Design Doc: Refactorization Overview

The goals of refactoring include:

1. Making it easy for external contributors to write new elementary computation operations.
2. Making the codebase clean and readable.
3. Designing a new computation representation – a computation graph of operators and variables.
4. Implementing auto-scalability and auto fault recoverable distributed computing with the help of computation graphs.

Computation Graphs

1. PaddlePaddle represents the computation, training and inference of Deep Learning models, by computation graphs.
2. Please refer to computation graphs for a concrete example.
3. Users write Python programs to describe the graphs and run them (locally or remotely).
4. A graph is composed of variables and operators.
5. The description of graphs must be serializable/deserializable, so that:
   1. It can be sent to the cloud for distributed execution, and
   2. It can be sent to clients for mobile or enterprise deployment.
6. The Python program does two things
   1. Compilation runs a Python program to generate a protobuf message representation of the graph and send it to
      1. the C++ library libpaddle.so for local execution,
      2. the master process of a distributed training job for training, or
      3. the server process of a Kubernetes serving job for distributed serving.
   2. Execution executes the graph by constructing instances of class Variable and OperatorBase, according to the protobuf message.

Description and Realization of Computation Graph

At compile time, the Python program generates a protobuf message representation of the graph, or a description of the graph.

At runtime, the C++ program realizes the graph and runs it.

| | Representation (protobuf messages) | Realization (C++ class objects) | |—|—|—| |Data|VarDesc|Variable| |Operation|OpDesc|Operator| |Block|BlockDesc|Block|

The word graph is interchangeable with block in this document. A graph consists of computation steps and local variables similar to a C++/Java program block, or a pair of parentheses({ and }).

Compilation and Execution

1. Run a Python program to describe the graph. In particular, the Python application program does the following:
   1. Create VarDesc to represent local/intermediate variables,
   2. Create operators and set attributes,
   3. Validate attribute values,
   4. Infer the type and the shape of variables,
   5. Plan memory-reuse for variables,
   6. Generate the backward graph
   7. Add optimization operators to the computation graph.
   8. Optionally, split the graph for distributed training.
2. The invocation of train or infer methods in the Python program does the following:
   1. Create a new Scope instance in the scope hierarchy for each run of a block,
      1. realize local variables defined in the BlockDesc message in the new scope,
      2. a scope is similar to the stack frame in programming languages,
   2. Create an instance of class Block, in which,
      1. realize operators in the BlockDesc message,
   3. Run the Block by calling
      1. Block::Eval(vector<Variable>* targets) for forward and backward computations, or
      2. Block::Eval(vector<Operator>* targets) for optimization.

Intermediate Representation (IR)

Compile Time -> IR -> Runtime

Benefits of IR

• Optimization

  Compile Time -> IR -> Optimized IR -> Runtime
  

• Automatically send partitioned IR to different nodes.

  • Automatic Data Parallelism

    Compile Time
    |-> Single GPU IR
        |-> [trainer-IR-0, trainer-IR-1, pserver-IR]
            |-> Node-0 (runs trainer-IR-0)
            |-> Node-1 (runs trainer-IR-1)
            |-> Node-2 (runs pserver-IR)
    

  • Automatic Model Parallelism (planned for future)

---

Operator/OpWithKernel/OpKernel

---

Operator

• Operator is the fundamental building block of the user interface.
  • Operator stores input/output variable names and attributes.
  • The InferShape interface is used to infer the shape of the output variables based on the shapes of the input variables.
  • Use Run to compute the output variables from the input variables.

---

OpWithKernel/Kernel

• OpWithKernel inherits Operator.
• OpWithKernel contains a Kernel map.
  • OpWithKernel::Run get device’s kernel, and invoke OpKernel::Compute.
  • OpKernelKey is the map key. Only device place now, but may be data type later.

---

Why separate Kernel and Operator

• Separate GPU and CPU code.
  • Make Paddle capable of running without GPU.
• Make one operator (which is a user interface) and create many implementations.
  • For example, same multiplication op can have different implementations kernels such as FP16 kernel, FP32 kernel, MKL, eigen kernel.

