool csc = A->is_csc ;
GrB_Type atype = A->type ;
size_t asize = atype->size ;
int nthreads_max = GB_Context_nthreads_max ( ) ;
double chunk = GB_Context_chunk ( ) ;
int64_t nouter = csc ? n : m ;
int64_t ninner = csc ? m : n ;
const int64_t *Tile_vdim = csc ? Tile_cols : Tile_rows ;
const int64_t *Tile_vlen = csc ? Tile_rows : Tile_cols ;
int64_t anvec = A->nvec ;
int64_t anz = GB_nnz (A) ;
GB_Ap_DECLARE (Ap, const) ; GB_Ap_PTR (Ap, A) ;
GB_Ah_DECLARE (Ah, const) ; GB_Ah_PTR (Ah, A) ;
GB_Ai_DECLARE (Ai, const) ; GB_Ai_PTR (Ai, A) ;
const bool A_iso = A->iso ;
const bool Ap_is_32 = A->p_is_32 ;
const bool Aj_is_32 = A->j_is_32 ;
const bool Ai_is_32 = A->i_is_32 ;
//--------------------------------------------------------------------------
// allocate workspace
//--------------------------------------------------------------------------
// FUTURE: Wp is allocated with the same integers as Ap, but it could be
// chosen based on anz instead.
GB_MDECL (Wp, , u) ; size_t Wp_size = 0 ;
size_t apsize = (Ap_is_32) ? sizeof (uint32_t) : sizeof (uint64_t) ;
Wp = GB_MALLOC_MEMORY (anvec, apsize, &Wp_size) ;
if (Wp == NULL)
{
// out of memory
GB_FREE_ALL ;
return (GrB_OUT_OF_MEMORY) ;
}
GB_memcpy (Wp, Ap, anvec * apsize, nthreads_max) ;
GB_IPTR (Wp, Ap_is_32) ;
//--------------------------------------------------------------------------
// split A into tiles
//--------------------------------------------------------------------------
int64_t akend = 0 ;
for (int64_t outer = 0 ; outer < nouter ; outer++)
{
//----------------------------------------------------------------------
// find the starting and ending vector of these tiles
//----------------------------------------------------------------------
// The tile appears in vectors avstart:avend-1 of A, and indices
// aistart:aiend-1.
const int64_t avstart = Tile_vdim [outer] ;
const int64_t avend = Tile_vdim [outer+1] ;
int64_t akstart = akend ;
if (A_is_hyper)
{
// A is hypersparse: look for vector avend in the A->h hyper list.
// The vectors to handle for this outer loop are in
// Ah [akstart:akend-1].
akend = akstart ;
int64_t pright = anvec - 1 ;
GB_split_binary_search (avend, Ah, Aj_is_32, &akend, &pright) ;
ASSERT (GB_IMPLIES (akstart <= akend-1,
GB_IGET (Ah, akend-1) < avend)) ;
}
else
{
// A is sparse; the vectors to handle are akstart:akend-1
akend = avend ;
}
// # of vectors in all tiles in this outer loop
int64_t cnvec = akend - akstart ;
int nth = GB_nthreads (cnvec, chunk, nthreads_max) ;
//----------------------------------------------------------------------
// create all tiles for vectors akstart:akend-1 in A
//----------------------------------------------------------------------
for (int64_t inner = 0 ; inner < ninner ; inner++)
{
//------------------------------------------------------------------
// allocate C, C->p, and C->h for this tile
//------------------------------------------------------------------
const int64_t aistart = Tile_vlen [inner] ;
const int64_t aiend = Tile_vlen [inner+1] ;
const int64_t cvdim = avend - avstart ;
const int64_t cvlen = aiend - aistart ;
// Assume this tile C can acquire all the entries of A to determine
// the p_is_32, j_is_32, and i_is_32 settings for the new Tile.
