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
|
/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/core/framework/dataset.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/resource_mgr.h"
#include "tensorflow/core/lib/core/threadpool.h"
#include "tensorflow/core/util/work_sharder.h"
namespace tensorflow {
namespace {
class ThreadPoolResource : public ResourceBase {
public:
ThreadPoolResource(Env* env, const ThreadOptions& thread_options,
const string& name, int num_threads, bool low_latency_hint,
int max_intra_op_parallelism)
: thread_pool_(env, thread_options, name, num_threads, low_latency_hint),
max_intra_op_parallelism_(max_intra_op_parallelism) {}
// Schedules fn() for execution in the pool of threads.
void Schedule(std::function<void()> fn) {
if (max_intra_op_parallelism_ < 0) {
thread_pool_.Schedule(std::move(fn));
} else {
thread_pool_.Schedule(std::bind(
[this](std::function<void()> bound_fn) {
// TODO(mrry): Consider moving this thread-local configuration to
// the threads themselves.
ScopedPerThreadMaxParallelism scope(max_intra_op_parallelism_);
bound_fn();
},
std::move(fn)));
}
}
string DebugString() override { return "ThreadPoolResource"; }
private:
thread::ThreadPool thread_pool_;
const int max_intra_op_parallelism_;
};
// Creates a handle to a ThreadPool resource. Note that we don't use
// ResourceOpKernel here because the ThreadPoolResource constructor requires
// access to `OpKernelContext::env()`, which isn't provided by
// `ResourceOpKernel<T>::CreateResource()`.
class ThreadPoolHandleOp : public OpKernel {
public:
explicit ThreadPoolHandleOp(OpKernelConstruction* ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("display_name", &display_name_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("num_threads", &num_threads_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("max_intra_op_parallelism",
&max_intra_op_parallelism_));
OP_REQUIRES(
ctx, num_threads_ > 0,
errors::InvalidArgument("`num_threads` must be greater than zero."));
}
// The resource is deleted from the resource manager only when it is private
// to kernel. Ideally the resource should be deleted when it is no longer held
// by anyone, but it would break backward compatibility.
~ThreadPoolHandleOp() override {
if (cinfo_.resource_is_private_to_kernel()) {
if (!cinfo_.resource_manager()
->Delete<ThreadPoolResource>(cinfo_.container(), cinfo_.name())
.ok()) {
// Do nothing; the resource can have been deleted by session resets.
}
}
}
void Compute(OpKernelContext* ctx) override LOCKS_EXCLUDED(mu_) {
mutex_lock l(mu_);
if (!initialized_) {
ResourceMgr* mgr = ctx->resource_manager();
OP_REQUIRES_OK(ctx, cinfo_.Init(mgr, def()));
ThreadPoolResource* resource;
OP_REQUIRES_OK(ctx, mgr->LookupOrCreate<ThreadPoolResource>(
cinfo_.container(), cinfo_.name(), &resource,
[this, ctx](ThreadPoolResource** ret)
EXCLUSIVE_LOCKS_REQUIRED(mu_) {
*ret = new ThreadPoolResource(
ctx->env(), {}, display_name_,
num_threads_, max_intra_op_parallelism_,
false /* low_latency_hint */);
return Status::OK();
}));
initialized_ = true;
}
OP_REQUIRES_OK(ctx, MakeResourceHandleToOutput(
ctx, 0, cinfo_.container(), cinfo_.name(),
MakeTypeIndex<ThreadPoolResource>()));
}
private:
mutex mu_;
ContainerInfo cinfo_ GUARDED_BY(mu_);
bool initialized_ GUARDED_BY(mu_) = false;
string display_name_;
int num_threads_;
int max_intra_op_parallelism_;
};
class ThreadPoolDatasetOp : public UnaryDatasetOpKernel {
public:
explicit ThreadPoolDatasetOp(OpKernelConstruction* ctx)
: UnaryDatasetOpKernel(ctx) {}
void MakeDataset(OpKernelContext* ctx, DatasetBase* input,
DatasetBase** output) override {
ThreadPoolResource* threadpool_resource;
OP_REQUIRES_OK(ctx, LookupResource(ctx, HandleFromInput(ctx, 1),
&threadpool_resource));
core::ScopedUnref unref_iterator(threadpool_resource);
*output = new Dataset(ctx, input, threadpool_resource);
}
private:
class Dataset : public DatasetBase {
public:
Dataset(OpKernelContext* ctx, const DatasetBase* input,
ThreadPoolResource* threadpool)
: DatasetBase(DatasetContext(ctx)),
input_(input),
threadpool_(threadpool) {
input_->Ref();
threadpool_->Ref();
}
~Dataset() override {
input_->Unref();
threadpool_->Unref();
}
std::unique_ptr<IteratorBase> MakeIteratorInternal(
const string& prefix) const override {
return std::unique_ptr<IteratorBase>(
new Iterator({this, strings::StrCat(prefix, "::ThreadPool")}));
}
const DataTypeVector& output_dtypes() const override {
return input_->output_dtypes();
}
const std::vector<PartialTensorShape>& output_shapes() const override {
return input_->output_shapes();
}
string DebugString() const override {
return "ThreadPoolDatasetOp::Dataset";
}
protected:
Status AsGraphDefInternal(SerializationContext* ctx,
DatasetGraphDefBuilder* b,
Node** output) const override {
return errors::Unimplemented("%s does not support serialization",
DebugString());
}
private:
class Iterator : public DatasetIterator<Dataset> {
public:
explicit Iterator(const Params& params)
: DatasetIterator<Dataset>(params) {}
Status Initialize(IteratorContext* ctx) override {
return dataset()->input_->MakeIterator(ctx, prefix(), &input_impl_);
}
Status GetNextInternal(IteratorContext* ctx,
std::vector<Tensor>* out_tensors,
bool* end_of_sequence) override {
ThreadPoolResource* pool = dataset()->threadpool_;
IteratorContext::Params params;
params.env = ctx->env();
params.runner = [pool](std::function<void()> c) {
pool->Schedule(std::move(c));
};
params.stats_aggregator_getter = ctx->stats_aggregator_getter();
params.lib = ctx->lib();
params.function_library = ctx->function_library();
params.allocator_getter = ctx->allocator_getter();
IteratorContext threadpool_ctx(params);
return input_impl_->GetNext(&threadpool_ctx, out_tensors,
end_of_sequence);
}
private:
std::unique_ptr<IteratorBase> input_impl_;
};
const DatasetBase* const input_;
ThreadPoolResource* const threadpool_;
};
};
REGISTER_KERNEL_BUILDER(Name("ThreadPoolHandle").Device(DEVICE_CPU),
ThreadPoolHandleOp);
REGISTER_KERNEL_BUILDER(Name("ThreadPoolDataset").Device(DEVICE_CPU),
ThreadPoolDatasetOp);
} // namespace
} // namespace tensorflow
|
