提交 1bbbc4e7 编写于 作者: Y Yi Wang

Merge branch 'develop' of http://github.com/paddlepaddle/paddle into...

Merge branch 'develop' of http://github.com/paddlepaddle/paddle into fix_cpplint_errors_operators_detail
repos:
- repo: https://github.com/Lucas-C/pre-commit-hooks.git
sha: v1.0.1
hooks:
......@@ -25,6 +26,14 @@
entry: bash ./.clang_format.hook -i
language: system
files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$
- repo: local
hooks:
- id: cpplint-cpp-source
name: cpplint
description: Check C++ code style using cpplint.py.
entry: bash ./tools/codestyle/cpplint_pre_commit.hook
language: system
files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx)$
- repo: https://github.com/PaddlePaddle/pre-commit-golang
sha: 8337620115c25ff8333f1b1a493bd031049bd7c0
hooks:
......
......@@ -146,6 +146,7 @@ include(external/cares)
include(external/grpc)
include(external/snappy) # download snappy
include(external/snappystream)
include(external/threadpool)
include(cudnn) # set cudnn libraries, must before configure
include(cupti)
......
INCLUDE(ExternalProject)
SET(THREADPOOL_SOURCE_DIR ${THIRD_PARTY_PATH}/threadpool)
SET(THREADPOOL_INCLUDE_DIR ${THREADPOOL_SOURCE_DIR}/src/extern_threadpool)
INCLUDE_DIRECTORIES(${THREADPOOL_INCLUDE_DIR})
ExternalProject_Add(
extern_threadpool
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/progschj/ThreadPool.git"
GIT_TAG 9a42ec1329f259a5f4881a291db1dcb8f2ad9040
PREFIX ${THREADPOOL_SOURCE_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
INSTALL_COMMAND ""
TEST_COMMAND ""
)
if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/threadpool_dummy.c)
file(WRITE ${dummyfile} "const char *dummy_threadpool = \"${dummyfile}\";")
add_library(simple_threadpool STATIC ${dummyfile})
else()
add_library(simple_threadpool INTERFACE)
endif()
add_dependencies(simple_threadpool extern_threadpool)
LIST(APPEND external_project_dependencies simple_threadpool)
add_custom_target(paddle_apis ALL
DEPENDS paddle_v2_apis paddle_fluid_apis)
add_custom_target(paddle_docs ALL
DEPENDS paddle_v2_docs paddle_v2_docs_cn
paddle_fluid_docs paddle_fluid_docs_cn)
add_subdirectory(v2)
add_subdirectory(fluid)
digraph G {
subgraph cluster_init {
label="Initialization"
startup_program [label="startup", shape=box]
node_w_g0 [label="W\nGPU0"]
startup_program -> node_w_g0 [label="Initialize"]
node_w_g1 [label="W\nGPU1"]
node_w_g0 -> node_w_g1 [label="broadcast"]
}
subgraph cluster_train {
label="forward_backward"
subgraph cluster_gpu0 {
label="GPU0"
fc_0 [label="fc\nGPU0", shape=box]
hidden_0 [label="hidden\nGPU0"]
node_w_g0 -> fc_0
fc_0 -> hidden_0
loss0 [label="loss\nGPU0"]
hidden_0 -> loss0 [label="many ops omitted"]
scale_loss_0 [label="scale_loss_gradient\nGPU0", shape=box]
loss_g0 [label="loss_grad\nGPU0"]
scale_loss_0->loss_g0
fc_g_0 [label="w_grad\nGPU0", shape=box]
loss0 -> fc_g_0
loss_g0 -> fc_g_0
hidden_0 -> fc_g_0
}
subgraph cluster_gpu1 {
label="GPU1"
fc_1 [label="fc\nGPU1", shape=box]
hidden_1 [label="hidden\nGPU1"]
node_w_g1 -> fc_1
fc_1 -> hidden_1
loss1 [label="loss\nGPU1"]
hidden_1 -> loss1 [label="many ops omitted"]
scale_loss_1 [label="scale_loss_gradient\nGPU1", shape=box]
loss_g1 [label="loss_grad\nGPU1"]
scale_loss_1->loss_g1
fc_g_1 [label="w_grad\nGPU1", shape=box]
loss1 -> fc_g_1
loss_g1 -> fc_g_1
hidden_1 -> fc_g_1
}
}
all_reduce_w [label="Merge Gradients(AllReduce)", shape=box]
fc_g_0 -> all_reduce_w
fc_g_1 -> all_reduce_w
fc_g_0_merged [label="w_grad\nMerged\nGPU0"]
fc_g_1_merged [label="w_grad\nMerged\nGPU1"]
all_reduce_w -> fc_g_0_merged
all_reduce_w -> fc_g_1_merged
subgraph cluster_optimization {
label="Optimization"
subgraph cluster_opt_gpu0 {
label="GPU0"
sgd_0 [label="SGD Op\nGPU0", shape=box]
fc_g_0_merged -> sgd_0
node_w_g0 -> sgd_0
optimized_w_0 [label="Optimized W\nGPU0"]
sgd_0 -> optimized_w_0
}
subgraph cluster_opt_gpu1 {
label="GPU1"
sgd_1 [label="SGD Op\nGPU1", shape=box]
fc_g_1_merged -> sgd_1
node_w_g1 -> sgd_1
optimized_w_1 [label="Optimized W\nGPU0"]
sgd_1 -> optimized_w_1
}
}
}
# ParallelExecutor
## Background
Neural network models are defined as a `ProgramDesc` in Fluid. The `ProgramDesc` can be executed by an interpreter(i.e. the `executor` concept in Fluid). The instructions or operators in a `Program` will be executed, and the results will be fetched in Python side.
