diff --git a/.travis.yml b/.travis.yml index b4b83fcdbc..e217c8f5a7 100644 --- a/.travis.yml +++ b/.travis.yml @@ -4,7 +4,6 @@ cache: - $HOME/.ccache - $HOME/.cache/pip - $TRAVIS_BUILD_DIR/build/third_party - - $TRAVIS_BUILD_DIR/build_android/third_party sudo: required dist: trusty os: @@ -12,7 +11,6 @@ os: env: - JOB=build_doc - JOB=check_style - - JOB=build_android addons: apt: packages: @@ -23,7 +21,6 @@ addons: - python - python-pip - python2.7-dev - - python-numpy - python-wheel - libboost-dev - curl @@ -37,8 +34,8 @@ before_install: - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python # protobuf version. - - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt - - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker + - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt + - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker - curl https://glide.sh/get | bash - eval "$(GIMME_GO_VERSION=1.8.3 gimme)" - go get -u github.com/alecthomas/gometalinter diff --git a/CMakeLists.txt b/CMakeLists.txt index ad559672ad..08237cd850 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -65,8 +65,8 @@ if(NOT CMAKE_BUILD_TYPE) endif() if(ANDROID) - if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21") - message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 21") + if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16") + message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16") endif() set(WITH_GPU OFF CACHE STRING diff --git a/Dockerfile.android b/Dockerfile.android index c0fa58c384..452aa15745 100644 --- a/Dockerfile.android +++ b/Dockerfile.android @@ -4,9 +4,15 @@ MAINTAINER PaddlePaddle Authors ARG UBUNTU_MIRROR RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi' +# ENV variables +ARG ANDROID_ABI + +ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"} + ENV HOME=/root \ ANDROID_NDK_HOME=/opt/android-ndk-linux \ - ANDROID_STANDALONE_TOOLCHAIN=/opt/android-toolchain-gcc + ANDROID_ARM_STANDALONE_TOOLCHAIN=/opt/arm-toolchain \ + ANDROID_ARM64_STANDALONE_TOOLCHAIN=/opt/arm64-toolchain RUN apt-get update && \ apt-get install -y \ @@ -15,12 +21,11 @@ RUN apt-get update && \ apt-get clean -y # Install Go and glide -RUN wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \ - tar -C /usr/local -xzf go.tgz && \ +RUN wget -qO- go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \ + tar -xz -C /usr/local && \ mkdir /root/gopath && \ mkdir /root/gopath/bin && \ - mkdir /root/gopath/src && \ - rm go.tgz + mkdir /root/gopath/src ENV GOROOT=/usr/local/go GOPATH=/root/gopath # should not be in the same line with GOROOT definition, otherwise docker build could not find GOROOT. ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin @@ -42,7 +47,8 @@ RUN mkdir /opt/android-ndk-tmp && \ wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \ unzip -q android-ndk-r14b-linux-x86_64.zip && \ mv android-ndk-r14b ${ANDROID_NDK_HOME} && \ - ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-21 --install-dir=${ANDROID_STANDALONE_TOOLCHAIN} && \ + ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-23 --install-dir=${ANDROID_ARM_STANDALONE_TOOLCHAIN} && \ + ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm64 --platform=android-23 --install-dir=${ANDROID_ARM64_STANDALONE_TOOLCHAIN} && \ rm -rf /opt/android-ndk-tmp && \ rm -rf ${ANDROID_NDK_HOME} diff --git a/cmake/cross_compiling/android.cmake b/cmake/cross_compiling/android.cmake index 5e3e437a8d..84219cfa55 100644 --- a/cmake/cross_compiling/android.cmake +++ b/cmake/cross_compiling/android.cmake @@ -20,6 +20,7 @@ # The supported variables are listed belows: # # ANDROID_STANDALONE_TOOLCHAIN +# ANDROID_TOOLCHAIN # ANDROID_ABI # ANDROID_NATIVE_API_LEVEL # ANDROID_ARM_MODE @@ -57,6 +58,10 @@ IF(NOT DEFINED CMAKE_SYSTEM_VERSION AND ANDROID_NATIVE_API_LEVEL) ENDIF() ENDIF() +IF(NOT DEFINED ANDROID_TOOLCHAIN) + SET(ANDROID_TOOLCHAIN clang) +ENDIF() + IF(NOT DEFINED ANDROID_ABI) SET(ANDROID_ABI "armeabi-v7a") ENDIF() @@ -82,6 +87,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") "${CMAKE_VERSION}), when cross-compiling for Android.") IF(ANDROID_STANDALONE_TOOLCHAIN) + # Use standalone toolchain SET(CMAKE_SYSROOT "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot") IF(NOT CMAKE_SYSTEM_VERSION) @@ -96,26 +102,44 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ENDIF() # Toolchain - SET(ANDROID_TOOLCHAIN "gcc") SET(ANDROID_TOOLCHAIN_ROOT ${ANDROID_STANDALONE_TOOLCHAIN}) - IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$") - SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi) - IF(ANDROID_ABI STREQUAL "armeabi") - SET(CMAKE_SYSTEM_PROCESSOR armv5te) - ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a") - SET(CMAKE_SYSTEM_PROCESSOR armv7-a) - ENDIF() - ENDIF() - IF(ANDROID_ABI STREQUAL "arm64-v8a") - SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android) - SET(CMAKE_SYSTEM_PROCESSOR aarch64) + ELSE(ANDROID_NDK) + # TODO: use android ndk + ENDIF() + + IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$") + SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi) + IF(ANDROID_ABI STREQUAL "armeabi") + SET(CMAKE_SYSTEM_PROCESSOR armv5te) + SET(ANDROID_CLANG_TRIPLE armv5te-none-linux-androideabi) + ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a") + SET(CMAKE_SYSTEM_PROCESSOR armv7-a) + SET(ANDROID_CLANG_TRIPLE armv7-none-linux-androideabi) ENDIF() - SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-") + ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a") + SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android) + SET(CMAKE_SYSTEM_PROCESSOR aarch64) + SET(ANDROID_CLANG_TRIPLE aarch64-none-linux-android) + ELSE() + MESSAGE(FATAL_ERROR "Invalid Android ABI: ${ANDROID_ABI}.") + ENDIF() + SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-") + + IF(ANDROID_TOOLCHAIN STREQUAL clang) + SET(ANDROID_C_COMPILER_NAME clang) + SET(ANDROID_CXX_COMPILER_NAME clang++) + SET(CMAKE_C_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE}) + SET(CMAKE_CXX_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE}) + ELSEIF(ANDROID_TOOLCHAIN STREQUAL gcc) + SET(ANDROID_C_COMPILER_NAME gcc) + SET(ANDROID_CXX_COMPILER_NAME g++) + ELSE() + MESSAGE(FATAL_ERROR "Invalid Android toolchain: ${ANDROID_TOOLCHAIN}") ENDIF() # C compiler IF(NOT CMAKE_C_COMPILER) - SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}gcc") + SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_C_COMPILER_NAME}") ELSE() GET_FILENAME_COMPONENT(ANDROID_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM) ENDIF() @@ -125,7 +149,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") # CXX compiler IF(NOT CMAKE_CXX_COMPILER) - SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}g++") + SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_CXX_COMPILER_NAME}") ELSE() GET_FILENAME_COMPONENT(ANDROID_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM) ENDIF() @@ -137,7 +161,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") SET(CMAKE_CXX_COMPILER ${ANDROID_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE) # Toolchain and ABI specific flags. - SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections -finline-limit=64") + SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections") SET(ANDROID_LINKER_FLAGS "-Wl,--gc-sections") IF(ANDROID_ABI STREQUAL "armeabi") @@ -145,8 +169,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") -march=armv5te -mtune=xscale -msoft-float) - ENDIF() - IF(ANDROID_ABI STREQUAL "armeabi-v7a") + ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a") LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv7-a -mfloat-abi=softfp) @@ -156,6 +179,8 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=vfpv3-d16) ENDIF() LIST(APPEND ANDROID_LINKER_FLAGS -Wl,--fix-cortex-a8) + ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a") + LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a) ENDIF() IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$") @@ -164,10 +189,18 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0") ELSE() LIST(APPEND ANDROID_COMPILER_FLAGS -mthumb) ENDIF() + IF(ANDROID_TOOLCHAIN STREQUAL clang) + # Disable integrated-as for better compatibility. + LIST(APPEND ANDROID_COMPILER_FLAGS -fno-integrated-as) + ENDIF() ENDIF() - IF(ANDROID_ABI STREQUAL "arm64-v8a") - LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a) + IF(ANDROID_TOOLCHAIN STREQUAL clang) + # CMake automatically forwards all compiler flags to the linker, + # and clang doesn't like having -Wa flags being used for linking. + # To prevent CMake from doing this would require meddling with + # the CMAKE__COMPILE_OBJECT rules, which would get quite messy. + LIST(APPEND ANDROID_LINKER_FLAGS -Qunused-arguments) ENDIF() STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}") diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index 0002a470d9..f9e05af59f 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -12,6 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. +IF(USE_EIGEN_FOR_BLAS) + return() +ENDIF(USE_EIGEN_FOR_BLAS) + INCLUDE(cblas) IF(NOT ${CBLAS_FOUND}) diff --git a/doc/design/functions_operators_layers.md b/doc/design/functions_operators_layers.md index 7a2e8fd0ac..d23ba56b57 100644 --- a/doc/design/functions_operators_layers.md +++ b/doc/design/functions_operators_layers.md @@ -86,12 +86,13 @@ def layer.fc(X): We'd like to have Python bindings to operators in package `paddle.operator`, and Python compositions of operators in package `paddle.layer`. So we have the following concepts in above illustrative example: -``` + | C++ functions/functors | mul | add | | | +|------------------------|--------------|--------------|-------------|----------| | C++ operator class | mulOp | addOp | FCOp | | | Python binding | operator.mul | operator.add | operator.fc | | | Python function | | | | layer.fc | -``` + This is how we differentiate layer and operators in PaddlePaddle: diff --git a/doc/design/graph.md b/doc/design/graph.md index 87f696f90f..51b7f87638 100644 --- a/doc/design/graph.md +++ b/doc/design/graph.md @@ -1,4 +1,4 @@ -# Design Doc: Computations as Graphs +# Design Doc: Computations as a Graph A primary goal of the refactorization of PaddlePaddle is a more flexible representation of deep learning computation, in particular, a graph of operators and variables, instead of sequences of layers as before. @@ -8,6 +8,8 @@ This document explains that the construction of a graph as three steps: - construct the backward part - construct the optimization part +## The Construction of a Graph + Let us take the problem of image classification as a simple example. The application program that trains the model looks like: ```python @@ -25,7 +27,9 @@ The first four lines of above program build the forward part of the graph. ![](images/graph_construction_example_forward_only.png) -In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x. `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b. +In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x. `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b, and the initialization operators. + +Initialization operators are kind of "run-once" operators -- the `Run` method increments a class data member counter so to run at most once. By doing so, a parameter wouldn't be initialized repeatedly, say, in every minibatch. In this example, all operators are created as `OpDesc` protobuf messages, and all variables are `VarDesc`. These protobuf messages are saved in a `BlockDesc` protobuf message. @@ -49,3 +53,18 @@ According to the chain rule of gradient computation, `ConstructBackwardGraph` wo For each parameter, like W and b created by `layer.fc`, marked as double circles in above graphs, `ConstructOptimizationGraph` creates an optimization operator to apply its gradient. Here results in the complete graph: ![](images/graph_construction_example_all.png) + +## Block and Graph + +The word block and graph are interchangable in the desgin of PaddlePaddle. A [Block[(https://github.com/PaddlePaddle/Paddle/pull/3708) is a metaphore of the code and local variables in a pair of curly braces in programming languages, where operators are like statements or instructions. A graph of operators and variables is a representation of the block. + +A Block keeps operators in an array `BlockDesc::ops` + +```protobuf +message BlockDesc { + repeated OpDesc ops = 1; + repeated VarDesc vars = 2; +} +``` + +in the order that there appear in user programs, like the Python program at the beginning of this article. We can imagine that in `ops`, we have some forward operators, followed by some gradient operators, and then some optimization operators. diff --git a/doc/design/images/graph_construction_example.dot b/doc/design/images/graph_construction_example.dot index bedb6de011..8d1b673abf 100644 --- a/doc/design/images/graph_construction_example.dot +++ b/doc/design/images/graph_construction_example.dot @@ -2,6 +2,8 @@ digraph ImageClassificationGraph { ///////// The forward part ///////// FeedX [label="Feed", color=blue, shape=box]; FeedY [label="Feed", color=blue, shape=box]; + InitW [label="Init", color=blue, shape=diamond]; + Initb [label="Init", color=blue, shape=diamond]; FC [label="FC", color=blue, shape=box]; MSE [label="MSE", color=blue, shape=box]; @@ -14,6 +16,8 @@ digraph ImageClassificationGraph { FeedX -> x -> FC -> y -> MSE -> cost [color=blue]; FeedY -> l [color=blue]; + InitW -> W [color=blue]; + Initb -> b [color=blue]; W -> FC [color=blue]; b -> FC [color=blue]; l -> MSE [color=blue]; diff --git a/doc/design/images/graph_construction_example_all.png b/doc/design/images/graph_construction_example_all.png index 18d8330b60..1811875034 100644 Binary files a/doc/design/images/graph_construction_example_all.png and b/doc/design/images/graph_construction_example_all.png differ diff --git a/doc/design/images/graph_construction_example_forward_backward.png b/doc/design/images/graph_construction_example_forward_backward.png index 61c3a02a04..3049a9315f 100644 Binary files a/doc/design/images/graph_construction_example_forward_backward.png and b/doc/design/images/graph_construction_example_forward_backward.png differ diff --git a/doc/design/images/graph_construction_example_forward_only.png b/doc/design/images/graph_construction_example_forward_only.png index 14805df11f..25d19088cb 100644 Binary files a/doc/design/images/graph_construction_example_forward_only.png and b/doc/design/images/graph_construction_example_forward_only.png differ diff --git a/doc/design/simple_op_design.md b/doc/design/simple_op_design.md index 5e07c29c56..fded4a6861 100644 --- a/doc/design/simple_op_design.md +++ b/doc/design/simple_op_design.md @@ -147,7 +147,7 @@ class CosineOp { struct CosineOpProtoMaker : public OpProtoMaker { CosineOpProtoMaker(OpProto* proto) : OpProtoMaker(proto) { AddInput("input", "input of cosine op"); - AddAttr("scale", "scale of cosine op", float).Default(1.0).LargerThan(0.0); + AddAttr("scale", "scale of cosine op", float).Default(1.0).GreaterThan(0.0); AddType("cos"); AddComment("This is cos op"); } diff --git a/doc/design/var_desc.md b/doc/design/var_desc.md new file mode 100644 index 0000000000..86a95c10d5 --- /dev/null +++ b/doc/design/var_desc.md @@ -0,0 +1,124 @@ +## Background +PaddlePaddle divides the description of neural network computation graph into two stages: compile time and runtime. + +PaddlePaddle use proto message to describe compile time graph for + +1. Computation graph should be able to be saved to a file. +1. In distributed training, the graph will be serialized and send to multiple workers. + +The computation graph is constructed by Data Node and Operation Node. The concept to represent them is in the table below. + +| |compile time|runtime| +|---|---|---| +|Data|VarDesc(proto)|Variable(cpp)| +|Operation|OpDesc(proto)|Operator(cpp)| + + +## Definition of VarDesc + +A VarDesc should have a name and value, in PaddlePaddle, the value will always be a tensor. Since we use LoDTensor most of the time. We add a LoDTesnorDesc to represent it. + +```proto +message VarDesc { + required string name = 1; + optional LoDTensorDesc lod_tensor = 2; +} +``` + +## Definition of LodTensorDesc + +```proto +enum DataType { + BOOL = 0; + INT16 = 1; + INT32 = 2; + INT64 = 3; + FP16 = 4; + FP32 = 5; + FP64 = 6; +} + +message LoDTensorDesc { + required DataType data_type = 1; + repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480] + optional int32 lod_level = 3 [default=0]; +} +``` + +## Definition of Variable in Python + +In Python API, layer will take Variable as Input, and return Variable as Output. There should be a class `Variable` in python to help create and manage Variable. + +```python +image = Variable(dims=[-1, 640, 480]) +# fc1 and fc2 are both Variable +fc1 = layer.fc(input=image, output_size=10) +fc2 = layer.fc(input=fc1, output_size=20) +``` +### what should class `Variable` Have +1. `name`.a name of string type is used to mark the value of the Variable. +1. `initializer`. Since our Tensor does not have value. we will always use some Operator to fullfill it when run. So we should have a initialize method to help add the init operator. +1. `operator`. Variable should record which operator produce itself. The reaon is: + - we use pd.eval(targets=[var1, var2]) to run the related ops to get the value of var1 and var2. var.op is used to trace the dependency of the current variable. + +In PaddlePaddle, we use Block to describe Computation Graph, so in the code we will use Block but not Graph. + +```python +import VarDesc +import LoDTensorDesc +import framework + +def AddInitialOperator(variable, initializer): + # add an initialize Operator to block to init this Variable + +class Variable(object): + def __init__(self, name, dims, type, initializer): + self._block = get_default_block() + self._name = name + self.op = None + + tensor_desc = LoDTensorDesc(data_type=type, dims=dims) + _var_desc = VarDesc(name=name, lod_tensor=tensor_desc) + self._var = framework.CreateVar(_var_desc) + self._block.add_var(self) + + # add initial op according to initializer + if initializer is not None: + AddInitialOperator(self, initializer) + + def dims(self): + return self._var.dims() + + def data_type(self): + return self._var.data_type() + + def to_proto(self): + pass +``` + +Then we can use this Variable to create a fc layer in Python. + +```python +import paddle as pd + +def flatten_size(X, num_flatten_dims): + prod = 1 # of last num_flatten_dims + for i in xrange(num_flatten_dims): + prod = prod * X.dims[-i-1] + return prod + +def layer.fc(X, output_size, num_flatten_dims): + W = Variable(pd.random_uniform(), type=FP32, dims=[flatten_size(X, num_flatten_dims), output_size]) + b = Variable(pd.random_uniform(), type=FP32, dims=[output_size]) + out = Variable(type=FP32) + y = operator.fc(X, W, b, output=out) # fc will put fc op input into out + pd.InferShape(y) + return out + +x = Variable(dims=[-1, 640, 480]) +y = layer.fc(x, output_size=100) +z = layer.fc(y, output_size=200) + +paddle.eval(targets=[z], ...) +print(z) +``` diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md index 3e71a0a592..58665e9f2b 100644 --- a/doc/howto/dev/new_op_cn.md +++ b/doc/howto/dev/new_op_cn.md @@ -23,17 +23,20 @@ - `framework::OperatorWithKernel`：继承自OperatorBase，Op有计算函数，称作有Kernel。 - `class OpProtoAndCheckerMaker`：描述该Op的输入、输出、属性、注释,主要用于Python API接口生成 -依据是否包含kernel，将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结Op需要包含的内容如下： +依据是否包含kernel，可以将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结Op需要包含的内容如下： - - 内容 | 定义位置 --------------- | :---------------------- + + 内容 | 定义位置 +-------------- | :---------------------- OpProtoMake定义 | `.cc`文件，Backward Op不需要定义OpProtoMake -Op定义 | `.cc`文件 -Kernel实现 | CPU、GPU共享Kernel在`.h`文件，否则，CPU可以在`.cc`文件，GPU可在`.cu`文件。 -注册Op | Op注册在`.cc`文件；Kernel注册CPU在`.cc`文件，GPU在`.cu`文件 - - +Op定义 | `.cc`文件 +Kernel实现 | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，GPU 实现在`.cu`文件中。 +注册Op | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中 + + +实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。 + + 下面以矩阵乘操作，即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。 @@ -42,9 +45,11 @@ Kernel实现 | CPU、GPU共享Kernel在`.h`文件，否则，CPU可以在` ### 1. 定义ProtoMaker类 -矩阵乘的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。首先定义`ProtoMaker`来描述该Op的输入、输出及注释： - -``` +矩阵乘法的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。 + +首先定义`ProtoMaker`来描述该Op的输入、输出，并添加注释： + +```cpp class MulOpMaker : public framework::OpProtoAndCheckerMaker { public: MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) @@ -59,20 +64,20 @@ The equation is: Out = X * Y } }; ``` - -[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`，构造函数包括2个： + +[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`，构造函数含有2个参数： - `framework::OpProto` ： 前者存储Op的输入输出和参数属性，将用于Python API接口的生成。 - `framework::OpAttrChecker` ：后者用于检查参数属性的合法性。 - -构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加该Op的注释，这些函数会将对应内容添加到`OpProto`中。 -在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，该命名尽可能的规范。 +构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加Op的注释。这些函数会将对应内容添加到`OpProto`中。 - -再举个[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)的例子： - -``` +上面的代码在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，命名请遵守命名规范。 + + +再以[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)为例： + +```cpp template class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { public: @@ -87,17 +92,19 @@ The equation is: Out = scale*X } }; ``` - - 在这个例子里，两处不同： - - - `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中。 - - `AddAttr("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。 - + +这个例子有两处不同： + +- `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中，如果Op的某个输入不参与反向梯度的计算，请显示地调用`.NotInGradient()`进行设置。 + +- `AddAttr("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。 + ### 2. 定义Operator类 +下面的点实现了MulOp的定义： -```c++ +```cpp class MulOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -121,33 +128,46 @@ class MulOp : public framework::OperatorWithKernel { ``` [`MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L22)继承自`OperatorWithKernel`。`public`成员： - -```c++ + +```cpp using framework::OperatorWithKernel::OperatorWithKernel; ``` 这句表示使用基类`OperatorWithKernel`的构造函数，也可写成： - -```c++ + +```cpp MulOp(const std::string &type, const framework::VariableNameMap &inputs, const framework::VariableNameMap &outputs, const framework::AttributeMap &attrs) : OperatorWithKernel(type, inputs, outputs, attrs) {} -``` - +``` + 还需要重写`InferShape`接口。`InferShape`为const函数，不能修改Op的成员变量，参数为`const framework::InferShapeContext &ctx`，通过该参数可获取到输入输出以及属性。它的功能是： - 1). 做检查， 尽早报错：检查输入数据维度、类型等是否合法。 - 2). 设置输出Tensor的形状。 -通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和要讲到的注册函数一起放在`.cc`中 +通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和下面将要介绍的注册函数一起放在`.cc`中 ### 3. 定义OpKernel类 -```C++ -template -class MulKernel : public framework::OpKernel { - public: +`MulKernel`继承自`framework::OpKernel`，带有下面两个模板参数: + +- `typename Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 + +- `typename T` : 表示数据类型，如`float`, `double`等。 + +需要为`MulKernel`类重写`Compute`接口。 +- `Compute`接受一个输入参数：`const framework::ExecutionContext& context`。 +- 与`InferShapeContext`相比，`ExecutionContext`增加了设备类型，同样可获取到输入输出和属性参数。 +- `Compute`函数里实现`OpKernel`的具体计算逻辑。 + +下面是 `MulKernel` `Compute`的实现： + + ```cpp + template + class MulKernel : public framework::OpKernel { + public: void Compute(const framework::ExecutionContext& context) const override { auto* X = context.Input("X"); auto* Y = context.Input("Y"); @@ -157,174 +177,197 @@ class MulKernel : public framework::OpKernel { const_cast(context.device_context_); math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); } -}; -``` + }; + ``` + +需要注意：**不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。** + +`MulOp`的CPU、GPU实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 + +为了使`OpKernel`的计算过程书写更加简单，并且CPU、GPU的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)。 + + +到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。 +反向Op类的定义，反向OpKernel的定义与前向Op类似，这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。 -`MulKernel`继承自`framework::OpKernel`，带有模板参数: - - - `typename Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 - - - `typename T` : 表示数据类型，如`float`, `double`等。 - -`MulKernel`需要重写`Compute`接口，该接口参数为`const framework::ExecutionContext& context`, `ExecutionContext`相比`InferShapeContext`增加了设备类型，同样可获取到输入输出和属性参数，`Compute`函数里写具体实现时。 - -注意，不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。`MulOp`的CPU、GPU实现共享同一个`Kernel`，`OpKernel`不共享的例子可以参考[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 - -到此前向Op实现完成，需要在`.cc`文件中注册该op和kernel。反向Op类的定义和Kernel定义与前向Op类似，这里不再重复。但注意，反向Op没有`ProtoMaker`。 - ### 4. 注册Operator -在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。 +- 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。 -```c++ -namespace ops = paddle::operators; -REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); -REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); -REGISTER_OP_CPU_KERNEL(mul_grad, - ops::MulGradKernel); -``` - - - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`， - - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。 - - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。 - -在 `.cu`文件中注册GPU Kernel。 - -```c++ -namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); -REGISTER_OP_GPU_KERNEL(mul_grad, - ops::MulGradKernel); -``` + ```cpp + namespace ops = paddle::operators; + REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad); + REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel); + REGISTER_OP_CPU_KERNEL(mul_grad, + ops::MulGradKernel); + ``` + + 在上面的代码中： + + - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`。 + - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。 + - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。 + + +- 在 `.cu`文件中注册GPU Kernel。 + - 请注意，如果GPU Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下： + + ```cpp + // if use Eigen unsupported module before include head files + #define EIGEN_USE_GPU + + namespace ops = paddle::operators; + REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel); + REGISTER_OP_GPU_KERNEL(mul_grad, + ops::MulGradKernel); + ``` ### 5. 编译 -在[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)文件中添加编译。 - -``` -op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function) -``` - -下面命令可以编译： - -``` -make mul_op -``` +- 简单**无特殊依赖**的OP无需修改CMakeList.txt文件。[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt) 会自动将 `paddle/operators` 目录下新增的 `*_op.cc` 文件加入编译。 +- 较为复杂、**有额外依赖** 的operator仍需要修改[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)。如，`mul_op` 依赖 `math_function`，需要在`CMakeLists.txt`中添加如下内容： + + ``` + op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function) + + ``` + +- 运行下面命令可以进行编译： + + ``` + make mul_op + ``` ## 绑定Python -- 绑定Python - - 在 [`paddle/pybind/pybind.cc -`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc)文件中添加该类： +- 绑定Python + + 在 [`paddle/pybind/pybind.cc +`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc) 使用`USE_OP`告知编译器需要链接的Op，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。 ``` USE_OP(mul); ``` 如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`: - + ``` USE_CPU_ONLY_OP(gather); ``` - + 如果OP不带Kernel，则使用`USE_NO_KENREL_OP`: - + ``` USE_NO_KENREL_OP(recurrent); ``` - - 使用`USE_OP`告知编译器需要链接该Op的目标文件，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。 - - + + - 生成库 - 在 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件添加类到`DEPS`中，使得该Op可以链接到生成的lib库中。 - - ``` - if(WITH_PYTHON) - cc_library(paddle_pybind SHARED - SRCS pybind.cc - DEPS pybind python backward - mul_op - minus_op) - endif(WITH_PYTHON) - ``` + 无需修改 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件，`paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。 ## 实现单元测试 -单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单测](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。 +单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。 -### 前向Operator单测 +### 前向Operator单元测试 -前向Op单测继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`，具体单测流程在`OpTestMeta`里完成。需在`setUp`函数定义输入输出和属性参数，以及Python对比的输出值。 +前向Op单元测试继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`。各项更加具体的单元测试在`OpTestMeta`里完成。测试前向Operator，需要： -``` -import unittest -import numpy as np -from gradient_checker import GradientChecker, create_op -from op_test_util import OpTestMeta +1. 在`setUp`函数定义输入、输出，以及相关的属性参数。 +2. 生成随机的输入数据。 +3. 在Python脚本中实现与前向operator相同的计算逻辑，得到输出值，与operator前向计算的输出进行对比。 + + + ```python + import unittest + import numpy as np + from gradient_checker import GradientChecker, create_op + from op_test_util import OpTestMeta -class TestMulOp(unittest.TestCase): - __metaclass__ = OpTestMeta + class TestMulOp(unittest.TestCase): + __metaclass__ = OpTestMeta + def setUp(self): + self.type = "mul" + self.inputs = { + 'X': np.random.random((32, 84)).astype("float32"), + 'Y': np.random.random((84, 100)).astype("float32") + } + self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} + ``` + +上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释： + +- `self.type = "mul" ` : 定义类型，与operator注册时注册的类型一致。 +- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。 +- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。 + + +### 反向Operator单元测试 + +反向Op单元测试继承自`GradientChecker`，而`GradientChecker`继承自`unittest.TestCase`，因此，**反向单元测试函数需要以`test_`开头**。 + +```python +class TestMulGradOp(GradientChecker): def setUp(self): - self.type = "mul" + self.op = create_op("mul") self.inputs = { 'X': np.random.random((32, 84)).astype("float32"), 'Y': np.random.random((84, 100)).astype("float32") } - self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} -``` - 首先需要`import`必要的包,下面详细解释其他值： - - - `self.type = "mul" ` : 定义类型，和注册的类型一致。 - - `self.inputs` : 定义输入，类型为Numpy.array，并初始化。 - - `self.outputs` : 定义输出，并得到Python结算结果。 - - -### 反向Operator单测 - -反向Op单测继承自`GradientChecker`，而`GradientChecker`集成自`unittest.TestCase`，所以反向单测函数需要`test_`开头。 - - ``` - class MulGradOpTest(GradientChecker): - def test_mul(self): - op = create_op("mul") - inputs = { - 'X': np.random.random((32, 84)).astype("float32"), - 'Y': np.random.random((84, 100)).astype("float32") - } - self.compare_grad(op, inputs) + + def test_cpu_gpu_compare(self): + self.compare_grad(self.op, self.inputs) + + def test_normal(self): # mul op will enlarge the relative error self.check_grad( - op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5) - ``` + self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5) + + def test_ignore_x(self): + self.check_grad( + self.op, + self.inputs, ["Y"], + "Out", + max_relative_error=0.5, + no_grad_set={"X"}) + + def test_ignore_y(self): + self.check_grad( + self.op, + self.inputs, ["X"], + "Out", + max_relative_error=0.5, + no_grad_set={"Y"}) +``` + +下面解释代码中一些关键的地方: - - 调用`create_op("mul")`创建反向Op对应的前向Op。 - - 定义输入`inputs`。 - - 调用`compare_grad`函数对比CPU、GPU计算结果。 - - 调用`check_grad`检查梯度稳定性，这里采用数值法检测梯度正确性。 - - 第一个参数`op` : 前向op。 - - 第二个参数`inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。 - - 第三个参数`set(["X", "Y"])` : 指定对输入变量`X`、`Y`做梯度检测。 - - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out` +- 调用`create_op("mul")`创建反向Op对应的前向Op。 +- 调用`compare_grad`函数对比CPU、GPU计算结果。 +- `test_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。 + - 第一个参数`self.op` : 前向Op。 + - 第二个参数`self.inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。 + - 第三个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。 + - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out` +- `test_ignore_x`和`test_ignore_y`分支用来测试只需要计算一个输入梯度的情况。 -### 编译和执行 +### 编译和执行单元测试 -单测完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)里添加编译： +单元测试编写完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)中添加以下内容，将单元测试加入工程： ``` py_test(test_mul_op SRCS test_mul_op.py) ``` -编译时需要打开`WITH_TESTING`, 即 `cmake paddle_dir -DWITH_TESTING=ON`，编译成功之后执行单测命令为： +请注意，**不同于Op的编译测试，运行单元测试测时需要编译整个工程**，并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON`。编译成功后，执行下面的命令来运行单元测试： -``` +```bash make test ARGS="-R test_mul_op -V" ``` + 或者: -``` +```bash ctest -R test_mul_op ``` diff --git a/doc/howto/dev/use_eigen_cn.md b/doc/howto/dev/use_eigen_cn.md new file mode 100644 index 0000000000..1367323b71 --- /dev/null +++ b/doc/howto/dev/use_eigen_cn.md @@ -0,0 +1,146 @@ +## 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) 和[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md) + + +### paddle::framework::Tensor + +Paddle Tensor定义在framework目录下，其主要接口如下： + +```cpp +class Tensor { + public: + /*! Return a pointer to mutable memory block. */ + template + inline T* data(); + + /** + * @brief Return a pointer to mutable memory block. + * @note If not exist, then allocation. + */ + template + inline T* mutable_data(platform::Place place); + + /** + * @brief Return a pointer to mutable memory block. + * + * @param[in] dims The dimensions of the memory block. + * @param[in] place The place of the memory block. + * + * @note If not exist, then allocation. + */ + template + inline T* mutable_data(DDim dims, platform::Place place); + + /*! Resize the dimensions of the memory block. */ + inline Tensor& Resize(const DDim& dims); + + /*! Return the dimensions of the memory block. */ + inline const DDim& dims() const; + + private: + /*! holds the memory block if allocated. */ + std::shared_ptr holder_; + + /*! points to dimensions of memory block. */ + DDim dim_; +}; +``` + +`Placeholder`的作用是延迟分配内存，即我们可以先定义一个Tensor，然后使用Resize接口设置Tensor的大小，最后再调用mutable_data接口分配实际的内存。 + +```cpp +paddle::framework::Tensor t; +paddle::platform::CPUPlace place; +// set size first +t.Resize({2, 3}); +// allocate memory on CPU later +t.mutable_data(place); +``` + +### paddle::framework::Tensor使用样例 +下面以AddOp为例说明Tensor的使用过程： + +- InferShape + +在运行神经网络计算图时，我们先调用每个`Operator`的`InferShape`接口，根据输入Tensor的大小来设置输出Tensor的大小，`Resize`接口会被调用。 + +```cpp +void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_EQ(ctx.Input("X")->dims(), + ctx.Input("Y")->dims(), + "Two input of Add Op's dimension must be same."); + ctx.Output("Out")->Resize(ctx.Input("X")->dims()); +} +``` + + +- Run + +`Operator`的`Run`接口最终会调用对应`OpKernel`的`Compute`接口，在这时真正的分配内存，`mutable_data`接口会被调用。 + +```cpp +void Compute(const framework::ExecutionContext& context) const override { + auto* input0 = context.Input("X"); + auto* input1 = context.Input("Y"); + auto* output = context.Output("Out"); + + output->mutable_data(context.GetPlace()); + + auto x = EigenVector::Flatten(*input0); + auto y = EigenVector::Flatten(*input1); + auto z = EigenVector::Flatten(*output); + + auto place = context.GetEigenDevice(); + + z.device(place) = x + y; +} +``` + + +### paddle::framework::Tensor到EigenTensor的转换 + +如上一小节所示，在具体的计算中，我们需要先把输入Tensor和输出Tensor转换为Eigen支持的格式。我们在[eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen.h)中提供了一些全局函数用来实现paddle::framework::Tensor到EigenTensor/EigenMatrix/EigenVector/EigenScalar的转换。 + +以EigenTensor为例，做一个介绍 + +```cpp +Tensor t; +float* p = t.mutable_data(make_ddim({1, 2, 3}), platform::CPUPlace()); +for (int i = 0; i < 1 * 2 * 3; i++) { + p[i] = static_cast(i); +} + +EigenTensor::Type et = EigenTensor::From(t); +``` + +From是EigenTensor模板提供的一个接口，可以实现从paddle::framework::Tensor到对EigenTensor的转换。由于Tensor的rank是模板参数，因此在转换时需要显示的指定。 + +在Eigen中，不同rank的Tensor是不同类型，Vector是rank为1的Tensor。需要额外注意的是，EigenVector::From方法是把paddle中的一维Tensor转为Eigen的一维Tensor，在这里用EigenVector来表示；而EigenVector::Flatten方法是把paddle中的一个Tensor进行reshape操作，压扁成为Eigen的一维Tensor，类型仍然为EigenVector。 + +更多的转换方法请参考eigen_test.cc中的[单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen_test.cc)。 + + + +### 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。 + +```cpp +auto x = EigenVector::Flatten(*input0); +auto y = EigenVector::Flatten(*input1); +auto z = EigenVector::Flatten(*output); +auto place = context.GetEigenDevice(); +z.device(place) = x + y; +``` + +在这段代码中，input0/input1/output可以是任意维度的Tensor。我们调用了EigenVector的Flatten接口，把任意维度的Tensor转为了一维的EigenVector。而在计算结束之后，input0/input1/output的原有shape信息不变。如果想改变原有Tensor的shape信息，可以调用Resize接口进行改变。 + +由于Eigen Tensor模块的文档较少，我们可以参考TensorFlow的[kernels](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/kernels)模块下的相关`OpKernel`的计算代码。 diff --git a/paddle/cuda/include/hl_cpu_gru.cuh b/paddle/cuda/include/hl_cpu_gru.cuh index c0a37ced2a..e4f6bf42c6 100644 --- a/paddle/cuda/include/hl_cpu_gru.cuh +++ b/paddle/cuda/include/hl_cpu_gru.cuh @@ -18,14 +18,6 @@ limitations under the License. */ #ifndef __NVCC__ -#include "paddle/math/MathFunctions.h" - -#ifndef PADDLE_TYPE_DOUBLE -#define CBLAS_GEMM paddle::gemm -#else -#define CBLAS_GEMM paddle::gemm -#endif - template void hl_naive_gru_forward_reset_output(OpResetOutput opResetOutput, real *gateValue, @@ -210,51 +202,6 @@ inline void forward_final_output(OpFinalOutput opFinalOutput, } } -template -void hl_cpu_gru_forward(OpResetOutput opResetOutput, - OpFinalOutput opFinalOutput, - hl_gru_value value, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate) { - if (value.prevOutValue) { - CBLAS_GEMM(CblasNoTrans, - CblasNoTrans, - batchSize, - 2 * frameSize, - frameSize, - 1, - value.prevOutValue, - frameSize, - value.gateWeight, - frameSize * 2, - 1, - value.gateValue, - frameSize * 3); - } - - forward_reset_output(opResetOutput, value, frameSize, batchSize, active_gate); - - if (value.prevOutValue) { - CBLAS_GEMM(CblasNoTrans, - CblasNoTrans, - batchSize, - frameSize, - frameSize, - 1, - value.resetOutputValue, - frameSize, - value.stateWeight, - frameSize, - 1, - value.gateValue + frameSize * 2, - frameSize * 3); - } - - forward_final_output(opFinalOutput, value, frameSize, batchSize, active_node); -} - template void hl_naive_gru_backward_state_grad(OpStateGrad opStateGrad, real *gateValue, @@ -525,86 +472,6 @@ inline void backward_reset_grad(OpResetGrad opResetGrad, } } -template -void hl_cpu_gru_backward(OpStateGrad opStateGrad, - OpResetGrad opResetGrad, - hl_gru_value value, - hl_gru_grad grad, - int frameSize, - int batchSize, - hl_activation_mode_t active_node, - hl_activation_mode_t active_gate) { - backward_state_grad(opStateGrad, value, grad, - frameSize, batchSize, active_node); - - if (value.prevOutValue && grad.prevOutGrad) { - CBLAS_GEMM(CblasNoTrans, - CblasTrans, - batchSize, - frameSize, - frameSize, - 1, - grad.gateGrad + frameSize * 2, - frameSize * 3, - value.stateWeight, - frameSize, - 0, - grad.resetOutputGrad, - frameSize); - - if (grad.stateWeightGrad) { - CBLAS_GEMM(CblasTrans, - CblasNoTrans, - frameSize, - frameSize, - batchSize, - 1, - value.resetOutputValue, - frameSize, - grad.gateGrad + frameSize * 2, - frameSize * 3, - 1, - grad.stateWeightGrad, - frameSize); - } - } - - backward_reset_grad(opResetGrad, value, grad, - frameSize, batchSize, active_gate); - - if (grad.prevOutGrad && value.prevOutValue) { - CBLAS_GEMM(CblasNoTrans, - CblasTrans, - batchSize, - frameSize, - frameSize * 2, - 1, - grad.gateGrad, - frameSize * 3, - value.gateWeight, - frameSize * 2, - 1, - grad.prevOutGrad, - frameSize); - - if (grad.gateWeightGrad) { - CBLAS_GEMM(CblasTrans, - CblasNoTrans, - frameSize, - frameSize * 2, - batchSize, - 1, - value.prevOutValue, - frameSize, - grad.gateGrad, - frameSize * 3, - 1, - grad.gateWeightGrad, - frameSize * 2); - } - } -} - #endif #endif // HL_CPU_GRU_CUH_ diff --git a/paddle/framework/attribute.cc b/paddle/framework/attribute.cc index 9eb07acdff..27132eaa0b 100644 --- a/paddle/framework/attribute.cc +++ b/paddle/framework/attribute.cc @@ -43,6 +43,10 @@ template <> AttrType AttrTypeID>() { return STRINGS; } +template <> +AttrType AttrTypeID>>() { + return INT_PAIRS; +} Attribute GetAttrValue(const OpDesc::Attr& attr_desc) { switch (attr_desc.type()) { @@ -76,6 +80,14 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc) { } return val; } + case paddle::framework::AttrType::INT_PAIRS: { + std::vector> val(attr_desc.int_pairs_size()); + for (int i = 0; i < attr_desc.int_pairs_size(); ++i) { + val[i].first = attr_desc.int_pairs(i).first(); + val[i].second = attr_desc.int_pairs(i).second(); + } + return val; + } } PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !"); return boost::blank(); diff --git a/paddle/framework/attribute.h b/paddle/framework/attribute.h index 08b47cabd4..2b788a76ca 100644 --- a/paddle/framework/attribute.h +++ b/paddle/framework/attribute.h @@ -28,7 +28,8 @@ namespace paddle { namespace framework { typedef boost::variant, - std::vector, std::vector> + std::vector, std::vector, + std::vector>> Attribute; typedef std::unordered_map AttributeMap; @@ -40,11 +41,23 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc); // check whether a value(attribute) fit a certain limit template -class LargerThanChecker { +class GreaterThanChecker { public: - explicit LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {} + explicit GreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {} void operator()(T& value) const { - PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail"); + PADDLE_ENFORCE(value > lower_bound_, "larger_than check fails."); + } + + private: + T lower_bound_; +}; + +template +class EqualGreaterThanChecker { + public: + explicit EqualGreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {} + void operator()(T& value) const { + PADDLE_ENFORCE_GE(value, lower_bound_, "equal_larger_than check fails."); } private: @@ -109,8 +122,13 @@ class TypedAttrChecker { return *this; } - TypedAttrChecker& LargerThan(const T& lower_bound) { - value_checkers_.push_back(LargerThanChecker(lower_bound)); + TypedAttrChecker& GreaterThan(const T& lower_bound) { + value_checkers_.push_back(GreaterThanChecker(lower_bound)); + return *this; + } + + TypedAttrChecker& EqualGreaterThan(const T& lower_bound) { + value_checkers_.push_back(EqualGreaterThanChecker(lower_bound)); return *this; } diff --git a/paddle/framework/backward.md b/paddle/framework/backward.md index 8aa6728a95..c762811dfc 100644 --- a/paddle/framework/backward.md +++ b/paddle/framework/backward.md @@ -2,20 +2,20 @@ ## Motivation -In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. - +In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the gradient operators/expressions together with the chain rule. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. + ## Backward Operator Registry -A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs and output gradients and then calculate its input gradients. +A backward network is built up with several backward operators. Backward operators take forward operators' inputs outputs, and output gradients and then calculate its input gradients. | | forward operator | backward operator | ---------------------- | ---------------- |------------------------- | | **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients | | **Operator::outputs_** | Outputs | InputGradients | - In most cases, there is a one-to-one correspondence between forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced. + In most cases, there is a one-to-one correspondence between the forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced. -For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro: +For example, we have got a `mul_op`, and we can register its information and corresponding backward operator by the following macro: ```cpp REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad); @@ -27,17 +27,17 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad); ## Backward Opeartor Creating -Given a certain forward operator, we can get its corresponding backward opeartor by calling: +Given a certain forward operator, we can get its corresponding backward operator by calling: ```cpp OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op); -``` +``` The function `BuildGradOp` will sequentially execute following processes: 1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`. -2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these are not necessary for gradient computing. +2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing. 3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`. @@ -49,31 +49,31 @@ A backward network is a series of backward operators. The main idea of building In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. -given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`. +given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`, `InputGradients`. 1. Op - when the input forward network is a Op, return its gradient Operator Immediately. + when the input forward network is an Op, return its gradient Operator Immediately. 2. NetOp - when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp. + when the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp. - **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable. + **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwrite their shared input variable.

-
+
- 1. shared variable in two operators. + 1. Shared variable in operators.

- Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links. + Share variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator replace the overwrite links.

-
+
- 2. replace shared variable gradient with `Add` Operator + 2. Replace shared variable's gradient with `Add` operator.

diff --git a/paddle/framework/ddim.cc b/paddle/framework/ddim.cc index cfd3e8dfde..fc3d508553 100644 --- a/paddle/framework/ddim.cc +++ b/paddle/framework/ddim.cc @@ -21,16 +21,16 @@ namespace framework { /// @cond HIDDEN template -Dim make_dim(const int* d) { +Dim make_dim(const int64_t* d) { return Dim(*d, make_dim(d + 1)); } template <> -Dim<1> make_dim<1>(const int* d) { +Dim<1> make_dim<1>(const int64_t* d) { return Dim<1>(*d); } -void make_ddim(DDim& ddim, const int* dims, int n) { +void make_ddim(DDim& ddim, const int64_t* dims, int n) { switch (n) { case 1: ddim = make_dim<1>(dims); @@ -67,13 +67,13 @@ void make_ddim(DDim& ddim, const int* dims, int n) { /// @endcond -DDim make_ddim(std::initializer_list dims) { +DDim make_ddim(std::initializer_list dims) { DDim result(make_dim(0)); make_ddim(result, dims.begin(), dims.size()); return result; } -DDim make_ddim(const std::vector& dims) { +DDim make_ddim(const std::vector& dims) { DDim result(make_dim(0)); make_ddim(result, &dims[0], dims.size()); return result; @@ -81,12 +81,12 @@ DDim make_ddim(const std::vector& dims) { /// @cond HIDDEN // XXX For some reason, putting this in an anonymous namespace causes errors -class DynamicMutableIndexer : public boost::static_visitor { +class DynamicMutableIndexer : public boost::static_visitor { public: explicit DynamicMutableIndexer(int idx) : idx_(idx) {} template - int& operator()(Dim& dim) const { + int64_t& operator()(Dim& dim) const { return dim[idx_]; } @@ -94,12 +94,12 @@ class DynamicMutableIndexer : public boost::static_visitor { int idx_; }; -class DynamicConstIndexer : public boost::static_visitor { +class DynamicConstIndexer : public boost::static_visitor { public: explicit DynamicConstIndexer(int idx) : idx_(idx) {} template - int operator()(const Dim& dim) const { + int64_t operator()(const Dim& dim) const { return dim[idx_]; } @@ -109,22 +109,22 @@ class DynamicConstIndexer : public boost::static_visitor { /// @endcond -int& DDim::operator[](int idx) { +int64_t& DDim::operator[](int idx) { return boost::apply_visitor(DynamicMutableIndexer(idx), var); } -int DDim::operator[](int idx) const { +int64_t DDim::operator[](int idx) const { return boost::apply_visitor(DynamicConstIndexer(idx), var); } -ssize_t DDim::size() const { return arity(*this); } +int64_t DDim::size() const { return arity(*this); } bool DDim::operator==(DDim d) const { if (var.which() != d.getVar().which()) { return false; } else { - std::vector v1 = vectorize(*this); - std::vector v2 = vectorize(d); + std::vector v1 = vectorize(*this); + std::vector v2 = vectorize(d); for (unsigned int i = 0; i < v1.size(); i++) { if (v1[i] != v2[i]) { @@ -139,10 +139,10 @@ bool DDim::operator==(DDim d) const { bool DDim::operator!=(DDim d) const { return !(*this == d); } DDim DDim::operator+(DDim d) const { - std::vector v1 = vectorize(*this); - std::vector v2 = vectorize(d); + std::vector v1 = vectorize(*this); + std::vector v2 = vectorize(d); - std::vector v3; + std::vector v3; assert(v1.size() == v2.size()); @@ -154,10 +154,10 @@ DDim DDim::operator+(DDim d) const { } DDim DDim::operator*(DDim d) const { - std::vector v1 = vectorize(*this); - std::vector v2 = vectorize(d); + std::vector v1 = vectorize(*this); + std::vector v2 = vectorize(d); - std::vector v3; + std::vector v3; assert(v1.size() == v2.size()); @@ -168,15 +168,15 @@ DDim DDim::operator*(DDim d) const { return make_ddim(v3); } -int get(const DDim& ddim, int idx) { return ddim[idx]; } +int64_t get(const DDim& ddim, int idx) { return ddim[idx]; } void set(DDim& ddim, int idx, int value) { ddim[idx] = value; } /// @cond HIDDEN struct VectorizeVisitor : public boost::static_visitor<> { - std::vector& vector; + std::vector& vector; - explicit VectorizeVisitor(std::vector& v) : vector(v) {} + explicit VectorizeVisitor(std::vector& v) : vector(v) {} template void operator()(const T& t) { @@ -188,31 +188,31 @@ struct VectorizeVisitor : public boost::static_visitor<> { }; /// @endcond -std::vector vectorize(const DDim& ddim) { - std::vector result; +std::vector vectorize(const DDim& ddim) { + std::vector result; VectorizeVisitor visitor(result); boost::apply_visitor(visitor, ddim); return result; } -struct ProductVisitor : public boost::static_visitor { +struct ProductVisitor : public boost::static_visitor { template - ssize_t operator()(const Dim& dim) { + int64_t operator()(const Dim& dim) { return product(dim); } }; -ssize_t product(const DDim& ddim) { +int64_t product(const DDim& ddim) { ProductVisitor visitor; return boost::apply_visitor(visitor, ddim); } struct SliceVectorizeVisitor : public boost::static_visitor<> { - std::vector& vector; + std::vector& vector; int begin; int end; - SliceVectorizeVisitor(std::vector& v, int b, int e) + SliceVectorizeVisitor(std::vector& v, int b, int e) : vector(v), begin(b), end(e) { PADDLE_ENFORCE(begin < end, "Begin index must be less than end index in ddim slice."); @@ -240,7 +240,7 @@ struct SliceVectorizeVisitor : public boost::static_visitor<> { }; DDim slice_ddim(const DDim& dim, int begin, int end) { - std::vector vec; + std::vector vec; vec.reserve(end - begin); SliceVectorizeVisitor visitor(vec, begin, end); boost::apply_visitor(visitor, dim); @@ -280,8 +280,17 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) { return os; } -DDim::DDim(std::initializer_list init_list) { +DDim::DDim(std::initializer_list init_list) { *this = make_ddim(init_list); } + +DDim flatten_to_2d(const DDim& src, int num_col_dims) { + int rank = src.size(); + return make_ddim({product(slice_ddim(src, 0, num_col_dims)), + product(slice_ddim(src, num_col_dims, rank))}); +} + +DDim flatten_to_1d(const DDim& src) { return make_ddim({product(src)}); } + } // namespace framework } // namespace paddle diff --git a/paddle/framework/ddim.h b/paddle/framework/ddim.h index 95f294b627..ca29e7e8c7 100644 --- a/paddle/framework/ddim.h +++ b/paddle/framework/ddim.h @@ -40,7 +40,7 @@ struct DDim { template explicit DDim(const Dim& in) : var(in) {} - /*implicit*/ DDim(std::initializer_list init_list); + /*implicit*/ DDim(std::initializer_list init_list); template DDim& operator=(const Dim& in) { @@ -48,8 +48,8 @@ struct DDim { return *this; } - int& operator[](int idx); - int operator[](int idx) const; + int64_t& operator[](int idx); + int64_t operator[](int idx) const; template typename Visitor::result_type apply_visitor(Visitor& visitor) { @@ -71,15 +71,15 @@ struct DDim { DDim operator*(DDim d) const; - ssize_t size() const; + int64_t size() const; }; /** - * \brief Make a DDim from std::vector + * \brief Make a DDim from std::vector * * \param dims An vector of ints. Must be sized between [1, 9] */ -DDim make_ddim(const std::vector& dims); +DDim make_ddim(const std::vector& dims); /** * \brief Make a DDim from an initializer list @@ -87,14 +87,14 @@ DDim make_ddim(const std::vector& dims); * \param dims An initializer list of ints. Must be sized between [1, 9] * */ -DDim make_ddim(std::initializer_list dims); +DDim make_ddim(std::initializer_list dims); -int get(const DDim& dim, int idx); +int64_t get(const DDim& dim, int idx); void set(DDim& dim, int idx, int val); -std::vector vectorize(const DDim& ddim); +std::vector vectorize(const DDim& ddim); -ssize_t product(const DDim& ddim); +int64_t product(const DDim& ddim); /** * \brief Slice a ddim @@ -115,6 +115,12 @@ int arity(const DDim& ddim); std::ostream& operator<<(std::ostream&, const DDim&); +// Reshape a tensor to a matrix. The matrix's first dimension(column length) +// will be the product of tensor's first `num_col_dims` dimensions. +DDim flatten_to_2d(const DDim& src, int num_col_dims); + +DDim flatten_to_1d(const DDim& src); + } // namespace framework } // namespace paddle diff --git a/paddle/framework/ddim_test.cc b/paddle/framework/ddim_test.cc index 9d18a2972c..756232b1b5 100644 --- a/paddle/framework/ddim_test.cc +++ b/paddle/framework/ddim_test.cc @@ -12,7 +12,7 @@ TEST(DDim, Equality) { EXPECT_EQ(ddim[2], 5); // construct a DDim from a vector - std::vector vec({9, 1, 5}); + std::vector vec({9, 1, 5}); paddle::framework::DDim vddim = paddle::framework::make_ddim(vec); EXPECT_EQ(ddim[0], 9); EXPECT_EQ(ddim[1], 1); @@ -25,7 +25,7 @@ TEST(DDim, Equality) { EXPECT_EQ(paddle::framework::get(ddim, 0), 6); // vectorize a DDim - std::vector res_vec = paddle::framework::vectorize(vddim); + std::vector res_vec = paddle::framework::vectorize(vddim); EXPECT_EQ(res_vec[0], 9); EXPECT_EQ(res_vec[1], 1); EXPECT_EQ(res_vec[2], 5); diff --git a/paddle/framework/dim.h b/paddle/framework/dim.h index 883fdc55eb..04d4b0e604 100644 --- a/paddle/framework/dim.h +++ b/paddle/framework/dim.h @@ -17,13 +17,13 @@ struct Dim { static constexpr int dimensions = i; template - HOSTDEVICE Dim(int _head, Args... _tail) : head(_head), tail(_tail...) { + HOSTDEVICE Dim(int64_t _head, Args... _tail) : head(_head), tail(_tail...) { static_assert(sizeof...