/
ProcessGroup.h
192 lines (159 loc) · 6.08 KB
/
ProcessGroup.h
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
// Copyright (c) 2022 PaddlePaddle 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.
#pragma once
#include <chrono>
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/distributed/collective/Types.h"
#include "paddle/fluid/eager/api/utils/tensor_utils.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/platform/enforce.h"
constexpr auto kWaitTimeout = std::chrono::milliseconds(0);
namespace paddle {
namespace distributed {
constexpr int IGNORE_ID = -1;
using Tensor = paddle::experimental::Tensor;
enum class CommType : std::uint8_t {
BROADCAST = 0,
ALLREDUCE = 1,
ALLREDUCE_SPARSE = 2, // TODO(shenliang03): to support sparse in allreduce
REDUCE = 3,
ALLGATHER = 4,
GATHER = 5,
SCATTER = 6,
REDUCE_SCATTER = 7,
ALLTOALL = 8,
SEND = 9,
RECV = 10,
BARRIER = 11,
UNKNOWN = 100,
};
class ProcessGroup {
public:
class Task {
public:
Task(int rank, const std::vector<Tensor>& inputTensors,
CommType opType = CommType::UNKNOWN);
virtual ~Task();
virtual bool IsCompleted();
virtual bool Wait(std::chrono::milliseconds timeout = kWaitTimeout);
virtual void Synchronize();
protected:
const int rank_;
CommType comm_type_;
std::mutex mutex_;
bool is_completed_ = false;
};
explicit ProcessGroup(int rank, int size, int gid);
virtual ~ProcessGroup() {}
int GetRank() const { return rank_; }
int GetSize() const { return size_; }
virtual const std::string GetBackendName() const = 0;
virtual std::shared_ptr<ProcessGroup::Task> AllReduce(
std::vector<Tensor>& /* tensors */,
const AllreduceOptions& = AllreduceOptions()) {
PADDLE_THROW(platform::errors::InvalidArgument(
"ProcessGroup%s does not support allreduce", GetBackendName()));
}
virtual std::shared_ptr<ProcessGroup::Task> Broadcast(
std::vector<Tensor>& /* tensors */,
const BroadcastOptions& = BroadcastOptions()) {
PADDLE_THROW(platform::errors::InvalidArgument(
"ProcessGroup%s does not support broadcast", GetBackendName()));
}
virtual void Broadcast(const phi::DenseTensor* in, phi::DenseTensor* out) {
PADDLE_THROW(platform::errors::Fatal(
"ProcessGroup%s does not support broadcast for static mode runtime",
GetBackendName()));
}
virtual std::shared_ptr<ProcessGroup::Task> Barrier(
const BarrierOptions& = BarrierOptions()) {
PADDLE_THROW(platform::errors::InvalidArgument(
"ProcessGroup%s does not support barrier", GetBackendName()));
}
virtual std::shared_ptr<ProcessGroup::Task> Send(
std::vector<Tensor>& tensors /* tensors */, int dst_rank) { // NOLINT
PADDLE_THROW(platform::errors::InvalidArgument(
"ProcessGroup%s does not support send", GetBackendName()));
}
virtual std::shared_ptr<ProcessGroup::Task> Recv(
std::vector<Tensor>& tensors /* tensors */, int src_rank) { // NOLINT
PADDLE_THROW(platform::errors::InvalidArgument(
"ProcessGroup%s does not support receive", GetBackendName()));
}
virtual std::shared_ptr<ProcessGroup::Task> AllGather(
std::vector<Tensor>& in_tensors /* tensors */, // NOLINT
std::vector<Tensor>& out_tensors /* tensors */) { // NOLINT
PADDLE_THROW(platform::errors::InvalidArgument(
"ProcessGroup%s does not support AllGather", GetBackendName()));
}
virtual std::shared_ptr<ProcessGroup::Task> AllToAll(
std::vector<Tensor>& in /* tensors */, // NOLINT
std::vector<Tensor>& out /* tensors */) { // NOLINT
PADDLE_THROW(platform::errors::InvalidArgument(
"ProcessGroup%s does not support AllToAll", GetBackendName()));
}
virtual std::shared_ptr<ProcessGroup::Task> Reduce(
std::vector<Tensor>& tensors /* tensors */, // NOLINT
const ReduceOptions& opts) { // NOLINT
PADDLE_THROW(platform::errors::InvalidArgument(
"ProcessGroup%s does not support Reduce", GetBackendName()));
}
virtual std::shared_ptr<ProcessGroup::Task> Scatter(
std::vector<Tensor>& in_tensors /* tensors */, // NOLINT
std::vector<Tensor>& out_tensors /* tensors */, // NOLINT
const ScatterOptions&) { // NOLINT
PADDLE_THROW(platform::errors::InvalidArgument(
"ProcessGroup%s does not support Scatter", GetBackendName()));
}
protected:
const int rank_;
const int size_;
const int gid_;
};
class ProcessGroupMapFromGid {
public:
bool has(int gid) {
auto it = map_.find(gid);
return it != map_.end();
}
void insert(int gid, ProcessGroup* pg) {
// TODO(sandyhouse): address ut and uncomment the following codes
// PADDLE_ENFORCE_EQ(has(gid), false,
// platform::errors::PreconditionNotMet(
// "The process group with id %d doesnot exist.",
// gid));
map_[gid] = pg;
}
ProcessGroup* get(int gid) {
// TODO(sandyhouse): address ut and uncomment the following codes
// PADDLE_ENFORCE_EQ(has(gid), true,
// platform::errors::PreconditionNotMet(
// "The process group with id %d doesnot exist.",
// gid));
return map_.find(gid)->second;
}
static std::shared_ptr<ProcessGroupMapFromGid> getInstance() {
static auto s_instance = std::make_shared<ProcessGroupMapFromGid>();
return s_instance;
}
ProcessGroupMapFromGid() = default;
~ProcessGroupMapFromGid() = default;
private:
std::unordered_map<int, ProcessGroup*> map_;
};
} // namespace distributed
} // namespace paddle