The executor is a very naive interpreter. It runs operators one by one. We can use `Parallel.Do` to support data parallelism, however, lacking device information in `ProgramDesc`; it is not possible to optimize the performance of `Parallel.Do`.
We want a `ProgramDesc` can be run on different nodes. It is better not to contain device information in `ProgramDesc`. However, we can write a high-performance interpreter, which can hold an alternative intermediate representation of `ProgramDesc`, to take full usage of Multi-GPUs.
ParallelExecutor is an interpreter of `ProgramDesc` which will [out-of-order execute](https://en.wikipedia.org/wiki/Out-of-order_execution) `Program` in data parallelism mode and maximise the utility of Multi-GPUs.
## Overview of MultiGPUs logic
The ParallelExecutor takes the startup program and main program as inputs. The parameters will be initialised on `GPU0` by startup program and will broadcast to multi-GPUs. The main program will be duplicated into multi-GPUs. The gradient will be merged during each iteration, and each device will optimize parameters independently. Since the gradients on each device will be merged before parameter optimization, the parameters will be the same on each device and it does not need to be broadcast the parameters.
![alt](images/parallel_executor_overview.png)
There are several optimizations for this logic.
1. We use an alternate representation in ParallelExecutor. It because the device information is critical for performance optimization.
2. The execution is out-of-order, i.e., an operator will be executed whenever the inputs of the operator are ready.
* GPU is a high-performance device; only one CPU thread cannot fulfil one GPU. So there is a thread pool to execute operators.
* Out-of-order also helps transpilers to generate `ProgramDesc`. It is no need to concern about the best order of performance when implementing a transpiler.
3. The streams of computation, merge gradients and fetch data are different.
The performance of `ResNeXt152` on `TitanX` which `batch_size=12` is shown below.
| Number of GPUs | 1 | 2 | 3 | 4|
| --- | --- | --- | --- | --- |
| Image/Sec | 17.9906 | 25.771 | 36.911 | 48.8428 |
| Speed Up | N/A | 1.43247029 | 2.05168255 | 2.71490667 |
## Static single assignment Graph
[Static single assignment form](https://en.wikipedia.org/wiki/Static_single_assignment_form)(`SSA` for short) is a common form for compiler optimization. To implement concurrent execution, we uses an `SSA` graph as an intermedia representation of `ProgramDesc`.
The `Program` is a directed acyclic graph, since a variable can be assigned multiple times. We enforce a variable will be assigned once, by adding version number to varaibles. We parsing the `Program` into a `SSA` graph. Also, ProgramExecutor duplicate `Program` into multi-devices. We also add a device number to varaibles and insert `NCCLAllReduce` into Graph.
The data structure of `SSA` graph is:
```c++
struct VarHandleBase {
OpHandleBase* generated_op_;
vector<OpHandleBase*> pending_ops_;
string name;
Place place;
size_t version;
};
struct OpHandleBase {
vector<OpHandleBase*> inputs_;
vector<OpHnadleBase*> outputs_;
};
struct SSAGraph {
// vars on each devices.
// * the vars in each map in vector is on different device.
// * the map is mapping a variable name to variable handles
// with different versions
vector<std::unordered_map<string, vector<VarHandleBase>>> vars_;
// All ops
vector<OpHandleBase> ops_;
};
```
The variable handles are the wrapper of `Variables`. The operator handles are the wrapper of `OperatorBase`. Some `OpHandle` is not an `OperatorBase`, such as `NCCLAllReduceOpHandle`, because `AllReduceOpHandle` will use new device contexts.
When the `ProgramDesc` converted into an `SSA` Graph, the [data hazard](https://en.wikipedia.org/wiki/Hazard_(computer_architecture)) problem is also need to be taken care. The dummy variables, which represent the dependency between operators, will be manually inserted into SSA graph to resolve the [data hazard](https://en.wikipedia.org/wiki/Hazard_(computer_architecture)) problem.
## Execute SSA Graph
The SSA graph can be out-of-order executed by an approximate [topological sorting](https://en.wikipedia.org/wiki/Topological_sorting) algorithm. The algorithm is
1. Maintaining a map of an operator and its needed input number.
2. If a variable is not generated by an operator, i.e., `var.generated_op == nullptr`, decrease the needed input number of its pending operators.
