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
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
|
//------------------------------------------------------------------------------
// GB_AxB_saxpy4_template: C+=A*B, C is full, A is sparse/hyper, B bitmap/full
//------------------------------------------------------------------------------
// SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2025, All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0
//------------------------------------------------------------------------------
// C is full. A is hyper/sparse, B is bitmap/full. M is not present.
// C += A*B is computed with the accumulator identical to the monoid.
// This template is used by mxm/template/GB_AxB_saxpy4_meta. It is not used
// for the generic case, nor for the ANY_PAIR case. It is only used for the
// pre-generated kernels, and for the JIT.
#ifndef GB_B_SIZE
#define GB_B_SIZE sizeof (GB_B_TYPE)
#endif
#ifndef GB_C_SIZE
#define GB_C_SIZE sizeof (GB_C_TYPE)
#endif
#ifndef GB_C_ISO
#define GB_C_ISO 0
#endif
{
if (use_coarse_tasks)
{
//----------------------------------------------------------------------
// C += A*B using coarse tasks
//----------------------------------------------------------------------
int64_t cvlenx = cvlen * GB_C_SIZE ;
int tid ;
#pragma omp parallel for num_threads(nthreads) schedule(dynamic,1)
for (tid = 0 ; tid < ntasks ; tid++)
{
//------------------------------------------------------------------
// determine the vectors of B and C for this coarse task
//------------------------------------------------------------------
int64_t jstart, jend ;
GB_PARTITION (jstart, jend, bvdim, tid, ntasks) ;
int64_t jtask = jend - jstart ;
int64_t jpanel = GB_IMIN (jtask, GB_SAXPY4_PANEL_SIZE) ;
//------------------------------------------------------------------
// get the workspace for this task
//------------------------------------------------------------------
// Hx workspace to compute the panel of C
GB_C_TYPE *restrict Hx = (GB_C_TYPE *)
(Wcx + H_slice [tid] * cvlenx) ;
//------------------------------------------------------------------
// C(:,jstart:jend-1) += A * B(:,jstart:jend-1) by panel
//------------------------------------------------------------------
for (int64_t j1 = jstart ; j1 < jend ; j1 += jpanel)
{
//--------------------------------------------------------------
// get the panel of np vectors j1:j2-1
//--------------------------------------------------------------
int64_t j2 = GB_IMIN (jend, j1 + jpanel) ;
int64_t np = j2 - j1 ;
//--------------------------------------------------------------
// G = B(:,j1:j2-1), of size bvlen-by-np, in column major order
//--------------------------------------------------------------
int8_t *restrict Gb = (int8_t *) (Bb + (j1 * bvlen)) ;
#if !GB_B_IS_PATTERN
GB_B_TYPE *restrict Gx = (GB_B_TYPE *)
(((GB_void *) (B->x)) +
(B_iso ? 0 : ((j1 * bvlen) * GB_B_SIZE))) ;
#endif
//--------------------------------------------------------------
// clear the panel H to compute C(:,j1:j2-1)
//--------------------------------------------------------------
if (np == 1)
{
// Make H an alias to C(:,j1)
int64_t j = j1 ;
int64_t pC_start = j * cvlen ; // get pointer to C(:,j)
// Hx is GB_C_TYPE, not GB_void, so pointer arithmetic on
// it is by units of size sizeof (GB_C_TYPE), not bytes.
Hx = Cx + pC_start ;
}
else
{
// C is full: set Hx = identity
int64_t nc = np * cvlen ;
#if GB_HAS_IDENTITY_BYTE
memset (Hx, GB_IDENTITY_BYTE, nc * GB_C_SIZE) ;
#else
for (int64_t i = 0 ; i < nc ; i++)
{
GB_HX_WRITE (i, zidentity) ; // Hx(i) = identity
}
#endif
}
#if GB_IS_PLUS_FC32_MONOID
float *restrict Hx_real = (float *) Hx ;
float *restrict Hx_imag = Hx_real + 1 ;
#elif GB_IS_PLUS_FC64_MONOID
double *restrict Hx_real = (double *) Hx ;
double *restrict Hx_imag = Hx_real + 1 ;
#endif
//--------------------------------------------------------------
// H += A*G for one panel
//--------------------------------------------------------------
// GB_HX_COMPUTE: computes H (i,jj) += A(i,k) * B(k,j)
// where H(i,jj) is located at Hx [pH+jj], aij = A(i,k),
// and gkj = B(k,j). gb is the bitmap value for B(k,j).
