Qiao Longfei
7 years ago
parent
02425b2f64
commit
3691a46fa3
@ -1,53 +0,0 @@
|
||||
/* Copyright (c) 2016 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 <cstdint>
|
||||
#include <cstring>
|
||||
#include <memory>
|
||||
#include <typeindex>
|
||||
#include <vector>
|
||||
#include "paddle/fluid/framework/data_layout.h"
|
||||
#include "paddle/fluid/framework/ddim.h"
|
||||
#include "paddle/fluid/framework/framework.pb.h"
|
||||
#include "paddle/fluid/memory/memory.h"
|
||||
#include "paddle/fluid/platform/device_context.h"
|
||||
#include "paddle/fluid/platform/enforce.h"
|
||||
#include "paddle/fluid/platform/place.h"
|
||||
|
||||
namespace paddle {
|
||||
|
||||
namespace framework {
|
||||
|
||||
class Communicator {
|
||||
public:
|
||||
Communicator() {}
|
||||
~Communicator() {}
|
||||
|
||||
// send grad
|
||||
void send() {}
|
||||
|
||||
void receive() {}
|
||||
|
||||
void prefetch() {}
|
||||
|
||||
void wait() {}
|
||||
|
||||
private:
|
||||
std::unique_ptr<std::thread> communicate_thread_;
|
||||
};
|
||||
|
||||
} // namespace framework
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,113 @@
|
||||
/* Copyright (c) 2019 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. */
|
||||
|
||||
#include "paddle/fluid/operators/distributed/communicator.h"
|
||||
|
||||
#include "paddle/fluid/framework/selected_rows.h"
|
||||
#include "paddle/fluid/framework/tensor_util.h"
|
||||
#include "paddle/fluid/framework/variable_helper.h"
|
||||
#include "paddle/fluid/operators/distributed/parameter_recv.h"
|
||||
#include "paddle/fluid/operators/distributed/parameter_send.h"
|
||||
#include "paddle/fluid/operators/math/selected_rows_functor.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
namespace distributed {
|
||||
|
||||
static void MergeVars(const std::string &var_name,
|
||||
const std::vector<std::shared_ptr<Variable>> &vars,
|
||||
Scope *scope) {
|
||||
PADDLE_ENFORCE(!vars.empty(), "should have value to merge!");
|
||||
auto cpu_place = platform::CPUPlace();
|
||||
auto &var0 = vars[0];
|
||||
auto *out_var = scope->Var(var_name);
|
||||
if (var0->IsType<framework::LoDTensor>()) {
|
||||
auto *out_t = out_var->GetMutable<framework::LoDTensor>();
|
||||
auto *out_ptr = out_t->mutable_data<float>(
|
||||
var0->Get<framework::LoDTensor>().dims(), cpu_place);
|
||||
auto numel = out_t->numel();
|
||||
for (auto i = 0; i < numel; ++i) {
|
||||
out_ptr[i] = 0;
|
||||
for (auto &var : vars) {
|
||||
auto &var_t = var->Get<framework::LoDTensor>();
|
||||
PADDLE_ENFORCE_EQ(var_t.numel(), numel, "should have the same dims");
|
||||
out_ptr[i] += var_t.data<float>()[i];
|
||||
}
|
||||
}
|
||||
} else if (var0->IsType<framework::SelectedRows>()) {
|
||||
auto *out_slr = out_var->GetMutable<framework::SelectedRows>();
|
||||
std::vector<const paddle::framework::SelectedRows *> inputs;
|
||||
inputs.reserve(vars.size());
|
||||
for (auto &var : vars) {
|
||||
inputs.push_back(&var->Get<framework::SelectedRows>());
|
||||
}
|
||||
math::scatter::MergeAdd<paddle::platform::CPUDeviceContext, float>
|
||||
merge_add;
|
||||
auto dev_ctx = paddle::platform::CPUDeviceContext();
|
||||
