ShareDataWith and ShareBufferWith are prohibited in OP(English Version)

ShareDataWith and ShareBufferWith are prohibited in OP(English Version)

Contents

• Section 1 Background
• Section 2 Modification of ShareDataWith or ShareBufferWith
• Section 3 Relevant Description of CI
  Additional Instructions: During the implementation process, there will be some specifications which haven't be considered before, and need to be continuously supplemented and improved during the process. Hopefully you can give us some feedbacks.

Background

At present, there are some RDs(Research and Development engineer) using method ShareDataWith on operator's input and output( Output = ShareDataWith(Input)). This operation may cause the following errors:

• This operation is equivalent to creating a hidden edge in the operator graph, and this edge links the input and the output. It is hard to express this hidden edge in graph analysis, causing the error of graph optimization.

• This operation is equivalent to operating inplace on internal Op, it possibly results in problems in memory release.(For more information, please)

Automatically, the Framework will check different Ops whether they can operate inplace or not. As a result, the operations of ShareDataWith or ShareBufferWith on input/output are not allowed in internal Op.

Modification of ShareDataWith or ShareBufferWith

Method ShareDateWith is using in some op files, such as lod_reset_op.h, here only the code content concerning the sharedatawith method is shown here.

template <typename DeviceContext, typename T>
class LoDResetGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const {
    auto* d_out = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
    auto* d_x = ctx.Output<framework::Tensor>(framework::GradVarName("X"));

    d_x->ShareDataWith(*d_out)；
  }
};

The operations of ShareDataWith or ShareBufferWith on input/output are not allowed in internal Op, but you can use the method framework::TensorCopy as a replacement. The above code can be changed to as follows:

template <typename DeviceContext, typename T>
class LoDResetGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const {
    auto* d_out = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
    auto* d_x = ctx.Output<framework::Tensor>(framework::GradVarName("X"));

    framework::TensorCopy(*d_out, d_out->place(), d_x);
  }
};

Relevant Description of CI

At present, this specification has been checked in PR_CI_CPU_Py2. This check would fail if the method ShareDataWith or ShareBufferWith were used in modified op files. thus appearing a similar error message in buildlog as follows:

**********************
Using ShareDataWith or ShareBufferWith is not recommended. You must have one RD's (zhhsplendid (Recommend), sneaxiy or luotao1 or lanxianghit) approval to use these methods. For more information, please refer to https://github.com/PaddlePaddle/Paddle/wiki/ShareDataWith-is-prohibited-in-OP. The error lines are as follows:
paddle/fluid/operators/fill_op.h
+      tensor.ShareDataWith(out)

These are 1 approval errors.
**********************

Please modified the relevant code according to the error message, in order to realize the specification of forbidding rules to tensor:: sharedatawith and tensor:: sharebufferwith methods in internal Op. If you want merge this pull request(PR), please check with the relevant approving officer (there is a list of approving officers in CI buildlog) and one approval is required at least.

Advice feedback

If you have any problem, please don't hesitate to contact @guofei