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
|
#include <ATen/ThreadLocalState.h>
#include <torch/csrc/distributed/c10d/Work.hpp>
namespace c10d {
Work::Work(
int rank,
OpType opType,
const char* profilingTitle,
const c10::optional<std::vector<at::Tensor>>& inputTensors)
: rank_(rank), opType_(opType) {
if (profilingTitle != nullptr) {
auto recordingFunction =
std::make_shared<at::RecordFunction>(at::RecordScope::USER_SCOPE);
if (recordingFunction->isActive()) {
// Work events follow a future like pattern and can potentially be marked
// as complete by different threads, so explicitly set as async event.
recordingFunction->_setAsync();
// Passing input tensor to recordFunction allows for shape information in
// profiling output.
std::vector<c10::IValue> inputs;
if (inputTensors) {
inputs.reserve(inputTensors->size());
for (const auto& tensor : *inputTensors) {
inputs.emplace_back(tensor);
}
}
recordingFunction->before(
profilingTitle,
c10::ArrayRef<const c10::IValue>(inputs.data(), inputs.size()));
std::function<void()> end_handler = [recordingFunction]() {
recordingFunction->end();
};
recordFunctionEndCallback_ = at::wrapPropagateTLSState(end_handler);
}
}
}
OpType Work::retrieveOpType() {
return opType_;
}
Work::~Work() = default;
bool Work::isCompleted() {
std::lock_guard<std::mutex> lock(mutex_);
return completed_;
}
bool Work::isSuccess() const {
std::lock_guard<std::mutex> lock(mutex_);
return !exception_;
}
std::exception_ptr Work::exception() const {
std::lock_guard<std::mutex> lock(mutex_);
return exception_;
}
int Work::sourceRank() const {
TORCH_CHECK(
false,
"sourceRank() may only be called on work objects "
"that correspond to a recv or recv-from-any call.");
}
std::vector<at::Tensor> Work::result() {
TORCH_CHECK(false, "result() not implemented.");
}
void Work::synchronize() {}
bool Work::wait(std::chrono::milliseconds timeout) {
std::unique_lock<std::mutex> lock(mutex_);
if (timeout == kNoTimeout) {
// This waits without a timeout.
cv_.wait(lock, [&] { return completed_; });
} else {
// Waits for the user-provided timeout.
cv_.wait_for(lock, timeout, [&] { return completed_; });
if (!completed_) {
// Throw exception if the wait operation timed out and the work was not
// completed.
TORCH_CHECK(false, "Operation timed out!");
}
}
if (exception_) {
std::rethrow_exception(exception_);
}
synchronize();
// Always return true, because abort API is not implemented.
return true;
}
void Work::abort() {
TORCH_CHECK(false, "Work::abort not implemented.");
}
c10::intrusive_ptr<c10::ivalue::Future> Work::getFuture() {
TORCH_CHECK(false, "Work::getFuture not implemented.")
}
void Work::finish(std::exception_ptr exception) {
std::unique_lock<std::mutex> lock(mutex_);
completed_ = true;
exception_ = exception;
if (recordFunctionEndCallback_) {
recordFunctionEndCallback_();
recordFunctionEndCallback_ = nullptr;
}
lock.unlock();
cv_.notify_all();
}
void Work::finishAndThrow(std::exception_ptr exception) {
std::unique_lock<std::mutex> lock(mutex_);
completed_ = true;
exception_ = exception;
if (recordFunctionEndCallback_) {
recordFunctionEndCallback_();
recordFunctionEndCallback_ = nullptr;
}
if (exception_) {
std::rethrow_exception(exception_);
}
}
class FutureWrappingWork : public Work {
public:
FutureWrappingWork(c10::intrusive_ptr<c10::ivalue::Future> fut)
: Work(), _fut(fut) {}
~FutureWrappingWork() {}
bool isCompleted() override {
return _fut->completed();
}
bool isSuccess() const override {
return _fut->hasValue();
}
std::exception_ptr exception() const override {
return _fut->exception_ptr();
}
int sourceRank() const override {
TORCH_CHECK(false, "FutureWrappingWork::sourceRank() not implemented");
}
std::vector<at::Tensor> result() override {
return _fut->value().toPyObjectHolder()->extractTensors();
}
bool wait(std::chrono::milliseconds timeout) override {
// FIXME
TORCH_CHECK(
timeout == kNoTimeout,
"FutureWrappingWork::wait() with finite timeout not implemented");
_fut->wait();
return true;
}
void abort() override {
TORCH_CHECK(false, "FutureWrappingWork::abort() not implemented");
}
c10::intrusive_ptr<c10::ivalue::Future> getFuture() override {
return _fut;
}
private:
c10::intrusive_ptr<c10::ivalue::Future> _fut;
};
c10::intrusive_ptr<Work> Work::create_from_future(
c10::intrusive_ptr<c10::ivalue::Future> future) {
return c10::make_intrusive<FutureWrappingWork>(future);
}
} // namespace c10d
|