(_tail) == i - 1, "Dim initialized with the wrong number of parameters"); } HOSTDEVICE - Dim(int _head, const Dim& _tail) : head(_head), tail(_tail) {} + Dim(int64_t _head, const Dim& _tail) : head(_head), tail(_tail) {} HOSTDEVICE Dim() : head(0), tail() {} @@ -31,12 +31,12 @@ struct Dim { /** Construct a Dim from a linear index and size. Uses Fortran order * indexing. */ HOSTDEVICE - Dim(int idx, const Dim& size) + Dim(int64_t idx, const Dim& size) : head(idx % size.head), tail(idx / size.head, size.tail) {} /** Construct a Dim with each dimension set to the given index */ HOSTDEVICE - Dim(int idx) : head(idx), tail(idx) {} + Dim(int64_t idx) : head(idx), tail(idx) {} HOSTDEVICE bool operator==(const Dim& o) const { @@ -47,13 +47,13 @@ struct Dim { bool operator!=(const Dim& o) const { return !(*this == o); } HOSTDEVICE - int& operator[](int idx); + int64_t& operator[](int idx); HOSTDEVICE - int operator[](int idx) const; + int64_t operator[](int idx) const; HOST std::string to_string() const; - int head; + int64_t head; Dim tail; }; @@ -63,7 +63,7 @@ struct Dim<1> { static constexpr int dimensions = 1; HOSTDEVICE - Dim(int _head) : head(_head) {} + Dim(int64_t _head) : head(_head) {} HOSTDEVICE Dim() : head(0) {} @@ -86,11 +86,11 @@ struct Dim<1> { bool operator!=(const Dim<1>& o) const { return !(*this == o); } HOSTDEVICE - int& operator[](int idx); + int64_t& operator[](int idx); HOSTDEVICE - int operator[](int idx) const; + int64_t operator[](int idx) const; - int head; + int64_t head; }; namespace { @@ -100,12 +100,12 @@ template struct DimGetter { // Return a copy if Dim is const template - HOSTDEVICE static int impl(const D& d) { + HOSTDEVICE static int64_t impl(const D& d) { return DimGetter::impl(d.tail); } // Return a reference if Dim is mutable template - HOSTDEVICE static int& impl(D& d) { + HOSTDEVICE static int64_t& impl(D& d) { return DimGetter::impl(d.tail); } }; @@ -115,18 +115,18 @@ template <> struct DimGetter<0> { // Return a copy if Dim is const template - HOSTDEVICE static int impl(const D& d) { + HOSTDEVICE static int64_t impl(const D& d) { return d.head; } // Return a reference if Dim is mutable template - HOSTDEVICE static int& impl(D& d) { + HOSTDEVICE static int64_t& impl(D& d) { return d.head; } }; template -HOSTDEVICE int& indexer(Dim& dim, int idx) { +HOSTDEVICE int64_t& indexer(Dim& dim, int idx) { #ifndef __CUDA_ARCH__ if (idx < 0) { throw std::invalid_argument("Tried to access a negative dimension"); @@ -141,7 +141,7 @@ HOSTDEVICE int& indexer(Dim& dim, int idx) { } template <> -HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) { +HOSTDEVICE int64_t& indexer<1>(Dim<1>& dim, int idx) { #ifndef __CUDA_ARCH__ if (idx != 0) { throw std::invalid_argument("Invalid index"); @@ -153,7 +153,7 @@ HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) { } template -HOSTDEVICE int indexer(const Dim& dim, int idx) { +HOSTDEVICE int64_t indexer(const Dim& dim, int idx) { #ifndef __CUDA_ARCH__ if (idx < 0) { throw std::invalid_argument("Tried to access a negative dimension"); @@ -168,7 +168,7 @@ HOSTDEVICE int indexer(const Dim& dim, int idx) { } template <> -HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) { +HOSTDEVICE int64_t indexer<1>(const Dim<1>& dim, int idx) { #ifndef __CUDA_ARCH__ if (idx != 0) { throw std::invalid_argument("Invalid index"); @@ -182,73 +182,76 @@ HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) { } // namespace // Static access to constant Dim template -HOSTDEVICE int get(const Dim& d) { +HOSTDEVICE int64_t get(const Dim& d) { return DimGetter::impl(d); } // Static access to mutable Dim template -HOSTDEVICE int& get(Dim& d) { +HOSTDEVICE int64_t& get(Dim& d) { return DimGetter::impl(d); } // Dynamic access to constant Dim template -HOSTDEVICE int Dim::operator[](int i) const { +HOSTDEVICE int64_t Dim::operator[](int i) const { return indexer(*this, i); } // Dynamic access to mutable Dim template -HOSTDEVICE int& Dim::operator[](int i) { +HOSTDEVICE int64_t& Dim::operator[](int i) { return indexer(*this, i); } // Dynamic access to constant Dim -inline HOSTDEVICE int Dim<1>::operator[](int i) const { +inline HOSTDEVICE int64_t Dim<1>::operator[](int i) const { return indexer(*this, i); } // Dynamic access to mutable Dim -inline HOSTDEVICE int& Dim<1>::operator[](int i) { return indexer(*this, i); } +inline HOSTDEVICE int64_t& Dim<1>::operator[](int i) { + return indexer(*this, i); +} // Dynamic access to constant Dim // without std::enable_if will try to instantiate this on get<0>(d) template -HOSTDEVICE typename std::enable_if<(l > 0), int>::type get(const Dim& d, - int i) { +HOSTDEVICE typename std::enable_if<(l > 0), int64_t>::type get(const Dim& d, + int i) { return d[i]; } // Dynamic access to mutable Dim template -HOSTDEVICE typename std::enable_if<(l > 0), int&>::type get(Dim& d, int i) { +HOSTDEVICE typename std::enable_if<(l > 0), int64_t&>::type get(Dim& d, + int i) { return d[i]; } // Dot product of two dims template -HOSTDEVICE int linearize(const Dim& a, const Dim& b) { +HOSTDEVICE int64_t linearize(const Dim& a, const Dim& b) { return a.head * b.head + linearize(a.tail, b.tail); } // Base case dot product of two Dims // Notice it is inline because it is no longer a template template <> -HOSTDEVICE inline int linearize(const Dim<1>& a, const Dim<1>& b) { +HOSTDEVICE inline int64_t linearize(const Dim<1>& a, const Dim<1>& b) { return a.head * b.head; } // Product of a Dim template -HOSTDEVICE int product(const Dim& a, int prod = 1) { +HOSTDEVICE int64_t product(const Dim& a, int prod = 1) { return prod * a.head * product(a.tail); } // Base case product of a Dim // Notice it is inline because it is no longer a template template <> -HOSTDEVICE inline int product(const Dim<1>& a, int prod) { +HOSTDEVICE inline int64_t product(const Dim<1>& a, int prod) { return prod * a.head; } diff --git a/paddle/framework/dim_test.cu b/paddle/framework/dim_test.cu index 3898d0a447..0a6a87669c 100644 --- a/paddle/framework/dim_test.cu +++ b/paddle/framework/dim_test.cu @@ -8,7 +8,7 @@ __global__ void test(paddle::framework::Dim<2>* o) { o[0] = paddle::framework::make_dim(5, 6); } -__global__ void dyn_idx_gpu(int* o) { +__global__ void dyn_idx_gpu(int64_t* o) { auto d = paddle::framework::make_dim(5, 6); o[0] = d[1]; } @@ -47,9 +47,9 @@ TEST(Dim, Equality) { EXPECT_EQ(b[1], 11); // dynamic access on GPU - thrust::device_vector r(1); + thrust::device_vector r(1); dyn_idx_gpu<<<1, 1>>>(thrust::raw_pointer_cast(r.data())); - int res = r[0]; + int64_t res = r[0]; EXPECT_EQ(res, 6); // ex_prefix_mul diff --git a/paddle/framework/eigen.h b/paddle/framework/eigen.h index a4667cc51f..54bbeafcab 100644 --- a/paddle/framework/eigen.h +++ b/paddle/framework/eigen.h @@ -28,7 +28,7 @@ struct EigenDim { static Type From(const DDim& dims) { PADDLE_ENFORCE(arity(dims) == D, "D must match arity(DDim)"); Type ret; - for (int d = 0; d < arity(dims); d++) { + for (int64_t d = 0; d < arity(dims); d++) { ret[d] = dims[d]; } return ret; @@ -63,20 +63,35 @@ struct EigenTensor { template -struct EigenMatrix : public EigenTensor {}; +struct EigenMatrix : public EigenTensor { + static typename EigenMatrix::Type Reshape(Tensor& tensor, int num_col_dims) { + int rank = tensor.dims_.size(); + PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank, + "`num_col_dims` must be between (0, rank_of_tensor)."); + return EigenMatrix::From(tensor, + flatten_to_2d(tensor.dims(), num_col_dims)); + } + + static typename EigenMatrix::ConstType Reshape(const Tensor& tensor, + int num_col_dims) { + int rank = tensor.dims_.size(); + PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank, + "`num_col_dims` must be between (0, rank_of_tensor)."); + return EigenMatrix::From(tensor, + flatten_to_2d(tensor.dims(), num_col_dims)); + } +}; template struct EigenVector : public EigenTensor { // Flatten reshapes a Tensor into an EigenVector. static typename EigenVector::Type Flatten(Tensor& tensor) { - return EigenVector::From( - tensor, make_ddim({static_cast(product(tensor.dims_))})); + return EigenVector::From(tensor, {product(tensor.dims_)}); } static typename EigenVector::ConstType Flatten(const Tensor& tensor) { - return EigenVector::From( - tensor, make_ddim({static_cast(product(tensor.dims_))})); + return EigenVector::From(tensor, {product(tensor.dims_)}); } }; diff --git a/paddle/framework/eigen_test.cc b/paddle/framework/eigen_test.cc index dc1957691b..bc4a2db32c 100644 --- a/paddle/framework/eigen_test.cc +++ b/paddle/framework/eigen_test.cc @@ -108,5 +108,24 @@ TEST(Eigen, Matrix) { } } +TEST(Eigen, MatrixReshape) { + Tensor t; + float* p = t.mutable_data({2, 3, 6, 4}, platform::CPUPlace()); + for (int i = 0; i < 2 * 3 * 6 * 4; ++i) { + p[i] = static_cast(i); + } + + EigenMatrix::Type em = EigenMatrix::Reshape(t, 2); + + ASSERT_EQ(2 * 3, em.dimension(0)); + ASSERT_EQ(6 * 4, em.dimension(1)); + + for (int i = 0; i < 2 * 3; i++) { + for (int j = 0; j < 6 * 4; j++) { + ASSERT_NEAR(i * 6 * 4 + j, em(i, j), 1e-6f); + } + } +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/framework.proto b/paddle/framework/framework.proto index ae44a1ffd4..dfcb5fb621 100644 --- a/paddle/framework/framework.proto +++ b/paddle/framework/framework.proto @@ -22,8 +22,14 @@ enum AttrType { INTS = 3; FLOATS = 4; STRINGS = 5; + INT_PAIRS = 6; } +message IntPair { + required int32 first = 1; + required int32 second = 2; +}; + // OpDesc describes an instance of a C++ framework::OperatorBase // derived class type. message OpDesc { @@ -37,6 +43,7 @@ message OpDesc { repeated int32 ints = 6; repeated float floats = 7; repeated string strings = 8; + repeated IntPair int_pairs = 9; }; message Var { @@ -80,3 +87,24 @@ message OpProto { repeated Attr attrs = 4; required string comment = 5; } + +enum DataType { + BOOL = 0; + INT16 = 1; + INT32 = 2; + INT64 = 3; + FP16 = 4; + FP32 = 5; + FP64 = 6; +} + +message LoDTensorDesc { + required DataType data_type = 1; + repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480] + optional int32 lod_level = 3 [ default = 0 ]; +} + +message VarDesc { + required string name = 1; + optional LoDTensorDesc lod_tensor = 2; +} diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc index 902c2655e9..9e3ca563c6 100644 --- a/paddle/framework/grad_op_builder_test.cc +++ b/paddle/framework/grad_op_builder_test.cc @@ -3,7 +3,7 @@ #include "paddle/framework/op_registry.h" #include "paddle/framework/operator.h" -USE_OP(add_two); +USE_OP(add); namespace paddle { namespace framework { @@ -41,7 +41,7 @@ namespace f = paddle::framework; TEST(GradOpBuilder, AddTwo) { std::shared_ptr add_op(f::OpRegistry::CreateOp( - "add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {})); + "add", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {})); std::shared_ptr grad_add_op = f::OpRegistry::CreateGradOp(*add_op); EXPECT_EQ(grad_add_op->Inputs().size(), 4UL); diff --git a/paddle/framework/images/duplicate_op2.graffle b/paddle/framework/images/duplicate_op2.graffle index 2b658085d6..ede3bca30a 100644 Binary files a/paddle/framework/images/duplicate_op2.graffle and b/paddle/framework/images/duplicate_op2.graffle differ diff --git a/paddle/framework/images/duplicate_op2.png b/paddle/framework/images/duplicate_op2.png index c5588015d1..4e872dc2ca 100644 Binary files a/paddle/framework/images/duplicate_op2.png and b/paddle/framework/images/duplicate_op2.png differ diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc index 71eac4a10b..908a1f2fd0 100644 --- a/paddle/framework/lod_tensor.cc +++ b/paddle/framework/lod_tensor.cc @@ -19,8 +19,8 @@ namespace paddle { namespace framework { -LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end) { - LOD new_lod; +LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) { + LoD new_lod; new_lod.reserve(level_end - level_begin); for (size_t i = level_begin; i < level_end; i++) { new_lod.emplace_back(in.at(i)); @@ -28,10 +28,10 @@ LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end) { return new_lod; } -LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin, +LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, size_t elem_end) { // slice the lod. - LOD new_lod; + LoD new_lod; new_lod.reserve(in.size() - level); auto start = in.at(level)[elem_begin]; auto end = in.at(level)[elem_end]; @@ -46,13 +46,13 @@ LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin, std::transform(new_lod.back().begin(), new_lod.back().end(), new_lod.back().begin(), [start](int v) { return v - start; }); - PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LOD"); + PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LoD"); } PADDLE_ENFORCE_LE(new_lod.size(), in.size()); return new_lod; } -bool operator==(const LOD& a, const LOD& b) { +bool operator==(const LoD& a, const LoD& b) { if (a.size() != b.size()) { return false; } @@ -72,12 +72,12 @@ bool operator==(const LOD& a, const LOD& b) { return true; } -void LODTensor::SliceLevels(size_t level_begin, size_t level_end) { +void LoDTensor::SliceLevels(size_t level_begin, size_t level_end) { auto new_lod = framework::SliceLevels(lod_, level_begin, level_end); lod_ = new_lod; } -void LODTensor::SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) { +void LoDTensor::SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) { PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, NumLevels()); PADDLE_ENFORCE(elem_begin < NumElements(level), diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h index 9e6b6b4aca..154068fef6 100644 --- a/paddle/framework/lod_tensor.h +++ b/paddle/framework/lod_tensor.h @@ -35,34 +35,34 @@ template using Vector = thrust::host_vector; #endif -using LOD = std::vector>; +using LoD = std::vector>; -LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end); +LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end); -LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin, +LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin, size_t elem_end); -bool operator==(const LOD& a, const LOD& b); +bool operator==(const LoD& a, const LoD& b); /* - * LODTensor (Level of details Tensor) + * LoDTensor (Level of details Tensor) * see https://en.wikipedia.org/wiki/Level_of_details for reference. */ -class LODTensor { +class LoDTensor { public: - LODTensor() {} - LODTensor(const LOD& lod, Tensor* t) : lod_(lod), tensor_(t) {} + LoDTensor() {} + LoDTensor(const LoD& lod, Tensor* t) : lod_(lod), tensor_(t) {} - void set_lod(const LOD& lod) { lod_ = lod; } + void set_lod(const LoD& lod) { lod_ = lod; } void set_tensor(Tensor* tensor) { tensor_ = tensor; } Tensor& tensor() { return *tensor_; } - LOD lod() { return lod_; } + LoD lod() { return lod_; } /* - * Get a element from LOD. + * Get a element from LoD. */ size_t lod_element(size_t level, size_t elem) const { PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level, @@ -74,7 +74,7 @@ class LODTensor { } /* - * Number of LODTensor's levels, each level has units of data, for example, + * Number of LoDTensor's levels, each level has units of data, for example, * in the sentence's view, article, paragraph, sentence are 3 levels. */ size_t NumLevels() const { return lod_.size(); } @@ -100,7 +100,7 @@ class LODTensor { void SliceInLevel(size_t level, size_t elem_begin, size_t elem_end); private: - LOD lod_; + LoD lod_; Tensor* tensor_; // not owned }; } // namespace framework diff --git a/paddle/framework/lod_tensor.md b/paddle/framework/lod_tensor.md index 8dfe3ee823..769b61f175 100644 --- a/paddle/framework/lod_tensor.md +++ b/paddle/framework/lod_tensor.md @@ -94,7 +94,7 @@ Let's go on slicing this slice. Its <1,1>-slice is ||| ``` -### The General Slicing Algorithm +### The Slicing Algorithm The algorithm, with over-simplified data structure, is defined as @@ -106,17 +106,41 @@ struct LoDTensor { float* tensor_; }; -LoDTensor Slice(const LoDTensor& lodt, int level, int sequence) { +LoDTensor Slice(const LoDTensor& lodt, int level, int sequence); +``` + +Let us revisit the example above -} +``` + 3 +3 1 2 +3 2 4 1 2 3 +||| || |||| | || ||| ``` -### Slicing the Top Level +Suppose that we want to retrieve the <1,2>-slice -Please be aware that an RNN operator only slices the top level of a LoD Tensor to get the step inputs. +``` +2 +2 3 +|| ||| +``` -```c++ -LoDTensor Slice(const LoDTensor& lodt, int sequence) { +we will need to find out the starting position of this slice by summing over all leaf nodes in `LoD` to the left of the slice, i.e., 3 + 2 + 4 + 1 = 10. + +To avoid the traversal of the LoD tree at slcing time, we can do it at the construction time -- instead of saving the lengths of the next level in the LoD tree, we can save the starting offset of the next level. For example, above LoD Tensor can be transformed into + +``` + 0 +0 9 10 +0 3 5 9 10 12 +||| || |||| | || ||| +``` + +We don't really need the 0 on top, so the LoD Tensor could be -} +``` +0 9 10 +0 3 5 9 10 12 +||| || |||| | || ||| ``` diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc index 9a351605ed..1da8553134 100644 --- a/paddle/framework/lod_tensor_test.cc +++ b/paddle/framework/lod_tensor_test.cc @@ -21,7 +21,7 @@ namespace paddle { namespace framework { -class LODTensorTester : public ::testing::Test { +class LoDTensorTester : public ::testing::Test { public: virtual void SetUp() override { // tensor's batch_size: 30 @@ -29,7 +29,7 @@ class LODTensorTester : public ::testing::Test { // 0 10 20 // 0 5 10 15 20 // 0 2 5 7 10 12 15 20 - LOD lod; + LoD lod; lod.push_back(std::vector{0, 10, 20}); lod.push_back(std::vector{0, 5, 10, 15, 20}); lod.push_back(std::vector{0, 2, 5, 7, 10, 12, 15, 17, 20}); @@ -47,21 +47,21 @@ class LODTensorTester : public ::testing::Test { protected: platform::CPUPlace place; Tensor tensor; - LODTensor lod_tensor; + LoDTensor lod_tensor; }; -TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); } +TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); } -TEST_F(LODTensorTester, NumElements) { +TEST_F(LoDTensorTester, NumElements) { ASSERT_EQ(lod_tensor.NumElements(0), 2UL); ASSERT_EQ(lod_tensor.NumElements(1), 4UL); ASSERT_EQ(lod_tensor.NumElements(2), 8UL); } -TEST_F(LODTensorTester, SliceLevels) { +TEST_F(LoDTensorTester, SliceLevels) { // slice 1 level for (size_t level = 0; level < 3UL; ++level) { - LODTensor new_lod_tensor = lod_tensor; + LoDTensor new_lod_tensor = lod_tensor; new_lod_tensor.SliceLevels(level, level + 1); ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL); ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level)); @@ -70,7 +70,7 @@ TEST_F(LODTensorTester, SliceLevels) { } // slice 2 level for (size_t level = 0; level < 2UL; ++level) { - LODTensor new_lod_tensor = lod_tensor; + LoDTensor new_lod_tensor = lod_tensor; new_lod_tensor.SliceLevels(level, level + 2); ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL); ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level)); @@ -80,9 +80,9 @@ TEST_F(LODTensorTester, SliceLevels) { } } -TEST_F(LODTensorTester, SliceInLevel) { +TEST_F(LoDTensorTester, SliceInLevel) { size_t level = 0; - LODTensor new_lod_tensor = lod_tensor; + LoDTensor new_lod_tensor = lod_tensor; new_lod_tensor.SliceInLevel(level, 0, 2); EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL); EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL); diff --git a/paddle/framework/op_registry_test.cc b/paddle/framework/op_registry_test.cc index 50c45919c5..e00c6e8d90 100644 --- a/paddle/framework/op_registry_test.cc +++ b/paddle/framework/op_registry_test.cc @@ -21,7 +21,7 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker { AddOutput("output", "output of cosine op"); AddAttr("scale", "scale of cosine op") .SetDefault(1.0) - .LargerThan(0.0); + .GreaterThan(0.0); AddComment("This is cos op"); } }; @@ -80,7 +80,7 @@ TEST(OpRegistry, CreateOp) { paddle::framework::Scope scope; paddle::platform::CPUDeviceContext dev_ctx; op->Run(scope, dev_ctx); - float scale_get = op->GetAttr("scale"); + float scale_get = op->Attr("scale"); ASSERT_EQ(scale_get, scale); } @@ -121,7 +121,7 @@ TEST(OpRegistry, DefaultValue) { paddle::framework::Scope scope; paddle::platform::CPUDeviceContext dev_ctx; op->Run(scope, dev_ctx); - ASSERT_EQ(op->GetAttr("scale"), 1.0); + ASSERT_EQ(op->Attr("scale"), 1.0); } TEST(OpRegistry, CustomChecker) { @@ -172,38 +172,6 @@ TEST(OpRegistry, CustomChecker) { paddle::platform::CPUDeviceContext dev_ctx; paddle::framework::Scope scope; op->Run(scope, dev_ctx); - int test_attr = op->GetAttr("test_attr"); + int test_attr = op->Attr("test_attr"); ASSERT_EQ(test_attr, 4); -} - -class TestAttrProtoMaker : public pd::OpProtoAndCheckerMaker { - public: - TestAttrProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddAttr("scale", "scale of test op"); - AddAttr("scale", "scale of test op"); - } -}; - -TEST(ProtoMaker, DuplicatedAttr) { - pd::OpProto op_proto; - pd::OpAttrChecker op_checker; - auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker); - ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); -} - -class TestInOutProtoMaker : public pd::OpProtoAndCheckerMaker { - public: - TestInOutProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("input", "input of test op"); - AddInput("input", "input of test op"); - } -}; - -TEST(ProtoMaker, DuplicatedInOut) { - pd::OpProto op_proto; - pd::OpAttrChecker op_checker; - auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker); - ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); -} +} \ No newline at end of file diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index da92220b04..9a98d4d3be 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -69,7 +69,7 @@ class OperatorBase { virtual ~OperatorBase() {} template - inline const T& GetAttr(const std::string& name) const { + inline const T& Attr(const std::string& name) const { PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap", name); return boost::get(attrs_.at(name)); @@ -238,8 +238,8 @@ class InferShapeContext { const Scope& scope() const { return scope_; } template - inline const T& GetAttr(const std::string& name) const { - return op_.GetAttr(name); + inline const T& Attr(const std::string& name) const { + return op_.Attr(name); } size_t InputSize(const std::string& name) const { diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc index f7c9e6b196..20bbb11896 100644 --- a/paddle/framework/operator_test.cc +++ b/paddle/framework/operator_test.cc @@ -102,7 +102,7 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker { AddOutput("y", "output of test op"); AddAttr("scale", "scale of cosine op") .SetDefault(1.0) - .LargerThan(0.0); + .GreaterThan(0.0); AddComment("This is test op"); } }; @@ -140,7 +140,7 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker AddOutput("ys", "outputs of test op").AsDuplicable(); AddAttr("scale", "scale of cosine op") .SetDefault(1.0) - .LargerThan(0.0); + .GreaterThan(0.0); AddComment("This is test op"); } }; @@ -263,4 +263,38 @@ TEST(Operator, Clone) { OperatorClone a("ABC", {}, {}, {}); auto b = a.Clone(); ASSERT_EQ(a.Type(), b->Type()); +} + +class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { + public: + TestAttrProtoMaker(paddle::framework::OpProto* proto, + paddle::framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddAttr("scale", "scale of test op"); + AddAttr("scale", "scale of test op"); + } +}; + +TEST(ProtoMaker, DuplicatedAttr) { + paddle::framework::OpProto op_proto; + paddle::framework::OpAttrChecker op_checker; + auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker); + ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); +} + +class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { + public: + TestInOutProtoMaker(paddle::framework::OpProto* proto, + paddle::framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("input", "input of test op"); + AddInput("input", "input of test op"); + } +}; + +TEST(ProtoMaker, DuplicatedInOut) { + paddle::framework::OpProto op_proto; + paddle::framework::OpAttrChecker op_checker; + auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker); + ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet); } \ No newline at end of file diff --git a/paddle/framework/tensor.h b/paddle/framework/tensor.h index 643f875491..ce938b2143 100644 --- a/paddle/framework/tensor.h +++ b/paddle/framework/tensor.h @@ -43,6 +43,9 @@ class Tensor { template friend struct EigenTensor; + template + friend struct EigenMatrix; + template friend struct EigenVector; diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h index 7893e233b7..637f04ae00 100644 --- a/paddle/framework/tensor_impl.h +++ b/paddle/framework/tensor_impl.h @@ -58,7 +58,7 @@ inline T* Tensor::mutable_data(platform::Place place) { "Tensor's numel must be larger than zero to call " "Tensor::mutable_data. Call Tensor::set_dim first."); /* some versions of boost::variant don't have operator!= */ - size_t size = product(dims_) * sizeof(T); + int64_t size = product(dims_) * sizeof(T); if (holder_ == nullptr || !