Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add tensor.h #2611

Merged
merged 3 commits into from
Jul 3, 2017
Merged
Changes from 1 commit
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
91 changes: 91 additions & 0 deletions paddle/framework/tensor.h
Original file line number Diff line number Diff line change
@@ -0,0 +1,91 @@
/*
Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Copy link
Member

@jacquesqiao jacquesqiao Jun 27, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The copyright message has a different style with former,what format should we use in the newly added code?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done. Switched back to the former one.

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

namespace paddle {
namespace framework {

class Tensor {
using paddle::platform::Place;
using paddle::platform::get_place;

public:
explicit Tensor(DDim dims) : dims_(dims), place_(get_place()) {}
explicit Tensor(DDim dims, Place place) : dims_(dims), place_(place) {}

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Missing GetDim() and GetPlace() ?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am not sure when would they be necessary, but we can add them in the future once we know we need them.

template <typename T>
const T* data() const {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cite from the link you pasted:

Accessors and mutators (get and set functions) may be named like variables. These often correspond to actual member variables, but this is not required. For example, int count() and void set_count(int count).

Here we use data and mutable_data since they are accessors of Tensor.

PADDLE_ASSERT(holder_ != nullptr);
PADDLE_ASSERT(holder_->Place() == place_);
PADDLE_ASSERT(holder_->Size() >= dims_.product() * sizeof(T));
return static_cast<const T*>(holder->Ptr());
}

template <typename T, // must be POD types
typename = std::enable_if<std::is_pod<T>::value>::type>
T* mutable_data() {
if (holder_ == nullptr || holder_->Place() != place_ ||
holder_->Size() < dims_.product() * sizeof(T)) {
holder_.reset(new PlaceholderImpl(place_, dims.product() * sizeof(T)));
}
return static_cast<T*>(holder_->Ptr());
}

template <typename T, // must be POD types
typename = std::enable_if<std::is_pod<T>::value>::type>
T* mutable_data(DDim dims) {
dims_ = dims;
return mutable_data<T>();
}

template <typename T, // must be POD types
typename = std::enable_if<std::is_pod<T>::value>::type>
T* mutable_data(DDim dims, Place place) {
dims_ = dims;
place_ = place;
return mutable_data<T>();
}

private:
// Placeholder hides type T, so it doesn't appear as a template
// parameter of Variable.
struct Placeholder {
virtual ~Placeholder() {}
virtual void* Ptr() const = 0;
virtual Place Place() const = 0;
virtual size_t Size() const = 0;
};

template <typename T>
struct PlaceholderImpl : public Placeholder {
PlaceholderImpl(Place pl, size_t size)
: ptr_(memory::Alloc(pl, size), paddle::memory::Deleter(pl)),
Copy link
Contributor

@gangliao gangliao Jun 27, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

we don't need to define Deleter, because we can use lamda function in shared_ptr's reset function.

holder_.reset(new PlaceholderImpl(place_, dims.product() * sizeof(T)), [&](){
   ....
   memory::Free(place_, ptr_.get());
});

Copy link
Collaborator Author

@wangkuiyi wangkuiyi Jun 27, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I knew that we can use lambda or anything that casts to std::function<void(T*)>, but I think we need a class here to save the place and size.

place_(pl),
size_(size) {}

virtual void* Ptr() const { return static_cast<void*>(ptr_.get()); }
virtual size_t Size() const { return size_; }
virtual Place Place() const { return place_; }

std::unique_ptr<T, memory::Deleter> ptr_;
Place place_; // record the place of ptr_.
size_t size_; // size of the memory block.
};

std::unique_ptr<Placeholder> holder_; // holds the memory block if allocated.
DDim dims_; // could be smallers than the holder_->Size().
paddle::platform::Place place_;
Copy link
Contributor

@qingqing01 qingqing01 Jun 27, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I wonder why the place_ is both located in Tensor and PlaceholderImpl at first. From the line 39, we can see that the Place in PlaceholderImpl is always equal to it in Tensor.

But it seems reasonable the user can use mutable_data(DDim dims, Place place) to get a pointer whose Place is different from the existed holder_.

Copy link
Collaborator Author

@wangkuiyi wangkuiyi Jun 27, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You are right. The current syntax of Tensor is that we can allocate memory in either of the following two ways:

Tensor t(dim, place);
int* p = t.mutable_data<int>();

or

Tensor t;
int* p = t.mutable_data(dim, place);

Should we remove one of them? @qingqing01

Copy link
Collaborator Author

@wangkuiyi wangkuiyi Jun 27, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done. Removed the first choice.

};

} // namespace framework
} // namespace paddle