1/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2
3Licensed under the Apache License, Version 2.0 (the "License");
4you may not use this file except in compliance with the License.
5You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9Unless required by applicable law or agreed to in writing, software
10distributed under the License is distributed on an "AS IS" BASIS,
11WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12See the License for the specific language governing permissions and
13limitations under the License.
14==============================================================================*/
15
16#ifndef TENSORFLOW_CORE_COMMON_RUNTIME_GPU_DEVICE_CONTEXT_H_
17#define TENSORFLOW_CORE_COMMON_RUNTIME_GPU_DEVICE_CONTEXT_H_
18
19#include "tensorflow/core/common_runtime/device.h"
20#include "tensorflow/core/framework/device_base.h"
21#include "tensorflow/core/lib/gtl/inlined_vector.h"
22
23namespace stream_executor {
24class Stream;
25} // namespace stream_executor
26
27namespace tensorflow {
28
29class GPUDeviceContext : public DeviceContext {
30 public:
31 // Does not take ownership of streams.
32 GPUDeviceContext(int stream_id, se::Stream* stream,
33#if TENSORFLOW_USE_ROCM
34 se::Stream* nccl_stream,
35#endif
36 se::Stream* host_to_device_stream,
37 se::Stream* device_to_host_stream,
38 gtl::InlinedVector<se::Stream*, 4> device_to_device_stream,
39 Allocator* host_memory_allocator)
40 : stream_id_(stream_id),
41 stream_(stream),
42#if TENSORFLOW_USE_ROCM
43 nccl_stream_(nccl_stream),
44#endif
45 host_to_device_stream_(host_to_device_stream),
46 device_to_host_stream_(device_to_host_stream),
47 device_to_device_stream_(device_to_device_stream),
48 host_memory_allocator_(host_memory_allocator) {
49 }
50
51 ~GPUDeviceContext() override {}
52
53 se::Stream* stream() const override { return stream_; }
54#if TENSORFLOW_USE_ROCM
55 se::Stream* nccl_stream() const { return nccl_stream_; }
56#endif
57 se::Stream* host_to_device_stream() const { return host_to_device_stream_; }
58 se::Stream* device_to_host_stream() const { return device_to_host_stream_; }
59 se::Stream* device_to_device_stream(int index) const {
60 return device_to_device_stream_[index % device_to_device_stream_.size()];
61 }
62 int stream_id() const { return stream_id_; }
63 Allocator* host_memory_allocator() const override {
64 return host_memory_allocator_;
65 }
66
67 void CopyCPUTensorToDevice(const Tensor* cpu_tensor, Device* device,
68 Tensor* device_tensor, StatusCallback done,
69 bool sync_dst_compute) const override;
70
71 void CopyDeviceTensorToCPU(const Tensor* device_tensor, StringPiece edge_name,
72 Device* device, Tensor* cpu_tensor,
73 StatusCallback done) override;
74
75 void CopyTensorInSameDevice(const Tensor* input_tensor, Device* device,
76 Tensor* output_tensor,
77 StatusCallback done) const override;
78
79 void MaintainLifetimeOnStream(const Tensor* t,
80 se::Stream* stream) const override {}
81
82 Status ThenExecute(Device* device, se::Stream* stream,
83 std::function<void()> func) override;
84
85 private:
86 int stream_id_;
87 // The default primary stream to use for this context.
88 // All the memory belongs to this stream.
89 se::Stream* stream_;
90#if TENSORFLOW_USE_ROCM
91 // The stream to use for nccl operations.
92 se::Stream* nccl_stream_;
93#endif
94 // The stream to use for copying data from host into GPU.
95 se::Stream* host_to_device_stream_;
96 // The stream to use for copying data from GPU to host.
97 se::Stream* device_to_host_stream_;
98 // Streams to use for copying data between GPUs.
99 gtl::InlinedVector<se::Stream*, 4> device_to_device_stream_;
100 // The allocator to use for allocating pinned host memory.
101 // Not owned.
102 Allocator* host_memory_allocator_;
103};
104
105} // namespace tensorflow
106
107#endif // TENSORFLOW_CORE_COMMON_RUNTIME_GPU_DEVICE_CONTEXT_H_
108