(holder_->place() == place) || holder_->size() < size + offset_) { if (platform::is_cpu_place(place)) { @@ -131,7 +131,7 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const { PADDLE_ENFORCE_LT(begin_idx, end_idx, "Begin index must be less than end index."); PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1."); - int base = product(dims_) / dims_[0]; + size_t base = product(dims_) / dims_[0]; Tensor dst; dst.holder_ = holder_; DDim dst_dims = dims_; @@ -148,5 +148,13 @@ inline Tensor& Tensor::Resize(const DDim& dims) { inline const DDim& Tensor::dims() const { return dims_; } +template +inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) { + Tensor res; + res.ShareDataWith(src); + res.Resize(flatten_to_2d(src.dims(), num_col_dims)); + return res; +} + } // namespace framework } // namespace paddle diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc index 7db38d5cae..55302ea471 100644 --- a/paddle/framework/tensor_test.cc +++ b/paddle/framework/tensor_test.cc @@ -262,3 +262,16 @@ TEST(Tensor, CopyFrom) { } #endif } + +TEST(Tensor, ReshapeToMatrix) { + using namespace paddle::framework; + using namespace paddle::platform; + Tensor src; + int* src_ptr = src.mutable_data({2, 3, 4, 9}, CPUPlace()); + for (int i = 0; i < 2 * 3 * 4 * 9; ++i) { + src_ptr[i] = i; + } + Tensor res = ReshapeToMatrix(src, 2); + ASSERT_EQ(res.dims()[0], 2 * 3); + ASSERT_EQ(res.dims()[1], 4 * 9); +} \ No newline at end of file diff --git a/paddle/function/CMakeLists.txt b/paddle/function/CMakeLists.txt index f43f15e5ca..4fd72d64a9 100644 --- a/paddle/function/CMakeLists.txt +++ b/paddle/function/CMakeLists.txt @@ -44,6 +44,7 @@ if(WITH_GPU) add_simple_unittest(RowConvOpTest) add_simple_unittest(BlockExpandOpTest) add_simple_unittest(CropOpTest) + add_simple_unittest(SwitchOpTest) endif() add_simple_unittest(Im2ColTest) diff --git a/paddle/function/EigenGemm.cpp b/paddle/function/EigenGemm.cpp index 674141ed39..b3e666e860 100644 --- a/paddle/function/EigenGemm.cpp +++ b/paddle/function/EigenGemm.cpp @@ -83,9 +83,9 @@ struct EigenBlasGemm { }; #ifdef PADDLE_TYPE_DOUBLE -template class EigenBlasGemm; +template struct EigenBlasGemm; #else -template class EigenBlasGemm; +template struct EigenBlasGemm; #endif } // namespace paddle diff --git a/paddle/function/GruFunctor.h b/paddle/function/GruFunctor.h new file mode 100644 index 0000000000..9f6392198e --- /dev/null +++ b/paddle/function/GruFunctor.h @@ -0,0 +1,159 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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 + +#include "GemmFunctor.h" +#include "hl_cpu_gru.cuh" + +namespace paddle { + +template +struct GruFunctor { + template + static void compute(OpResetOutput opResetOutput, + OpFinalOutput opFinalOutput, + hl_gru_value value, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate) { +#ifndef __NVCC__ + if (value.prevOutValue) { + BlasGemm::compute(false, + false, + batchSize, + 2 * frameSize, + frameSize, + 1, + value.prevOutValue, + frameSize, + value.gateWeight, + frameSize * 2, + 1, + value.gateValue, + frameSize * 3); + } + + forward_reset_output( + opResetOutput, value, frameSize, batchSize, active_gate); + + if (value.prevOutValue) { + BlasGemm::compute(false, + false, + batchSize, + frameSize, + frameSize, + 1, + value.resetOutputValue, + frameSize, + value.stateWeight, + frameSize, + 1, + value.gateValue + frameSize * 2, + frameSize * 3); + } + + forward_final_output( + opFinalOutput, value, frameSize, batchSize, active_node); +#endif + } +}; + +template +struct GruGradFunctor { + template + static void compute(OpStateGrad opStateGrad, + OpResetGrad opResetGrad, + hl_gru_value value, + hl_gru_grad grad, + int frameSize, + int batchSize, + hl_activation_mode_t active_node, + hl_activation_mode_t active_gate) { +#ifndef __NVCC__ + backward_state_grad( + opStateGrad, value, grad, frameSize, batchSize, active_node); + + if (value.prevOutValue && grad.prevOutGrad) { + BlasGemm::compute(false, + true, + batchSize, + frameSize, + frameSize, + 1, + grad.gateGrad + frameSize * 2, + frameSize * 3, + value.stateWeight, + frameSize, + 0, + grad.resetOutputGrad, + frameSize); + + if (grad.stateWeightGrad) { + BlasGemm::compute(true, + false, + frameSize, + frameSize, + batchSize, + 1, + value.resetOutputValue, + frameSize, + grad.gateGrad + frameSize * 2, + frameSize * 3, + 1, + grad.stateWeightGrad, + frameSize); + } + } + + backward_reset_grad( + opResetGrad, value, grad, frameSize, batchSize, active_gate); + + if (grad.prevOutGrad && value.prevOutValue) { + BlasGemm::compute(false, + true, + batchSize, + frameSize, + frameSize * 2, + 1, + grad.gateGrad, + frameSize * 3, + value.gateWeight, + frameSize * 2, + 1, + grad.prevOutGrad, + frameSize); + + if (grad.gateWeightGrad) { + BlasGemm::compute(true, + false, + frameSize, + frameSize * 2, + batchSize, + 1, + value.prevOutValue, + frameSize, + grad.gateGrad, + frameSize * 3, + 1, + grad.gateWeightGrad, + frameSize * 2); + } + } +#endif + } +}; + +} // namespace paddle diff --git a/paddle/function/Im2Col.h b/paddle/function/Im2Col.h index 9b91e223a6..1e0cff436f 100644 --- a/paddle/function/Im2Col.h +++ b/paddle/function/Im2Col.h @@ -94,95 +94,4 @@ public: int paddingWidth); }; -template -struct Padding { - static void run(const T* src, - T* dest, - int channels, - int inputHeight, - int inputWidth, - int paddingHeight, - int paddingWidth) { - const int destWidth = inputWidth + 2 * paddingWidth; - for (int c = 0; c < channels; c++) { - if (paddingHeight > 0) { - memset(dest, 0, destWidth * paddingHeight * sizeof(T)); - dest += destWidth * paddingHeight; - } - - for (int i = 0; i < inputHeight; i++) { - // padding head - for (int j = 0; j < paddingWidth; j++) { - *dest++ = T(0); - } - - memcpy(dest, src, inputWidth * sizeof(T)); - dest += inputWidth; - src += inputWidth; - - // padding tail - for (int j = 0; j < paddingWidth; j++) { - *dest++ = T(0); - } - } - - if (paddingHeight > 0) { - memset(dest, 0, destWidth * paddingHeight * sizeof(T)); - dest += destWidth * paddingHeight; - } - } - } -}; - -#if defined(__ARM_NEON__) || defined(__ARM_NEON) -template <> -struct Padding { - static void run(const float* src, - float* dest, - int channels, - int inputHeight, - int inputWidth, - int paddingHeight, - int paddingWidth) { - const int destWidth = inputWidth + 2 * paddingWidth; - for (int c = 0; c < channels; c++) { - if (paddingHeight > 0) { - memset(dest, 0, destWidth * paddingHeight * sizeof(float)); - dest += destWidth * paddingHeight; - } - - for (int i = 0; i < inputHeight; i++) { - // padding head - for (int j = 0; j < paddingWidth; j++) { - *dest++ = float(0); - } - - int step = inputWidth >> 2; - int remain = inputWidth & 3; - for (int s = 0; s < step; s++) { - float32x4_t s0 = vld1q_f32(src); - vst1q_f32(dest, s0); - src += 4; - dest += 4; - } - for (int r = 0; r < remain; r++) { - *dest++ = *src++; - } - - // padding tail - for (int j = 0; j < paddingWidth; j++) { - *dest++ = float(0); - } - } - - if (paddingHeight > 0) { - memset(dest, 0, destWidth * paddingHeight * sizeof(float)); - dest += destWidth * paddingHeight; - } - } - } -}; - -#endif - } // namespace paddle diff --git a/paddle/function/MulOp.cpp b/paddle/function/MulOp.cpp index 91b4b8ed91..25e41edad5 100644 --- a/paddle/function/MulOp.cpp +++ b/paddle/function/MulOp.cpp @@ -13,18 +13,10 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "MulOp.h" -/// todo(tianbing), delete it -#include -#include "paddle/math/MathFunctions.h" +#include "GemmFunctor.h" #include "paddle/math/SIMDFunctions.h" #include "paddle/utils/ThreadLocal.h" -#ifndef PADDLE_TYPE_DOUBLE -#define GEMM paddle::gemm -#else -#define GEMM paddle::gemm -#endif - namespace { inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) { for (unsigned int i = 0; i < len; ++i) { @@ -114,19 +106,20 @@ void MulOp(CpuMatrix& out, real scaleT, bool aTrans, bool bTrans) { - GEMM(aTrans ? CblasTrans : CblasNoTrans, - bTrans ? CblasTrans : CblasNoTrans, - out.getHeight(), - out.getWidth(), - !aTrans ? a.getWidth() : a.getHeight(), - scaleAB, - a.getData(), - a.getStride(), - b.getData(), - b.getStride(), - scaleT, - out.getData(), - out.getStride()); + BlasGemm::compute( + aTrans, + bTrans, + out.getHeight(), + out.getWidth(), + !aTrans ? a.getWidth() : a.getHeight(), + scaleAB, + a.getData(), + a.getStride(), + b.getData(), + b.getStride(), + scaleT, + out.getData(), + out.getStride()); } /// dense matrix (+)= sparse matrix * dense matrix diff --git a/paddle/function/SwitchOp.cpp b/paddle/function/SwitchOp.cpp new file mode 100644 index 0000000000..01e252a8dc --- /dev/null +++ b/paddle/function/SwitchOp.cpp @@ -0,0 +1,140 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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. */ + +#include "SwitchOp.h" +#include "paddle/math/Vector.h" + +namespace paddle { + +template <> +void NCHW2NHWC(real* outputs, + const real* inputs, + const int num, + const int inC, + const int inH, + const int inW, + const int argType) { + for (int n = 0; n < num; ++n) { + for (int c = 0; c < inC; ++c) { + for (int h = 0; h < inH; ++h) { + for (int w = 0; w < inW; ++w) { + if (argType == ADD_TO) { + outputs[((n * inH + h) * inW + w) * inC + c] += *(inputs++); + } else { + outputs[((n * inH + h) * inW + w) * inC + c] = *(inputs++); + } + } + } + } + } +} + +template <> +void NHWC2NCHW(real* outputs, + const real* inputs, + const int num, + const int inH, + const int inW, + const int inC, + const int argType) { + for (int n = 0; n < num; ++n) { + for (int h = 0; h < inH; ++h) { + for (int w = 0; w < inW; ++w) { + for (int c = 0; c < inC; ++c) { + if (argType == ADD_TO) { + outputs[((n * inC + c) * inH + h) * inW + w] += *(inputs++); + } else { + outputs[((n * inC + c) * inH + h) * inW + w] = *(inputs++); + } + } + } + } + } +} + +/** + * \brief Switch dimension order of image input. + * The input and output is a 4D tensor. Switch order + * 'batch_size,channels, height, width' to + * order 'batch_size, height, width, channels'. + * + * Argument in this Function: + * \param inputs input data with order 'batch_size,channels, height, width'. + * \param outputs output data with order 'batch_size, height, width, channels'. + */ +template +class NCHW2NHWCFunc : public FunctionBase { +public: + void init(const FuncConfig& config) override {} + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + + size_t num = inputs[0].shape()[0]; + size_t inC = inputs[0].shape()[1]; + size_t inH = inputs[0].shape()[2]; + size_t inW = inputs[0].shape()[3]; + NCHW2NHWC(outputs[0].data(), + inputs[0].data(), + num, + inC, + inH, + inW, + outputs[0].getArgType()); + } +}; + +/** + * \brief Switch dimension order of image input. + * The input and output is a 4D tensor. Switch order + * 'batch_size, height, width, channels' to + * order 'batch_size, channels, height, width'. + * + * Argument in this Function: + * \param inputs input data with order 'batch_size, height, width, channels'. + * \param outputs output data with order 'batch_size, channels, height, width'. + */ +template +class NHWC2NCHWFunc : public FunctionBase { +public: + void init(const FuncConfig& config) override {} + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(1UL, inputs.size()); + CHECK_EQ(1UL, outputs.size()); + + size_t num = inputs[0].shape()[0]; + size_t inH = inputs[0].shape()[1]; + size_t inW = inputs[0].shape()[2]; + size_t inC = inputs[0].shape()[3]; + + NHWC2NCHW(outputs[0].data(), + inputs[0].data(), + num, + inH, + inW, + inC, + outputs[0].getArgType()); + } +}; + +REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc); +REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc); +#ifndef PADDLE_ONLY_CPU +REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc); +REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc); +#endif + +} // namespace paddle diff --git a/paddle/function/SwitchOp.h b/paddle/function/SwitchOp.h new file mode 100644 index 0000000000..e4c1c3ac92 --- /dev/null +++ b/paddle/function/SwitchOp.h @@ -0,0 +1,66 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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 + +#include "Function.h" + +namespace paddle { + +/** + * \brief This funtion switch dimension order of image input. + * The input and output is a 4D tensor. Switch order 'batch_size, + *channels, height, width' to + * order 'batch_size, height, width, channels'. + * + * \param[out] outputs save results. + * \param[in] inputs input data. + * \param[in] num batch size of input data. + * \param[in] inC channel number of input data. + * \param[in] inH height of input data. + * \param[in] inH with of input data. + * \param[in] argType type of output argument. + */ +template +void NCHW2NHWC(real* outputs, + const real* inputs, + const int num, + const int inC, + const int inH, + const int inW, + const int argtype); + +/** + * \brief This funtion switch dimension order of image input. + * The input and output is a 4D tensor. Switch order 'batch_size, + *height, width, channels' to + * order 'batch_size, channels, height, width'. + * + * \param[out] inGrad gradients of previous layer. + * \param[in] outGrad output gradients. + * \param[in] num batch size of input data. + * \param[in] inH height of input data. + * \param[in] inW with of input data. + * \param[in] inC channel number of input data. + * \param[in] argType type of output argument. + */ +template +void NHWC2NCHW(real* inGrad, + const real* outGrad, + const int num, + const int inH, + const int inW, + const int inC, + const int argType); +} // namespace paddle diff --git a/paddle/function/SwitchOpGpu.cu b/paddle/function/SwitchOpGpu.cu new file mode 100644 index 0000000000..45390a56c3 --- /dev/null +++ b/paddle/function/SwitchOpGpu.cu @@ -0,0 +1,98 @@ +/* Copyright (c) 2016 Paddle + +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. */ + +#include "SwitchOp.h" +#include "hl_base.h" + +namespace paddle { + +__global__ void KeNCHW2NHWC(real* outputs, + const real* inputs, + int inC, + int inH, + int inW, + int nthreads, + int argType) { + const int idx = threadIdx.x + blockIdx.x * blockDim.x; + if (idx < nthreads) { + const int w = idx % inW; + const int h = (idx / inW) % inH; + const int c = (idx / inW / inH) % inC; + const int n = idx / inW / inH / inC; + + const int off = ((n * inH + h) * inW + w) * inC + c; + if (argType == ADD_TO) { + outputs[off] += inputs[idx]; + } else { + outputs[off] = inputs[idx]; + } + } +} + +template <> +void NCHW2NHWC(real* outputs, + const real* inputs, + const int num, + const int inC, + const int inH, + const int inW, + const int argType) { + size_t nth = num * inC * inH * inW; + int blockSize = 1024; + int gridSize = (nth + 1024 - 1) / 1024; + KeNCHW2NHWC<<>>( + outputs, inputs, inC, inH, inW, nth, argType); + CHECK_SYNC("NCHW2NHWC"); +} + +__global__ void KeNHWC2NCHW(real* outputs, + const real* inputs, + int inH, + int inW, + int inC, + int nthreads, + int argType) { + const int idx = threadIdx.x + blockIdx.x * blockDim.x; + if (idx < nthreads) { + const int c = idx % inC; + const int w = (idx / inC) % inW; + const int h = (idx / inC / inW) % inH; + const int n = idx / inW / inH / inC; + + const int off = ((n * inC + c) * inH + h) * inW + w; + if (argType == ADD_TO) { + outputs[off] += inputs[idx]; + } else { + outputs[off] = inputs[idx]; + } + } +} + +template <> +void NHWC2NCHW(real* outputs, + const real* inputs, + const int num, + const int inH, + const int inW, + const int inC, + const int argType) { + int nth = num * inC * inH * inW; + int blockSize = 1024; + int gridSize = (nth + 1024 - 1) / 1024; + KeNHWC2NCHW<<>>( + outputs, inputs, inH, inW, inC, nth, argType); + CHECK_SYNC("NHWC2NCHW"); +} + +} // namespace paddle diff --git a/paddle/function/SwitchOpTest.cpp b/paddle/function/SwitchOpTest.cpp new file mode 100644 index 0000000000..03b0dd66dd --- /dev/null +++ b/paddle/function/SwitchOpTest.cpp @@ -0,0 +1,44 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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. */ + +#include +#include "FunctionTest.h" + +namespace paddle { + +TEST(Pad, real) { + for (size_t numSamples : {1, 4, 8, 16}) { + for (size_t channels : {1, 4, 8, 16}) { + for (size_t imgSizeH : {1, 4, 8, 16}) { + for (size_t imgSizeW : {1, 4, 8, 16}) { + VLOG(3) << " numSamples=" << numSamples << " channels=" << channels + << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW; + for (bool test_grad : {true, false}) { + CpuGpuFuncCompare compare(test_grad ? "NHWC2NCHW" : "NCHW2NHWC", + FuncConfig()); + TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW}; + TensorShape outDims{numSamples, imgSizeH, imgSizeW, channels}; + compare.addInputs( + BufferArg(VALUE_TYPE_FLOAT, test_grad ? outDims : inDims)); + compare.addOutputs(BufferArg( + VALUE_TYPE_FLOAT, test_grad ? inDims : outDims, ASSIGN_TO)); + compare.run(); + } + } + } + } + } +} + +} // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConv.cpp b/paddle/function/neon/NeonDepthwiseConv.cpp index f09e98587d..18126152ea 100644 --- a/paddle/function/neon/NeonDepthwiseConv.cpp +++ b/paddle/function/neon/NeonDepthwiseConv.cpp @@ -12,468 +12,13 @@ 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. */ -#include "neon_util.h" +#include "NeonDepthwiseConv.h" #include "paddle/function/ConvOp.h" -#include "paddle/function/Im2Col.h" namespace paddle { -namespace neon { - #if defined(__ARM_NEON__) || defined(__ARM_NEON) -template -struct DepthwiseConvKernel {}; - -inline float32_t conv3x3(float32x4_t r0, - float32x4_t r1, - float32x4_t r2, - float32x4_t k0, - float32x4_t k1, - float32x4_t k2) { - float32x4_t tmp; - tmp = vmulq_f32(r0, k0); - tmp = vmlaq_f32(tmp, r1, k1); - tmp = vmlaq_f32(tmp, r2, k2); - return vaddvq_f32(tmp); -} - -inline float32_t conv4x4(float32x4_t r0, - float32x4_t r1, - float32x4_t r2, - float32x4_t r3, - float32x4_t k0, - float32x4_t k1, - float32x4_t k2, - float32x4_t k3) { - float32x4_t tmp; - tmp = vmulq_f32(r0, k0); - tmp = vmlaq_f32(tmp, r1, k1); - tmp = vmlaq_f32(tmp, r2, k2); - tmp = vmlaq_f32(tmp, r3, k3); - return vaddvq_f32(tmp); -} - -/** - * Each step calculates four elements of the output. - * First step: - * R0[0, 1, 2, 3...] * K[0][0] - * R0[1, 2, 3, 4...] * K[0][1] - * R0[2, 3, 4, 5...] * K[0][2] - * R1[0, 1, 2, 3...] * K[1][0] - * R1[1, 2, 3, 4...] * K[1][1] - * R1[2, 3, 4, 5...] * K[1][2] - * R2[0, 1, 2, 3...] * K[2][0] - * R2[1, 2, 3, 4...] * K[2][1] - * + R2[2, 3, 4, 5...] * K[2][2] - * ------------------------------ - * Output[0, 1, 2, 3] - */ -template <> -struct DepthwiseConvKernel<3, 1> { - static void run(const float* inputData, - const float* filterData, - int inputHeight, - int inputWidth, - int outputChannels, - int outputHeight, - int outputWidth, - int filterMultiplier, - float* outputData) { - const int steps = outputWidth >> 2; - const int remain = outputWidth & 3; - for (int c = 0; c < outputChannels; c++, filterData += 9) { - // Load the filters - float32x4_t k[3]; - k[0] = vld1q_f32(filterData); - k[1] = vld1q_f32(filterData + 3); - k[2] = vld1q_f32(filterData + 6); - k[0] = vsetq_lane_f32(0.f, k[0], 3); - k[1] = vsetq_lane_f32(0.f, k[1], 3); - k[2] = vsetq_lane_f32(0.f, k[2], 3); - - const float* r0 = - inputData + (c / filterMultiplier) * (inputHeight * inputWidth); - const float* r1 = r0 + inputWidth; - const float* r2 = r0 + inputWidth * 2; - float32x4_t input[3][3]; - for (int h = 0; h < outputHeight; h++) { - for (int s = 0; s < steps; s++) { - // Load the inputs - float32x4_t tmp; - input[0][0] = vld1q_f32(r0); - tmp = vld1q_f32(r0 + 4); - input[0][1] = vextq_f32(input[0][0], tmp, 1); - input[0][2] = vextq_f32(input[0][0], tmp, 2); - input[1][0] = vld1q_f32(r1); - tmp = vld1q_f32(r1 + 4); - input[1][1] = vextq_f32(input[1][0], tmp, 1); - input[1][2] = vextq_f32(input[1][0], tmp, 2); - input[2][0] = vld1q_f32(r2); - tmp = vld1q_f32(r2 + 4); - input[2][1] = vextq_f32(input[2][0], tmp, 1); - input[2][2] = vextq_f32(input[2][0], tmp, 2); - - float32x4_t tmp1 = vdupq_n_f32(0.f); - float32x4_t tmp2 = vdupq_n_f32(0.f); - tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0); - tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1); - tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2); - tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); - tmp1 = vaddq_f32(tmp1, tmp2); - - vst1q_f32(outputData, tmp1); - r0 += 4; - r1 += 4; - r2 += 4; - outputData += 4; - } - - for (int r = 0; r < remain; r++) { - float32x4_t i0 = vld1q_f32(r0); - float32x4_t i1 = vld1q_f32(r1); - float32x4_t i2 = vld1q_f32(r2); - *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); - r0++; - r1++; - r2++; - outputData++; - } - - r0 += 2; - r1 += 2; - r2 += 2; - } - } - } -}; - -/** - * Each step calculates four elements of the output. - * First step: - * R0[0, 2, 4, 6...] * K[0][0] - * R0[1, 3, 5, 7...] * K[0][1] - * R0[2, 4, 6, 8...] * K[0][2] - * R1[0, 2, 4, 6...] * K[1][0] - * R1[1, 3, 5, 7...] * K[1][1] - * R1[2, 4, 6, 8...] * K[1][2] - * R2[0, 2, 4, 6...] * K[2][0] - * R2[1, 3, 5, 7...] * K[2][1] - * R2[2, 4, 6, 8...] * K[2][2] - * ------------------------------ - * Output[0, 1, 2, 3] - */ -template <> -struct DepthwiseConvKernel<3, 2> { - static void run(const float* inputData, - const float* filterData, - int inputHeight, - int inputWidth, - int outputChannels, - int outputHeight, - int outputWidth, - int filterMultiplier, - float* outputData) { - const int steps = outputWidth >> 2; - const int remain = outputWidth & 3; - for (int c = 0; c < outputChannels; c++, filterData += 9) { - // Load the filters - float32x4_t k[3]; - k[0] = vld1q_f32(filterData); - k[1] = vld1q_f32(filterData + 3); - k[2] = vld1q_f32(filterData + 6); - k[0] = vsetq_lane_f32(0.f, k[0], 3); - k[1] = vsetq_lane_f32(0.f, k[1], 3); - k[2] = vsetq_lane_f32(0.f, k[2], 3); - - const float* start = - inputData + (c / filterMultiplier) * (inputHeight * inputWidth); - float32x4_t input[3][3]; - for (int h = 0; h < outputHeight; h++) { - const float* r0 = start + 2 * h * inputWidth; - const float* r1 = start + (2 * h + 1) * inputWidth; - const float* r2 = start + (2 * h + 2) * inputWidth; - for (int s = 0; s < steps; s++) { - // Load the inputs - float32x4_t data1; - float32x4x2_t data2; - - data2 = vld2q_f32(r0); - input[0][0] = data2.val[0]; - input[0][1] = data2.val[1]; - data1 = vld1q_f32(r0 + 8); - input[0][2] = vextq_f32(data2.val[0], data1, 1); - - data2 = vld2q_f32(r1); - input[1][0] = data2.val[0]; - input[1][1] = data2.val[1]; - data1 = vld1q_f32(r1 + 8); - input[1][2] = vextq_f32(data2.val[0], data1, 1); - - data2 = vld2q_f32(r2); - input[2][0] = data2.val[0]; - input[2][1] = data2.val[1]; - data1 = vld1q_f32(r2 + 8); - input[2][2] = vextq_f32(data2.val[0], data1, 1); - - float32x4_t tmp1 = vdupq_n_f32(0.f); - float32x4_t tmp2 = vdupq_n_f32(0.f); - tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0); - tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1); - tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2); - tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); - tmp1 = vaddq_f32(tmp1, tmp2); - - vst1q_f32(outputData, tmp1); - r0 += 8; - r1 += 8; - r2 += 8; - outputData += 4; - } - - for (int r = 0; r < remain; r++) { - float32x4_t i0 = vld1q_f32(r0); - float32x4_t i1 = vld1q_f32(r1); - float32x4_t i2 = vld1q_f32(r2); - *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); - r0 += 2; - r1 += 2; - r2 += 2; - outputData++; - } - } - } - } -}; - -/** - * Each step calculates four elements of the output. - */ -template <> -struct DepthwiseConvKernel<4, 1> { - static void run(const float* inputData, - const float* filterData, - int inputHeight, - int inputWidth, - int outputChannels, - int outputHeight, - int outputWidth, - int filterMultiplier, - float* outputData) { - const int steps = outputWidth >> 2; - const int remain = outputWidth & 3; - for (int c = 0; c < outputChannels; c++, filterData += 16) { - // Load the filters - float32x4_t k[4]; - k[0] = vld1q_f32(filterData); - k[1] = vld1q_f32(filterData + 4); - k[2] = vld1q_f32(filterData + 8); - k[3] = vld1q_f32(filterData + 12); - - const float* r0 = - inputData + (c / filterMultiplier) * (inputHeight * inputWidth); - const float* r1 = r0 + inputWidth; - const float* r2 = r0 + inputWidth * 2; - const float* r3 = r0 + inputWidth * 3; - float32x4_t input[4][4]; - for (int h = 0; h < outputHeight; h++) { - for (int s = 0; s < steps; s++) { - // Load the inputs - float32x4_t tmp; - input[0][0] = vld1q_f32(r0); - tmp = vld1q_f32(r0 + 4); - input[0][1] = vextq_f32(input[0][0], tmp, 1); - input[0][2] = vextq_f32(input[0][0], tmp, 2); - input[0][3] = vextq_f32(input[0][0], tmp, 3); - - input[1][0] = vld1q_f32(r1); - tmp = vld1q_f32(r1 + 4); - input[1][1] = vextq_f32(input[1][0], tmp, 1); - input[1][2] = vextq_f32(input[1][0], tmp, 2); - input[1][3] = vextq_f32(input[1][0], tmp, 3); - - input[2][0] = vld1q_f32(r2); - tmp = vld1q_f32(r2 + 4); - input[2][1] = vextq_f32(input[2][0], tmp, 1); - input[2][2] = vextq_f32(input[2][0], tmp, 2); - input[2][3] = vextq_f32(input[2][0], tmp, 3); - - input[3][0] = vld1q_f32(r3); - tmp = vld1q_f32(r3 + 4); - input[3][1] = vextq_f32(input[3][0], tmp, 1); - input[3][2] = vextq_f32(input[3][0], tmp, 2); - input[3][3] = vextq_f32(input[3][0], tmp, 3); - - float32x4_t tmp1 = vdupq_n_f32(0.f); - float32x4_t tmp2 = vdupq_n_f32(0.f); - tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3); - tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3); - tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3); - tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3); - tmp1 = vaddq_f32(tmp1, tmp2); - - vst1q_f32(outputData, tmp1); - r0 += 4; - r1 += 4; - r2 += 4; - r3 += 4; - outputData += 4; - } - - for (int r = 0; r < remain; r++) { - float32x4_t i0 = vld1q_f32(r0); - float32x4_t i1 = vld1q_f32(r1); - float32x4_t i2 = vld1q_f32(r2); - float32x4_t i3 = vld1q_f32(r3); - *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]); - r0++; - r1++; - r2++; - r3++; - outputData++; - } - - r0 += 3; - r1 += 3; - r2 += 3; - r3 += 3; - } - } - } -}; - -/** - * Each step calculates four elements of the output. - */ -template <> -struct DepthwiseConvKernel<4, 2> { - static void run(const float* inputData, - const float* filterData, - int inputHeight, - int inputWidth, - int outputChannels, - int outputHeight, - int outputWidth, - int filterMultiplier, - float* outputData) { - const int steps = outputWidth >> 2; - const int remain = outputWidth & 3; - for (int c = 0; c < outputChannels; c++, filterData += 16) { - // Load the filters - float32x4_t k[4]; - k[0] = vld1q_f32(filterData); - k[1] = vld1q_f32(filterData + 4); - k[2] = vld1q_f32(filterData + 8); - k[3] = vld1q_f32(filterData + 12); - - const float* start = - inputData + (c / filterMultiplier) * (inputHeight * inputWidth); - float32x4_t input[4][4]; - for (int h = 0; h < outputHeight; h++) { - const float* r0 = start + 2 * h * inputWidth; - const float* r1 = start + (2 * h + 1) * inputWidth; - const float* r2 = start + (2 * h + 2) * inputWidth; - const float* r3 = start + (2 * h + 3) * inputWidth; - for (int s = 0; s < steps; s++) { - // Load the inputs - float32x4x2_t data1; - float32x4x2_t data2; - - data1 = vld2q_f32(r0); - data2 = vld2q_f32(r0 + 8); - input[0][0] = data1.val[0]; - input[0][1] = data1.val[1]; - input[0][2] = vextq_f32(data1.val[0], data2.val[0], 1); - input[0][3] = vextq_f32(data1.val[1], data2.val[1], 1); - - data1 = vld2q_f32(r1); - data2 = vld2q_f32(r1 + 8); - input[1][0] = data1.val[0]; - input[1][1] = data1.val[1]; - input[1][2] = vextq_f32(data1.val[0], data2.val[0], 1); - input[1][3] = vextq_f32(data1.val[1], data2.val[1], 1); - - data1 = vld2q_f32(r2); - data2 = vld2q_f32(r2 + 8); - input[2][0] = data1.val[0]; - input[2][1] = data1.val[1]; - input[2][2] = vextq_f32(data1.val[0], data2.val[0], 1); - input[2][3] = vextq_f32(data1.val[1], data2.val[1], 1); - - data1 = vld2q_f32(r3); - data2 = vld2q_f32(r3 + 8); - input[3][0] = data1.val[0]; - input[3][1] = data1.val[1]; - input[3][2] = vextq_f32(data1.val[0], data2.val[0], 1); - input[3][3] = vextq_f32(data1.val[1], data2.val[1], 1); - - float32x4_t tmp1 = vdupq_n_f32(0.f); - float32x4_t tmp2 = vdupq_n_f32(0.f); - tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3); - tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3); - tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3); - tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0); - tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1); - tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2); - tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3); - tmp1 = vaddq_f32(tmp1, tmp2); - - vst1q_f32(outputData, tmp1); - r0 += 8; - r1 += 8; - r2 += 8; - r3 += 8; - outputData += 4; - } - - for (int r = 0; r < remain; r++) { - float32x4_t i0 = vld1q_f32(r0); - float32x4_t i1 = vld1q_f32(r1); - float32x4_t i2 = vld1q_f32(r2); - float32x4_t i3 = vld1q_f32(r3); - *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]); - r0 += 2; - r1 += 2; - r2 += 2; - r3 += 2; - outputData++; - } - } - } - } -}; - template class NeonDepthwiseConvFunction : public ConvFunctionBase { public: @@ -497,16 +42,16 @@ public: const TensorShape& filter = inputs[1].shape(); const TensorShape& output = outputs[0].shape(); - size_t batchSize = input[0]; - size_t inputChannels = input[1]; - size_t inputHeight = input[2]; - size_t