merge_add(dev_ctx, inputs, out_slr, false);
|
||||
} else {
|
||||
PADDLE_THROW("unsupported var type!");
|
||||
}
|
||||
}
|
||||
|
||||
void Communicator::SendThread() {
|
||||
for (auto &iter : send_varname_to_queue_) {
|
||||
auto &var_name = iter.first;
|
||||
VLOG(3) << "merge var " << var_name << " and send";
|
||||
auto &var_queue = iter.second;
|
||||
std::vector<std::shared_ptr<Variable>> vars;
|
||||
const size_t max_merge_var_num = 20;
|
||||
size_t merged_var_num = 0;
|
||||
while (var_queue->Size() > 0 && merged_var_num < max_merge_var_num) {
|
||||
vars.push_back(var_queue->Pop());
|
||||
merged_var_num++;
|
||||
}
|
||||
MergeVars(var_name, vars, send_scope_.get());
|
||||
auto send_functor = distributed::ParameterSend<float>();
|
||||
// send_functor(var_name, send_varname_to_ctx_[var_name], exe_ctx,
|
||||
// send_scope_, true);
|
||||
}
|
||||
}
|
||||
|
||||
void Communicator::RecvThread() {
|
||||
// parallel run recv graph
|
||||
for (auto &iter : recv_varname_to_ctx_) {
|
||||
auto &var_name = iter.first;
|
||||
VLOG(3) << "recv var " << iter.first;
|
||||
auto recv_functor = distributed::ParameterRecv<float>();
|
||||
// recv_functor(var_name, iter.second, exe_ctx, recv_scope_);
|
||||
}
|
||||
}
|
||||
|
||||
void Communicator::Send(const std::string &var_name,
|
||||
const framework::Scope &scope) {
|
||||
// push var into send queue by var_name
|
||||
auto *grad_var = scope.FindVar(var_name);
|
||||
PADDLE_ENFORCE(grad_var->IsInitialized(), "grad var should be inited");
|
||||
auto tmp_grad_var = std::make_shared<Variable>();
|
||||
framework::CopyVariable(*grad_var, tmp_grad_var.get());
|
||||
send_varname_to_queue_[var_name]->Push(tmp_grad_var);
|
||||
}
|
||||
|
||||
void Communicator::Start() {
|
||||
// start send and recv thread
|
||||
send_thread_.reset(
|
||||
new std::thread(std::bind(&Communicator::SendThread, this)));
|
||||
recv_thread_.reset(
|
||||
new std::thread(std::bind(&Communicator::RecvThread, this)));
|
||||
}
|
||||
|
||||
} // namespace distributed
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,129 @@
|
||||
/* Copyright (c) 2019 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 <deque>
|
||||
#include <memory>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#include "paddle/fluid/framework/scope.h"
|
||||
#include "paddle/fluid/framework/variable.h"
|
||||
#include "paddle/fluid/operators/distributed/rpc_common.h"
|
||||
#include "paddle/fluid/platform/device_context.h"
|
||||
#include "paddle/fluid/platform/enforce.h"
|
||||
#include "paddle/fluid/platform/place.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
namespace distributed {
|
||||
|
||||
using Scope = framework::Scope;
|
||||
using Variable = framework::Variable;
|
||||
|
||||
template <typename T>
|
||||
class BlockingQueue {
|
||||
public:
|
||||
explicit BlockingQueue(size_t capacity) : capacity_(capacity) {
|
||||
PADDLE_ENFORCE_GT(capacity_, 0, "The capacity must be greater than 0.");
|
||||
}
|
||||
|
||||
bool Push(const T& elem) {
|
||||
std::unique_lock<std::mutex> lock(mutex_);
|
||||
send_cv_.wait(lock, [&] { return queue_.size() < capacity_; });
|
||||
PADDLE_ENFORCE_LT(queue_.size(), capacity_);
