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
|
//------------------------------------------------------------------------------
// GB_pslice: partition Ap for a set of tasks
//------------------------------------------------------------------------------
// SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2022, All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0
//------------------------------------------------------------------------------
// Ap [0..n] is an array with monotonically increasing entries. This function
// slices Ap so that each chunk has the same number of total values of its
// entries. Ap can be A->p for a matrix and then n = A->nvec. Or it can be
// the work needed for computing each vector of a matrix (see GB_ewise_slice
// and GB_subref_slice, for example).
// If Ap is NULL then the matrix A (not provided here) is full or bitmap,
// which this function handles (Ap is implicit).
#include "GB.h"
//------------------------------------------------------------------------------
// GB_pslice_worker: partition Ap for a set of tasks
//------------------------------------------------------------------------------
static void GB_pslice_worker
(
int64_t *restrict Slice, // size ntasks+1
const int64_t *restrict Ap, // array size n+1
int tlo, // assign to Slice [(tlo+1):(thi-1)]
int thi
)
{
//--------------------------------------------------------------------------
// check inputs
//--------------------------------------------------------------------------
#ifdef GB_DEBUG
ASSERT (Ap != NULL) ;
ASSERT (Slice != NULL) ;
ASSERT (0 <= tlo && tlo < thi - 1) ;
for (int t = tlo+1 ; t <= thi-1 ; t++)
{
ASSERT (Slice [t] == -1) ;
}
#endif
//--------------------------------------------------------------------------
// assign work to Slice [(tlo+1):(thi-1)]
//--------------------------------------------------------------------------
// klo = Slice [tlo] and khi = Slice [thi] are defined on input, where
// tlo < thi - 1. This determines the task boundaries for tasks
// tlo+1 to thi-1, which defines Slice [(tlo+1):(thi-1)].
int64_t klo = Slice [tlo] ;
int64_t khi = Slice [thi] ; ASSERT (0 <= klo && klo <= khi) ;
int64_t p1 = Ap [klo] ;
int64_t p2 = Ap [khi] ; ASSERT (p1 <= p2) ;
if (p1 == p2 || klo == khi)
{
//----------------------------------------------------------------------
// no work is left so simply fill in with empty tasks
//----------------------------------------------------------------------
int64_t k = klo ;
for (int64_t t = tlo+1 ; t <= thi-1 ; t++)
{
Slice [t] = k ;
}
}
else // p1 < p2 && klo < khi
{
//----------------------------------------------------------------------
// find task t that evenly partitions the work p1:p2 to tasks tlo:thi
//----------------------------------------------------------------------
int64_t k = (klo + khi) / 2 ; ASSERT (klo <= k && k <= khi) ;
int64_t p = Ap [k] ; ASSERT (p1 <= p && p <= p2) ;
double ntasks = thi - tlo ;
double ratio = (((double) (p - p1)) / ((double) (p2 - p1))) ;
int t = tlo + (int) floor (ratio * ntasks) ;
t = GB_IMAX (t, tlo+1) ;
t = GB_IMIN (t, thi-1) ; ASSERT (tlo < t && t < thi) ;
//----------------------------------------------------------------------
// assign work to task t
//----------------------------------------------------------------------
ASSERT (Slice [t] == -1) ;
Slice [t] = k ;
//----------------------------------------------------------------------
// recursively partition for tasks (tlo+1):(t-1) and (t+1):(thi-1)
//----------------------------------------------------------------------
if (tlo < t-1)
{
GB_pslice_worker (Slice, Ap, tlo, t) ;
}
if (t < thi-1)
{
GB_pslice_worker (Slice, Ap, t, thi) ;
}
}
}
//------------------------------------------------------------------------------
// GB_pslice: partition Ap for a set of tasks
//------------------------------------------------------------------------------
GB_PUBLIC
void GB_pslice // slice Ap
(
int64_t *restrict Slice, // size ntasks+1
const int64_t *restrict Ap, // array size n+1 (NULL if full or bitmap)
const int64_t n,
const int ntasks, // # of tasks
const bool perfectly_balanced
)
{
//--------------------------------------------------------------------------
// check inputs
//--------------------------------------------------------------------------
ASSERT (Slice != NULL) ;
#ifdef GB_DEBUG
for (int taskid = 0 ; taskid <= ntasks ; taskid++)
{
Slice [taskid] = -1 ;
}
#endif
//--------------------------------------------------------------------------
// slice the work
//--------------------------------------------------------------------------
if (Ap == NULL)
{
//----------------------------------------------------------------------
// A is full or bitmap: slice 0:n equally for all tasks
//----------------------------------------------------------------------
GB_eslice (Slice, n, ntasks) ;
}
else
{
//----------------------------------------------------------------------
// A is sparse or hypersparse
//----------------------------------------------------------------------
if (n == 0 || ntasks <= 1 || Ap [n] == 0)
{
// matrix is empty, or a single thread is used
memset ((void *) Slice, 0, ntasks * sizeof (int64_t)) ;
Slice [ntasks] = n ;
}
else
{
// slice Ap by # of entries
Slice [0] = 0 ;
Slice [ntasks] = n ;
if (perfectly_balanced)
{
// this method is costly, and should only be used if the
// work is to be perfectly balanced (in particular, when there
// is just one task per thread, with static scheduling)
const double work = (double) (Ap [n]) ;
int64_t k = 0 ;
for (int taskid = 1 ; taskid < ntasks ; taskid++)
{
// binary search to find k so that Ap [k] == (taskid*work) /
// ntasks. The exact value will not typically not be found;
// just pick what the binary search comes up with.
int64_t wtask = (int64_t) GB_PART (taskid, work, ntasks) ;
int64_t pright = n ;
GB_TRIM_BINARY_SEARCH (wtask, Ap, k, pright) ;
Slice [taskid] = k ;
}
}
else
{
// this is much faster, and results in good load balancing if
// there is more than one task per thread, and dynamic
// scheduling is used.
GB_pslice_worker (Slice, Ap, 0, ntasks) ;
}
}
}
//--------------------------------------------------------------------------
// check result
//--------------------------------------------------------------------------
#ifdef GB_DEBUG
ASSERT (Slice [0] == 0) ;
ASSERT (Slice [ntasks] == n) ;
for (int taskid = 0 ; taskid < ntasks ; taskid++)
{
ASSERT (Slice [taskid] <= Slice [taskid+1]) ;
}
#endif
}
|