inputWidth = input[3]; - size_t filterHeight = getFilterHeight(filter); - size_t filterWidth = getFilterWidth(filter); - size_t outputChannels = output[1]; - size_t outputHeight = output[2]; - size_t outputWidth = output[3]; - size_t filterMultiplier = outputChannels / groups_; + int batchSize = input[0]; + int inputChannels = input[1]; + int inputHeight = input[2]; + int inputWidth = input[3]; + int filterHeight = getFilterHeight(filter); + int filterWidth = getFilterWidth(filter); + int outputChannels = output[1]; + int outputHeight = output[2]; + int outputWidth = output[3]; + int filterMultiplier = outputChannels / groups_; CHECK_EQ(inputChannels, groups_); // only support strideH() == strideW() and filterHeight == filterWidth. @@ -519,22 +64,19 @@ public: // padding the input float* inputPadding = inputData; + int padInputHeight = inputHeight + 2 * paddingH(); + int padInputWidth = inputWidth + 2 * paddingW(); if (paddingH() > 0 || paddingW() > 0) { - int newSize = batchSize * inputChannels * (inputHeight + 2 * paddingH()) * - (inputWidth + 2 * paddingW()); + int newSize = batchSize * inputChannels * padInputHeight * padInputWidth; resizeBuffer(newSize); inputPadding = reinterpret_cast(memory_->getBuf()); - Padding::run(inputData, - inputPadding, - batchSize * inputChannels, - inputHeight, - inputWidth, - paddingH(), - paddingW()); - - // height and width of padding data - inputHeight += 2 * paddingH(); - inputWidth += 2 * paddingW(); + neon::Padding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); } std::function::run; + DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run; } else if (filterWidth == 3 && strideW() == 2) { - DepthWiseConv = DepthwiseConvKernel<3, 2>::run; + DepthWiseConv = neon::DepthwiseConvKernel<3, 2>::run; } else if (filterWidth == 4 && strideW() == 1) { - DepthWiseConv = DepthwiseConvKernel<4, 1>::run; + DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run; } else if (filterWidth == 4 && strideW() == 2) { - DepthWiseConv = DepthwiseConvKernel<4, 2>::run; + DepthWiseConv = neon::DepthwiseConvKernel<4, 2>::run; } else { LOG(FATAL) << "Not supported"; } - for (size_t i = 0; i < batchSize; i++) { + for (int i = 0; i < batchSize; i++) { DepthWiseConv(inputPadding, filterData, - inputHeight, - inputWidth, + padInputHeight, + padInputWidth, outputChannels, outputHeight, outputWidth, filterMultiplier, outputData); - inputPadding += inputChannels * inputHeight * inputWidth; + inputPadding += inputChannels * padInputHeight * padInputWidth; outputData += outputChannels * outputHeight * outputWidth; } } }; +#ifndef PADDLE_TYPE_DOUBLE REGISTER_TYPED_FUNC(NeonDepthwiseConv, CPU, NeonDepthwiseConvFunction); +#endif #endif -} // namespace neon } // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConv.h b/paddle/function/neon/NeonDepthwiseConv.h new file mode 100644 index 0000000000..aefeea78ba --- /dev/null +++ b/paddle/function/neon/NeonDepthwiseConv.h @@ -0,0 +1,631 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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 + +#include +#include "neon_util.h" + +namespace paddle { + +namespace neon { + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +template +struct DepthwiseConvKernel {}; + +inline float32_t conv3x3(float32x4_t r0, + float32x4_t r1, + float32x4_t r2, + float32x4_t k0, + float32x4_t k1, + float32x4_t k2) { + float32x4_t tmp; + tmp = vmulq_f32(r0, k0); + tmp = vmlaq_f32(tmp, r1, k1); + tmp = vmlaq_f32(tmp, r2, k2); + return vaddvq_f32(tmp); +} + +inline float32_t conv4x4(float32x4_t r0, + float32x4_t r1, + float32x4_t r2, + float32x4_t r3, + float32x4_t k0, + float32x4_t k1, + float32x4_t k2, + float32x4_t k3) { + float32x4_t tmp; + tmp = vmulq_f32(r0, k0); + tmp = vmlaq_f32(tmp, r1, k1); + tmp = vmlaq_f32(tmp, r2, k2); + tmp = vmlaq_f32(tmp, r3, k3); + return vaddvq_f32(tmp); +} + +/** + * Each step calculates four elements of the output. + * First step: + * R0[0, 1, 2, 3...] * K[0][0] + * R0[1, 2, 3, 4...] * K[0][1] + * R0[2, 3, 4, 5...] * K[0][2] + * R1[0, 1, 2, 3...] * K[1][0] + * R1[1, 2, 3, 4...] * K[1][1] + * R1[2, 3, 4, 5...] * K[1][2] + * R2[0, 1, 2, 3...] * K[2][0] + * R2[1, 2, 3, 4...] * K[2][1] + * + R2[2, 3, 4, 5...] * K[2][2] + * ------------------------------ + * Output[0, 1, 2, 3] + */ +template <> +struct DepthwiseConvKernel<3, 1> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 9) { + // Load the filters + float32x4_t k[3]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 3); + k[2] = vld1q_f32(filterData + 6); + k[0] = vsetq_lane_f32(0.f, k[0], 3); + k[1] = vsetq_lane_f32(0.f, k[1], 3); + k[2] = vsetq_lane_f32(0.f, k[2], 3); + + const float* r0 = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + const float* r1 = r0 + inputWidth; + const float* r2 = r0 + inputWidth * 2; + float32x4_t input[3][3]; + for (int h = 0; h < outputHeight; h++) { + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4_t tmp; + input[0][0] = vld1q_f32(r0); + tmp = vld1q_f32(r0 + 4); + input[0][1] = vextq_f32(input[0][0], tmp, 1); + input[0][2] = vextq_f32(input[0][0], tmp, 2); + input[1][0] = vld1q_f32(r1); + tmp = vld1q_f32(r1 + 4); + input[1][1] = vextq_f32(input[1][0], tmp, 1); + input[1][2] = vextq_f32(input[1][0], tmp, 2); + input[2][0] = vld1q_f32(r2); + tmp = vld1q_f32(r2 + 4); + input[2][1] = vextq_f32(input[2][0], tmp, 1); + input[2][2] = vextq_f32(input[2][0], tmp, 2); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 4; + r1 += 4; + r2 += 4; + outputData += 4; + } + + for (int r = 0; r < remain; r++) { + float32x4_t i0 = vld1q_f32(r0); + float32x4_t i1 = vld1q_f32(r1); + float32x4_t i2 = vld1q_f32(r2); + *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); + r0++; + r1++; + r2++; + outputData++; + } + + r0 += 2; + r1 += 2; + r2 += 2; + } + } + } +}; + +/** + * Each step calculates four elements of the output. + * First step: + * R0[0, 2, 4, 6...] * K[0][0] + * R0[1, 3, 5, 7...] * K[0][1] + * R0[2, 4, 6, 8...] * K[0][2] + * R1[0, 2, 4, 6...] * K[1][0] + * R1[1, 3, 5, 7...] * K[1][1] + * R1[2, 4, 6, 8...] * K[1][2] + * R2[0, 2, 4, 6...] * K[2][0] + * R2[1, 3, 5, 7...] * K[2][1] + * R2[2, 4, 6, 8...] * K[2][2] + * ------------------------------ + * Output[0, 1, 2, 3] + */ +template <> +struct DepthwiseConvKernel<3, 2> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 9) { + // Load the filters + float32x4_t k[3]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 3); + k[2] = vld1q_f32(filterData + 6); + k[0] = vsetq_lane_f32(0.f, k[0], 3); + k[1] = vsetq_lane_f32(0.f, k[1], 3); + k[2] = vsetq_lane_f32(0.f, k[2], 3); + + const float* start = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + float32x4_t input[3][3]; + for (int h = 0; h < outputHeight; h++) { + const float* r0 = start + 2 * h * inputWidth; + const float* r1 = start + (2 * h + 1) * inputWidth; + const float* r2 = start + (2 * h + 2) * inputWidth; + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4_t data1; + float32x4x2_t data2; + + data2 = vld2q_f32(r0); + input[0][0] = data2.val[0]; + input[0][1] = data2.val[1]; + data1 = vld1q_f32(r0 + 8); + input[0][2] = vextq_f32(data2.val[0], data1, 1); + + data2 = vld2q_f32(r1); + input[1][0] = data2.val[0]; + input[1][1] = data2.val[1]; + data1 = vld1q_f32(r1 + 8); + input[1][2] = vextq_f32(data2.val[0], data1, 1); + + data2 = vld2q_f32(r2); + input[2][0] = data2.val[0]; + input[2][1] = data2.val[1]; + data1 = vld1q_f32(r2 + 8); + input[2][2] = vextq_f32(data2.val[0], data1, 1); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][0], k[1], 0); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][1], k[1], 1); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][2], k[1], 2); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 8; + r1 += 8; + r2 += 8; + outputData += 4; + } + + for (int r = 0; r < remain; r++) { + float32x4_t i0 = vld1q_f32(r0); + float32x4_t i1 = vld1q_f32(r1); + float32x4_t i2 = vld1q_f32(r2); + *outputData = conv3x3(i0, i1, i2, k[0], k[1], k[2]); + r0 += 2; + r1 += 2; + r2 += 2; + outputData++; + } + } + } + } +}; + +/** + * Each step calculates four elements of the output. + */ +template <> +struct DepthwiseConvKernel<4, 1> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 16) { + // Load the filters + float32x4_t k[4]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 4); + k[2] = vld1q_f32(filterData + 8); + k[3] = vld1q_f32(filterData + 12); + + const float* r0 = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + const float* r1 = r0 + inputWidth; + const float* r2 = r0 + inputWidth * 2; + const float* r3 = r0 + inputWidth * 3; + float32x4_t input[4][4]; + for (int h = 0; h < outputHeight; h++) { + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4_t tmp; + input[0][0] = vld1q_f32(r0); + tmp = vld1q_f32(r0 + 4); + input[0][1] = vextq_f32(input[0][0], tmp, 1); + input[0][2] = vextq_f32(input[0][0], tmp, 2); + input[0][3] = vextq_f32(input[0][0], tmp, 3); + + input[1][0] = vld1q_f32(r1); + tmp = vld1q_f32(r1 + 4); + input[1][1] = vextq_f32(input[1][0], tmp, 1); + input[1][2] = vextq_f32(input[1][0], tmp, 2); + input[1][3] = vextq_f32(input[1][0], tmp, 3); + + input[2][0] = vld1q_f32(r2); + tmp = vld1q_f32(r2 + 4); + input[2][1] = vextq_f32(input[2][0], tmp, 1); + input[2][2] = vextq_f32(input[2][0], tmp, 2); + input[2][3] = vextq_f32(input[2][0], tmp, 3); + + input[3][0] = vld1q_f32(r3); + tmp = vld1q_f32(r3 + 4); + input[3][1] = vextq_f32(input[3][0], tmp, 1); + input[3][2] = vextq_f32(input[3][0], tmp, 2); + input[3][3] = vextq_f32(input[3][0], tmp, 3); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 4; + r1 += 4; + r2 += 4; + r3 += 4; + outputData += 4; + } + + for (int r = 0; r < remain; r++) { + float32x4_t i0 = vld1q_f32(r0); + float32x4_t i1 = vld1q_f32(r1); + float32x4_t i2 = vld1q_f32(r2); + float32x4_t i3 = vld1q_f32(r3); + *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]); + r0++; + r1++; + r2++; + r3++; + outputData++; + } + + r0 += 3; + r1 += 3; + r2 += 3; + r3 += 3; + } + } + } +}; + +/** + * Each step calculates four elements of the output. + */ +template <> +struct DepthwiseConvKernel<4, 2> { + static void run(const float* inputData, + const float* filterData, + int inputHeight, + int inputWidth, + int outputChannels, + int outputHeight, + int outputWidth, + int filterMultiplier, + float* outputData) { + const int steps = outputWidth >> 2; + const int remain = outputWidth & 3; + for (int c = 0; c < outputChannels; c++, filterData += 16) { + // Load the filters + float32x4_t k[4]; + k[0] = vld1q_f32(filterData); + k[1] = vld1q_f32(filterData + 4); + k[2] = vld1q_f32(filterData + 8); + k[3] = vld1q_f32(filterData + 12); + + const float* start = + inputData + (c / filterMultiplier) * (inputHeight * inputWidth); + float32x4_t input[4][4]; + for (int h = 0; h < outputHeight; h++) { + const float* r0 = start + 2 * h * inputWidth; + const float* r1 = start + (2 * h + 1) * inputWidth; + const float* r2 = start + (2 * h + 2) * inputWidth; + const float* r3 = start + (2 * h + 3) * inputWidth; + for (int s = 0; s < steps; s++) { + // Load the inputs + float32x4x2_t data1; + float32x4x2_t data2; + + data1 = vld2q_f32(r0); + data2 = vld2q_f32(r0 + 8); + input[0][0] = data1.val[0]; + input[0][1] = data1.val[1]; + input[0][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[0][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + data1 = vld2q_f32(r1); + data2 = vld2q_f32(r1 + 8); + input[1][0] = data1.val[0]; + input[1][1] = data1.val[1]; + input[1][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[1][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + data1 = vld2q_f32(r2); + data2 = vld2q_f32(r2 + 8); + input[2][0] = data1.val[0]; + input[2][1] = data1.val[1]; + input[2][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[2][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + data1 = vld2q_f32(r3); + data2 = vld2q_f32(r3 + 8); + input[3][0] = data1.val[0]; + input[3][1] = data1.val[1]; + input[3][2] = vextq_f32(data1.val[0], data2.val[0], 1); + input[3][3] = vextq_f32(data1.val[1], data2.val[1], 1); + + float32x4_t tmp1 = vdupq_n_f32(0.f); + float32x4_t tmp2 = vdupq_n_f32(0.f); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][0], k[0], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][1], k[0], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[0][2], k[0], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[0][3], k[0], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][0], k[1], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][1], k[1], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[1][2], k[1], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[1][3], k[1], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][0], k[2], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][1], k[2], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[2][2], k[2], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[2][3], k[2], 3); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][0], k[3], 0); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][1], k[3], 1); + tmp1 = vmlaq_laneq_f32(tmp1, input[3][2], k[3], 2); + tmp2 = vmlaq_laneq_f32(tmp2, input[3][3], k[3], 3); + tmp1 = vaddq_f32(tmp1, tmp2); + + vst1q_f32(outputData, tmp1); + r0 += 8; + r1 += 8; + r2 += 8; + r3 += 8; + outputData += 4; + } + + for (int r = 0; r < remain; r++) { + float32x4_t i0 = vld1q_f32(r0); + float32x4_t i1 = vld1q_f32(r1); + float32x4_t i2 = vld1q_f32(r2); + float32x4_t i3 = vld1q_f32(r3); + *outputData = conv4x4(i0, i1, i2, i3, k[0], k[1], k[2], k[3]); + r0 += 2; + r1 += 2; + r2 += 2; + r3 += 2; + outputData++; + } + } + } + } +}; + +template +struct Padding { + static void run(const T* input, + T* inputPadding, + int channels, + int inputHeight, + int inputWidth, + int padInputHeight, + int padInputWidth) { + const int paddingHeight = (padInputHeight - inputHeight) / 2; + const int paddingWidth = (padInputWidth - inputWidth) / 2; + for (int c = 0; c < channels; c++) { + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(T)); + inputPadding += padInputWidth * paddingHeight; + } + + for (int i = 0; i < inputHeight; i++) { + // padding head + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = T(0); + } + + memcpy(inputPadding, input, inputWidth * sizeof(T)); + inputPadding += inputWidth; + input += inputWidth; + + // padding tail + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = T(0); + } + } + + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(T)); + inputPadding += padInputWidth * paddingHeight; + } + } + } +}; + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) +template <> +struct Padding { + static void run(const float* input, + float* inputPadding, + int channels, + int inputHeight, + int inputWidth, + int padInputHeight, + int padInputWidth) { + const int paddingHeight = (padInputHeight - inputHeight) / 2; + const int paddingWidth = (padInputWidth - inputWidth) / 2; + for (int c = 0; c < channels; c++) { + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(float)); + inputPadding += padInputWidth * paddingHeight; + } + + for (int i = 0; i < inputHeight; i++) { + // padding head + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = float(0); + } + + int step = inputWidth >> 2; + int remain = inputWidth & 3; + for (int s = 0; s < step; s++) { + float32x4_t s0 = vld1q_f32(input); + vst1q_f32(inputPadding, s0); + input += 4; + inputPadding += 4; + } + for (int r = 0; r < remain; r++) { + *inputPadding++ = *input++; + } + + // padding tail + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = float(0); + } + } + + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(float)); + inputPadding += padInputWidth * paddingHeight; + } + } + } +}; + +// for stride is 2 +struct StridePadding { + static void run(const float* input, + float* inputPadding, + int channels, + int inputHeight, + int inputWidth, + int padInputHeight, + int padInputWidth) { + const int paddingHeight = (padInputHeight - (inputHeight * 2 - 1)) / 2; + const int paddingWidth = (padInputWidth - (inputWidth * 2 - 1)) / 2; + for (int c = 0; c < channels; c++) { + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(float)); + inputPadding += padInputWidth * paddingHeight; + } + + for (int i = 0; i < inputHeight; i++) { + // padding head + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = float(0); + } + + int step = inputWidth >> 2; + int remain = inputWidth & 3; + float32x4_t s1 = vdupq_n_f32(0.f); + for (int s = 0; s < step; s++) { + float32x4_t s0 = vld1q_f32(input); + float32x4x2_t v = {s0, s1}; + vst2q_f32(inputPadding, v); + input += 4; + inputPadding += 8; + } + for (int r = 0; r < remain; r++) { + *inputPadding++ = *input++; + *inputPadding++ = float(0); + } + inputPadding--; + + // padding tail + for (int j = 0; j < paddingWidth; j++) { + *inputPadding++ = float(0); + } + if (i != inputHeight - 1) { + memset(inputPadding, 0, padInputWidth * sizeof(float)); + inputPadding += padInputWidth; + } + } + + if (paddingHeight > 0) { + memset(inputPadding, 0, padInputWidth * paddingHeight * sizeof(float)); + inputPadding += padInputWidth * paddingHeight; + } + } + } +}; + +#endif + +#endif + +} // namespace neon +} // namespace paddle diff --git a/paddle/function/neon/NeonDepthwiseConvTranspose.cpp b/paddle/function/neon/NeonDepthwiseConvTranspose.cpp new file mode 100644 index 0000000000..49ca4bc8a0 --- /dev/null +++ b/paddle/function/neon/NeonDepthwiseConvTranspose.cpp @@ -0,0 +1,136 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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. */ + +#include "NeonDepthwiseConv.h" +#include "paddle/function/ConvOp.h" + +namespace paddle { + +#if defined(__ARM_NEON__) || defined(__ARM_NEON) + +template +class NeonDepthwiseConvTransposeFunction : public ConvFunctionBase { +public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + } + + void check(const BufferArgs& inputs, const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + check(inputs, outputs); + + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + int batchSize = input[0]; + int inputChannels = input[1]; + int inputHeight = input[2]; + int inputWidth = input[3]; + int filterHeight = getFilterHeight(filter); + int filterWidth = getFilterWidth(filter); + int outputChannels = output[1]; + int outputHeight = output[2]; + int outputWidth = output[3]; + int filterMultiplier = outputChannels / groups_; + CHECK_EQ(inputChannels, groups_); + + // only support strideH() == strideW() and filterHeight == filterWidth. + CHECK_EQ(strideH(), strideW()); + CHECK_EQ(paddingH(), paddingW()); + CHECK_EQ(filterHeight, filterWidth); + + float* inputData = inputs[0].data(); + float* filterData = inputs[1].data(); + float* outputData = outputs[0].data(); + + // padding the input, input -> inputPadding + float* inputPadding = inputData; + int padInputHeight = + (inputHeight - 1) * strideH() + 2 * filterHeight - 1 - 2 * paddingH(); + int padInputWidth = + (inputWidth - 1) * strideW() + 2 * filterWidth - 1 - 2 * paddingW(); + + if (padInputHeight > inputHeight || padInputWidth > inputWidth) { + int newSize = batchSize * inputChannels * padInputHeight * padInputWidth; + resizeBuffer(newSize); + inputPadding = reinterpret_cast(memory_->getBuf()); + if (strideH() == 1) { + neon::Padding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); + } else if (strideH() == 2) { + neon::StridePadding::run(inputData, + inputPadding, + batchSize * inputChannels, + inputHeight, + inputWidth, + padInputHeight, + padInputWidth); + } else { + LOG(FATAL) << "Not supported"; + } + } + + std::function + DepthWiseConv; + + if (filterWidth == 3) { + DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run; + } else if (filterWidth == 4) { + DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run; + } else { + LOG(FATAL) << "Not supported"; + } + + for (int i = 0; i < batchSize; i++) { + DepthWiseConv(inputPadding, + filterData, + padInputHeight, + padInputWidth, + outputChannels, + outputHeight, + outputWidth, + filterMultiplier, + outputData); + inputPadding += inputChannels * padInputHeight * padInputWidth; + outputData += outputChannels * outputHeight * outputWidth; + } + } +}; + +#ifndef PADDLE_TYPE_DOUBLE + +REGISTER_TYPED_FUNC(NeonDepthwiseConvTranspose, + CPU, + NeonDepthwiseConvTransposeFunction); + +#endif + +#endif + +} // namespace paddle diff --git a/paddle/function/neon/neon_util.h b/paddle/function/neon/neon_util.h index 56b3febe2d..e2db045067 100644 --- a/paddle/function/neon/neon_util.h +++ b/paddle/function/neon/neon_util.h @@ -33,12 +33,8 @@ inline float32_t vaddvq_f32(float32x4_t a) { return vget_lane_f32(vpadd_f32(v, v), 0); } -inline float32x4_t vmlaq_laneq_f32(float32x4_t a, - float32x4_t b, - float32x4_t v, - const int lane) { - return vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane)); -} +#define vmlaq_laneq_f32(a, b, v, lane) \ + vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane)) #endif } // namespace neon diff --git a/paddle/gserver/layers/Conv3DLayer.cpp b/paddle/gserver/layers/Conv3DLayer.cpp index 7cc9937cce..9deda2de98 100644 --- a/paddle/gserver/layers/Conv3DLayer.cpp +++ b/paddle/gserver/layers/Conv3DLayer.cpp @@ -42,10 +42,10 @@ bool Conv3DLayer::init(const LayerMap &layerMap, if (sharedBiases_) { CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); biases_ = - std::unique_ptr(new Weight(1, numFilters_, biasParameter_)); + std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); } else { biases_ = - std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); } } return true; @@ -83,8 +83,8 @@ void Conv3DLayer::forward(PassType passType) { int outWidth = getSize(); resetOutput(batchSize, outWidth); + REGISTER_TIMER_INFO("FwdConv3D", getName().c_str()); for (size_t i = 0; i != inputLayers_.size(); ++i) { - REGISTER_TIMER_INFO("FwdConv3D", getName().c_str()); const MatrixPtr &inMat = getInputValue(i); const MatrixPtr &outMat = getOutputValue(); int M = M_[i]; @@ -120,7 +120,6 @@ void Conv3DLayer::forward(PassType passType) { } } if (nullptr != this->biasParameter_) { - REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str()); this->addBias(); } forwardActivation(); @@ -134,15 +133,14 @@ void Conv3DLayer::backward(const UpdateCallback &callback) { biases_->getParameterPtr()->incUpdate(callback); } + REGISTER_TIMER_INFO("BwdConv3D", getName().c_str()); for (size_t i = 0; i != inputLayers_.size(); ++i) { - REGISTER_TIMER_INFO("BwdConv3D", getName().c_str()); if (weights_[i]->getWGrad()) { bpropWeights(i); } if (getInputGrad(i)) { bpropData(i); } - REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); weights_[i]->getParameterPtr()->incUpdate(callback); } } @@ -224,20 +222,31 @@ void Conv3DLayer::bpropData(int i) { } void Conv3DLayer::bpropBiases() { + MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(), + 1, + biases_->getWGrad()->getElementCnt(), + false, + useGpu_); MatrixPtr outGradMat = getOutputGrad(); + if (this->sharedBiases_) { - biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f); + biases->collectSharedBias(*outGradMat, 1.0f); } else { - biases_->getWGrad()->collectBias(*outGradMat, 1.0f); + biases->collectBias(*outGradMat, 1.0f); } } void Conv3DLayer::addBias() { MatrixPtr outMat = getOutputValue(); + MatrixPtr bias = Matrix::create(biases_->getW()->getData(), + 1, + biases_->getW()->getElementCnt(), + false, + useGpu_); if (this->sharedBiases_) { - outMat->addSharedBias(*(biases_->getW()), 1.0f); + outMat->addSharedBias(*(bias), 1.0f); } else { - outMat->addBias(*(biases_->getW()), 1.0f); + outMat->addBias(*(bias), 1.0f); } } diff --git a/paddle/gserver/layers/DeConv3DLayer.cpp b/paddle/gserver/layers/DeConv3DLayer.cpp index 7d5c772c89..1b59ed60c5 100644 --- a/paddle/gserver/layers/DeConv3DLayer.cpp +++ b/paddle/gserver/layers/DeConv3DLayer.cpp @@ -42,10 +42,10 @@ bool DeConv3DLayer::init(const LayerMap &layerMap, if (sharedBiases_) { CHECK_EQ((size_t)numFilters_, biasParameter_->getSize()); biases_ = - std::unique_ptr(new Weight(1, numFilters_, biasParameter_)); + std::unique_ptr(new Weight(numFilters_, 1, biasParameter_)); } else { biases_ = - std::unique_ptr(new Weight(1, getSize(), biasParameter_)); + std::unique_ptr(new Weight(getSize(), 1, biasParameter_)); } } return true; @@ -84,8 +84,8 @@ void DeConv3DLayer::forward(PassType passType) { resetOutput(batchSize, outWidth); const MatrixPtr outMat = getOutputValue(); + REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str()); for (size_t i = 0; i != inputLayers_.size(); ++i) { - REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str()); const MatrixPtr &inMat = getInputValue(i); int M = M_[i]; int N = N_[i]; @@ -120,7 +120,6 @@ void DeConv3DLayer::forward(PassType passType) { } } if (nullptr != this->biasParameter_) { - REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str()); this->addBias(); } forwardActivation(); @@ -133,12 +132,12 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) { bpropBiases(); biases_->getParameterPtr()->incUpdate(callback); } + REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str()); for (size_t i = 0; i < inputLayers_.size(); ++i) { if (weights_[i]->getWGrad() || this->needGradient_) { int M = M_[i]; int N = N_[i]; int K = K_[i]; - REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str()); Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_); const MatrixPtr &inMat = getInputValue(i); for (int n = 0; n < batchSize; ++n) { @@ -182,7 +181,6 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) { } } } - REGISTER_TIMER_INFO("WeightUpdate", getName().c_str()); weights_[i]->getParameterPtr()->incUpdate(callback); } } @@ -191,21 +189,31 @@ void DeConv3DLayer::bpropWeights(int i) {} void DeConv3DLayer::bpropData(int i) {} void DeConv3DLayer::bpropBiases() { + MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(), + 1, + biases_->getWGrad()->getElementCnt(), + false, + useGpu_); const MatrixPtr &outGradMat = getOutputGrad(); if (this->sharedBiases_) { - biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f); + biases->collectSharedBias(*outGradMat, 1.0f); } else { - biases_->getWGrad()->collectBias(*outGradMat, 1.0f); + biases->collectBias(*outGradMat, 1.0f); } } void DeConv3DLayer::addBias() { MatrixPtr outMat = getOutputValue(); + MatrixPtr bias = Matrix::create(biases_->getW()->getData(), + 1, + biases_->getW()->getElementCnt(), + false, + useGpu_); if (this->sharedBiases_) { - outMat->addSharedBias(*(biases_->getW()), 1.0f); + outMat->addSharedBias(*(bias), 1.0f); } else { - outMat->addBias(*(biases_->getW()), 1.0f); + outMat->addBias(*(bias), 1.0f); } } diff --git a/paddle/gserver/layers/GruCompute.cpp b/paddle/gserver/layers/GruCompute.cpp index 06907768e9..148516391c 100644 --- a/paddle/gserver/layers/GruCompute.cpp +++ b/paddle/gserver/layers/GruCompute.cpp @@ -14,6 +14,7 @@ limitations under the License. */ #include "GruCompute.h" #include "hl_recurrent_apply.cuh" +#include "paddle/function/GruFunctor.h" #include "paddle/utils/Util.h" namespace paddle { @@ -25,13 +26,13 @@ void GruCompute::init(LayerConfig &config) { template <> void