3. If there is an operator which needed input number is decreased to zero, just run this operator.
4. After run this operator, just mark the variables are generated and repeat step 2 until all variables are generated.
Running an operator can be asynchronized. There is a thread pool to execute an `SSA` graph.
## Synchronize GPU Kernels
The GPU is a non-blocking device. The different streams need be synchronized when switing streams. In current implementation, the synchronization based on the following algorithm:
1. `OpHandle` will record `DeviceContext` that it is used.
2. In `OpHandle::Run`, if the `DeviceContext` of current operator is different from `DeviceContext` of any input variable, just wait the generate operator of this input variable.
The `wait` are implemented by two strategies:
1. Invoke `DeviceContext->Wait()`, It will wait all operators on this device contexts complete.
2. Uses `cudaStreamWaitEvent` to sending a event to the stream. It is a non-blocking call. The wait operators will be executed in GPU.
Generally, the `cudaStreamWaitEvent` will have a better perforamnce. However, `DeviceContext->Wait()` strategy is easier to debug. The strategy can be changed in runtime.
## What's next?
* Merging gradient of dense parameters has been done. However, the merging of sparse parameters has not been done.
* The CPU version of Parallel Executor has not been implemented. The out-of-order logic will make CPU compuatation faster, too.
* A better strategy to merge gradients can be introduced. We can shrink the gradients from `float32` to `int8` or `int4` while merging. It will significantly speed up multi-GPUs training without much loss of precision.
* Combine multi-Nodes implementation. By the benifit of out-of-order, sending and recving operator can be an blocking operator, and the transpiler does not need to concern about the best position of operator.
......@@ -27,6 +27,8 @@ sphinx_add_target(paddle_fluid_docs
${CMAKE_CURRENT_SOURCE_DIR}
${SPHINX_HTML_DIR_EN})
add_dependencies(paddle_fluid_docs gen_proto_py)
# configured documentation tools and intermediate build results
set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build")
......@@ -47,3 +49,7 @@ sphinx_add_target(paddle_fluid_docs_cn
${SPHINX_CACHE_DIR_CN}
${CMAKE_CURRENT_SOURCE_DIR}
${SPHINX_HTML_DIR_CN})
add_dependencies(paddle_fluid_docs_cn gen_proto_py)
add_subdirectory(api)
# configured documentation tools and intermediate build results
set(BINARY_BUILD_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_build")
# Sphinx cache with pickled ReST documents
set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees")
# HTML output director
set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html")
configure_file(
"${CMAKE_CURRENT_SOURCE_DIR}/../../templates/conf.py.en.in"
"${BINARY_BUILD_DIR_EN}/conf.py"
@ONLY)
sphinx_add_target(paddle_fluid_apis
html
${BINARY_BUILD_DIR_EN}
${SPHINX_CACHE_DIR_EN}
${CMAKE_CURRENT_SOURCE_DIR}
${SPHINX_HTML_DIR_EN})
add_dependencies(paddle_fluid_apis gen_proto_py framework_py_proto copy_paddle_pybind)
../../v2/build_and_install/build_from_source_cn.rst
\ No newline at end of file
../../v2/build_and_install/build_from_source_en.rst
\ No newline at end of file
../../v2/build_and_install/docker_install_cn.rst
\ No newline at end of file
../../v2/build_and_install/docker_install_en.rst
\ No newline at end of file
../../v2/build_and_install/index_cn.rst
\ No newline at end of file
../../v2/build_and_install/index_en.rst
\ No newline at end of file
../../v2/build_and_install/pip_install_cn.rst
\ No newline at end of file
../../v2/build_and_install/pip_install_en.rst
\ No newline at end of file
梯度更新算法
------------
.. toctree::
:maxdepth: 1
parameter_average.md
Gradient Update Algorithm
--------------------------------------
.. toctree::
:maxdepth: 1
parameter_average.md
......@@ -2,7 +2,7 @@ A few months ago when we were trying to replace CMake with Bazel, @emailweixu su
Here are some initial thoughts. Your comments are welcome!
### Required CMake Function
# Required CMake Function
I think we need only the following few CMake functions to make a project description mean and clean:
......@@ -25,7 +25,7 @@ Also,
- to describe external dependencies, we need `external_library`.
- to build shared libraries, we need `shared_library`.
### An Example Project
## An Example Project
Suppose that we have aforementioned functions defined in our `/cmake` directory. The following example `CMakeLists.txt` describes a project including the following source files:
......@@ -102,11 +102,11 @@ shared_library(api
```
### Implementation
## Implementation
As above example CMakeLists.txt executes, each function invocation adds "nodes" to a dependency graph. It also use this graph to generate CMake commands including `add_executable`, `add_dependencies`, `target_link_libraries`, and `add_test`.