bool Cp_is_32, Cj_is_32, Ci_is_32 ;
GB_determine_pji_is_32 (&Cp_is_32, &Cj_is_32, &Ci_is_32,
A_sparsity, anz, cvlen, cvdim, Werk) ;
C = NULL ;
GB_OK (GB_new (&C, // new header
atype, cvlen, cvdim, GB_ph_malloc, csc, A_sparsity,
hyper_switch, cnvec, Cp_is_32, Cj_is_32, Ci_is_32)) ;
C->sparsity_control = sparsity_control ;
C->hyper_switch = hyper_switch ;
C->nvec = cnvec ;
GB_Cp_DECLARE (Cp, ) ; GB_Ap_PTR (Cp, C) ;
GB_Ch_DECLARE (Ch, ) ; GB_Ah_PTR (Ch, C) ;
//------------------------------------------------------------------
// determine the boundaries of this tile
//------------------------------------------------------------------
int64_t k ;
#pragma omp parallel for num_threads(nth) schedule(static)
for (k = akstart ; k < akend ; k++)
{
const int64_t pC_start = GB_IGET (Wp, k) ;
int64_t pA = pC_start ;
const int64_t pA_end = GB_IGET (Ap, k+1) ;
const int64_t aknz = pA_end - pA ;
if (aknz == 0 || GB_IGET (Ai, pA) >= aiend)
{
// this vector of C is empty
}
else if (aknz > 256)
{
// use binary search to find aiend
int64_t pright = pA_end - 1 ;
GB_split_binary_search (aiend, Ai, Ai_is_32, &pA, &pright) ;
#ifdef GB_DEBUG
// check the results with a linear search
int64_t p2 = pC_start ;
for ( ; p2 < pA_end ; p2++)
{
if (GB_IGET (Ai, p2) >= aiend) break ;
}
ASSERT (pA == p2) ;
#endif
}
else
{
// use a linear-time search to find aiend
for ( ; pA < pA_end ; pA++)
{
if (GB_IGET (Ai, pA) >= aiend) break ;
}
#ifdef GB_DEBUG
// check the results with a binary search
int64_t p2 = pC_start ;
int64_t p2_end = pA_end - 1 ;
GB_split_binary_search (aiend, Ai, Ai_is_32, &p2, &p2_end) ;
ASSERT (pA == p2) ;
#endif
}
int64_t kC = k - akstart ;
int64_t cknz = pA - pC_start ; // # entries in C(:,kC)
GB_ISET (Cp, kC, cknz) ; // Cp [kC] = cknz ;
if (A_is_hyper)
{
int64_t jC = GB_IGET (Ah, k) - avstart ;
GB_ISET (Ch, kC, jC) ; // Ch [kC] = jC ;
}
}
int64_t nvec_nonempty ;
GB_cumsum (Cp, Cp_is_32, cnvec, &nvec_nonempty, nth, Werk) ;
GB_nvec_nonempty_set (C, nvec_nonempty) ;
int64_t cnz = GB_IGET (Cp, cnvec) ;
//------------------------------------------------------------------
// allocate C->i and C->x for this tile
//------------------------------------------------------------------
GB_OK (GB_bix_alloc (C, cnz, GxB_SPARSE, false, true, A_iso)) ;
GB_Ci_DECLARE (Ci, ) ; GB_Ci_PTR (Ci, C) ;
C->nvals = cnz ;
C->magic = GB_MAGIC ; // for GB_nnz_held(C), to slice C
//------------------------------------------------------------------
// copy the tile from A into C
//------------------------------------------------------------------
int C_ntasks, C_nthreads ;
GB_SLICE_MATRIX (C, 8) ;
info = GrB_NO_VALUE ;
if (A_iso)
{
//--------------------------------------------------------------
// split an iso matrix A into an iso tile C
//--------------------------------------------------------------
// A is iso and so is C; copy the iso entry
GBURBLE ("(iso sparse split) ") ;
memcpy (C->x, A->x, asize) ;
#define GB_ISO_SPLIT
#define GB_COPY(pC,pA) ;
#include "split/template/GB_split_sparse_template.c"
info = GrB_SUCCESS ;
}
else
{
//--------------------------------------------------------------
// split a non-iso matrix A into an non-iso tile C
//--------------------------------------------------------------