#undef GB_HX_COMPUTE
#if GB_B_IS_BITMAP
#define GB_HX_COMPUTE(pH,i,gkj,gb,jj) \
if (gb) \
{ \
GB_MULTADD2 (Hx, pH+jj, aik, gkj, i, k, j1+jj) ; \
}
#else
#define GB_HX_COMPUTE(pH,i,gkj,gb,jj) \
{ \
GB_MULTADD2 (Hx, pH+jj, aik, gkj, i, k, j1+jj) ; \
}
#endif
#include "template/GB_AxB_saxpy4_panel.c"
#undef GB_HX_COMPUTE
//--------------------------------------------------------------
// C(:,j1:j2-1) = H
//--------------------------------------------------------------
if (np == 1)
{
// Hx is already aliased to Cx; no more work to do
continue ;
}
for (int64_t jj = 0 ; jj < np ; jj++)
{
//----------------------------------------------------------
// C(:,j) = H (:,jj)
//----------------------------------------------------------
int64_t j = j1 + jj ;
int64_t pC_start = j * cvlen ; // get pointer to C(:,j)
for (int64_t i = 0 ; i < cvlen ; i++)
{
int64_t pC = pC_start + i ; // pointer to C(i,j)
int64_t pH = i * np + jj ; // pointer to H(i,jj)
// C(i,j) = H(i,jj)
GB_CIJ_GATHER_UPDATE (pC, pH) ;
}
}
}
}
}
else if (use_atomics)
{
//----------------------------------------------------------------------
// C += A*B using fine tasks and atomics
//----------------------------------------------------------------------
// allocate workspace to implement the atomic update
#if !GB_Z_HAS_ATOMIC_UPDATE
int8_t *restrict Wf = NULL ; size_t Wf_size = 0 ;
Wf = GB_CALLOC_MEMORY (C->vlen * C->vdim, sizeof (int8_t), &Wf_size) ;
if (Wf == NULL)
{
return (GrB_OUT_OF_MEMORY) ;
}
#endif
int tid ;
#pragma omp parallel for num_threads(nthreads) schedule(dynamic,1)
for (tid = 0 ; tid < ntasks ; tid++)
{
//------------------------------------------------------------------
// determine the vector of B and C for this fine task
//------------------------------------------------------------------
// The fine task operates on C(:,j) and B(:,j). Its fine task
// id ranges from 0 to nfine_tasks_per_vector-1, and determines
// which slice of A to operate on.