### Using Package Manager For Go
## Using Package Manager For Go
Building Go binaries and libraries need to satisfy their dependencies, generally
we can do `go get ./...` to download and compile all external dependencies. The
......@@ -122,7 +122,7 @@ problems are:
at many cloud file hosting, so users what to compile paddle by themselves can
download this "vendor" package from a mirror site.
#### Choose A Suitable Tool
### Choose A Suitable Tool
As mentioned by @wangkuiyi, [Here](https://github.com/golang/go/wiki/PackageManagementTools)
list dozens of Go package managers. We choose the tool using following principles:
......@@ -140,7 +140,7 @@ management tool has been started at: https://github.com/golang/dep to resolve
such problems, but it's currently at Alpha stage. So the best choice now is
glide obviously.
#### Manage Go Packages
### Manage Go Packages
- Dependencies: `go/glide.yaml` will store the dependencies and their versions which
is directly imported by paddle. `go/glide.lock` will store all dependencies recursively
......
核心概念
-------------
.. toctree::
:maxdepth: 1
README.md
cpp_data_feeding.md
functions_operators_layers.md
program.md
variable.md
var_desc.md
tensor.md
tensor_array.md
lod_tensor.md
block.md
scope.md
executor.md
Core Concepts
--------------------------------------
.. toctree::
:maxdepth: 1
README.md
cpp_data_feeding.md
functions_operators_layers.md
program.md
variable.md
var_desc.md
tensor.md
tensor_array.md
lod_tensor.md
block.md
scope.md
executor.md
......@@ -78,7 +78,7 @@ In `Scope` class, there is a private data member called `parent_`. `parent_` is
A local scope is very useful when we implement Recurrent Neural Network. Each timestep of an RNN should be a `Net`. Each `Net` of timestep (`StepNet` for short) should use an independent local scope. Just like variables in a while loop is inside a local scope in programming languages. By using a single `StepNet` and changing local scope, we can implement an RNN easily.
# Interface Design
## Interface Design
```cpp
class Variable {
......
# Design Doc: Var_desc
## Background
PaddlePaddle divides the description of neural network computation into two stages: compile time and runtime. At compile time, the neural network computation is described as a `ProgramDesc` whereas at runtime an `Executor` interprets the `ProgramDesc` to compute the operations.
......
并发编程
------------
.. toctree::
:maxdepth: 1
concurrent_programming.md
parallel_do.md
Concurrent Programming
-------------------------
.. toctree::
:maxdepth: 1
concurrent_programming.md
parallel_do.md
数据类型
------------
.. toctree::
:maxdepth: 1
float16.md
Data Type
------------
.. toctree::
:maxdepth: 1
float16.md
## Design Doc: Distributed Lookup Table Operator
# Design Doc: Distributed Lookup Table Operator
A lookup table operator in PaddlePaddle where the table could be out
of the memory of a computer.
......
分布式训练
------------
.. toctree::
:maxdepth: 1
distributed_architecture.md
distributed_lookup_table_design.md
parameter_server.md
Distributed Training
---------------------
.. toctree::
:maxdepth: 1
distributed_architecture.md
distributed_lookup_table_design.md
parameter_server.md
动态RNN
------------
.. toctree::
:maxdepth: 1
rnn.md
rnn_design.md
Dynamic RNN
------------
.. toctree::
:maxdepth: 1
rnn.md
rnn_design.md
......@@ -233,7 +233,10 @@ x x
- 将每个序列concat 为规则的mini-batch表示
## 参考文献
1. [Tensorflow Bucketing](https://www.tensorflow.org/versions/r0.12/api_docs/python/contrib.training/bucketing)
2. [mxnet Bucketing](http://mxnet.io/how_to/bucketing.html)
3. [variable length input in RNN scenario](https://discuss.pytorch.org/t/about-the-variable-length-input-in-rnn-scenario/345/5)
4. [Level of details](https://en.wikipedia.org/wiki/Level_of_detail)
[Tensorflow Bucketing](https://www.tensorflow.org/versions/r0.12/api_docs/python/contrib.training/bucketing)
[mxnet Bucketing](http://mxnet.io/how_to/bucketing.html)
[variable length input in RNN scenario](https://discuss.pytorch.org/t/about-the-variable-length-input-in-rnn-scenario/345/5)
[Level of details](https://en.wikipedia.org/wiki/Level_of_detail)
执行流程
-------------
.. toctree::
:maxdepth: 1
switch.md
if_else_op.md
Execution Process
--------------------------------------
.. toctree::
:maxdepth: 1
switch.md
if_else_op.md
### Design Doc: Switch
# Design Doc: Switch
### Background
## Background
Many programming languages provide `switch` as a generalization of `if-elif-else`. We want to add it to Fluid.
