[WIP] Using LoDTensor instead of Tensor in every operator. #4048

qingqing01 · 2017-09-12T13:45:27Z

解决的问题：

LoDTensor管理成员变量Tensor* tensor_的构造和释放
- 详细描述： Replace Tensor by LoDTensor in each operator. #4047 (comment)
如何在不感知LoD的operators用LoDTensor替换Tensor
- 详细描述： Replace Tensor by LoDTensor in each operator. #4047 (comment)
如何传递LoD
- 详细描述：Replace Tensor by LoDTensor in each operator. #4047 (comment)

@wangkuiyi @reyoung @Superjom @QiJune @hedaoyuan

这里只写了示例程序，还没有全局替换。
和大家讨论下，这种方式是否可以。
如果可行，再全局替换。

qingqing01 · 2017-09-12T13:50:25Z

paddle/framework/variable.h

+                     "Variable must be type LoDTensor or Tensor");
+      return *static_cast<const Tensor*>(holder_->Ptr());
+    }
+  }


为了使得不用LoD信息的operators比较方便的从LoDTensor里获取Tensor：

这里特化了Variable::Get<Tensor>接口使得可以

也可通过方式 https://github.com/PaddlePaddle/Paddle/pull/4048/files#diff-91a47df6a639afa5ad046a8d43c640e0R298 在特化InferShapeContext::Input<Tensor>(const std::string& name) 和 InferShapeContext::Output<Tensor>(const std::string& name) 接口。

qingqing01 · 2017-09-12T13:51:10Z

paddle/operators/mul_op.cc

+        {x_mat_dims[0], y_mat_dims[1]});
+    // Only the forward operator needs to pass the lod.
+    // pass the lod of Input(X) to output.
+    ctx.CopyLoD(/*in = */ "X", /* out = */ "Out");


在每个Op里显示调用LoD传递函数，解释： #4047 (comment)

qingqing01 · 2017-09-12T13:52:24Z

paddle/framework/lod_tensor.h


-  void set_tensor(Tensor* tensor) { tensor_ = tensor; }
+  ~LoDTensor() { delete tensor_; }


LoDTensor管理成员变量Tensor* tensor_的构造与释放，原因：#4047 (comment)

maybe LoDTensor should

Disable copy

Disable assign

Disable move

The simplest way to implement it is make tensor_ as unique_ptr

reyoung · 2017-09-12T17:10:25Z

paddle/framework/lod_tensor.h


-  void set_lod(const LoD& lod) { lod_ = lod; }
+  LoDTensor(const LoD& lod) : lod_(lod), tensor_(new Tensor()) {}


explicit LoDTensor

reyoung · 2017-09-12T17:11:40Z

paddle/framework/operator.h

@@ -289,13 +289,26 @@ class InferShapeContext {
  }

  template <typename T>
-  const T* Input(const std::string& name) const {
+  virtual const T* Input(const std::string& name) const {


A template method cannot be virtual.

reyoung · 2017-09-12T17:11:59Z

paddle/framework/operator.h

    auto* var = InputVar(name);
    return var == nullptr ? nullptr : &var->Get<T>();
  }

+  // template <>


??
either delete it or uncomment it

Superjomn · 2017-09-12T17:04:56Z

paddle/framework/operator.h

+  // template <>
+  // virtual const Tensor* Input<Tensor>(const std::string& name) const {
+  //   auto* var = InputVar(name);
+  //   if (var == nullptr) return nullptr;


delete these

Superjomn · 2017-09-12T17:38:55Z

paddle/framework/operator.h

@@ -289,13 +289,26 @@ class InferShapeContext {
  }

  template <typename T>
-  const T* Input(const std::string& name) const {
+  virtual const T* Input(const std::string& name) const {
    auto* var = InputVar(name);
    return var == nullptr ? nullptr : &var->Get<T>();


var ? var : &var->Get<T>()

var is false only if it equals nullptr.

Superjomn · 2017-09-12T17:39:33Z

paddle/framework/operator.h

  template <typename T>
-  T* Output(const std::string& name) const {
+  virtual T* Output(const std::string& name) const {
    auto var = OutputVar(name);
    return var == nullptr ? nullptr : var->GetMutable<T>();


same as above

Superjomn · 2017-09-12T17:42:19Z

paddle/framework/operator.h

@@ -363,6 +384,27 @@ class ExecutionContext : public InferShapeContext {
    return device_context_;
  }

+  template <typename T>
+  T* Output(const std::string& name) const override {


Output be a const T& better?

const T& Output T* mutable_output or T* MutableOutput

borrowed from Protobuf.

Superjomn · 2017-09-12T17:42:41Z

paddle/framework/operator.h

+  }
+
+  template <typename T>
+  std::vector<T*> MultiOutput(const std::string& name) const override {


same as above.

wangkuiyi · 2017-09-12T22:31:51Z

paddle/framework/operator.h

+  T* Output(const std::string& name) const override {
+    auto var = OutputVar(name);
+    // Different from InferShapeContext, call Get instread of GetMutable.
+    return var == nullptr ? nullptr : var->Get<T>();


Variable::Get 返回的是一个const reference。Variable::GetMutable返回的是一个mutable pointer。也就是说如要要返回mutable pointer，应该调用的是 GetMutable，而不是Get。这里为什么要调用Get而不是GetMutable呢？

Using LoDTensor instead of Tensor in every operator.

c60a027

qingqing01 commented Sep 12, 2017

View reviewed changes

qingqing01 requested review from Superjomn, reyoung, wangkuiyi, hedaoyuan and QiJune September 12, 2017 13:53

reyoung reviewed Sep 12, 2017

View reviewed changes

Superjomn reviewed Sep 12, 2017

View reviewed changes

wangkuiyi reviewed Sep 12, 2017

View reviewed changes

qingqing01 closed this Sep 15, 2017

qingqing01 deleted the lod_tensor branch November 14, 2019 05:20

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[WIP] Using LoDTensor instead of Tensor in every operator. #4048

[WIP] Using LoDTensor instead of Tensor in every operator. #4048

qingqing01 commented Sep 12, 2017 •

edited

Loading

qingqing01 Sep 12, 2017

qingqing01 Sep 12, 2017

qingqing01 Sep 12, 2017

reyoung Sep 12, 2017

reyoung Sep 12, 2017

reyoung Sep 12, 2017

reyoung Sep 12, 2017

Superjomn Sep 12, 2017

Superjomn Sep 12, 2017

Superjomn Sep 12, 2017

Superjomn Sep 12, 2017

Superjomn Sep 12, 2017

wangkuiyi Sep 12, 2017


		void set_tensor(Tensor* tensor) { tensor_ = tensor; }
		~LoDTensor() { delete tensor_; }


		void set_lod(const LoD& lod) { lod_ = lod; }
		LoDTensor(const LoD& lod) : lod_(lod), tensor_(new Tensor()) {}

[WIP] Using LoDTensor instead of Tensor in every operator. #4048

[WIP] Using LoDTensor instead of Tensor in every operator. #4048

Conversation

qingqing01 commented Sep 12, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

qingqing01 commented Sep 12, 2017 •

edited

Loading