int64_t j = tid / nfine_tasks_per_vector ;
int fine_tid = tid % nfine_tasks_per_vector ;
int64_t kfirst = A_slice [fine_tid] ;
int64_t klast = A_slice [fine_tid + 1] ;
int64_t pB_start = j * bvlen ; // pointer to B(:,j)
int64_t pC_start = j * cvlen ; // pointer to C(:,j)
GB_GET_T_FOR_SECONDJ ; // t = j or j+1 for SECONDJ*
// for Hx Gustavason workspace: use C(:,j) in-place:
GB_C_TYPE *restrict Hx = (GB_C_TYPE *)
(((GB_void *) Cx) + (pC_start * GB_C_SIZE)) ;
#if !GB_Z_HAS_ATOMIC_UPDATE
int8_t *restrict Hf = Wf + pC_start ;
#endif
#if GB_IS_PLUS_FC32_MONOID
float *restrict Hx_real = (float *) Hx ;
float *restrict Hx_imag = Hx_real + 1 ;
#elif GB_IS_PLUS_FC64_MONOID
double *restrict Hx_real = (double *) Hx ;
double *restrict Hx_imag = Hx_real + 1 ;
#endif
//------------------------------------------------------------------
// C(:,j) += A(:,k1:k2) * B(k1:k2,j)
//------------------------------------------------------------------
for (int64_t kk = kfirst ; kk < klast ; kk++)
{
//--------------------------------------------------------------
// C(:,j) += A(:,k) * B(k,j)
//--------------------------------------------------------------
int64_t k = GBh_A (Ah, kk) ; // k in range k1:k2
int64_t pB = pB_start + k ; // get pointer to B(k,j)
#if GB_B_IS_BITMAP
if (!GBb_B (Bb, pB)) continue ;
#endif
int64_t pA = GB_IGET (Ap, kk) ;
int64_t pA_end = GB_IGET (Ap, kk+1) ;
GB_GET_B_kj ; // bkj = B(k,j)
for ( ; pA < pA_end ; pA++)
{
//----------------------------------------------------------
// get A(i,k)
//----------------------------------------------------------
int64_t i = GB_IGET (Ai, pA) ; // get A(i,k) index
//----------------------------------------------------------
// C(i,j) += A(i,k) * B(k,j)
//----------------------------------------------------------
GB_MULT_A_ik_B_kj ; // t = A(i,k) * B(k,j)
#if GB_Z_HAS_ATOMIC_UPDATE
{
// the monoid has an atomic update
GB_Z_ATOMIC_UPDATE_HX (i, t) ; // C(i,j) += t
}
#else
{
// the update must be done in a critical section using
// the mutex byte Hf (i,j), located at Hf [i]. If
// zero, the mutex is unlocked.
int8_t f ;
do
{
// do this atomically:
// { f = Hf [i] ; Hf [i] = 1 ; }
GB_ATOMIC_CAPTURE_INT8 (f, Hf [i], 1) ;
}
while (f == 1) ;
GB_Z_ATOMIC_UPDATE_HX (i, t) ; // C(i,j) += t
GB_ATOMIC_WRITE
Hf [i] = 0 ;
}
#endif
}
}
}
// free workspace for atomic update
#if !GB_Z_HAS_ATOMIC_UPDATE
GB_FREE_MEMORY (&Wf, Wf_size) ;
#endif
}
else
{
//----------------------------------------------------------------------
// C += A*B using fine tasks and workspace, with no atomics
//----------------------------------------------------------------------
// Each fine task is given size-cvlen workspace to compute its result
// in the first phase, W(:,tid) = A(:,k1:k2) * B(k1:k2,j), where k1:k2
// is defined by the fine_tid of the task. The workspaces are then
// summed into C in the second phase.
//----------------------------------------------------------------------
// first phase: W (:,tid) = A (:,k1:k2) * B (k2:k2,j) for each fine task
//----------------------------------------------------------------------
int tid ;
#pragma omp parallel for num_threads(nthreads) schedule(dynamic,1)
for (tid = 0 ; tid < ntasks ; tid++)
{
//------------------------------------------------------------------
// determine the vector of B and C for this fine task
//------------------------------------------------------------------
// The fine task operates on C(:,j) and B(:,j). Its fine task
// id ranges from 0 to nfine_tasks_per_vector-1, and determines
// which slice of A to operate on.