......@@ -19,7 +19,7 @@ with switch() as switch:
fluid.print("Case 3")
```
### The Semantics
## The Semantics
1. A `switch` control-flow checks cases one-by-one.
1. The condition of each case is a boolean value, which is a scalar, and differs from the `fluid.if_else` control-flow, which condition could be a vector of boolean values.
......
设计思想
------------
.. toctree::
:maxdepth: 1
motivation/index_cn.rst
execution/index_cn.rst
concepts/index_cn.rst
data_type/index_cn.rst
memory/index_cn.rst
muti_devices/index_cn.rst
dynamic_rnn/index_cn.rst
concurrent/index_cn.rst
algorithm/index_cn.rst
network/index_cn.rst
modules/index_cn.rst
interface/index_cn.rst
dist_train/index_cn.rst
Design
------------
.. toctree::
:maxdepth: 1
motivation/index_en.rst
execution/index_en.rst
concepts/index_en.rst
data_type/index_en.rst
memory/index_en.rst
muti_devices/index_en.rst
dynamic_rnn/index_en.rst
concurrent/index_en.rst
algorithm/index_en.rst
network/index_en.rst
modules/index_en.rst
interface/index_en.rst
dist_train/index_en.rst
多语言接口
------------
TBD
Multi-Language Interface
-----------------------
TBD
内存管理
------------
.. toctree::
:maxdepth: 1
memory_optimization.md
Memory Management
-------------------
.. toctree::
:maxdepth: 1
memory_optimization.md
## Evaluator Design
# Evaluator Design
### Problem Statement
## Problem Statement
During training or inference, we provide an evaluation function to measure the model performance, for example, accuracy, precision, etc. In the operator based framework design, the data passes through the network pipeline batch by batch. As a result, inside the operator, we only calculate the metrics for one minibatch. Thus, we need to provide a mechanism to calculate the metrics for each N pass/batch the user wants.
### Evaluator Design
## Evaluator Design
Currently, every operation is expressed in the graph. We divide the evaluator process into three steps.
1. Initialize the metric state and add it into the block.
......@@ -14,7 +14,7 @@ Currently, every operation is expressed in the graph. We divide the evaluator pr
3. Merge the mini-batch statistics to form the evaluation result for multiple mini-batches. When it comes to distributed training/Multi-GPU training, aggregate the value from different devices.
### Implementation
## Implementation
This design is shown in the Python API.
Each metric operator needs to caculate the metric statistic and return the batch-aware states. Python side is responsible for accumulating the states for each pass.
......
代码结构和重要模块
-----------------
.. toctree::
:maxdepth: 1
backward.md
python_api.md
regularization.md
infer_var_type.md
optimizer.md
prune.md
register_grad_op.md
net_op_design.md
Code Structure and Important Modules
-------------------------------------
.. toctree::
:maxdepth: 1
backward.md
python_api.md
regularization.md
infer_var_type.md
optimizer.md
prune.md
register_grad_op.md
net_op_design.md
......@@ -8,9 +8,9 @@ A network object knows all Operators belonging to this network. Variables,
which are inputs and outputs of these operators,
are created and managed by a hierarchy of Scope objects.
# API
## API
## Net
### Net
To make the `Network` extendable, a base class is defined like this
```c++
......@@ -80,7 +80,7 @@ if (net) {
}
```
## `PlainNet` as a simple implementation of `BaseNet`
### `PlainNet` as a simple implementation of `BaseNet`
A very basic implementation is as follows. All it does is simply to run every operators in sequence.
......@@ -211,7 +211,7 @@ class NetBuilder final {
}
```
## Compatibility with RNN
### Compatibility with RNN
Benefitting from the decoupling of `PlainNet.Run` and `Scope`, `PlainNet` is compatible with future RNN design,
for example we can implement a simple recurrent neural network as follows
......
## Optimizer Design
# Optimizer Design
### The Problem
## The Problem
A PaddlePaddle program, or a block, is a sequence of operators operating variables. A training program needs to do three kinds of works:
......@@ -19,7 +19,7 @@ It's true that users should be able to create all these operators manually by ca
In this design, we propose a high-level API that automatically derives the optimisation pass and operators from the forward pass.
### High-level Python API to describe the training process
## High-level Python API to describe the training process
1. User write code to describe the network:
......@@ -54,7 +54,7 @@ In this design, we propose a high-level API that automatically derives the optim
sess.run(target= opt_op_list, ...)
```
#### Optimizer Python interface:
### Optimizer Python interface:
```python
class Optimizer(object):
......
设计动机和目标
-------------
.. toctree::
:maxdepth: 1
api.md
refactorization.md
fluid.md
fluid_compiler.md
Design Motivations and Goals
--------------------------------------
.. toctree::
:maxdepth: 1
api.md
refactorization.md
fluid.md
fluid_compiler.md
......@@ -97,13 +97,13 @@ Compile Time -> IR -> Runtime
---
# Operator/OpWithKernel/OpKernel
## Operator/OpWithKernel/OpKernel
![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/49caf1fb70820fb4a6c217634317c9306f361f36/op_op_with_kern_class_diagram.dot)
---
# Operator
## Operator
![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot)
* `Operator` is the fundamental building block of the user interface.