GruCompute::forward<0>(hl_gru_value value, int frameSize, int batchSize) { - hl_cpu_gru_forward(hppl::forward::gru_resetOutput(), - hppl::forward::gru_finalOutput(), - value, - frameSize, - batchSize, - activeNode_, - activeGate_); + GruFunctor::compute(hppl::forward::gru_resetOutput(), + hppl::forward::gru_finalOutput(), + value, + frameSize, + batchSize, + activeNode_, + activeGate_); } template <> @@ -39,14 +40,15 @@ void GruCompute::backward<0>(hl_gru_value value, hl_gru_grad grad, int frameSize, int batchSize) { - hl_cpu_gru_backward(hppl::backward::gru_stateGrad(), - hppl::backward::gru_resetGrad(), - value, - grad, - frameSize, - batchSize, - activeNode_, - activeGate_); + GruGradFunctor::compute( + hppl::backward::gru_stateGrad(), + hppl::backward::gru_resetGrad(), + value, + grad, + frameSize, + batchSize, + activeNode_, + activeGate_); } } // namespace paddle diff --git a/paddle/gserver/layers/SwitchOrderLayer.cpp b/paddle/gserver/layers/SwitchOrderLayer.cpp new file mode 100644 index 0000000000..92cd61cdd5 --- /dev/null +++ b/paddle/gserver/layers/SwitchOrderLayer.cpp @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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. */ + +#include "SwitchOrderLayer.h" +#include "paddle/utils/Stat.h" + +namespace paddle { + +REGISTER_LAYER(switch_order, SwitchOrderLayer); + +bool SwitchOrderLayer::init(const LayerMap& layerMap, + const ParameterMap& parameterMap) { + /* Initialize the basic parent class */ + Layer::init(layerMap, parameterMap); + auto& img_conf = config_.inputs(0).image_conf(); + size_t inH = + img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size(); + size_t inW = img_conf.img_size(); + size_t inC = img_conf.channels(); + inDims_ = TensorShape({0, inC, inH, inW}); + outDims_ = TensorShape(4); + + auto& reshape_conf = config_.reshape_conf(); + for (int i = 0; i < reshape_conf.height_axis_size(); i++) { + heightAxis_.push_back(reshape_conf.height_axis(i)); + } + for (int i = 0; i < reshape_conf.width_axis_size(); i++) { + widthAxis_.push_back(reshape_conf.width_axis(i)); + } + createFunction(nchw2nhwc_, "NCHW2NHWC", FuncConfig()); + createFunction(nhwc2nchw_, "NHWC2NCHW", FuncConfig()); + return true; +} + +void SwitchOrderLayer::setOutDims() { + outDims_.setDim(0, inDims_[0]); + outDims_.setDim(1, inDims_[2]); + outDims_.setDim(2, inDims_[3]); + outDims_.setDim(3, inDims_[1]); + reshapeHeight_ = 1; + for (size_t i = 0; i < heightAxis_.size(); i++) { + reshapeHeight_ *= outDims_[heightAxis_[i]]; + } + output_.setFrameHeight(reshapeHeight_); + reshapeWidth_ = 1; + for (size_t i = 0; i < widthAxis_.size(); i++) { + reshapeWidth_ *= outDims_[widthAxis_[i]]; + } + output_.setFrameWidth(reshapeWidth_); +} + +void SwitchOrderLayer::setInDims() { + MatrixPtr input = inputLayers_[0]->getOutputValue(); + size_t batchSize = input->getHeight(); + inDims_.setDim(0, batchSize); + + int h = inputLayers_[0]->getOutput().getFrameHeight(); + if (h != 0) inDims_.setDim(2, h); + int w = inputLayers_[0]->getOutput().getFrameWidth(); + if (w != 0) inDims_.setDim(3, w); + int totalCount = input->getElementCnt(); + int channels = totalCount / (inDims_[0] * inDims_[2] * inDims_[3]); + if (channels != 0) inDims_.setDim(1, channels); +} + +void SwitchOrderLayer::forward(PassType passType) { + Layer::forward(passType); + setInDims(); + setOutDims(); + resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]); + if (heightAxis_.size() > 0) { + getOutputValue()->reshape(reshapeHeight_, reshapeWidth_); + getOutputGrad()->reshape(reshapeHeight_, reshapeWidth_); + } + + // switch NCHW to NHWC + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getInputValue(0), inDims_); + outputs.addArg(*getOutputValue(), outDims_); + nchw2nhwc_[0]->calc(inputs, outputs); + forwardActivation(); +} + +void SwitchOrderLayer::backward(const UpdateCallback& callback) { + (void)callback; + backwardActivation(); + + // switch NHWC to NCHW + BufferArgs inputs; + BufferArgs outputs; + inputs.addArg(*getOutputGrad(), outDims_); + outputs.addArg(*getInputGrad(0), inDims_, ADD_TO); + nhwc2nchw_[0]->calc(inputs, outputs); +} +} // namespace paddle diff --git a/paddle/gserver/layers/SwitchOrderLayer.h b/paddle/gserver/layers/SwitchOrderLayer.h new file mode 100644 index 0000000000..47b1f7f73e --- /dev/null +++ b/paddle/gserver/layers/SwitchOrderLayer.h @@ -0,0 +1,47 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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 + +#include "Layer.h" + +namespace paddle { + +/** + * \brief This layer calculate softmax in image channel dimension. + */ +class SwitchOrderLayer : public Layer { +public: + explicit SwitchOrderLayer(const LayerConfig& config) : Layer(config) {} + + ~SwitchOrderLayer() {} + + bool init(const LayerMap& layerMap, + const ParameterMap& parameterMap) override; + void forward(PassType passType) override; + void backward(const UpdateCallback& callback = nullptr) override; + void setInDims(); + void setOutDims(); + +protected: + std::vector> nchw2nhwc_; + std::vector> nhwc2nchw_; + TensorShape inDims_; + TensorShape outDims_; + std::vector heightAxis_; + std::vector widthAxis_; + size_t reshapeHeight_; + size_t reshapeWidth_; +}; +} // namespace paddle diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp index a831ffbc73..d1f3bc241f 100644 --- a/paddle/gserver/tests/test_LayerGrad.cpp +++ b/paddle/gserver/tests/test_LayerGrad.cpp @@ -2008,6 +2008,31 @@ TEST(Layer, CropLayer) { } } +TEST(Layer, SwitchOrderLayer) { + TestConfig config; + // config input_0 + config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0}); + LayerInputConfig* input = config.layerConfig.add_inputs(); + ImageConfig* img = input->mutable_image_conf(); + img->set_channels(4); + img->set_img_size(16); + img->set_img_size_y(16); + + ReshapeConfig* reshape = config.layerConfig.mutable_reshape_conf(); + reshape->add_height_axis(0); + reshape->add_height_axis(1); + reshape->add_height_axis(2); + reshape->add_width_axis(3); + + // config softmax layer + config.layerConfig.set_type("switch_order"); + config.layerConfig.set_name("switchOrderLayer"); + + for (auto useGpu : {false, true}) { + testLayerGrad(config, "switch_order", 100, false, useGpu, true); + } +} + vector randSampling(real range, int n) { CHECK_GE(range, n); vector num(range); diff --git a/paddle/math/MathFunctions.cpp b/paddle/math/MathFunctions.cpp index c8ba1074a1..c2f17beeb8 100644 --- a/paddle/math/MathFunctions.cpp +++ b/paddle/math/MathFunctions.cpp @@ -84,6 +84,7 @@ LAPACK_ROUTINE_EACH(DYNAMIC_LOAD_LAPACK_WRAP) namespace paddle { +#ifndef PADDLE_USE_EIGEN_FOR_BLAS template <> void gemm(const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB, @@ -143,6 +144,7 @@ void gemm(const CBLAS_TRANSPOSE transA, C, ldc); } +#endif template <> int getrf(const CBLAS_ORDER order, @@ -182,6 +184,7 @@ int getri(const CBLAS_ORDER order, return dynload::PADDLE_DGETRI(order, N, A, lda, ipiv); } +#ifndef PADDLE_USE_EIGEN_FOR_BLAS template <> void axpy(const int n, const float alpha, const float* x, float* y) { cblas_saxpy(n, alpha, x, 1, y, 1); @@ -201,6 +204,7 @@ template <> double dotProduct(const int n, const double* x, const double* y) { return cblas_ddot(n, x, 1, y, 1); } +#endif #if defined(PADDLE_USE_MKL) || defined(PADDLE_USE_MKLML) diff --git a/paddle/math/MathFunctions.h b/paddle/math/MathFunctions.h index 637643838f..e8ea6e37ac 100644 --- a/paddle/math/MathFunctions.h +++ b/paddle/math/MathFunctions.h @@ -40,7 +40,14 @@ extern "C" { #ifndef LAPACK_FOUND extern "C" { +#ifndef PADDLE_USE_EIGEN_FOR_BLAS #include +#else +typedef enum CBLAS_ORDER { + CblasRowMajor = 101, + CblasColMajor = 102 +} CBLAS_ORDER; +#endif int LAPACKE_sgetrf( int matrix_layout, int m, int n, float* a, int lda, int* ipiv); int LAPACKE_dgetrf( @@ -56,6 +63,7 @@ int LAPACKE_dgetri( namespace paddle { +#ifndef PADDLE_USE_EIGEN_FOR_BLAS template void gemm(const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB, @@ -70,6 +78,7 @@ void gemm(const CBLAS_TRANSPOSE transA, const T beta, T* C, const int ldc); +#endif template int getrf(const CBLAS_ORDER Order, @@ -84,10 +93,21 @@ int getri( const CBLAS_ORDER Order, const int N, T* A, const int lda, const int* ipiv); template -void axpy(const int n, const T alpha, const T* x, T* y); +void axpy(const int n, const T alpha, const T* x, T* y) { + /// y = y + alpha * x + for (int i = 0; i < n; i++) { + y[i] = y[i] + alpha * x[i]; + } +} template -T dotProduct(const int n, const T* x, const T* y); +T dotProduct(const int n, const T* x, const T* y) { + T result = static_cast(0); + for (int i = 0; i < n; i++) { + result += x[i] * y[i]; + } + return result; +} template void vExp(const int n, const T* a, T* r); diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp index 8bc42571f7..4a2132c8d1 100644 --- a/paddle/math/Matrix.cpp +++ b/paddle/math/Matrix.cpp @@ -28,6 +28,7 @@ limitations under the License. */ #include "hl_top_k.h" #include "paddle/utils/Logging.h" +#include "paddle/function/GemmFunctor.h" #include "paddle/utils/ThreadLocal.h" #include "SIMDFunctions.h" @@ -2773,24 +2774,24 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) { CHECK(!isTransposed()) << "Not supported"; size_t a_col, b_col, a_row, b_row; - CBLAS_TRANSPOSE a_trans, b_trans; + bool a_trans, b_trans; if (!a->isTransposed()) { a_col = a->getWidth(); a_row = a->getHeight(); - a_trans = CblasNoTrans; + a_trans = false; } else { a_col = a->getHeight(); a_row = a->getWidth(); - a_trans = CblasTrans; + a_trans = true; } if (!b->isTransposed()) { b_col = b->getWidth(); b_row = b->getHeight(); - b_trans = CblasNoTrans; + b_trans = false; } else { b_col = b->getHeight(); b_row = b->getWidth(); - b_trans = CblasTrans; + b_trans = true; } CHECK_EQ(a_col, b_row); @@ -2807,7 +2808,7 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) { int lda = a->getStride(); int ldb = b->getStride(); int ldc = getStride(); - gemm( + BlasGemm::compute( a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc); } diff --git a/paddle/math/Matrix.h b/paddle/math/Matrix.h index 431d4e0710..44180bca8b 100644 --- a/paddle/math/Matrix.h +++ b/paddle/math/Matrix.h @@ -1616,6 +1616,10 @@ public: }; class CpuMatrix : public Matrix { +private: + MatrixPtr sftmaxSum_; + MatrixPtr sftmaxDot_; + public: CpuMatrix(size_t height, size_t width, bool trans = false); CpuMatrix(real* data, size_t height, size_t width, bool trans = false) diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index e5efcccb0e..f9ea25ab04 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -14,27 +14,31 @@ function(op_library TARGET) cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - foreach(src ${op_library_SRCS}) - if (${src} MATCHES ".*\\.cu$") - list(APPEND cu_srcs ${src}) - elseif(${src} MATCHES ".*\\.cc$") - list(APPEND cc_srcs ${src}) - else() - message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu") + list(LENGTH op_library_SRCS op_library_SRCS_len) + if (${op_library_SRCS_len} EQUAL 0) + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc) + list(APPEND cc_srcs ${TARGET}.cc) endif() - endforeach() + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu) + list(APPEND cu_srcs ${TARGET}.cu) + endif() + else() + foreach(src ${op_library_SRCS}) + if (${src} MATCHES ".*\\.cu$") + list(APPEND cu_srcs ${src}) + elseif(${src} MATCHES ".*\\.cc$") + list(APPEND cc_srcs ${src}) + else() + message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu") + endif() + endforeach() + endif() list(LENGTH cc_srcs cc_srcs_len) if (${cc_srcs_len} EQUAL 0) message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file") endif() - list(LENGTH cu_srcs cu_srcs_len) - list(LENGTH op_library_DEPS dep_len) - if (${cu_srcs_len} EQUAL 0 AND ${dep_len} EQUAL 0) - message(WARNING "The op library ${TARGET} not support GPU!") - endif() - if (WITH_GPU) nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS} ${op_common_deps}) @@ -46,22 +50,22 @@ endfunction() add_subdirectory(math) -list(REMOVE_ITEM GENERAL_OPS - net_op - minus_op - mul_op - recurrent_op - scale_op) - -op_library(net_op SRCS net_op.cc) -op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op) -op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function) -op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc +set(DEPS_OPS + identity_op + minus_op + mul_op + recurrent_op + scale_op) +op_library(identity_op DEPS scale_op) +op_library(minus_op DEPS scale_op) +op_library(mul_op DEPS math_function) +op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc DEPS framework_proto tensor operator net_op) -op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op) +op_library(scale_op DEPS net_op) +list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS}) foreach(src ${GENERAL_OPS}) - op_library(${src} SRCS ${src}.cc ${src}.cu) + op_library(${src}) endforeach() set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library") diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc index 8ab748ed71..8dbd47cf0d 100644 --- a/paddle/operators/add_op.cc +++ b/paddle/operators/add_op.cc @@ -57,7 +57,6 @@ class AddOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker, add_two_grad, ops::AddOpGrad); +REGISTER_OP(add, ops::AddOp, ops::AddOpMaker, add_grad, ops::AddOpGrad); -REGISTER_OP_CPU_KERNEL(add_two, - ops::AddKernel); +REGISTER_OP_CPU_KERNEL(add, ops::AddKernel); diff --git a/paddle/operators/add_op.cu b/paddle/operators/add_op.cu index cec5f558cb..d9c6d20a6c 100644 --- a/paddle/operators/add_op.cu +++ b/paddle/operators/add_op.cu @@ -12,10 +12,7 @@ See the License for the specific language governing permissions and limitations under the License. */ -#define EIGEN_USE_GPU -#include "paddle/framework/op_registry.h" #include "paddle/operators/add_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(add_two, - ops::AddKernel); +REGISTER_OP_GPU_KERNEL(add, ops::AddKernel); diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc new file mode 100644 index 0000000000..c033af3b74 --- /dev/null +++ b/paddle/operators/cos_sim_op.cc @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + 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. */ + +#include "paddle/operators/cos_sim_op.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class CosSimOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null."); + PADDLE_ENFORCE_EQ(ctx.Input("X")->dims(), + ctx.Input("Y")->dims(), + "Dimensions of Input(X) and Input(Y) must be the same."); + + auto dims = ctx.Input("X")->dims(); + ctx.Output("Out")->Resize({dims[0], 1}); + ctx.Output("XNorm")->Resize({dims[0], 1}); + ctx.Output("YNorm")->Resize({dims[0], 1}); + } +}; + +class CosSimOpMaker : public framework::OpProtoAndCheckerMaker { + public: + CosSimOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The first input of cos_sim op."); + AddInput("Y", "The second input of cos_sim op."); + AddOutput("Out", "The output of cos_sim op."); + AddOutput("XNorm", "Row norm of the first input.").AsIntermediate(); + AddOutput("YNorm", "Row norm of the second input.").AsIntermediate(); + + AddComment(R"DOC( +Cosine Similarity Operator. + +The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)) +)DOC"); + } +}; + +class CosSimOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null."); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null."); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("XNorm"), + "Input(XNorm) must not be null."); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("YNorm"), + "Input(YNorm) must not be null."); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), + "Input(Out@GRAD) must not be null."); + + auto x_dims = ctx.Input("X")->dims(); + auto y_dims = ctx.Input("Y")->dims(); + auto xnorm_dims = ctx.Input("XNorm")->dims(); + auto ynorm_dims = ctx.Input("YNorm")->dims(); + auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); + PADDLE_ENFORCE_EQ(x_dims, y_dims, + "Dimensions of Input(X) and Input(Y) must be the same."); + PADDLE_ENFORCE_EQ(xnorm_dims[0], x_dims[0], + "1st dimension of XNorm must equal that of Input(X)."); + PADDLE_ENFORCE_EQ(xnorm_dims[1], 1, "2st dimension of XNorm must be one."); + PADDLE_ENFORCE_EQ(ynorm_dims[0], y_dims[0], + "1st dimension of YNorm must equal that of Input(Y)."); + PADDLE_ENFORCE_EQ(ynorm_dims[1], 1, "2st dimension of YNorm must be one."); + PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0], + "1st dimension of Out@GRAD must equal that of Input(X)"); + PADDLE_ENFORCE_EQ(out_dims[1], 1, "1st dimension of Out@GRAD must be one."); + + auto *x_grad = ctx.Output(framework::GradVarName("X")); + auto *y_grad = ctx.Output(framework::GradVarName("Y")); + if (x_grad) x_grad->Resize(x_dims); + if (y_grad) y_grad->Resize(y_dims); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(cos_sim, ops::CosSimOp, ops::CosSimOpMaker, cos_sim_grad, + ops::CosSimOpGrad); +REGISTER_OP_CPU_KERNEL(cos_sim, + ops::CosSimKernel); +REGISTER_OP_CPU_KERNEL( + cos_sim_grad, ops::CosSimGradKernel); diff --git a/paddle/operators/gather_op.cu b/paddle/operators/cos_sim_op.cu similarity index 72% rename from paddle/operators/gather_op.cu rename to paddle/operators/cos_sim_op.cu index 3f04a7b3f8..0cb8fd26de 100644 --- a/paddle/operators/gather_op.cu +++ b/paddle/operators/cos_sim_op.cu @@ -13,8 +13,10 @@ limitations under the License. */ #define EIGEN_USE_GPU -#include "paddle/operators/gather_op.h" +#include "paddle/operators/cos_sim_op.h" namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(gather, - ops::GatherOpKernel); +REGISTER_OP_GPU_KERNEL(cos_sim, + ops::CosSimKernel); +REGISTER_OP_GPU_KERNEL( + cos_sim_grad, ops::CosSimGradKernel); diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h new file mode 100644 index 0000000000..9e2bcebe3b --- /dev/null +++ b/paddle/operators/cos_sim_op.h @@ -0,0 +1,107 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + 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 +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenMatrix = framework::EigenMatrix; +template +using EigenVector = framework::EigenVector; + +template +class CosSimKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* input_x = context.Input("X"); + auto* input_y = context.Input("Y"); + auto* output_z = context.Output("Out"); + auto* output_x_norm = context.Output("XNorm"); + auto* output_y_norm = context.Output("YNorm"); + + output_z->mutable_data(context.GetPlace()); + output_x_norm->mutable_data(context.GetPlace()); + output_y_norm->mutable_data(context.GetPlace()); + + auto dims = input_x->dims(); + int size = static_cast(framework::product(dims)); + auto new_dims = framework::make_ddim({dims[0], size / dims[0]}); + auto x = EigenMatrix::From(*input_x, new_dims); + auto y = EigenMatrix::From(*input_y, new_dims); + auto z = EigenVector::Flatten(*output_z); + auto x_norm = EigenVector::Flatten(*output_x_norm); + auto y_norm = EigenVector::Flatten(*output_y_norm); + + auto place = context.GetEigenDevice(); + auto xy = (x * y).sum(Eigen::array({{1}})); + x_norm.device(place) = x.square().sum(Eigen::array({{1}})).sqrt(); + y_norm.device(place) = y.square().sum(Eigen::array({{1}})).sqrt(); + z.device(place) = xy / x_norm / y_norm; + } +}; + +template +class CosSimGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* input_x = context.Input("X"); + auto* input_y = context.Input("Y"); + auto* input_z = context.Input("Out"); + auto* input_x_norm = context.Input("XNorm"); + auto* input_y_norm = context.Input("YNorm"); + auto* output_grad_x = context.Output(framework::GradVarName("X")); + auto* output_grad_y = context.Output(framework::GradVarName("Y")); + auto* input_grad_z = context.Input(framework::GradVarName("Out")); + + auto dims = input_x->dims(); + int size = static_cast(framework::product(dims)); + auto new_dims = framework::make_ddim({dims[0], size / dims[0]}); + auto x = EigenMatrix::From(*input_x, new_dims); + auto y = EigenMatrix::From(*input_y, new_dims); + auto z = EigenMatrix::From(*input_z); + auto x_norm = EigenMatrix::From(*input_x_norm); + auto y_norm = EigenMatrix::From(*input_y_norm); + auto dz = EigenMatrix::From(*input_grad_z); + + Eigen::DSizes bcast(1, new_dims[1]); + auto z_bcast = z.broadcast(bcast); + auto dz_bcast = dz.broadcast(bcast); + auto place = context.GetEigenDevice(); + auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast); + auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast); + auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast); + if (output_grad_x) { + output_grad_x->mutable_data(context.GetPlace()); + auto dx = EigenMatrix::From(*output_grad_x, new_dims); + dx.device(place) = + dz_bcast * (y / norm_prod_bcast - z_bcast * x / x_snorm_bcast); + } + if (output_grad_y) { + output_grad_y->mutable_data(context.GetPlace()); + auto dy = EigenMatrix::From(*output_grad_y, new_dims); + dy.device(place) = + dz_bcast * (x / norm_prod_bcast - z_bcast * y / y_snorm_bcast); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc index 056447901d..6574880c0e 100644 --- a/paddle/operators/gaussian_random_op.cc +++ b/paddle/operators/gaussian_random_op.cc @@ -19,20 +19,20 @@ template class CPUGaussianRandomKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - float mean = context.GetAttr("mean"); - float std = context.GetAttr("std"); + float mean = context.Attr("mean"); + float std = context.Attr("std"); auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); - unsigned int seed = static_cast(context.GetAttr("seed")); + unsigned int seed = static_cast(context.Attr("seed")); std::minstd_rand engine; if (seed == 0) { seed = std::random_device()(); } engine.seed(seed); std::normal_distribution dist(mean, std); - ssize_t size = framework::product(tensor->dims()); - for (ssize_t i = 0; i < size; ++i) { + int64_t size = framework::product(tensor->dims()); + for (int64_t i = 0; i < size; ++i) { data[i] = dist(engine); } } @@ -45,10 +45,15 @@ class GaussianRandomOp : public framework::OperatorWithKernel { protected: void InferShape(const framework::InferShapeContext& context) const override { auto* tensor = context.Output("Out"); - auto dims = GetAttr>("dims"); + auto dims = Attr>("dims"); + std::vector temp; + temp.reserve(dims.size()); + for (auto dim : dims) { + temp.push_back(static_cast(dim)); + } PADDLE_ENFORCE(dims.size() > 0UL, "dims can be one int or array. dims must be set."); - tensor->Resize(framework::make_ddim(dims)); + tensor->Resize(framework::make_ddim(temp)); } }; diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu index 833a82bbf2..d9dbc1dcfe 100644 --- a/paddle/operators/gaussian_random_op.cu +++ b/paddle/operators/gaussian_random_op.cu @@ -42,13 +42,13 @@ class GPUGaussianRandomKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& context) const override { auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); - unsigned int seed = static_cast(context.GetAttr("seed")); + unsigned int seed = static_cast(context.Attr("seed")); if (seed == 0) { std::random_device rd; seed = rd(); } - T mean = static_cast(context.GetAttr("mean")); - T std = static_cast(context.GetAttr("std")); + T mean = static_cast(context.Attr("mean")); + T std = static_cast(context.Attr("std")); thrust::counting_iterator index_sequence_begin(0); ssize_t N = framework::product(tensor->dims()); thrust::transform(index_sequence_begin, index_sequence_begin + N, diff --git a/paddle/operators/identity_op.cc b/paddle/operators/identity_op.cc new file mode 100644 index 0000000000..7d9d4fa519 --- /dev/null +++ b/paddle/operators/identity_op.cc @@ -0,0 +1,57 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + 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. */ + +#include "paddle/operators/net_op.h" +#include "paddle/operators/scale_op.h" + +namespace paddle { +namespace operators { + +// The identity operator is an alias of the scale operator. This is also an +// example for creating an alias for an existing operator. +template +class IdentityOpMaker : public framework::OpProtoAndCheckerMaker { + public: + IdentityOpMaker(framework::OpProto *proto, + framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input tensor of identity operator."); + AddOutput("Out", "The output tensor of identity operator."); + AddComment(R"DOC( +The identity operator is an alias of the scale operator +with the attribute scale fixed to 1.0. +)DOC"); + } +}; + +template +class IdentityOp : public NetOp { + public: + IdentityOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : NetOp(type, inputs, outputs, attrs) { + AppendOp(framework::OpRegistry::CreateOp( + "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}}, + {{"scale", static_cast(1)}})); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_WITHOUT_GRADIENT(identity, ops::IdentityOp, + ops::IdentityOpMaker); diff --git a/paddle/operators/lookup_table_op.h b/paddle/operators/lookup_table_op.h index 4da8079b91..877b36cef4 100644 --- a/paddle/operators/lookup_table_op.h +++ b/paddle/operators/lookup_table_op.h @@ -30,12 +30,12 @@ class LookupTableKernel : public framework::OpKernel { auto ids_t = context.Input("Ids"); // int tensor auto output_t = context.Output("Out"); // float tensor - size_t N = table_t->dims()[0]; - size_t D = table_t->dims()[1]; + int N = table_t->dims()[0]; + int D = table_t->dims()[1]; auto ids = ids_t->data(); auto table = table_t->data(); auto output = output_t->mutable_data(context.GetPlace()); - for (size_t i = 0; i < product(ids_t->dims()); ++i) { + for (ssize_t i = 0; i < product(ids_t->dims()); ++i) { PADDLE_ENFORCE_LT(ids[i], N); PADDLE_ENFORCE_GE(ids[i], 0); memcpy(output + i * D, table + ids[i] * D, D * sizeof(T)); @@ -51,8 +51,8 @@ class LookupTableGradKernel : public framework::OpKernel { auto d_output_t = context.Input(framework::GradVarName("Out")); auto d_table_t = context.Output(framework::GradVarName("W")); - size_t N = d_table_t->dims()[0]; - size_t D = d_table_t->dims()[1]; + int N = d_table_t->dims()[0]; + int D = d_table_t->dims()[1]; auto ids = ids_t->data(); const T* d_output = d_output_t->data(); T* d_table = d_table_t->mutable_data(context.GetPlace()); @@ -61,10 +61,10 @@ class LookupTableGradKernel : public framework::OpKernel { t.device(context.GetEigenDevice()) = t.constant(static_cast(0)); - for (size_t i = 0; i < product(ids_t->dims()); ++i) { + for (ssize_t i = 0; i < product(ids_t->dims()); ++i) { PADDLE_ENFORCE_LT(ids[i], N); PADDLE_ENFORCE_GE(ids[i], 0); - for (size_t j = 0; j < D; ++j) { + for (int j = 0; j < D; ++j) { d_table[ids[i] * D + j] += d_output[i * D + j]; } } diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt index ed51d416ed..f8333f34f7 100644 --- a/paddle/operators/math/CMakeLists.txt +++ b/paddle/operators/math/CMakeLists.txt @@ -1,8 +1,10 @@ if(WITH_GPU) - nv_library(math_function SRCS math_function.cc math_function.cu DEPS cblas device_context) + nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc + im2col.cu DEPS cblas device_context) else() - cc_library(math_function SRCS math_function.cc DEPS cblas device_context) + cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context) endif() nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor) +cc_test(im2col_test