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
|
//------------------------------------------------------------------------------
// GB_AxB_saxpy3_cumsum: finalize nnz(C(:,j)) and find cumulative sum of Cp
//------------------------------------------------------------------------------
// SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2025, All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0
//------------------------------------------------------------------------------
// phase3: fine tasks finalize their computation nnz(C(:,j))
// phase4: cumulative sum of C->p
#include "GB.h"
GB_CALLBACK_SAXPY3_CUMSUM_PROTO (GB_AxB_saxpy3_cumsum)
{
//--------------------------------------------------------------------------
// get C
//--------------------------------------------------------------------------
ASSERT (!GB_IS_BITMAP (C)) ;
ASSERT (!GB_IS_FULL (C)) ;
GB_Cp_DECLARE (Cp, ) ; GB_Cp_PTR (Cp, C) ;
const int64_t cvlen = C->vlen ;
const int64_t cnvec = C->nvec ;
const bool Cp_is_32 = C->p_is_32 ;
ASSERT (Cp != NULL) ;
//==========================================================================
// phase3: count nnz(C(:,j)) for fine tasks
//==========================================================================
int taskid ;
#pragma omp parallel for num_threads(nthreads) schedule(dynamic,1)
for (taskid = 0 ; taskid < nfine ; taskid++)
{
//----------------------------------------------------------------------
// get the task descriptor
//----------------------------------------------------------------------
// int64_t kk = SaxpyTasks [taskid].vector ;
uint64_t hash_size = SaxpyTasks [taskid].hsize ;
bool use_Gustavson = (hash_size == cvlen) ;
int team_size = SaxpyTasks [taskid].team_size ;
int leader = SaxpyTasks [taskid].leader ;
int my_teamid = taskid - leader ;
int64_t my_cjnz = 0 ;
if (use_Gustavson)
{
//------------------------------------------------------------------
// phase3: fine Gustavson task, C=A*B, C<M>=A*B, or C<!M>=A*B
//------------------------------------------------------------------
// Hf [i] == 2 if C(i,j) is an entry in C(:,j)
int8_t *restrict Hf ;
Hf = (int8_t *restrict) SaxpyTasks [taskid].Hf ;
int64_t istart, iend ;
GB_PARTITION (istart, iend, cvlen, my_teamid, team_size) ;
for (int64_t i = istart ; i < iend ; i++)
{
if (Hf [i] == 2)
{
my_cjnz++ ;
}
}
}
else
{
//------------------------------------------------------------------
// phase3: fine hash task, C=A*B, C<M>=A*B, or C<!M>=A*B
//------------------------------------------------------------------
// (Hf [hash] & 3) == 2 if C(i,j) is an entry in C(:,j),
// and the index i of the entry is (Hf [hash] >> 2) - 1.
uint64_t *restrict Hf = (uint64_t *restrict) SaxpyTasks [taskid].Hf;
uint64_t mystart, myend ;
GB_PARTITION (mystart, myend, hash_size, my_teamid, team_size) ;
for (uint64_t hash = mystart ; hash < myend ; hash++)
{
if ((Hf [hash] & 3) == 2)
{
my_cjnz++ ;
}
}
}
SaxpyTasks [taskid].my_cjnz = my_cjnz ; // count this task's nnz(C(:,j))
}
//==========================================================================
// phase4: compute Cp with cumulative sum
//==========================================================================
//--------------------------------------------------------------------------
// sum nnz (C (:,j)) for fine tasks
//--------------------------------------------------------------------------
// SaxpyTasks [taskid].my_cjnz is the # of unique entries found in C(:,j) by
// that task. Sum these terms to compute total # of entries in C(:,j).