......@@ -113,7 +113,7 @@ Compile Time -> IR -> Runtime
---
# OpWithKernel/Kernel
## OpWithKernel/Kernel
![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/9d7f4eba185cf41c8e2fbfb40ae21890dbddcd39/op_with_kernel.dot)
......@@ -124,7 +124,7 @@ Compile Time -> IR -> Runtime
---
# Why separate Kernel and Operator
## Why separate Kernel and Operator
* Separate GPU and CPU code.
* Make Paddle capable of running without GPU.
......@@ -132,7 +132,7 @@ Compile Time -> IR -> Runtime
* For example, same multiplication op can have different implementations kernels such as FP16 kernel, FP32 kernel, MKL, eigen kernel.
---
# Libraries for Kernel development
## Libraries for Kernel development
* `Eigen::Tensor` contains basic math and element-wise functions.
* Note that `Eigen::Tensor` has broadcast implementation.
......@@ -143,16 +143,16 @@ Compile Time -> IR -> Runtime
* 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
## Operator Registration
## Why is registration necessary?
### Why is registration necessary?
We need a method to build mappings between Op type names and Op classes.
## How is registration implemented?
### How is registration implemented?
Maintaining a map, whose key is the type name and the value is the corresponding Op constructor.
---
# The Registry Map
## The Registry Map
### `OpInfoMap`
......@@ -166,7 +166,7 @@ Maintaining a map, whose key is the type name and the value is the corresponding
- **`checker`**: Used to check attributes.
---
# Related Concepts
## Related Concepts
### Op_Maker
It's constructor takes `proto` and `checker`. They are completed during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37))
......@@ -178,7 +178,7 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
```
---
# Registration Process
## 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
......@@ -186,13 +186,13 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
2. Using the completed `proto` and `checker`, it will add a new key-value pair to the `OpInfoMap`
---
# Backward Module (1/2)
## Backward Module (1/2)
### Create Backward Operator
- Mapping from forward Op to backward Op
![backward](https://gist.githubusercontent.com/dzhwinter/a6fbd4623ee76c459f7f94591fd1abf0/raw/61026ab6e518e66bde66a889bc42557a1fccff33/backward.png)
---
# Backward Module (2/2)
## Backward Module (2/2)
### Build Backward Network
- **Input**: a graph of forward operators
- **Output**: a graph of backward operators
......@@ -205,7 +205,7 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
---
# Scope, Variable, Tensor
## Scope, Variable, Tensor
* `Tensor` is an n-dimension array with type.
* Only dims and data pointers are stored in `Tensor`.
......@@ -218,8 +218,8 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
* `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
## 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.
......@@ -227,7 +227,7 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
- When graph executes, a Block with `BlockDesc` is passed. It then creates `Op` and `Var` instances and then invokes `Run`.
---
# Milestone
## 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,
......@@ -240,7 +240,7 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
- Accept imperfection, concentrate on solving the specific problem at the right price.
---
# Control the migration quality
## 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.
......
多设备支持
------------
.. toctree::
:maxdepth: 1
operator_kernel_type.md
kernel_selection.md
kernel_hint_design.md
Multi-Device Support
----------------------
.. toctree::
:maxdepth: 1
operator_kernel_type.md
kernel_selection.md
kernel_hint_design.md
## Problem
# Problem
In PaddlePaddle's [Design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md), one Operator may have multiple kernels. Users may have some personal preference to choose a certain type of kernel for an operator, such as `force_cpu` to choose a CPU kernel, `use_cudnn` to choose a CUDNN kernel, we need to provide a way for users to do this.
In the current design, we use KernelType to describe one kernel.
......
## Background
# Background
Every operator has many kernels because there are multiple data types, places, data layout, library type that Fluid supports. We use the `OpKernelType ` to describe kernel types that operators can hold.
The `OpKernelType ` is as follows:
......
复杂网络设计
------------
.. toctree::
:maxdepth: 1
sequence_decoder.md
Complex Network Design
------------------------
.. toctree::
:maxdepth: 1
sequence_decoder.md
开发标准
------------
.. toctree::
:maxdepth: 1
new_op_en.md
new_op_kernel_en.md
use_eigen_en.md
name_convention.md
support_new_device.md
releasing_process.md
op_markdown_format.md
Development
------------
This is Development page
.. toctree::
:maxdepth: 1
new_op_en.md
new_op_kernel_en.md
use_eigen_en.md
name_convention.md
support_new_device.md
releasing_process.md
op_markdown_format.md
## Operator's Parameter Name Convention
# Operator's Parameter Name Convention
To make the operator document itself more clear, we recommend operator names obey the listing conventions.