#ifndef GBCOMPACT
GB_IF_FACTORY_KERNELS_ENABLED
{
// no typecasting needed
switch (asize)
{
#undef GB_COPY
#define GB_COPY(pC,pA) Cx [pC] = Ax [pA] ;
case GB_1BYTE : // uint8, int8, bool, or 1-byte user
#define GB_C_TYPE uint8_t
#define GB_A_TYPE uint8_t
#include "split/template/GB_split_sparse_template.c"
info = GrB_SUCCESS ;
break ;
case GB_2BYTE : // uint16, int16, or 2-byte user-defined
#define GB_C_TYPE uint16_t
#define GB_A_TYPE uint16_t
#include "split/template/GB_split_sparse_template.c"
info = GrB_SUCCESS ;
break ;
case GB_4BYTE : // uint32, int32, float, or 4-byte user
#define GB_C_TYPE uint32_t
#define GB_A_TYPE uint32_t
#include "split/template/GB_split_sparse_template.c"
info = GrB_SUCCESS ;
break ;
case GB_8BYTE : // uint64, int64, double, float complex,
// or 8-byte user defined
#define GB_C_TYPE uint64_t
#define GB_A_TYPE uint64_t
#include "split/template/GB_split_sparse_template.c"
info = GrB_SUCCESS ;
break ;
case GB_16BYTE : // double complex or 16-byte user
#define GB_C_TYPE GB_blob16
#define GB_A_TYPE GB_blob16
#include "split/template/GB_split_sparse_template.c"
info = GrB_SUCCESS ;
break ;
default:;
}
}
#endif
}
//------------------------------------------------------------------
// via the JIT or PreJIT kernel
//------------------------------------------------------------------
if (info == GrB_NO_VALUE)
{
struct GB_UnaryOp_opaque op_header ;
GB_Operator op = GB_unop_identity (atype, &op_header) ;
ASSERT_OP_OK (op, "identity op for split sparse", GB0) ;
info = GB_split_sparse_jit (C, op, A, akstart, aistart, Wp,
C_ek_slicing, C_ntasks, C_nthreads) ;
}
//------------------------------------------------------------------
// via the generic kernel
//------------------------------------------------------------------
if (info == GrB_NO_VALUE)
{
GBURBLE ("(generic split) ") ;
#define GB_C_TYPE GB_void
#define GB_A_TYPE GB_void
#undef GB_COPY
#define GB_COPY(pC,pA) \
memcpy (Cx + (pC)*asize, Ax +(pA)*asize, asize) ;
#include "split/template/GB_split_sparse_template.c"
info = GrB_SUCCESS ;
}
//------------------------------------------------------------------
// free workspace
//------------------------------------------------------------------
GB_WERK_POP (C_ek_slicing, int64_t) ;
GB_OK (info) ;
//------------------------------------------------------------------
// advance to the next tile
//------------------------------------------------------------------
if (inner < ninner - 1)
{
int64_t k ;
#pragma omp parallel for num_threads(nth) schedule(static)
for (k = akstart ; k < akend ; k++)
{
int64_t ck = k - akstart ;
int64_t cknz = GB_IGET (Cp, ck+1) - GB_IGET (Cp, ck) ;
GB_IINC (Wp, k, cknz) ; // Wp [k] += cknz ;
}
}
//------------------------------------------------------------------
// conform the tile and save it in the Tiles array
//------------------------------------------------------------------
ASSERT_MATRIX_OK (C, "C for GB_split", GB0) ;
GB_OK (GB_hyper_prune (C, Werk)) ;
GB_OK (GB_conform (C, Werk)) ;
if (csc)
{
GB_TILE (Tiles, inner, outer) = C ;
}
else
{
GB_TILE (Tiles, outer, inner) = C ;
}
ASSERT_MATRIX_OK (C, "final tile C for GB_split", GB0) ;
C = NULL ;
}
}
GB_FREE_WORKSPACE ;
return (GrB_SUCCESS) ;
}
|