|
||||
queue_.push_back(elem);
|
||||
recv_cv_.notify_one();
|
||||
return true;
|
||||
}
|
||||
|
||||
bool Push(T&& elem) {
|
||||
std::unique_lock<std::mutex> lock(mutex_);
|
||||
send_cv_.wait(lock, [&] { return queue_.size() < capacity_; });
|
||||
PADDLE_ENFORCE_LT(queue_.size(), capacity_);
|
||||
queue_.emplace_back(std::move(elem));
|
||||
recv_cv_.notify_one();
|
||||
return true;
|
||||
}
|
||||
|
||||
T Pop() {
|
||||
std::unique_lock<std::mutex> lock(mutex_);
|
||||
recv_cv_.wait(lock, [=] { return !queue_.empty(); });
|
||||
T rc(std::move(queue_.front()));
|
||||
queue_.pop_front();
|
||||
return rc;
|
||||
}
|
||||
|
||||
size_t Cap() const {
|
||||
std::lock_guard<std::mutex> lock(mutex_);
|
||||
return capacity_;
|
||||
}
|
||||
|
||||
size_t Size() const {
|
||||
std::lock_guard<std::mutex> lock(mutex_);
|
||||
return queue_.size();
|
||||
}
|
||||
|
||||
private:
|
||||
const size_t capacity_;
|
||||
std::deque<T> queue_;
|
||||
|
||||
mutable std::mutex mutex_;
|
||||
std::condition_variable recv_cv_;
|
||||
std::condition_variable send_cv_;
|
||||
};
|
||||
|
||||
class Communicator {
|
||||
public:
|
||||
Communicator(
|
||||
const std::unordered_map<std::string, RpcContext>& send_varname_to_ctx,
|
||||
const std::unordered_map<std::string, RpcContext>& recv_varname_to_ctx,
|
||||
Scope* recv_scope)
|
||||
: send_varname_to_ctx_(send_varname_to_ctx),
|
||||
recv_varname_to_ctx_(recv_varname_to_ctx),
|
||||
recv_scope_(recv_scope) {
|
||||
// get all send information from graph, build vars_to_send
|
||||
send_scope_.reset(new Scope());
|
||||
for (auto& iter : send_varname_to_ctx_) {
|
||||
send_varname_to_queue_[iter.first] =
|
||||
std::make_shared<BlockingQueue<std::shared_ptr<Variable>>>(10);
|
||||
}
|
||||
}
|
||||
|
||||
~Communicator() {}
|
||||
|
||||
void Start();
|
||||
|
||||
// send grad
|
||||
void Send(const std::string& var_name, const framework::Scope& scope);
|
||||
|
||||
private:
|
||||
void SendThread();
|
||||
void RecvThread();
|
||||
|
||||
std::unordered_map<std::string,
|
||||
std::shared_ptr<BlockingQueue<std::shared_ptr<Variable>>>>
|
||||
send_varname_to_queue_;
|
||||
std::unordered_map<std::string, RpcContext> send_varname_to_ctx_;
|
||||
std::unordered_map<std::string, RpcContext> recv_varname_to_ctx_;
|
||||
std::unique_ptr<std::thread> send_thread_;
|
||||
std::unique_ptr<std::thread> recv_thread_;
|
||||
Scope* recv_scope_; // should be global scope
|
||||
std::unique_ptr<Scope> send_scope_; // an independent scope
|
||||
};
|
||||
|
||||
} // namespace distributed
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,33 @@
|
||||
/* Copyright (c) 2019 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 <string>
|
||||
#include <vector>
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
namespace distributed {
|
||||
|
||||
struct RpcContext {
|
||||
std::string var_name;
|
||||
std::vector<std::string> splited_var_names;
|
||||
std::vector<std::string> epmap;
|
||||
std::vector<int64_t> height_sections;
|
||||
};
|
||||
|
||||
} // namespace distributed
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
Loading…
Reference in new issue