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
|
/* 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/compiler/xla/service/backend.h"
#include <algorithm>
#include <string>
#include <utility>
#define EIGEN_USE_THREADS
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/compiler/xla/service/compiler.h"
#include "tensorflow/compiler/xla/service/platform_util.h"
#include "tensorflow/compiler/xla/status_macros.h"
#include "tensorflow/compiler/xla/statusor.h"
#include "tensorflow/compiler/xla/types.h"
#include "tensorflow/compiler/xla/util.h"
#include "tensorflow/core/common_runtime/eigen_thread_pool.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/threadpool.h"
#include "tensorflow/core/platform/cpu_info.h"
#include "tensorflow/core/platform/env.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/stream_executor_no_cuda.h"
namespace se = ::perftools::gputools;
namespace xla {
BackendOptions& BackendOptions::set_platform(
perftools::gputools::Platform* platform) {
platform_ = platform;
return *this;
}
perftools::gputools::Platform* BackendOptions::platform() const {
return platform_;
}
BackendOptions& BackendOptions::set_intra_op_parallelism_threads(
int num_threads) {
intra_op_parallelism_threads_ = num_threads;
return *this;
}
int BackendOptions::intra_op_parallelism_threads() const {
return intra_op_parallelism_threads_;
}
// Define this in .cc file to avoid having to include eigen or forward declare
// these types in the header.
struct Backend::EigenThreadPoolWrapper {
explicit EigenThreadPoolWrapper(const int num_threads)
: pool(new tensorflow::thread::ThreadPool(tensorflow::Env::Default(),
"XLAEigen", num_threads)),
wrapper(new tensorflow::EigenThreadPoolWrapper(pool.get())),
device(new Eigen::ThreadPoolDevice(wrapper.get(),
wrapper->NumThreads())) {}
std::unique_ptr<tensorflow::thread::ThreadPool> pool;
std::unique_ptr<tensorflow::EigenThreadPoolWrapper> wrapper;
std::unique_ptr<Eigen::ThreadPoolDevice> device;
};
/* static */ StatusOr<std::unique_ptr<Backend>> Backend::CreateBackend(
const BackendOptions& options) {
perftools::gputools::Platform* platform = options.platform();
TF_ASSIGN_OR_RETURN(auto compiler, Compiler::GetForPlatform(platform));
TF_ASSIGN_OR_RETURN(auto stream_executors,
PlatformUtil::GetStreamExecutors(platform));
TF_ASSIGN_OR_RETURN(auto transfer_manager,
TransferManager::GetForPlatform(platform));
TF_ASSIGN_OR_RETURN(auto computation_placer,
ComputationPlacer::GetForPlatform(platform));
std::unique_ptr<Backend> backend(
new Backend(platform, compiler, stream_executors, transfer_manager,
computation_placer, options.intra_op_parallelism_threads()));
return std::move(backend);
}
/* static */ StatusOr<std::unique_ptr<Backend>>
Backend::CreateDefaultBackend() {
TF_ASSIGN_OR_RETURN(se::Platform * platform,
PlatformUtil::GetDefaultPlatform());
BackendOptions backend_options;
backend_options.set_platform(platform);
return CreateBackend(backend_options);
}
StatusOr<Backend::StreamPtr> Backend::BorrowStream(int device_ordinal) {
TF_ASSIGN_OR_RETURN(auto exec, stream_executor(device_ordinal));
return BorrowStream(exec);
}
StatusOr<Backend::StreamPtr> Backend::BorrowStream(
se::StreamExecutor* executor) {
tensorflow::mutex_lock l(mu_);
if (0 == stream_pools_.count(executor)) {
stream_pools_.emplace(std::piecewise_construct,
std::forward_as_tuple(executor),
std::forward_as_tuple([executor]() {
auto stream = MakeUnique<se::Stream>(executor);
stream->Init();
return stream;
}));
}
return stream_pools_.at(executor).Allocate();
}
Backend::Backend(
perftools::gputools::Platform* platform, Compiler* compiler,
tensorflow::gtl::ArraySlice<se::StreamExecutor*> stream_executors,
TransferManager* transfer_manager, ComputationPlacer* computation_placer,
int intra_op_parallelism_threads)
: platform_(platform),
compiler_(compiler),
transfer_manager_(transfer_manager),
computation_placer_(computation_placer) {
// The given set of stream executors set may include invalid executors.
for (se::StreamExecutor* exec : stream_executors) {
if (exec != nullptr) {
stream_executors_.push_back(exec);
}
}
// Create a memory allocator for the valid stream executors.
memory_allocator_ =
MakeUnique<StreamExecutorMemoryAllocator>(platform, stream_executors);
CHECK(!stream_executors_.empty())
<< "Service found no devices for backend " << platform_->Name() << '.';
if (platform->id() == se::host::kHostPlatformId) {
inter_op_thread_pool_.reset(new tensorflow::thread::ThreadPool(
tensorflow::Env::Default(), "xla_inter_op",
tensorflow::port::NumSchedulableCPUs()));
const int num_threads = intra_op_parallelism_threads > 0
? intra_op_parallelism_threads
: tensorflow::port::NumSchedulableCPUs();
intra_op_thread_pool_wrapper_.reset(
new EigenThreadPoolWrapper(num_threads));
}
}
Backend::~Backend() {}
int Backend::default_device_ordinal() const {
return default_stream_executor()->device_ordinal();
}
tensorflow::thread::ThreadPool* Backend::inter_op_thread_pool() const {
return inter_op_thread_pool_.get();
}
const Eigen::ThreadPoolDevice* Backend::eigen_intra_op_thread_pool_device()
const {
if (intra_op_thread_pool_wrapper_ == nullptr) {
return nullptr;
}
return intra_op_thread_pool_wrapper_->device.get();
}
tensorflow::thread::ThreadPool* Backend::eigen_intra_op_thread_pool() const {
if (intra_op_thread_pool_wrapper_ == nullptr) {
return nullptr;
}
return intra_op_thread_pool_wrapper_->pool.get();
}
StatusOr<perftools::gputools::StreamExecutor*> Backend::stream_executor(
int device_ordinal) const {
if (device_ordinal < 0 ||
device_ordinal > stream_executors_.back()->device_ordinal()) {
return InvalidArgument(
"Invalid device ordinal value (%d). Valid range is [0, %d].",
device_ordinal, stream_executors_.back()->device_ordinal());
}
for (auto* executor : stream_executors_) {
if (executor->device_ordinal() == device_ordinal) {
return executor;
}
}
return InvalidArgument("device %s not supported by XLA service",
device_name(device_ordinal).c_str());
}
StatusOr<bool> Backend::devices_equivalent(int device_ordinal_a,
int device_ordinal_b) {
// Use the name from device description to determine equivalence. This is a
// bit crude but works for GPUs which is the important case where we compile
// an executable for one GPU and want to know if it will run (well) on
// another.
TF_ASSIGN_OR_RETURN(perftools::gputools::StreamExecutor * executor_a,
stream_executor(device_ordinal_a));
TF_ASSIGN_OR_RETURN(perftools::gputools::StreamExecutor * executor_b,
stream_executor(device_ordinal_b));
return (executor_a->GetDeviceDescription().name() ==
executor_b->GetDeviceDescription().name());
}
Status Backend::ResetDevices() {
return transfer_manager_->ResetDevices(stream_executors_);
}
} // namespace xla
|
