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
|
//------------------------------------------------------------------------------
// GB_transpose_bucket_template: transpose and typecast and/or apply operator
//------------------------------------------------------------------------------
// SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2025, All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0
//------------------------------------------------------------------------------
// Symbolic phase for GB_transpose_bucket.
{
GB_Cp_TYPE *restrict Cp = C->p ;
int64_t nvec_nonempty ;
if (nthreads == 1)
{
//----------------------------------------------------------------------
// sequential method: A is not sliced
//----------------------------------------------------------------------
// Only requires a single workspace of size avlen for a single thread.
// The resulting C matrix is not jumbled.
GBURBLE ("(1-thread bucket transpose) ") ;
// compute the row counts of A. No need to scan the A->p pointers
ASSERT (nworkspaces == 1) ;
GB_Cp_TYPE *restrict workspace = Workspaces [0] ;
memset (workspace, 0, (avlen + 1) * sizeof (GB_Cp_TYPE)) ;
for (int64_t p = 0 ; p < anz ; p++)
{
int64_t i = GB_IGET (Ai, p) ;
workspace [i]++ ;
}
// cumulative sum of the workspace, and copy back into C->p
GB_cumsum (workspace, Cp_is_32, avlen, &nvec_nonempty, 1, NULL) ;
memcpy (Cp, workspace, (avlen + 1) * sizeof (GB_Cp_TYPE)) ;
}
else if (nworkspaces == 1)
{
//----------------------------------------------------------------------
// atomic method: A is sliced but workspace is shared
//----------------------------------------------------------------------
// Only requires a single workspace of size avlen, shared by all
// threads. Scales well, but requires atomics. If the # of rows is
// very small and the average row degree is high, this can be very slow
// because of contention on the atomic workspace. Otherwise, it is
// typically faster than the non-atomic method. The resulting C matrix
// is jumbled.
GBURBLE ("(%d-thread atomic bucket transpose) ", nthreads) ;
// compute the row counts of A. No need to scan the A->p pointers
GB_Cp_TYPE *restrict workspace = Workspaces [0] ;
GB_memset (workspace, 0, (avlen + 1) * sizeof (GB_Cp_TYPE), nth) ;
int64_t p ;
#pragma omp parallel for num_threads(nthreads) schedule(static)
for (p = 0 ; p < anz ; p++)
{
int64_t i = GB_IGET (Ai, p) ;
// update workspace [i]++ automically:
GB_ATOMIC_UPDATE
workspace [i]++ ;
}
C->jumbled = true ; // atomic transpose leaves C jumbled
// cumulative sum of the workspace, and copy back into C->p
GB_cumsum (workspace, Cp_is_32, avlen, &nvec_nonempty, nth, Werk) ;
GB_memcpy (Cp, workspace, (avlen + 1) * sizeof (GB_Cp_TYPE), nth) ;
}
else
{
//----------------------------------------------------------------------
// non-atomic method
//----------------------------------------------------------------------
// compute the row counts of A for each slice, one per thread; This
// method is parallel, but not highly scalable. Each thread requires
// workspace of size avlen, but no atomics are required. The resulting
// C matrix is not jumbled, so this can save work if C needs to be
// unjumbled later.
GBURBLE ("(%d-thread non-atomic bucket transpose) ", nthreads) ;
ASSERT (nworkspaces == nthreads) ;
int tid ;
#pragma omp parallel for num_threads(nthreads) schedule(static)
for (tid = 0 ; tid < nthreads ; tid++)
{
// get the row counts for this slice, of size A->vlen
GB_Cp_TYPE *restrict workspace = Workspaces [tid] ;
memset (workspace, 0, (avlen + 1) * sizeof (GB_Cp_TYPE)) ;
for (int64_t k = A_slice [tid] ; k < A_slice [tid+1] ; k++)
{
// iterate over the entries in A(:,j); j not itself not needed
int64_t pA_start = GB_IGET (Ap, k) ;
int64_t pA_end = GB_IGET (Ap, k+1) ;
for (int64_t pA = pA_start ; pA < pA_end ; pA++)
{
// count one more entry in C(i,:) for this slice
int64_t i = GB_IGET (Ai, pA) ;
workspace [i]++ ;
}
}
}
// cumulative sum of the workspaces across the slices
int64_t i ;
#pragma omp parallel for num_threads(nth) schedule(static)
for (i = 0 ; i < avlen ; i++)
{
GB_Cp_TYPE s = 0 ;
for (int tid = 0 ; tid < nthreads ; tid++)
{
GB_Cp_TYPE *restrict workspace = Workspaces [tid] ;
GB_Cp_TYPE c = workspace [i] ;
workspace [i] = s ;
s += c ;
}
Cp [i] = s ;
}
Cp [avlen] = 0 ;
//----------------------------------------------------------------------
// compute the vector pointers for C
//----------------------------------------------------------------------
GB_cumsum (Cp, Cp_is_32, avlen, &nvec_nonempty, nth, Werk) ;
//----------------------------------------------------------------------
// add Cp back to all Workspaces
//----------------------------------------------------------------------
#pragma omp parallel for num_threads(nth) schedule(static)
for (i = 0 ; i < avlen ; i++)
{
GB_Cp_TYPE s = Cp [i] ;
GB_Cp_TYPE *restrict workspace = Workspaces [0] ;
workspace [i] = s ;
for (int tid = 1 ; tid < nthreads ; tid++)
{
GB_Cp_TYPE *restrict workspace = Workspaces [tid] ;
workspace [i] += s ;
}
}
}
GB_nvec_nonempty_set (C, nvec_nonempty) ;
}
#undef GB_Cp_TYPE
|