### OpProtoMaker names
## OpProtoMaker names
When defining an operator in Paddle, a corresponding [OpProtoMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L170) (TODO: OpProtoMaker Doc)need to be defined. All the Input/Output and Attributes will write into the [OpProto](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L61) , and will be used in client language to create operator.
......@@ -20,7 +20,7 @@ When defining an operator in Paddle, a corresponding [OpProtoMaker](https://gith
- Order.
- Follow the order of Input/Output, then Attribute, then Comments. See the example in best practice.
### Best Practice
## Best Practice
Here we give some examples to show how these rules will be used.
......
## Add Kernels for a New Device
# Add Kernels for a New Device
### Background
## Background
PaddlePaddle Fluid have hundreds of operators. Each operator could have one or more kernels. A kernel is an implementation of the operator for a certain device, which could be a hardware device, e.g., the CUDA GPU, or a library that utilizes a device, e.g., Intel MKL that makes full use of the Xeon CPU.
[This document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_en.md) explains how to add an operator, and its kernels. The kernels of an operator are indexed by a C++ type [`OpKernelType`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md). An operator chooses the right kernel at runtime. This choosing mechanism is described [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md).
### Write Kernels for A New Device
## Write Kernels for A New Device
#### Add A New Device
### Add A New Device
For some historical reaons, we misuse the word *library* for *device*. For example, we call the deivce type by *library type*. An example is the header file [`library_type.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/library_type.h#L24). We will correct this ASAP.
......@@ -23,7 +23,7 @@ enum class LibraryType {
```
#### Add A New [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L53)
### Add A New [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L53)
If you have a new kind of Device, firstly you need to add a new kind of [`Place`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L53). For example `CUDAPlace`:
......@@ -45,7 +45,7 @@ struct CUDAPlace {
typedef boost::variant<CUDAPlace, CPUPlace> Place;
```
#### Add [device context]((https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L37))
### Add [device context]((https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L37))
After a new kind of Device is added, you should add a corresponding [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L37) for it.
```cpp
......@@ -58,7 +58,7 @@ class DeviceContext {
};
```
#### Implement new [OpKernel](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L351) for your Device.
### Implement new [OpKernel](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L351) for your Device.
A detailed documentation can be found in [`new_op_and_kernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_en.md)
......@@ -85,7 +85,7 @@ class OpKernel : public OpKernelBase {
```
#### Register the OpKernel to framework
### Register the OpKernel to framework
After writing the components described above, we should register the kernel to the framework.
......
......@@ -15,26 +15,26 @@ The signature of the operator.
Each section mentioned above has been covered in further detail in the rest of the document.
# PaddlePaddle Operator Name
## PaddlePaddle Operator Name
This should be in all small letters, in case of multiple words, we separate them with an underscore. For example:
`array to lod tensor` should be written as `array_to_lod_tensor`.
This naming convention should be standard across all PaddlePaddle operators.
# Standard Operator Name
## Standard Operator Name
This is the standard name of the operator as used in the community. The general standard is usually:
- Standard abbreviations like `SGD` are written in all capital letters.
- Operator names that have multiple words inside a single word use `camelCase` (capitalize word boundaries inside of a word).
- Keep numbers inside a word as is, with no boundary delimiters.
- Follow the name of the operator with the keyword: `Activation Operator.`
# Operator description
## Operator description
This section should contain the description of what the operator does, including the operation performed, the literature from where it comes and was introduced first, and other important details. The relevant paper/article including the hyperlink should be cited in this section.
# LaTeX equation
## LaTeX equation
This section should contain an overall equation of the update or operation that the operator performs. The variables used in the equation should follow the naming convention of operators as described [here](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/name_convention.md). Two words in the same word should be separated by an underscore (`_`).
# The signature
## The signature
This section describes the signature of the operator. A list of Inputs and Outputs, each of which have a small description of what the variable represents and the type of variable. The variable names follow the `CamelCase` naming convention. The proposed format for this is:
`Section :
VariableName : (VariableType) VariableDescription
......
## 在Paddle中如何使用Eigen
# 在Paddle中如何使用Eigen
神经网络本质上是一个计算图,计算需要的数据存放在`Tensor`中,而计算过程是由`Operartor`来描述的。在执行时,`Operator`调用对应`OpKernel`中的`Compute`接口,实现对`Tensor`的操作。
### Eigen Tensor模块
## Eigen Tensor模块
Eigen Tensor模块对element-wise计算提供了强大的支持,并且书写一份代码,可以同时在CPU、GPU执行。但Eigen Tensor是一个正在开发中的模块,因此可能测试不够完备,文档较少。
关于Eigen Tensor模块的详细介绍请参考[文档1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md)[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md)
### paddle::framework::Tensor
## paddle::framework::Tensor
Paddle Tensor定义在framework目录下,其主要接口如下:
......@@ -129,7 +129,7 @@ From是EigenTensor模板提供的一个接口,可以实现从paddle::framework
### 实现计算
## 实现计算
当需要完成计算时,我们需要等式左边的EigenTensor调用device接口。在这里需要注意的是,这里的EigenTensor之间的运算只是改变了原有Tensor中的数据,而不会改变原有Tensor的shape信息。
......
