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
|
/* Copyright 2015 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/common_runtime/copy_tensor.h"
#include <atomic>
#include <utility>
#include <vector>
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/tracing.h"
namespace tensorflow {
namespace {
struct RegistrationInfo {
RegistrationInfo(DeviceType s, DeviceType r, CopyTensor::CopyFunction cf)
: sender_device_type(std::move(s)),
receiver_device_type(r),
copy_function(cf) {}
DeviceType sender_device_type;
DeviceType receiver_device_type;
CopyTensor::CopyFunction copy_function;
};
// We use a vector instead of a map since we expect there to be very
// few registrations.
std::vector<RegistrationInfo>* MutableRegistry() {
static std::vector<RegistrationInfo>* registry =
new std::vector<RegistrationInfo>;
return registry;
}
} // namespace
// static
void CopyTensor::ViaDMA(StringPiece edge_name, DeviceContext* send_dev_context,
DeviceContext* recv_dev_context, Device* src,
Device* dst, const AllocatorAttributes src_alloc_attr,
const AllocatorAttributes dst_alloc_attr,
const Tensor* input, Tensor* output,
StatusCallback done) {
port::Tracing::ScopedAnnotation annotation(edge_name);
VLOG(1) << "Copy " << edge_name;
const DeviceType src_device_type(
src_alloc_attr.on_host() ? DEVICE_CPU : src->attributes().device_type());
const DeviceType dst_device_type(
dst_alloc_attr.on_host() ? DEVICE_CPU : dst->attributes().device_type());
const bool non_cpu_src = src_device_type != DeviceType(DEVICE_CPU);
const bool non_cpu_dst = dst_device_type != DeviceType(DEVICE_CPU);
// E.g., gpu -> gpu
if (non_cpu_src && non_cpu_dst) {
// Device to device copy. Look through registry for an appropriate
// CopyFunction.
std::vector<RegistrationInfo>* registry = MutableRegistry();
for (const RegistrationInfo& ri : *registry) {
if (ri.sender_device_type == src_device_type &&
ri.receiver_device_type == dst_device_type) {
ri.copy_function(send_dev_context, recv_dev_context, src, dst,
src_alloc_attr, dst_alloc_attr, input, output, done);
return;
}
}
// TODO(josh11b): If no CopyFunction is found, we currently fail
// but we could copy between devices via CPU.
done(errors::Unimplemented(
"No function registered to copy from devices of type ",
src_device_type.type(), " to devices of type ",
dst_device_type.type()));
return;
}
// E.g., gpu -> cpu
if (non_cpu_src && !non_cpu_dst) {
// Device to host copy.
send_dev_context->CopyDeviceTensorToCPU(input, edge_name, src, output,
done);
return;
}
// E.g., cpu -> gpu
if (!non_cpu_src && non_cpu_dst) {
// Host to Device copy.
recv_dev_context->CopyCPUTensorToDevice(input, dst, output, done);
return;
}
// cpu -> cpu
CHECK(!non_cpu_src && !non_cpu_dst);
*output = *input;
done(Status::OK());
}
// static
Status CopyTensor::Register(DeviceType sender_device_type,
DeviceType receiver_device_type,
CopyFunction copy_function) {
std::vector<RegistrationInfo>* registry = MutableRegistry();
registry->emplace_back(sender_device_type, receiver_device_type,
copy_function);
return Status::OK();
}
} // namespace tensorflow
|