---

Libraries for Kernel development

• Eigen::Tensor contains basic math and element-wise functions.
  • Note that Eigen::Tensor has broadcast implementation.
  • Limit the number of tensor.device(dev) = in your code.
• thrust::transform and std::transform.
  • thrust has the same API as C++ standard library. Using transform, one can quickly implement customized element-wise kernels.
  • thrust, in addition, supports more complex APIs, like scan, reduce, reduce_by_key.
• Hand-writing GPUKernel and CPU code
  • Do not write in header (.h) files. CPU Kernel should be in cpp source (.cc) and GPU kernels should be in cuda (.cu) files. (GCC cannot compile GPU code.)

---

Operator Registration

Why is registration necessary?

We need a method to build mappings between Op type names and Op classes.

How is registration implemented?

Maintaining a map, whose key is the type name and the value is the corresponding Op constructor.

---

The Registry Map

OpInfoMap

op_type(string) -> OpInfo

OpInfo:

• creator: The Op constructor.
• grad_op_type: The type of the gradient Op.
• proto: The Op’s Protobuf, including inputs, outputs and required attributes.
• checker: Used to check attributes.

---

Related Concepts

Op_Maker

It’s constructor takes proto and checker. They are completed during Op_Maker’s construction. (ScaleOpMaker)

Register Macros

REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, grad_op_class)
REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)

USE Macros

Make sure the registration process is executed and linked.

---

Registration Process

1. Write an Op class and its gradient Op class, if required.
2. Write an Op maker class. In the constructor of this class, describe the inputs, outputs and attributes of the operator.
3. Invoke the macro REGISTER_OP. This macro will
   1. Call maker class to complete proto and checker
   2. Using the completed proto and checker, it will add a new key-value pair to the OpInfoMap
4. Invoke the USE macro in which the Op is used to make sure that it is linked.

---

Backward Module (1/2)

Create Backward Operator

• Mapping from forward Op to backward Op

---

Backward Module (2/2)

Build Backward Network

• Input: a graph of forward operators
• Output: a graph of backward operators
• Corner cases in construction
  • Shared Variables => insert an Add operator to combine gradients
  • No Gradient => insert a fill_zero_grad operator
  • Recursive NetOp => call Backward recursively
  • RNN Op => recursively call Backward on stepnet
  • RNN Op => recursively call Backward on stepnet

---

Scope, Variable, Tensor

• Tensor is an n-dimension array with type.
  • Only dims and data pointers are stored in Tensor.
  • All operations on Tensor are written in Operator or global functions.
  • Variable length Tensor design LoDTensor
• Variable instances are the inputs and the outputs of an operator, not just Tensor.
  • step_scopes in RNN is a variable and not a tensor.
• Scope is where variables are stored.
  • map<string var name, Variable>
  • Scope has a hierarchical structure. The local scope can get variables from its parent scope.

---

Block (in design)

the difference between original RNNOp and Block

• As an operator is more intuitive than RNNOp,
• Offers a new interface Eval(targets) to deduce the minimal block to Run,
• Fits the compile-time/ runtime separation design paradigm.
  • During the compilation, SymbolTable stores VarDescs and OpDescs and serialize to a BlockDesc
  • When graph executes, a Block with BlockDesc is passed. It then creates Op and Var instances and then invokes Run.

---

Milestone

• Take Paddle/books as the main line, the requirement of the models motivates framework refactoring,
• Model migration
  • Framework development gives priority support to model migration, for example,
    • the MNIST demo needs a Python interface,
    • the RNN models require the framework to support LoDTensor.
  • Determine some timelines,
  • Frequently used Ops need to be migrated first,
  • Different models can be migrated in parallel.
• Improve the framework at the same time
• Accept imperfection, concentrate on solving the specific problem at the right price.

---

Control the migration quality

• Compare the performance of migrated models with old ones.
• Follow the google C++ style guide.
• Build the automatic workflow of generating Python/C++ documentations.
  • The documentation of layers and ops should be written inside the code.
  • Take the documentation quality into account when submitting pull requests.
  • Preview the documentations, read and improve them from a user’s perspective.