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
|
/*
* Copyright 2014 Google Inc.
*
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
*/
#include "SkTaskGroup.h"
#include "SkCondVar.h"
#include "SkRunnable.h"
#include "SkTDArray.h"
#include "SkThread.h"
#include "SkThreadUtils.h"
#if defined(SK_BUILD_FOR_WIN32)
static inline int num_cores() {
SYSTEM_INFO sysinfo;
GetSystemInfo(&sysinfo);
return sysinfo.dwNumberOfProcessors;
}
#else
#include <unistd.h>
static inline int num_cores() {
return (int) sysconf(_SC_NPROCESSORS_ONLN);
}
#endif
namespace {
class ThreadPool : SkNoncopyable {
public:
static void Add(SkRunnable* task, SkAtomic<int32_t>* pending) {
if (!gGlobal) { // If we have no threads, run synchronously.
return task->run();
}
gGlobal->add(&CallRunnable, task, pending);
}
static void Add(void (*fn)(void*), void* arg, SkAtomic<int32_t>* pending) {
if (!gGlobal) {
return fn(arg);
}
gGlobal->add(fn, arg, pending);
}
static void Batch(void (*fn)(void*), void* args, int N, size_t stride,
SkAtomic<int32_t>* pending) {
if (!gGlobal) {
for (int i = 0; i < N; i++) { fn((char*)args + i*stride); }
return;
}
gGlobal->batch(fn, args, N, stride, pending);
}
static void Wait(SkAtomic<int32_t>* pending) {
if (!gGlobal) { // If we have no threads, the work must already be done.
SkASSERT(pending->load(sk_memory_order_relaxed) == 0);
return;
}
// Acquire pairs with decrement release here or in Loop.
while (pending->load(sk_memory_order_acquire) > 0) {
// Lend a hand until our SkTaskGroup of interest is done.
Work work;
{
AutoLock lock(&gGlobal->fReady);
if (gGlobal->fWork.isEmpty()) {
// Someone has picked up all the work (including ours). How nice of them!
// (They may still be working on it, so we can't assert *pending == 0 here.)
continue;
}
gGlobal->fWork.pop(&work);
}
// This Work isn't necessarily part of our SkTaskGroup of interest, but that's fine.
// We threads gotta stick together. We're always making forward progress.
work.fn(work.arg);
work.pending->fetch_add(-1, sk_memory_order_release); // Pairs with load above.
}
}
private:
struct AutoLock {
AutoLock(SkCondVar* c) : fC(c) { fC->lock(); }
~AutoLock() { fC->unlock(); }
private:
SkCondVar* fC;
};
static void CallRunnable(void* arg) { static_cast<SkRunnable*>(arg)->run(); }
struct Work {
void (*fn)(void*); // A function to call,
void* arg; // its argument,
SkAtomic<int32_t>* pending; // then decrement pending afterwards.
};
explicit ThreadPool(int threads) : fDraining(false) {
if (threads == -1) {
threads = num_cores();
}
for (int i = 0; i < threads; i++) {
fThreads.push(SkNEW_ARGS(SkThread, (&ThreadPool::Loop, this)));
fThreads.top()->start();
}
}
~ThreadPool() {
SkASSERT(fWork.isEmpty()); // All SkTaskGroups should be destroyed by now.
{
AutoLock lock(&fReady);
fDraining = true;
fReady.broadcast();
}
for (int i = 0; i < fThreads.count(); i++) {
fThreads[i]->join();
}
SkASSERT(fWork.isEmpty()); // Can't hurt to double check.
fThreads.deleteAll();
}
void add(void (*fn)(void*), void* arg, SkAtomic<int32_t>* pending) {
Work work = { fn, arg, pending };
pending->fetch_add(+1, sk_memory_order_relaxed); // No barrier needed.
{
AutoLock lock(&fReady);
fWork.push(work);
fReady.signal();
}
}
void batch(void (*fn)(void*), void* arg, int N, size_t stride, SkAtomic<int32_t>* pending) {
pending->fetch_add(+N, sk_memory_order_relaxed); // No barrier needed.
{
AutoLock lock(&fReady);
Work* batch = fWork.append(N);
for (int i = 0; i < N; i++) {
Work work = { fn, (char*)arg + i*stride, pending };
batch[i] = work;
}
fReady.broadcast();
}
}
static void Loop(void* arg) {
ThreadPool* pool = (ThreadPool*)arg;
Work work;
while (true) {
{
AutoLock lock(&pool->fReady);
while (pool->fWork.isEmpty()) {
if (pool->fDraining) {
return;
}
pool->fReady.wait();
}
pool->fWork.pop(&work);
}
work.fn(work.arg);
work.pending->fetch_add(-1, sk_memory_order_release); // Pairs with load in Wait().
}
}
SkTDArray<Work> fWork;
SkTDArray<SkThread*> fThreads;
SkCondVar fReady;
bool fDraining;
static ThreadPool* gGlobal;
friend struct SkTaskGroup::Enabler;
};
ThreadPool* ThreadPool::gGlobal = NULL;
} // namespace
SkTaskGroup::Enabler::Enabler(int threads) {
SkASSERT(ThreadPool::gGlobal == NULL);
if (threads != 0 && SkCondVar::Supported()) {
ThreadPool::gGlobal = SkNEW_ARGS(ThreadPool, (threads));
}
}
SkTaskGroup::Enabler::~Enabler() {
SkDELETE(ThreadPool::gGlobal);
}
SkTaskGroup::SkTaskGroup() : fPending(0) {}
void SkTaskGroup::wait() { ThreadPool::Wait(&fPending); }
void SkTaskGroup::add(SkRunnable* task) { ThreadPool::Add(task, &fPending); }
void SkTaskGroup::add(void (*fn)(void*), void* arg) { ThreadPool::Add(fn, arg, &fPending); }
void SkTaskGroup::batch (void (*fn)(void*), void* args, int N, size_t stride) {
ThreadPool::Batch(fn, args, N, stride, &fPending);
}
|