## How to use Eigen in Paddle
# How to use Eigen in Paddle
Essentially, a neural network is a compute graph. T data needed for the computation is stored in `Tensor`s and its computation procedure is described by `Operator`s. An `Operator` calls the `Compute` interface in its corresponding `OpKernel` and operates on the `Tensor`.
### Eigen Tensor Module
## Eigen Tensor Module
The Eigen Tensor module supports powerful element-wise computation. In addition, a piece of code written using it can be run on both the CPU and the GPU.
......@@ -12,7 +12,7 @@ Note that Eigen Tensor is still being actively developed, so its tests are not c
For details on Eigen Tensor module, please see [doc 1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) and [doc 2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md).
### paddle::framework::Tensor
## paddle::framework::Tensor
Paddle Tensor's is defined in the framework directory with the following interface:
......@@ -105,7 +105,7 @@ void Compute(const framework::ExecutionContext& context) const override {
```
### paddle::framework::Tensor到EigenTensor的转换
## paddle::framework::Tensor到EigenTensor的转换
As shown above, in actual computation, we need to transform the input and output `Tensor`s into formats Eigen supports. We show some functions in [eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/eigen.h) to implement the transformation from `paddle::framework::Tensor`to `EigenTensor/EigenMatrix/EigenVector/EigenScalar`.
......@@ -129,7 +129,7 @@ For more transformations, see the [unit tests](https://github.com/PaddlePaddle/P
### Implementing Computation
## Implementing Computation
While computing, the device interface is needed from the EigenTensors on the left hand side of the assignments. Note that the computation between EigenTensors only changes the data originally inthe Tensor and does not change all the shape information associated with the Tensor.
......
基本使用概念
============
TBD
新手入门
------------
============
新手入门
如果需要快速了解PaddlePaddle的使用,可以参考以下指南。
.. toctree::
:maxdepth: 1
quickstart_cn.rst
在使用PaddlePaddle构建应用时,需要了解一些基本概念。
这里以一个线性回归为例子,详细介绍了PaddlePaddle的使用流程,包括数据格式,模型配置与训练等。
.. toctree::
:maxdepth: 1
concepts/use_concepts_cn.rst
GET STARTED
------------
============
This is get started page
If you want to quickly know how to use PaddlePaddle, please refer to the following guide:
.. toctree::
:maxdepth: 1
quickstart_en.rst
While using PaddlePaddle to build applications, please understand some basic concepts.
Here is an example of linear regression. It introduces workflow of PaddlePaddle, including data format, model configuration and training, etc.
.. toctree::
:maxdepth: 1
concepts/index_en.rst
../../v2/getstarted/quickstart_cn.rst
\ No newline at end of file
../../v2/getstarted/quickstart_en.rst
\ No newline at end of file
进阶使用
------------
.. toctree::
:maxdepth: 1
optimization/index_cn.rst
HOW TO
------------
This is how to page
.. toctree::
:maxdepth: 1
optimization/index_en.rst
../../../../../benchmark/cluster/README.md
\ No newline at end of file
基准
------------
.. toctree::
:maxdepth: 1
vgg16/README.md
README.md
Benchmark
------------
.. toctree::
:maxdepth: 1
vgg16/README.md
README.md
../../../../../../benchmark/cluster/vgg16/README.md
\ No newline at end of file
......@@ -8,7 +8,7 @@ PaddlePaddle 用户一般通过调用 Python API 编写深度学习程序。大
* Python 与 C++ 混合代码的性能分析
## Python代码的性能分析
# Python代码的性能分析
### 生成性能分析文件
......
......@@ -14,7 +14,7 @@ the profiling and tuning of
1. the Python code and
1. the mixture of Python and C++ code.
## Profiling the Python Code
# Profiling the Python Code
### Generate the Performance Profiling File
......
性能优化
------------
.. toctree::
:maxdepth: 1
timeline.md
cpu_profiling_cn.md
benchmark/index_cn.rst
Performance Optimization
---------------------------
.. toctree::
:maxdepth: 1
timeline.md
cpu_profiling_en.md
benchmark/index_en.rst
## how to use timeline tool to do profile
# how to use timeline tool to do profile
1. Add `with profiler.profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number.
......
......@@ -5,8 +5,8 @@
:maxdepth: 1
getstarted/index_cn.rst
design/index_cn.rst
build_and_install/index_cn.rst
design/index_cn.rst
howto/index_cn.rst
dev/index_cn.rst
faq/index_cn.rst
......@@ -5,8 +5,8 @@
:maxdepth: 1
getstarted/index_en.rst
design/index_en.rst
build_and_install/index_en.rst
design/index_en.rst
howto/index_en.rst
dev/index_en.rst
faq/index_en.rst
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册