int64_t j = tid / nfine_tasks_per_vector ;
int fine_tid = tid % nfine_tasks_per_vector ;
int64_t kfirst = A_slice [fine_tid] ;
int64_t klast = A_slice [fine_tid + 1] ;
int64_t pB_start = j * bvlen ; // pointer to B(:,j)
int64_t pW_start = tid * cvlen ; // pointer to W for this thread
GB_GET_T_FOR_SECONDJ ; // t = j or j+1 for SECONDJ*
GB_C_TYPE *restrict Hx = (GB_C_TYPE *)
(Wcx + (pW_start * GB_C_SIZE)) ;
#if GB_IS_PLUS_FC32_MONOID
float *restrict Hx_real = (float *) Hx ;
float *restrict Hx_imag = Hx_real + 1 ;
#elif GB_IS_PLUS_FC64_MONOID
double *restrict Hx_real = (double *) Hx ;
double *restrict Hx_imag = Hx_real + 1 ;
#endif
//------------------------------------------------------------------
// clear the panel
//------------------------------------------------------------------
// Hx = identity
#if GB_HAS_IDENTITY_BYTE
{
memset (Hx, GB_IDENTITY_BYTE, cvlen * GB_C_SIZE) ;
}
#else
{
for (int64_t i = 0 ; i < cvlen ; i++)
{
GB_HX_WRITE (i, zidentity) ; // Hx(i) = identity
}
}
#endif
//------------------------------------------------------------------
// W = A(:,k1:k2) * B(k1:k2,j)
//------------------------------------------------------------------
for (int64_t kk = kfirst ; kk < klast ; kk++)
{
//--------------------------------------------------------------
// W(:,k) += A(:,k) * B(k,j)
//--------------------------------------------------------------
int64_t k = GBh_A (Ah, kk) ; // k in range k1:k2
int64_t pB = pB_start + k ; // get pointer to B(k,j)
#if GB_B_IS_BITMAP
if (!GBb_B (Bb, pB)) continue ;
#endif
int64_t pA = GB_IGET (Ap, kk) ;
int64_t pA_end = GB_IGET (Ap, kk+1) ;
GB_GET_B_kj ; // bkj = B(k,j)
for ( ; pA < pA_end ; pA++)
{
int64_t i = GB_IGET (Ai, pA) ; // get A(i,k) index
// W(i,k) += A(i,k) * B(k,j)
GB_MULT_A_ik_B_kj ; // t = A(i,k)*B(k,j)
GB_HX_UPDATE (i, t) ; // Hx(i) += t
}
}
}
//----------------------------------------------------------------------
// second phase: C += reduce (W)
//----------------------------------------------------------------------
#pragma omp parallel for num_threads(nthreads) schedule(dynamic,1)
for (tid = 0 ; tid < ntasks ; tid++)
{
//------------------------------------------------------------------
// determine the W and C for this fine task
//------------------------------------------------------------------
// The fine task operates on C(i1:i2,j) and W(i1:i2,w1:w2), where
// i1:i2 is defined by the fine task id. Its fine task id ranges
// from 0 to nfine_tasks_per_vector-1.
// w1:w2 are the updates to C(:,j), where w1:w2 =
// [j*nfine_tasks_per_vector : (j+1)*nfine_tasks_per_vector-1].
int64_t j = tid / nfine_tasks_per_vector ;
int fine_tid = tid % nfine_tasks_per_vector ;
int64_t istart, iend ;
GB_PARTITION (istart, iend, cvlen, fine_tid,
nfine_tasks_per_vector) ;
int64_t pC_start = j * cvlen ; // pointer to C(:,j)
int64_t wstart = j * nfine_tasks_per_vector ;
int64_t wend = (j + 1) * nfine_tasks_per_vector ;
// Hx = (typecasted) Wcx workspace
GB_C_TYPE *restrict Hx = ((GB_C_TYPE *) Wcx) ;
#if GB_IS_PLUS_FC32_MONOID
float *restrict Hx_real = (float *) Hx ;
float *restrict Hx_imag = Hx_real + 1 ;
#elif GB_IS_PLUS_FC64_MONOID
double *restrict Hx_real = (double *) Hx ;
double *restrict Hx_imag = Hx_real + 1 ;
#endif
//------------------------------------------------------------------
// C(i1:i2,j) += reduce (W (i2:i2, wstart:wend))
//------------------------------------------------------------------
for (int64_t w = wstart ; w < wend ; w++)
{
//--------------------------------------------------------------
// C(i1:i2,j) += W (i1:i2,w)
//--------------------------------------------------------------
int64_t pW_start = w * cvlen ; // pointer to W (:,w)
for (int64_t i = istart ; i < iend ; i++)
{
int64_t pW = pW_start + i ; // pointer to W(i,w)
int64_t pC = pC_start + i ; // pointer to C(i,j)
// C(i,j) += W(i,w)
GB_CIJ_GATHER_UPDATE (pC, pW) ;
}
}
}
}
}
|
