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
|
/* 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/computation_placer.h"
#include <string>
#include <utility>
#include <vector>
#include "tensorflow/compiler/xla/literal_util.h"
#include "tensorflow/compiler/xla/ptr_util.h"
#include "tensorflow/compiler/xla/shape_util.h"
#include "tensorflow/compiler/xla/status.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/lib/core/errors.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/stream_executor_no_cuda.h"
namespace se = ::perftools::gputools;
namespace xla {
Status DeviceAssignment::Serialize(DeviceAssignmentProto* proto) const {
proto->set_replica_count(replica_count());
proto->set_computation_count(computation_count());
for (int computation = 0; computation < computation_count(); ++computation) {
DeviceAssignmentProto::ComputationDevice* computation_device =
proto->add_computation_devices();
for (int replica = 0; replica < replica_count(); ++replica) {
computation_device->add_replica_device_ids((*this)(replica, computation));
}
}
return Status::OK();
}
/* static */ StatusOr<DeviceAssignment> DeviceAssignment::Deserialize(
const DeviceAssignmentProto& proto) {
TF_RET_CHECK(proto.computation_devices_size() == proto.computation_count());
DeviceAssignment assignment(proto.replica_count(), proto.computation_count());
for (int computation = 0; computation < proto.computation_count();
++computation) {
const auto& computation_device = proto.computation_devices(computation);
TF_RET_CHECK(computation_device.replica_device_ids_size() ==
proto.replica_count());
for (int replica = 0; replica < proto.replica_count(); ++replica) {
assignment(replica, computation) =
computation_device.replica_device_ids(replica);
}
}
return std::move(assignment);
}
StatusOr<int> ComputationPlacer::DeviceId(int replica, int computation,
int replica_count,
int computation_count) {
TF_RET_CHECK(replica < replica_count);
TF_RET_CHECK(computation < computation_count);
return computation * replica_count + replica;
}
StatusOr<DeviceAssignment> ComputationPlacer::AssignDevices(
int replica_count, int computation_count) {
DeviceAssignment assignment(replica_count, computation_count);
for (int replica = 0; replica < replica_count; ++replica) {
for (int computation = 0; computation < computation_count; ++computation) {
TF_ASSIGN_OR_RETURN(
int device_id,
DeviceId(replica, computation, replica_count, computation_count));
assignment(replica, computation) = device_id;
}
}
return std::move(assignment);
}
/* static */ void ComputationPlacer::RegisterComputationPlacer(
se::Platform::Id platform_id,
ComputationPlacerCreationFunction creation_function) {
tensorflow::mutex_lock lock(
*ComputationPlacer::platform_computation_placer_mutex());
auto* computation_placers = GetPlatformComputationPlacers();
CHECK(computation_placers->find(platform_id) == computation_placers->end());
(*computation_placers)[platform_id].creation_function = creation_function;
}
/* static */ StatusOr<ComputationPlacer*> ComputationPlacer::GetForPlatform(
const se::Platform* platform) {
tensorflow::mutex_lock lock(
*ComputationPlacer::platform_computation_placer_mutex());
auto* computation_placers = GetPlatformComputationPlacers();
auto it = computation_placers->find(platform->id());
if (it == computation_placers->end()) {
return NotFound(
"could not find registered computation placer for platform %s -- check "
"target linkage",
platform->Name().c_str());
}
if (it->second.placer == nullptr) {
// Lazily create the computation placer the first time it is needed.
it->second.placer = (*it->second.creation_function)();
}
return it->second.placer.get();
}
/* static */ tensorflow::mutex*
ComputationPlacer::platform_computation_placer_mutex() {
static tensorflow::mutex* m = new tensorflow::mutex;
return m;
}
/* static */ std::map<perftools::gputools::Platform::Id,
ComputationPlacer::State>*
ComputationPlacer::GetPlatformComputationPlacers() {
static auto* r =
new std::map<perftools::gputools::Platform::Id, ComputationPlacer::State>;
return r;
}
} // namespace xla
static std::unique_ptr<xla::ComputationPlacer> CreateComputationPlacer() {
return xla::MakeUnique<xla::ComputationPlacer>();
}
static bool InitModule() {
xla::ComputationPlacer::RegisterComputationPlacer(se::host::kHostPlatformId,
&CreateComputationPlacer);
xla::ComputationPlacer::RegisterComputationPlacer(se::cuda::kCudaPlatformId,
&CreateComputationPlacer);
return true;
}
static bool module_initialized = InitModule();
|