for (taskid = 0 ; taskid < nfine ; taskid++)
{
int64_t kk = SaxpyTasks [taskid].vector ;
GB_ISET (Cp, kk, 0) ; // Cp [kk] = 0 ;
}
for (taskid = 0 ; taskid < nfine ; taskid++)
{
int64_t kk = SaxpyTasks [taskid].vector ;
int64_t my_cjnz = SaxpyTasks [taskid].my_cjnz ;
GB_IINC (Cp, kk, my_cjnz) ; // Cp [kk] += my_cjnz ;
ASSERT (my_cjnz <= cvlen) ;
}
//--------------------------------------------------------------------------
// cumulative sum for Cp (fine and coarse tasks)
//--------------------------------------------------------------------------
// Cp [kk] is now nnz (C (:,j)), for all vectors j, whether computed by
// fine tasks or coarse tasks, and where j == GBh_B (Bh, kk)
#ifdef GBCOVER
// tell GB_cumsum to fake a failure and return ok as false:
if (GB_Global_hack_get (4)) GB_Global_hack_set (5, 1) ;
#endif
int nth = GB_nthreads (cnvec, chunk, nthreads) ;
int64_t nvec_nonempty ;
bool ok = GB_cumsum (Cp, Cp_is_32, cnvec, &nvec_nonempty, nth, Werk) ;
if (ok)
{
GB_nvec_nonempty_set (C, nvec_nonempty) ;
}
#ifdef GBCOVER
// restore the hack (for test coverage only)
if (GB_Global_hack_get (4)) GB_Global_hack_set (5, 0) ;
#endif
#ifdef GB_DEBUG
int64_t cnz1 = 0, cnz2 = 0 ;
if (Cp_is_32)
{
uint32_t *Cp_debug = C->p ;
if (ok) cnz1 = Cp_debug [cnvec] ;
for (int k = 0 ; k <= cnvec ; k++)
{
if (!ok && k < cnvec) cnz1 += Cp_debug [k] ;
}
}
else
{
uint64_t *Cp_debug = C->p ;
if (ok) cnz1 = Cp_debug [cnvec] ;
for (int k = 0 ; k <= cnvec ; k++)
{
if (!ok && k < cnvec) cnz1 += Cp_debug [k] ;
}
}
#endif
if (!ok)
{
// convert Cp to uint64_t and redo the cumulative sum
ASSERT (Cp_is_32) ;
ASSERT (!C->p_shallow) ;
void *Cp_new = NULL ;
size_t Cp_new_size = 0 ;
Cp_new = GB_MALLOC_MEMORY (cnvec+1, sizeof (uint64_t), &Cp_new_size) ;
if (Cp_new == NULL)
{
return (GrB_OUT_OF_MEMORY) ;
}
// Cp_new = (uint64_t) Cp, casting from 32-bit to 64-bit
GB_cast_int (Cp_new, GB_UINT64_code, Cp, GB_UINT32_code, cnvec+1, nth) ;
GB_FREE_MEMORY (&Cp, C->p_size) ;
C->p = Cp_new ;
C->p_size = Cp_new_size ;
C->p_is_32 = false ;
// redo the cumsum (this will always succeed)
GB_cumsum (C->p, false, cnvec, &nvec_nonempty, nth, Werk) ;
GB_nvec_nonempty_set (C, nvec_nonempty) ;
}
#ifdef GB_DEBUG
if (C->p_is_32)
{
uint32_t *Cp_debug = C->p ;
cnz2 = Cp_debug [cnvec] ;
}
else
{
uint64_t *Cp_debug = C->p ;
cnz2 = Cp_debug [cnvec] ;
}
ASSERT (cnz1 == cnz2) ;
#endif
//--------------------------------------------------------------------------
// cumulative sum of nnz (C (:,j)) for each team of fine tasks
//--------------------------------------------------------------------------
int64_t cjnz_sum = 0 ;
for (taskid = 0 ; taskid < nfine ; taskid++)
{
if (taskid == SaxpyTasks [taskid].leader)
{
cjnz_sum = 0 ;
}
int64_t my_cjnz = SaxpyTasks [taskid].my_cjnz ;
SaxpyTasks [taskid].my_cjnz = cjnz_sum ;
cjnz_sum += my_cjnz ;
}
return (GrB_SUCCESS) ;
}
|
