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

Add tensor.h #2611

merged 3 commits into from
Jul 3, 2017

Conversation

wangkuiyi
Copy link
Collaborator

This is a work in progress. It's primary purpose is to verify my ideas presented in the design of tensor at #2579 (review).

@wangkuiyi wangkuiyi mentioned this pull request Jun 26, 2017
@QiJune QiJune changed the title Add tesnor.h Add tensor.h Jun 27, 2017
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.

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

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.


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

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.

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.

@@ -0,0 +1,76 @@
/*
Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Copy link
Member

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.

@JiayiFeng
Copy link
Collaborator

Here is my implementation draft based on current design: #2645. Please review it.

Copy link
Collaborator

@JiayiFeng JiayiFeng left a comment

Choose a reason for hiding this comment

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

LGTM

@JiayiFeng JiayiFeng merged commit bad85f9 into PaddlePaddle:develop Jul 3, 2017
@gangliao gangliao moved this from Doing to Done in PaddlePaddle Refactoring: Phase 1 Aug 2, 2017
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
No open projects
Development

Successfully merging this pull request may close these issues.

None yet

6 participants