SRCS im2col_test.cc DEPS math_function tensor) diff --git a/paddle/operators/math/im2col.cc b/paddle/operators/math/im2col.cc new file mode 100644 index 0000000000..5727c1cab1 --- /dev/null +++ b/paddle/operators/math/im2col.cc @@ -0,0 +1,260 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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. */ + +#include "paddle/operators/math/im2col.h" + +namespace paddle { +namespace operators { +namespace math { + +/* + * im = [input_channels, input_height, input_width] + * col = + * [input_channels, filter_height, filter_width, output_height, output_width] + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[1]; + int filter_width = col.dims()[2]; + int output_height = col.dims()[3]; + int output_width = col.dims()[4]; + int channels_col = input_channels * filter_height * filter_width; + + const T* im_data = im.data(); + T* col_data = col.data(); + + for (int c = 0; c < channels_col; ++c) { + int w_offset = c % filter_width; + int h_offset = (c / filter_width) % filter_height; + int c_im = c / filter_width / filter_height; + for (int h = 0; h < output_height; ++h) { + for (int w = 0; w < output_width; ++w) { + int im_row_idx = h * stride_height + h_offset; + int im_col_idx = w * stride_width + w_offset; + if ((im_row_idx - padding_height) < 0 || + (im_row_idx - padding_height) >= input_height || + (im_col_idx - padding_width) < 0 || + (im_col_idx - padding_width) >= input_width) { + col_data[(c * output_height + h) * output_width + w] = T(0); + } else { + im_row_idx += c_im * input_height - padding_height; + im_col_idx -= padding_width; + col_data[(c * output_height + h) * output_width + w] = + im_data[im_row_idx * input_width + im_col_idx]; + } + } + } + } + } +}; + +/* + * im = [input_channels, input_height, input_width] + * col = + * [input_channels, filter_height, filter_width, output_height, output_width] + */ +template +class Col2ImFunctor { + public: + void operator()(framework::Tensor& im, const framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[1]; + int filter_width = col.dims()[2]; + int output_height = col.dims()[3]; + int output_width = col.dims()[4]; + int channels_col = input_channels * filter_height * filter_width; + + T* im_data = im.data(); + const T* col_data = col.data(); + + for (int c = 0; c < channels_col; ++c) { + int w_offset = c % filter_width; + int h_offset = (c / filter_width) % filter_height; + int c_im = c / filter_width / filter_height; + for (int h = 0; h < output_height; ++h) { + for (int w = 0; w < output_width; ++w) { + int im_row_idx = h * stride_height + h_offset; + int im_col_idx = w * stride_width + w_offset; + if ((im_row_idx - padding_height) >= 0 && + (im_row_idx - padding_height) < input_height && + (im_col_idx - padding_width) >= 0 && + (im_col_idx - padding_width) < input_width) { + im_row_idx += c_im * input_height - padding_height; + im_col_idx -= padding_width; + im_data[im_row_idx * input_width + im_col_idx] += + col_data[(c * output_height + h) * output_width + w]; + } + } + } + } + } +}; + +template class Im2ColFunctor; +template class Im2ColFunctor; +template class Col2ImFunctor; +template class Col2ImFunctor; + +/* + * im = [input_channels, input_height, input_width] + * col = + * [output_height, output_width, input_channels, filter_height, filter_width] + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[3]; + int filter_width = col.dims()[4]; + int output_height = col.dims()[0]; + int output_width = col.dims()[1]; + + const T* im_data = im.data(); + T* col_data = col.data(); + + for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) { + for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) { + for (int channel = 0; channel < input_channels; ++channel) { + for (int filter_row_idx = 0; filter_row_idx < filter_height; + ++filter_row_idx) { + for (int filter_col_idx = 0; filter_col_idx < filter_width; + ++filter_col_idx) { + int im_row_offset = + col_row_idx * stride_height + filter_row_idx - padding_height; + int im_col_offset = + col_col_idx * stride_width + filter_col_idx - padding_width; + int col_offset = (((col_row_idx * output_width + col_col_idx) * + input_channels + + channel) * + filter_height + + filter_row_idx) * + filter_width + + filter_col_idx; + if (im_row_offset < 0 || im_row_offset >= input_height || + im_col_offset < 0 || im_col_offset >= input_width) { + col_data[col_offset] = T(0); + } else { + int im_offset = + (channel * input_height + im_row_offset) * input_width + + im_col_offset; + col_data[col_offset] = im_data[im_offset]; + } + } + } + } + } + } + } +}; + +/* + * im = [input_channels, input_height, input_width] + * col = + * [output_height, output_width, input_channels, filter_height, filter_width] + */ +template +class Col2ImFunctor { + public: + void operator()(framework::Tensor& im, const framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[3]; + int filter_width = col.dims()[4]; + int output_height = col.dims()[0]; + int output_width = col.dims()[1]; + + T* im_data = im.data(); + const T* col_data = col.data(); + + for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) { + for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) { + for (int channel = 0; channel < input_channels; ++channel) { + for (int filter_row_idx = 0; filter_row_idx < filter_height; + ++filter_row_idx) { + for (int filter_col_idx = 0; filter_col_idx < filter_width; + ++filter_col_idx) { + int im_row_offset = + col_row_idx * stride_height + filter_row_idx - padding_height; + int im_col_offset = + col_col_idx * stride_width + filter_col_idx - padding_width; + int col_offset = (((col_row_idx * output_width + col_col_idx) * + input_channels + + channel) * + filter_height + + filter_row_idx) * + filter_width + + filter_col_idx; + if (im_row_offset >= 0 && im_row_offset < input_height && + im_col_offset >= 0 && im_col_offset < input_width) { + int im_offset = + (channel * input_height + im_row_offset) * input_width + + im_col_offset; + im_data[im_offset] += col_data[col_offset]; + } + } + } + } + } + } + } +}; + +template class Im2ColFunctor; +template class Im2ColFunctor; +template class Col2ImFunctor; +template class Col2ImFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/im2col.cu b/paddle/operators/math/im2col.cu new file mode 100644 index 0000000000..9bff7bee3c --- /dev/null +++ b/paddle/operators/math/im2col.cu @@ -0,0 +1,374 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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. */ + +#include "paddle/operators/math/im2col.h" +#include "paddle/platform/cuda_helper.h" + +namespace paddle { +namespace operators { +namespace math { + +template +__global__ void im2col(const T* data_im, int num_outs, int height, int width, + int filter_height, int filter_width, int stride_height, + int stride_width, int padding_height, int padding_width, + int output_height, int output_width, T* data_col) { + int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + if (index < num_outs) { + int w_out = index % output_width; + index /= output_width; + int h_out = index % output_height; + int channel_in = index / output_height; + int channel_out = channel_in * filter_height * filter_width; + int h_in = h_out * stride_height; + int w_in = w_out * stride_width; + + data_col += (channel_out * output_height + h_out) * output_width + w_out; + for (int i = 0; i < filter_height; ++i) { + for (int j = 0; j < filter_width; ++j) { + int rIdx = int(h_in + i); + int cIdx = int(w_in + j); + if ((rIdx - (int)padding_height) >= (int)height || + (rIdx - (int)padding_height) < 0 || + (cIdx - (int)padding_width) >= (int)width || + (cIdx - (int)padding_width) < 0) { + *data_col = 0; + } else { + rIdx = rIdx + channel_in * height - padding_height; + cIdx = cIdx - padding_width; + *data_col = data_im[rIdx * width + cIdx]; + } + data_col += output_height * output_width; + } + } + } +} + +/* + * im = [input_channels, input_height, input_width] + * col = + * [input_channels, filter_height, filter_width, output_height, output_width] + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[1]; + int filter_width = col.dims()[2]; + int output_height = col.dims()[3]; + int output_width = col.dims()[4]; + + int num_outputs = input_channels * output_height * output_width; + int blocks = (num_outputs + 1024 - 1) / 1024; + int block_x = 512; + int block_y = (blocks + 512 - 1) / 512; + dim3 threads(1024, 1); + dim3 grid(block_x, block_y); + im2col<<< + grid, threads, 0, + reinterpret_cast(context)->stream()>>>( + im.data(), num_outputs, input_height, input_width, filter_height, + filter_width, stride_height, stride_width, padding_height, + padding_width, output_height, output_width, col.data()); + } +}; + +template +__global__ void col2im(size_t n, const T* data_col, size_t height, size_t width, + size_t channels, size_t filter_height, + size_t filter_width, size_t stride_height, + size_t stride_width, size_t padding_height, + size_t padding_width, size_t output_height, + size_t output_width, T* data_im) { + size_t index = + (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x; + if (index < n) { + T val = 0; + int w = int(index % width); + int h = int((index / width) % height); + int c = int(index / (width * height)); + if ((w - (int)padding_width) >= 0 && + (w - (int)padding_width) < (width - 2 * padding_width) && + (h - (int)padding_height) >= 0 && + (h - padding_height) < (height - 2 * padding_height)) { + // compute the start and end of the output + int w_col_start = (w < (int)filter_width) + ? 0 + : (w - int(filter_width)) / (int)stride_width + 1; + int w_col_end = + min((int)(w / (int)stride_width + 1), (int)(output_width)); + int h_col_start = (h < (int)filter_height) + ? 0 + : (h - (int)filter_height) / (int)stride_height + 1; + int h_col_end = min(int(h / stride_height + 1), int(output_height)); + for (int h_col = h_col_start; h_col < h_col_end; ++h_col) { + for (int w_col = w_col_start; w_col < w_col_end; ++w_col) { + // the col location: [c * width * height + h_out, w_out] + int c_col = int(c * filter_height * filter_width) + + (h - h_col * (int)stride_height) * (int)filter_width + + (w - w_col * (int)stride_width); + val += + data_col[(c_col * output_height + h_col) * output_width + w_col]; + } + } + h -= padding_height; + w -= padding_width; + data_im[c * ((width - 2 * padding_width) * + (height - 2 * padding_height)) + + h * (width - 2 * padding_width) + w] += val; + } + } +} + +/* + * im = [input_channels, input_height, input_width] + * col = + * [input_channels, filter_height, filter_width, output_height, output_width] + */ +template +class Col2ImFunctor { + public: + void operator()(framework::Tensor& im, const framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[1]; + int filter_width = col.dims()[2]; + int output_height = col.dims()[3]; + int output_width = col.dims()[4]; + + size_t num_kernels = input_channels * (input_height + 2 * padding_height) * + (input_width + 2 * padding_width); + + size_t blocks = (num_kernels + 1024 - 1) / 1024; + size_t block_x = 512; + size_t block_y = (blocks + 512 - 1) / 512; + dim3 threads(1024, 1); + dim3 grid(block_x, block_y); + + // To avoid involving atomic operations, we will launch one kernel per + // bottom dimension, and then in the kernel add up the top dimensions. + col2im<<< + grid, threads, 0, + reinterpret_cast(context)->stream()>>>( + num_kernels, col.data(), input_height + 2 * padding_height, + input_width + 2 * padding_width, input_channels, filter_height, + filter_width, stride_height, stride_width, padding_height, + padding_width, output_height, output_width, im.data()); + } +}; + +template class Im2ColFunctor; +template class Im2ColFunctor; +template class Col2ImFunctor; +template class Col2ImFunctor; + +template +__global__ void im2colOCF(const T* im_data, T* col_data, int input_channels, + int input_height, int input_width, int filter_height, + int filter_width, int stride_height, int stride_width, + int padding_height, int padding_width, + int output_height, int output_width) { + int swid = blockIdx.x; + int shid = blockIdx.y; + for (int channelid = threadIdx.z; channelid < input_channels; + channelid += blockDim.z) { + for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) { + for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) { + int width_offset = idx + swid * stride_width - padding_width; + int height_offset = idy + shid * stride_height - padding_height; + int im_offset = width_offset + height_offset * input_width + + channelid * input_height * input_width; + + int col_offset = idx + idy * filter_width + + channelid * filter_height * filter_width + + (shid * output_width + swid) * + (input_channels * filter_height * filter_width); + + if (height_offset >= input_height || height_offset < 0 || + width_offset >= input_width || width_offset < 0) { + col_data[col_offset] = T(0); + } else { + col_data[col_offset] = im_data[im_offset]; + } + } + } + } +} + +/* + * im = [input_channels, input_height, input_width] + * col = + * [output_height, output_width, input_channels, filter_height, filter_width] + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[3]; + int filter_width = col.dims()[4]; + int output_height = col.dims()[0]; + int output_width = col.dims()[1]; + + int block_dim_x = 0; + int block_dim_y = 0; + if (filter_height <= 4 && filter_width <= 4) { + block_dim_x = 4; + block_dim_y = 4; + } else if (filter_height <= 8 && filter_width <= 8) { + block_dim_x = 8; + block_dim_y = 8; + } else if (filter_height <= 16 && filter_width <= 16) { + block_dim_x = 16; + block_dim_y = 16; + } else { + block_dim_x = 32; + block_dim_y = 32; + } + + int block_dim_z = 1024 / block_dim_x / block_dim_y; + dim3 threads(block_dim_x, block_dim_y, + std::min(block_dim_z, input_channels)); + dim3 grid(output_width, output_height); + im2colOCF<<< + grid, threads, 0, + reinterpret_cast(context)->stream()>>>( + im.data(), col.data(), input_channels, input_height, input_width, + filter_height, filter_width, stride_height, stride_width, + padding_height, padding_width, output_height, output_width); + } +}; + +template +__global__ void col2imOCF(T* im_data, const T* col_data, int input_channels, + int input_height, int input_width, int filter_height, + int filter_width, int stride_height, int stride_width, + int padding_height, int padding_width, + int output_height, int output_width) { + int swid = blockIdx.x; + int shid = blockIdx.y; + for (int channelid = threadIdx.z; channelid < input_channels; + channelid += blockDim.z) { + for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) { + for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) { + int width_offset = idx + swid * stride_width - padding_width; + int height_offset = idy + shid * stride_height - padding_height; + int im_offset = width_offset + height_offset * input_width + + channelid * input_height * input_width; + + int col_offset = idx + idy * filter_width + + channelid * filter_height * filter_width + + (shid * output_width + swid) * + (input_channels * filter_height * filter_width); + + if (height_offset >= 0 && height_offset < input_height && + width_offset >= 0 && width_offset < input_width) { + paddle::platform::CudaAtomicAdd(im_data + im_offset, + col_data[col_offset]); + } + } + } + } +} + +/* + * im = [input_channels, input_height, input_width] + * col = + * [output_height, output_width, input_channels, filter_height, filter_width] + */ +template +class Col2ImFunctor { + public: + void operator()(framework::Tensor& im, const framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context) { + PADDLE_ENFORCE(im.dims().size() == 3); + PADDLE_ENFORCE(col.dims().size() == 5); + int input_channels = im.dims()[0]; + int input_height = im.dims()[1]; + int input_width = im.dims()[2]; + int filter_height = col.dims()[3]; + int filter_width = col.dims()[4]; + int output_height = col.dims()[0]; + int output_width = col.dims()[1]; + + int block_dim_x = 0; + int block_dim_y = 0; + if (filter_height <= 4 && filter_width <= 4) { + block_dim_x = 4; + block_dim_y = 4; + } else if (filter_height <= 8 && filter_width <= 8) { + block_dim_x = 8; + block_dim_y = 8; + } else if (filter_height <= 16 && filter_width <= 16) { + block_dim_x = 16; + block_dim_y = 16; + } else { + block_dim_x = 32; + block_dim_y = 32; + } + + int block_dim_z = 1024 / block_dim_x / block_dim_y; + dim3 threads(block_dim_x, block_dim_y, + std::min(block_dim_z, input_channels)); + dim3 grid(output_width, output_height); + col2imOCF<<< + grid, threads, 0, + reinterpret_cast(context)->stream()>>>( + im.data(), col.data(), input_channels, input_height, input_width, + filter_height, filter_width, stride_height, stride_width, + padding_height, padding_width, output_height, output_width); + } +}; + +template class Im2ColFunctor; +template class Im2ColFunctor; +template class Col2ImFunctor; +template class Col2ImFunctor; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/im2col.h b/paddle/operators/math/im2col.h new file mode 100644 index 0000000000..8958c5457c --- /dev/null +++ b/paddle/operators/math/im2col.h @@ -0,0 +1,90 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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 + +#include "paddle/framework/tensor.h" +#include "paddle/platform/device_context.h" + +namespace paddle { +namespace operators { +namespace math { + +/* The storage format of the coldata in the Im2ColFunctor and Col2ImFunctor. */ +enum class ColFormat { kCFO = 0, kOCF = 1 }; + +/* + * \brief Converts the image data of three dimensions(CHW) into a colData of + * five dimensions in the Im2ColFunctor calculation, + * And in the Col2ImFunctor calculation, it is reversed. + * + * \param imData Image data. + * \param imShape The shape of imData, + * [input_channels, input_height, input_width]. + * \param colData Column data. + * \param colShape The shape of colData. + * + * If the template argument Format is kCFO, the shape of colData is: + * [input_channels, filter_height, filter_width, output_height, output_width] + * So, it is easy to reshape into a convolution matrix for convolution + * calculation based on matrix multiplication. + * The shape of convolution matrix is [height, width], where the height is equal + * input_channels * filter_height * filter_width, and the width is equal + * output_height * output_width. + * + * Reshape: + * shape of colData shape of convolution matrix + * [input_channels, + * filter_height, + * filter_width, ======> [height, width] + * output_height, + * output_width] + * + * If the template argument Format is kOCF, the shape of colData is: + * [output_height, output_width, input_channels, filter_height, filter_width] + * So, it is easy to reshape into a sequence matrix for rnn calculation. + * The shape of sequence matrix is [seq_length, step_size], where the seq_length + * is equal output_height * output_width, and the step_size is equal + * input_channels * filter_height * filter_width. + * + * Reshape: + * shape of colData shape of sequence matrix + * [output_height, + * output_width, + * input_channels, ======> [seqLength, stepSize] + * filter_height, + * filter_width] + * + * \note The caller needs to ensure that imShape.inputChannels is equal to + * colShape.inputChannels. + */ +template +class Im2ColFunctor { + public: + void operator()(const framework::Tensor& im, framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context); +}; + +template +class Col2ImFunctor { + public: + void operator()(framework::Tensor& im, const framework::Tensor& col, + int stride_height, int stride_width, int padding_height, + int padding_width, platform::DeviceContext* context); +}; + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc new file mode 100644 index 0000000000..186a33edce --- /dev/null +++ b/paddle/operators/math/im2col_test.cc @@ -0,0 +1,122 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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. */ + +#include "paddle/operators/math/im2col.h" +#include +#include + +template +void testIm2col() { + paddle::framework::Tensor input_tmp; + paddle::framework::Tensor input; + paddle::framework::Tensor output_cfo; + paddle::framework::Tensor output_ocf; + paddle::framework::Tensor output_tmp; + + /** + * input = [0, 1, 2, + * 3, 4, 5] + * + * output_cfo = [0, 1 + * 1, 2 + * 3, 4 + * 4, 5] + * + * output_ocf = [0, 1, 3, 4 + * 1, 2, 4, 5] + */ + int input_height = 2; + int input_width = 3; + int filter_size = 2; + int stride = 1; + int padding = 0; + int output_height = (input_height - filter_size + 2 * padding) / stride + 1; + int output_width = (input_width - filter_size + 2 * padding) / stride + 1; + float* input_ptr = input_tmp.mutable_data( + {1, input_height, input_width}, paddle::platform::CPUPlace()); + float arr[6] = {0, 1, 2, 3, 4, 5}; + memcpy(input_ptr, arr, 6 * sizeof(float)); + + auto* place = new Place(); + if (paddle::platform::is_cpu_place(*place)) { + input = input_tmp; + } else { + input.CopyFrom(input_tmp, *place); + } + output_cfo.mutable_data( + {1, filter_size, filter_size, output_height, output_width}, *place); + output_ocf.mutable_data( + {output_height, output_width, 1, filter_size, filter_size}, *place); + + paddle::operators::math::Im2ColFunctor< + paddle::operators::math::ColFormat::kCFO, Place, float> + im2col; + paddle::operators::math::Im2ColFunctor< + paddle::operators::math::ColFormat::kOCF, Place, float> + im2col_ocf; + + paddle::platform::DeviceContext* context; + if (paddle::platform::is_cpu_place(*place)) { + context = + new paddle::platform::CPUDeviceContext(paddle::platform::CPUPlace()); + } else { +#ifndef PADDLE_ONLY_CPU + context = + new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace()); +#else + PADDLE_THROW("no GPU support"); +#endif // PADDLE_ONLY_CPU + } + im2col(input, output_cfo, stride, stride, padding, padding, context); + im2col_ocf(input, output_ocf, stride, stride, padding, padding, context); + + float* out_cfo_ptr; + if (paddle::platform::is_cpu_place(*place)) { + out_cfo_ptr = output_cfo.data(); + } else { + output_tmp.CopyFrom(output_cfo, paddle::platform::CPUPlace()); + out_cfo_ptr = output_tmp.data(); + } + EXPECT_EQ(out_cfo_ptr[0], 0); + EXPECT_EQ(out_cfo_ptr[1], 1); + EXPECT_EQ(out_cfo_ptr[2], 1); + EXPECT_EQ(out_cfo_ptr[3], 2); + EXPECT_EQ(out_cfo_ptr[4], 3); + EXPECT_EQ(out_cfo_ptr[5], 4); + EXPECT_EQ(out_cfo_ptr[6], 4); + EXPECT_EQ(out_cfo_ptr[7], 5); + + float* out_ocf_ptr; + if (paddle::platform::is_cpu_place(*place)) { + out_ocf_ptr = output_ocf.data(); + } else { + output_tmp.CopyFrom(output_ocf, paddle::platform::CPUPlace()); + out_ocf_ptr = output_tmp.data(); + } + EXPECT_EQ(out_ocf_ptr[0], 0); + EXPECT_EQ(out_ocf_ptr[1], 1); + EXPECT_EQ(out_ocf_ptr[2], 3); + EXPECT_EQ(out_ocf_ptr[3], 4); + EXPECT_EQ(out_ocf_ptr[4], 1); + EXPECT_EQ(out_ocf_ptr[5], 2); + EXPECT_EQ(out_ocf_ptr[6], 4); + EXPECT_EQ(out_ocf_ptr[7], 5); +} + +TEST(math, im2col) { + testIm2col(); +#ifndef PADDLE_ONLY_CPU + testIm2col(); +#endif +} \ No newline at end of file diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc index 5b8b5f6c11..710a56a0e8 100644 --- a/paddle/operators/mul_op.cc +++ b/paddle/operators/mul_op.cc @@ -25,18 +25,27 @@ class MulOp : public framework::OperatorWithKernel { protected: void InferShape(const framework::InferShapeContext &ctx) const override { - auto dim0 = ctx.Input("X")->dims(); - auto dim1 = ctx.Input("Y")->dims(); - PADDLE_ENFORCE_EQ(dim0.size(), 2, - "input X(%s) should be a tensor with 2 dims, a matrix", - ctx.op().Input("X")); - PADDLE_ENFORCE_EQ(dim1.size(), 2, - "input Y(%s) should be a tensor with 2 dims, a matrix", - ctx.op().Input("Y")); + auto x_dims = ctx.Input("X")->dims(); + auto y_dims = ctx.Input("Y")->dims(); + int x_num_col_dims = Attr("x_num_col_dims"); + int y_num_col_dims = Attr("y_num_col_dims"); + + PADDLE_ENFORCE(x_dims.size() > x_num_col_dims, + "The rank of input tensor X(%s) should be larger than " + "`mul_op`'s `x_num_col_dims`.", + ctx.op().Input("X")); + PADDLE_ENFORCE(y_dims.size() > y_num_col_dims, + "The rank of input tensor Y(%s) should be larger than " + "`mul_op`'s `y_num_col_dims`.", + ctx.op().Input("Y")); + + auto x_mat_dims = framework::flatten_to_2d(x_dims, x_num_col_dims); + auto y_mat_dims = framework::flatten_to_2d(y_dims, y_num_col_dims); + PADDLE_ENFORCE_EQ( - dim0[1], dim1[0], + x_mat_dims[1], y_mat_dims[0], "First matrix's width must be equal with second matrix's height."); - ctx.Output("Out")->Resize({dim0[0], dim1[1]}); + ctx.Output("Out")->Resize({x_mat_dims[0], y_mat_dims[1]}); } }; @@ -47,6 +56,23 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("X", "The first input of mul op"); AddInput("Y", "The second input of mul op"); AddOutput("Out", "The output of mul op"); + AddAttr( + "x_num_col_dims", + R"DOC(mul_op can take tensors with more than two dimensions as input `X`, + in that case, tensors will be reshaped to a matrix. The matrix's first + dimension(column length) will be the product of tensor's last + `num_col_dims` dimensions, and the matrix's second dimension(row length) + will be the product of tensor's first `rank - num_col_dims` dimensions. + )DOC") + .SetDefault(1) + .EqualGreaterThan(1); + AddAttr( + "y_num_col_dims", + R"DOC(mul_op can take tensors with more than two dimensions as input `Y`, + in that case, tensors will be reshaped to a matrix. Just like input `X`. + )DOC") + .SetDefault(1) + .EqualGreaterThan(1); AddComment(R"DOC( Two Element Mul Operator. @@ -70,13 +96,23 @@ class MulOpGrad : public framework::OperatorWithKernel { auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); auto *x_grad = ctx.Output(framework::GradVarName("X")); auto *y_grad = ctx.Output(framework::GradVarName("Y")); - PADDLE_ENFORCE(x_dims[0] == out_dims[0], - "Out@GRAD M X N must equal to X dims 0, M "); - PADDLE_ENFORCE(y_dims[1] == out_dims[1], - "Out@GRAD M X N must equal to Y dims 1, N "); - x_grad->Resize(x_dims); - y_grad->Resize(y_dims); + auto x_mat_dims = + framework::flatten_to_2d(x_dims, Attr("x_num_col_dims")); + auto y_mat_dims = + framework::flatten_to_2d(y_dims, Attr("y_num_col_dims")); + + PADDLE_ENFORCE_EQ( + x_mat_dims[0], out_dims[0], + "The first dimension of Out@GRAD must equal to the first dimension of " + "the first operand."); + PADDLE_ENFORCE_EQ( + y_mat_dims[1], out_dims[1], + "The second dimension of Out@GRAD must equal to the second " + "dimension of the second operand."); + + if (x_grad) x_grad->Resize(x_dims); + if (y_grad) y_grad->Resize(y_dims); } }; diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h index 8facc02814..3c01f868bd 100644 --- a/paddle/operators/mul_op.h +++ b/paddle/operators/mul_op.h @@ -1,7 +1,7 @@ /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with 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 @@ -31,13 +31,25 @@ template class MulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* X = context.Input("X"); - auto* Y = context.Input("Y"); - auto* Z = context.Output("Out"); - Z->mutable_data(context.GetPlace()); + const Tensor* x = context.Input("X"); + const Tensor* y = context.Input("Y"); + Tensor* z = context.Output("Out"); + const Tensor x_matrix = + x->dims().size() > 2 + ? framework::ReshapeToMatrix( + *x, context.template Attr("x_num_col_dims")) + : *x; + const Tensor y_matrix = + y->dims().size() > 2 + ? framework::ReshapeToMatrix( + *y, context.template Attr("y_num_col_dims")) + : *y; + + z->mutable_data(context.GetPlace()); auto* device_context = const_cast(context.device_context_); - math::matmul(*X, false, *Y, false, 1, Z, 0, device_context); + math::matmul(x_matrix, false, y_matrix, false, 1, z, 0, + device_context); } }; @@ -45,20 +57,40 @@ template class MulGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - auto* X = ctx.Input("X"); - auto* Y = ctx.Input("Y"); - auto* dOut = ctx.Input(framework::GradVarName("Out")); + int x_num_col_dims = ctx.template Attr("x_num_col_dims"); + int y_num_col_dims = ctx.template Attr("y_num_col_dims"); + const Tensor* x = ctx.Input("X"); + const Tensor* y = ctx.Input("Y"); + const Tensor x_matrix = + x->dims().size() > 2 ? framework::ReshapeToMatrix(*x, x_num_col_dims) + : *x; + const Tensor y_matrix = + y->dims().size() > 2 ? framework::ReshapeToMatrix(*y, y_num_col_dims) + : *y; + const Tensor* dout = ctx.Input(framework::GradVarName("Out")); - auto* dX = ctx.Output(framework::GradVarName("X")); - auto* dY = ctx.Output(framework::GradVarName("Y")); - dX->mutable_data(ctx.GetPlace()); - dY->mutable_data(ctx.GetPlace()); + Tensor* dx = ctx.Output(framework::GradVarName("X")); + Tensor* dy = ctx.Output(framework::GradVarName("Y")); auto* device_context = const_cast(ctx.device_context_); - // dX = dOut * Y'. dX: M x K, dOut : M x N, Y : K x N - math::matmul(*dOut, false, *Y, true, 1, dX, 0, device_context); - // dY = X' * dOut. dY: K x N, dOut : M x N, X : M x K - math::matmul(*X, true, *dOut, false, 1, dY, 0, device_context); + if (dx) { + dx->mutable_data(ctx.GetPlace()); + Tensor dx_matrix = dx->dims().size() > 2 ? framework::ReshapeToMatrix( + *dx, x_num_col_dims) + : *dx; + // dx = dout * y'. dx: M x K, dout : M x N, y : K x N + math::matmul(*dout, false, y_matrix, true, 1, &dx_matrix, 0, + device_context); + } + if (dy) { + dy->mutable_data(ctx.GetPlace()); + Tensor dy_matrix = dy->dims().size() > 2 ? framework::ReshapeToMatrix( + *dy, y_num_col_dims) + : *dy; + // dy = x' * dout. dy K x N, dout : M x N, x : M x K + math::matmul(x_matrix, true, *dout, false, 1, &dy_matrix, 0, + device_context); + } } }; diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc index a9b65c30f2..97872c67ac 100644 --- a/paddle/operators/rnn/recurrent_op_utils.cc +++ b/paddle/operators/rnn/recurrent_op_utils.cc @@ -61,7 +61,7 @@ void ConcatOutputs(const std::vector& step_scopes, PADDLE_ENFORCE(step_scope_var != nullptr, "%s not in scope", outlinks[i].internal); f::DDim step_dims = step_scope_var->template GetMutable()->dims(); - std::vector dims_vec = vectorize(step_dims); + std::vector dims_vec = vectorize(step_dims); dims_vec.insert(dims_vec.begin(), seq_len); output->Resize(f::make_ddim(dims_vec)); } else { @@ -109,7 +109,7 @@ void InitArgument(const ArgumentName& name, Argument* arg, arg->step_scopes = op.Output(name.step_scopes); auto inlinks = op.Inputs(name.inlinks); - auto inlink_alias = op.GetAttr>(name.inlink_alias); + auto inlink_alias = op.Attr>(name.inlink_alias); PADDLE_ENFORCE(inlinks.size() == inlink_alias.size(), "the size of inlinks and inlink_alias don't match:%d,%d", inlinks.size(), inlink_alias.size()); @@ -121,7 +121,7 @@ void InitArgument(const ArgumentName& name, Argument* arg, } auto outlinks = op.Outputs(name.outlinks); - auto outlink_alias = op.GetAttr>(name.outlink_alias); + auto outlink_alias = op.Attr>(name.outlink_alias); PADDLE_ENFORCE(outlinks.size() == outlink_alias.size(), "the size of outlinks and outlink_alias don't match:%d,%d", outlinks.size(), outlink_alias.size()); @@ -135,8 +135,8 @@ void InitArgument(const ArgumentName& name, Argument* arg, auto boot_memories = op.Inputs(name.boot_memories); // attributes - auto memories = op.GetAttr>(name.memories); - auto pre_memories = op.GetAttr>(name.pre_memories); + auto memories = op.Attr>(name.memories); + auto pre_memories = op.Attr>(name.pre_memories); PADDLE_ENFORCE(memories.size() == boot_memories.size(), "the size of memories, boot_memories don't match:%d,%d", diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc index 6825dce332..fa8f0ff1a8 100644 --- a/paddle/operators/rowwise_add_op.cc +++ b/paddle/operators/rowwise_add_op.cc @@ -25,14 +25,19 @@ class RowwiseAddOp : public framework::OperatorWithKernel { protected: void InferShape(const framework::InferShapeContext &ctx) const override { - auto dim0 = ctx.Input("X")->dims(); - auto dim1 = ctx.Input("b")->dims(); - - PADDLE_ENFORCE(dim0.size() == 2, "Input 0 must be matrix"); - PADDLE_ENFORCE(dim1.size() == 1, "The second input must be vector"); - PADDLE_ENFORCE(dim0[1] == dim1[0], "The width of two input must be same"); - PADDLE_ENFORCE(ctx.OutputSize("Out") == 1, "The output size must be 1"); - ctx.Output("Out")->Resize(ctx.Input("X")->dims()); + auto x_dims = ctx.Input("X")->dims(); + auto b_dims = ctx.Input("b")->dims(); + PADDLE_ENFORCE_GT( + x_dims.size(), b_dims.size(), + "The rank of input `X` must be larger than the one of input `b`."); + + int num_col_dims = x_dims.size() - b_dims.size(); + + PADDLE_ENFORCE_EQ( + framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims, + "The width of two operands must be same"); + PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1"); + ctx.Output("Out")->Resize(x_dims); } }; @@ -61,11 +66,20 @@ class RowwiseAddGradOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"), "b should not be null"); PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), "Input(Out@GRAD) should not be null"); - auto dims0 = ctx.Input("X")->dims(); - auto dims1 = ctx.Input("b")->dims(); - PADDLE_ENFORCE_EQ(1, dims1.size(), "b dims should be 1") - ctx.Output(framework::GradVarName("X"))->Resize(dims0); - ctx.Output(framework::GradVarName("b"))->Resize(dims1); + auto x_dims = ctx.Input("X")->dims(); + auto b_dims = ctx.Input("b")->dims(); + PADDLE_ENFORCE_GT( + x_dims.size(), b_dims.size(), + "The rank of input `X` must be larger than the one of input `b`."); + + int num_col_dims = x_dims.size() - b_dims.size(); + PADDLE_ENFORCE_EQ( + framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims, + "The width of two operands must be same"); + auto *dx = ctx.Output(framework::GradVarName("X")); + auto *db = ctx.Output(framework::GradVarName("b")); + if (dx) dx->Resize(x_dims); + if (db) db->Resize(b_dims); } }; diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h index 1cbd8bb31a..35774b9409 100644 --- a/paddle/operators/rowwise_add_op.h +++ b/paddle/operators/rowwise_add_op.h @@ -33,10 +33,12 @@ class RowwiseAddKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& context) const override { auto out = context.Output("Out"); out->mutable_data(context.GetPlace()); - - auto input = EigenMatrix::From(*context.Input("X")); - auto bias = EigenVector::From(*context.Input("b")); - auto output = EigenMatrix::From(*out); + int num_col_dims = context.Input("X")->dims().size() - + context.Input("b")->dims().size(); + auto input = + EigenMatrix::Reshape(*context.Input("X"), num_col_dims); + auto bias = EigenVector::Flatten(*context.Input("b")); + auto output = EigenMatrix::Reshape(*out, num_col_dims); const int bias_size = bias.dimension(0); const int rest_size = input.size() / bias_size; @@ -51,20 +53,27 @@ template class RowwiseAddGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* dOut = context.Input(framework::GradVarName("Out")); - auto* dX = context.Output(framework::GradVarName("X")); + auto* dout = context.Input(framework::GradVarName("Out")); + auto* dx = context.Output(framework::GradVarName("X")); auto* db = context.Output(framework::GradVarName("b")); - dX->mutable_data(context.GetPlace()); - db->mutable_data(context.GetPlace()); + int num_col_dims = context.Input("X")->dims().size() - + context.Input("b")->dims().size(); - auto OutGrad = EigenMatrix::From(*dOut); + auto out_grad = EigenMatrix::Reshape(*dout, num_col_dims); auto place = context.GetEigenDevice(); - EigenMatrix::From(*dX).device(place) = OutGrad; - // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html - // colwise add - Eigen::array dims{{0}}; /* dimension to reduce */ - EigenVector::Flatten(*db).device(place) = OutGrad.sum(dims); + if (dx) { + dx->mutable_data(context.GetPlace()); + EigenMatrix::Reshape(*dx, num_col_dims).device(place) = out_grad; + } + + if (db) { + db->mutable_data(context.GetPlace()); + // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html + // colwise add + Eigen::array dims{{0}}; /* dimension to reduce */ + EigenVector::Flatten(*db).device(place) = out_grad.sum(dims); + } } }; } // namespace operators diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc index 8e96a74c94..ea991f683d 100644 --- a/paddle/operators/scale_op.cc +++ b/paddle/operators/scale_op.cc @@ -44,11 +44,13 @@ class ScaleOpMaker : public framework::OpProtoAndCheckerMaker { The equation is: Out = scale*X )DOC"); - AddAttr("scale", "scale of scale operator.").SetDefault(1.0); + AddAttr("scale", "The scaling factor of the scale operator.") + .SetDefault(1.0); } }; -// Identity Op's gradient is identity op, too. +// The operator to calculate gradients of a scale operator is just the scale +// operator itself. // Grad(Out=scale(X)) => Grad(X) = scale(Grad(Out)) template class ScaleGradOp : public NetOp { @@ -60,38 +62,11 @@ class ScaleGradOp : public NetOp { AppendOp(framework::OpRegistry::CreateOp( "scale", {{"X", {Input(framework::GradVarName("Out"))}}}, {{"Out", {Output(framework::GradVarName("X"))}}}, - {{"scale", GetAttr("scale")}})); + {{"scale", Attr("scale")}})); CompleteAddOp(false); } }; -// identity is a alias of scale op. This is also a example for creating a alias -// operator. -template -class IdentityOpMaker : public framework::OpProtoAndCheckerMaker { - public: - IdentityOpMaker(framework::OpProto *proto, - framework::OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "input tensor of identity op"); - AddOutput("Out", "output tensor of identity op"); - AddComment("identity operator. Just a alias of scale op which scale = 1.0"); - } -}; - -template -class IdentityOp : public NetOp { - public: - IdentityOp(const std::string &type, const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : NetOp(type, inputs, outputs, attrs) { - AppendOp(framework::OpRegistry::CreateOp( - "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}}, - {{"scale", static_cast(1)}})); - } -}; - } // namespace operators } // namespace paddle @@ -101,5 +76,3 @@ REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker, scale_grad, ops::ScaleGradOp); REGISTER_OP_CPU_KERNEL(scale, ops::ScaleKernel); -REGISTER_OP_WITHOUT_GRADIENT(identity, ops::IdentityOp, - ops::IdentityOpMaker); diff --git a/paddle/operators/scale_op.h b/paddle/operators/scale_op.h index 65fb77eefa..02fbdc52bb 100644 --- a/paddle/operators/scale_op.h +++ b/paddle/operators/scale_op.h @@ -27,7 +27,7 @@ class ScaleKernel : public framework::OpKernel { auto* in = context.Input("X"); tensor->mutable_data(in->place()); - auto scale = static_cast(context.GetAttr("scale")); + auto scale = static_cast(context.Attr("scale")); auto eigen_out = framework::EigenVector::Flatten(*tensor); auto eigen_in = framework::EigenVector::Flatten(*in); diff --git a/paddle/operators/scatter_op.cu b/paddle/operators/scatter_op.cu deleted file mode 100644 index 6716b47883..0000000000 --- a/paddle/operators/scatter_op.cu +++ /dev/null @@ -1,20 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -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. */ - -#define EIGEN_USE_GPU -#include "paddle/operators/scatter_op.h" - -namespace ops = paddle::operators; -REGISTER_OP_GPU_KERNEL(scatter, - ops::ScatterOpKernel); diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h index 8422b622ee..f8888f9c36 100644 --- a/paddle/operators/sgd_op.h +++ b/paddle/operators/sgd_op.h @@ -31,7 +31,7 @@ class SGDOpKernel : public framework::OpKernel { auto param = ctx.Input("param"); auto grad = ctx.Input("grad"); auto param_out = ctx.Output("param_out"); - float lr = ctx.GetAttr("learning_rate"); + float lr = ctx.Attr("learning_rate"); param_out->mutable_data(ctx.GetPlace()); diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc index 40c51a64c4..7166b2f60b 100644 --- a/paddle/operators/softmax_op.cc +++ b/paddle/operators/softmax_op.cc @@ -24,7 +24,7 @@ class SoftmaxOp : public framework::OperatorWithKernel { protected: void InferShape(const framework::InferShapeContext &ctx) const override { PADDLE_ENFORCE(ctx.Input("X")->dims().size() == 2UL, - "The input of softmax op must be matrix"); + "The input of softmax op must be a matrix."); ctx.Output("Y")->Resize(ctx.Input("X")->dims()); } }; @@ -34,9 +34,27 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { SoftmaxOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "input of softmax"); - AddOutput("Y", "output of softmax"); - AddComment("Softmax Op"); + AddInput("X", + "The input tensor of softmax. " + "2-D with shape [batch_size, input_feature_dimensions]."); + AddOutput("Y", "The normalized values with the same shape as X."); + AddComment(R"DOC( +The input of softmax operator is a 2-D tensor with shape N x K (N is the +batch_size, K is the dimension of input feature). The output tensor has the +same shape as the input tensor. + +For each row of the input tensor, the softmax operator squashes the +K-dimensional vector of arbitrary real values to a K-dimensional vector of real +values in the range [0, 1] that add up to 1. Specifically, it computes the +exponential of the given dimension and the sum of exponential values of all +the other dimensions in the K-dimensional vector input. Then the ratio of the +exponential of the given dimension and the sum of exponential values of all +the other dimensions is the output of the softmax operator. + +For each row `i` and each column `j` in input X, we have: + Y[i, j] = exp(X[i, j]) / sum_j(exp(X[i, j])) + +)DOC"); } }; @@ -46,14 +64,15 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { protected: void InferShape(const framework::InferShapeContext &ctx) const override { - PADDLE_ENFORCE(ctx.InputVar("Y") != nullptr, "Input(Y) should not be null"); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should be not null."); PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")), - "Input(Y@GRAD) should not be null"); - PADDLE_ENFORCE(ctx.Input("Y")->dims() == - ctx.Input(framework::GradVarName("Y"))->dims(), - "the shape of Input(0) and Input(1) should be the same"); + "Input(Y@GRAD) should be not null."); + PADDLE_ENFORCE_EQ(ctx.Input("Y")->dims(), + ctx.Input(framework::GradVarName("Y"))->dims(), + "Input(Y) and its gradients should have a same shape."); + ctx.Output(framework::GradVarName("X")) - ->Resize(ctx.Input("Y")->dims()); + ->Resize(ctx.Input("X")->dims()); } }; diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h index 4fa6b59540..8a3a5ab927 100644 --- a/paddle/operators/softmax_op.h +++ b/paddle/operators/softmax_op.h @@ -28,12 +28,12 @@ template class SoftmaxKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto input = context.Input("X"); - auto output = context.Output("Y"); - output->mutable_data(context.GetPlace()); + auto X = context.Input("X"); + auto Y = context.Output("Y"); + Y->mutable_data(context.GetPlace()); - auto logits = EigenMatrix::From(*input); - auto softmax = EigenMatrix::From(*output); + auto logits = EigenMatrix::From(*X); + auto softmax = EigenMatrix::From(*Y); const int kBatchDim = 0; const int kClassDim = 1; diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc new file mode 100644 index 0000000000..dc30644a5e --- /dev/null +++ b/paddle/operators/squared_l2_distance_op.cc @@ -0,0 +1,118 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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. */ + +#include "paddle/operators/squared_l2_distance_op.h" + +namespace paddle { +namespace operators { + +class SquaredL2DistanceOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext& ctx) const override { + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), + "Input of SquaredL2DistanceOp " + "must be initialized."); + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), + "Target of SquaredL2DistanceOp " + "must be initialized."); + + auto* x = ctx.Input("X"); + auto x_dims = x->dims(); + auto* y = ctx.Input("Y"); + auto y_dims = y->dims(); + + PADDLE_ENFORCE_EQ(framework::arity(x_dims), framework::arity(y_dims), + "Tensor rank of both SquaredL2DistanceOp's " + "inputs must be same."); + + int rank = framework::arity(x_dims); + PADDLE_ENFORCE_GE(rank, 2, "Tensor rank should be at least equal to 2."); + PADDLE_ENFORCE_EQ(framework::product(x_dims) / x_dims[0], + framework::product(y_dims) / y_dims[0], + "Product of dimensions expcet the first dimension of " + "input and target must be equal."); + PADDLE_ENFORCE(y_dims[0] == 1 || y_dims[0] == x_dims[0], + "First dimension of target must be equal to input " + "or to 1."); + + ctx.Output("sub_result") + ->Resize({static_cast(x_dims[0]), + static_cast(framework::product(x_dims) / x_dims[0])}); + ctx.Output("Out")->Resize({x_dims[0], 1}); + } +}; + +class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SquaredL2DistanceOpMaker(framework::OpProto* proto, + framework::OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of SquaredL2DistanceOp."); + AddInput("Y", "Target of SquaredL2DistanceOp."); + AddOutput("sub_result", + "Buffering substraction result which " + "will be reused in backward.") + .AsIntermediate(); + AddOutput("Out", "Squared l2 distance between input and target."); + AddComment(R"DOC( + SquaredL2DistanceOp will cacluate the squared L2 distance for + input and target. Number of distance value equals to the + first dimension of input. First dimension of target could be equal to + input or to 1. If the first dimension of target is 1, SquaredL2DistanceOp + will broadcast target's first dimension to input's first dimension. + You can decide whether calculate the gradient of input and target. + )DOC"); + } +}; + +class SquaredL2DistanceGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext& ctx) const override { + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")), + "Gradient of Out should not be null"); + auto out_dims = ctx.Input(framework::GradVarName("Out"))->dims(); + auto x_dims = ctx.Input("X")->dims(); + auto y_dims = ctx.Input("Y")->dims(); + PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0], + "First dimension of output gradient and " + "input value must be equal."); + PADDLE_ENFORCE_EQ(out_dims[1], 1, + "Second dimension of output gradient " + "must be 1."); + auto* x_grad = ctx.Output(framework::GradVarName("X")); + auto* y_grad = ctx.Output(framework::GradVarName("Y")); + if (x_grad) x_grad->Resize(x_dims); + if (y_grad) y_grad->Resize(y_dims); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP(squared_l2_distance, ops::SquaredL2DistanceOp, + ops::SquaredL2DistanceOpMaker, squared_l2_distance_grad, + ops::SquaredL2DistanceGradOp); +REGISTER_OP_CPU_KERNEL( + squared_l2_distance, + ops::SquaredL2DistanceKernel); +REGISTER_OP_CPU_KERNEL( + squared_l2_distance_grad, + ops::SquaredL2DistanceGradKernel); diff --git a/paddle/operators/squared_l2_distance_op.cu b/paddle/operators/squared_l2_distance_op.cu new file mode 100644 index 0000000000..3fe62f1a9c --- /dev/null +++ b/paddle/operators/squared_l2_distance_op.cu @@ -0,0 +1,25 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + 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. */ + +#define EIGEN_USE_GPU + +#include "paddle/operators/squared_l2_distance_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_GPU_KERNEL( + squared_l2_distance, + ops::SquaredL2DistanceKernel); +REGISTER_OP_GPU_KERNEL( + squared_l2_distance_grad, + ops::SquaredL2DistanceGradKernel); diff --git a/paddle/operators/squared_l2_distance_op.h b/paddle/operators/squared_l2_distance_op.h new file mode 100644 index 0000000000..ad3347a0b3 --- /dev/null +++ b/paddle/operators/squared_l2_distance_op.h @@ -0,0 +1,123 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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 +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenVector = framework::EigenVector; +template +using EigenMatrix = framework::EigenMatrix; + +template +class SquaredL2DistanceKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* in0 = context.Input("X"); + auto* in1 = context.Input("Y"); + auto* out0 = context.Output("sub_result"); + auto* out1 = context.Output("Out"); + + auto in0_dims = in0->dims(); + auto in1_dims = in1->dims(); + + int cols = framework::product(in0_dims) / in0_dims[0]; + // reduce dimensions except the first + auto x = + EigenMatrix::From(*in0, framework::make_ddim({in0_dims[0], cols})); + auto y = + EigenMatrix::From(*in1, framework::make_ddim({in1_dims[0], cols})); + + out0->mutable_data(context.GetPlace()); + out1->mutable_data(context.GetPlace()); + auto sub_result = EigenMatrix::From(*out0); + auto z = EigenVector::Flatten(*out1); + + auto place = context.GetEigenDevice(); + auto x_dims = x.dimensions(); + auto y_dims = y.dimensions(); + // buffer the substraction result + if (y_dims[0] == 1 && x_dims[0] > y_dims[0]) { + sub_result.device(place) = + x - + y.broadcast(Eigen::array({{static_cast(x_dims[0]), 1}})); + } else { + sub_result.device(place) = x - y; + } + auto sub_res_pow2 = sub_result * sub_result; + z.device(place) = sub_res_pow2.sum(Eigen::array({{1}})); + } +}; + +template +class SquaredL2DistanceGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* in0 = context.Input("sub_result"); + auto* in1 = context.Input(framework::GradVarName("Out")); + auto* x_g = context.Output(framework::GradVarName("X")); + auto* y_g = context.Output(framework::GradVarName("Y")); + + auto sub_result = EigenMatrix::From(*in0); + auto out_grad = EigenMatrix::From(*in1); + + auto x_dims = x_g->dims(); + auto y_dims = y_g->dims(); + + int cols = framework::product(x_dims) / x_dims[0]; + // calculate gradient + auto grad_mat = 2 * + (out_grad.broadcast(Eigen::array({{1, cols}}))) * + sub_result; + + // propagate back to input + auto eigen_place = context.GetEigenDevice(); + if (x_g) { + x_g->mutable_data(context.GetPlace()); + // eigen matrix + auto x_grad = + EigenMatrix::From(*x_g, framework::make_ddim({x_dims[0], cols})); + // dimensions are same with subResult + x_grad.device(eigen_place) = grad_mat; + } + + if (y_g) { + y_g->mutable_data(context.GetPlace()); + + PADDLE_ENFORCE_GE(sub_result.dimensions()[0], y_dims[0], + "First dimension of gradient must be greater or " + "equal than first dimension of target."); + + if (sub_result.dimensions()[0] == y_dims[0]) { + auto y_grad = + EigenMatrix::From(*y_g, framework::make_ddim({y_dims[0], cols})); + y_grad.device(eigen_place) = -1 * grad_mat; + } else { + auto col_sum_res = -1 * (grad_mat.sum(Eigen::array({{0}}))); + auto y_grad = EigenVector::Flatten(*y_g); + y_grad.device(eigen_place) = col_sum_res; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/top_k_op.cc b/paddle/operators/top_k_op.cc new file mode 100644 index 0000000000..38d2f0a09a --- /dev/null +++ b/paddle/operators/top_k_op.cc @@ -0,0 +1,67 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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. */ + +#include "paddle/operators/top_k_op.h" + +namespace paddle { +namespace operators { + +class TopkOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(const framework::InferShapeContext &ctx) const override { + PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), + "Input of TopkOP must be initialized."); + auto *input = ctx.Input("X"); + const int k = static_cast(ctx.Attr("k")); + + PADDLE_ENFORCE_GE(k, 1, "k must >= 1"); + PADDLE_ENFORCE_GE(input->dims().size(), 1, "input must have >= 1d shape"); + PADDLE_ENFORCE_GE(input->dims()[input->dims().size() - 1], k, + "input must have >= k columns"); + + framework::DDim dims = input->dims(); + dims[dims.size() - 1] = k; + ctx.Output("Out")->Resize(dims); + ctx.Output("Indices")->Resize(dims); + } +}; + +class TopkOpMaker : public framework::OpProtoAndCheckerMaker { + public: + TopkOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input of Topk op"); + AddOutput("Out", "The output tensor of Topk op"); + AddOutput("Indices", "The indices of Topk elements of input"); + AddComment( + R"DOC(If the input is a vector (1d tensor), finds the k largest entries in the vector and outputs their values and indices as vectors. Thus values[j] is the j-th largest entry in input, and its index is indices[j]. + + For matrices, computes the top k entries in each row. )DOC"); + AddAttr("k", + "Number of top elements to look for along the last " + "dimension (along each row for matrices).") + .SetDefault(1); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_WITHOUT_GRADIENT(top_k, ops::TopkOp, ops::TopkOpMaker); +REGISTER_OP_CPU_KERNEL(top_k, + ops::TopkKernel); diff --git a/paddle/operators/top_k_op.cu b/paddle/operators/top_k_op.cu new file mode 100644 index 0000000000..afe4d149c5 --- /dev/null +++ b/paddle/operators/top_k_op.cu @@ -0,0 +1,318 @@ +/* Copyright (c) 2016 PaddlePaddle Authors All Rights Reserve. + + 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. */ + +#include "paddle/framework/op_registry.h" +#include "paddle/platform/assert.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +struct Pair { + __device__ __forceinline__ Pair() {} + __device__ __forceinline__ Pair(T value, int id) : v(value), id(id) {} + + __device__ __forceinline__ void set(T value, int id) { + v = value; + id = id; + } + + __device__ __forceinline__ void operator=(const Pair& in) { + v = in.v; + id = in.id; + } + + __device__ __forceinline__ bool operator<(const T value) const { + return (v < value); + } + + __device__ __forceinline__ bool operator<(const Pair& in) const { + return (v < in.v) || ((v == in.v) && (id > in.id)); + } + + __device__ __forceinline__ bool operator>(const Pair& in) const { + return (v > in.v) || ((v == in.v) && (id < in.id)); + } + + T v; + int id; +}; + +template +__device__ __forceinline__ void AddTo(Pair topk[], const Pair& p, + int beam_size) { + for (int k = beam_size - 2; k >= 0; k--) { + if (topk[k] < p) { + topk[k + 1] = topk[k]; + } else { + topk[k + 1] = p; + return; + } + } + topk[0] = p; +} + +template +__device__ __forceinline__ void AddTo(Pair topk[], const Pair& p) { + for (int k = beam_size - 2; k >= 0; k--) { + if (topk[k] < p) { + topk[k + 1] = topk[k]; + } else { + topk[k + 1] = p; + return; + } + } + topk[0] = p; +} + +template +__device__ __forceinline__ void GetTopK(Pair topk[], const T* src, int idx, + int dim, int beam_size) { + while (idx < dim) { + if (topk[beam_size - 1] < src[idx]) { + Pair tmp(src[idx], idx); + AddTo(topk, tmp, beam_size); + } + idx += BlockSize; + } +} + +template +__device__ __forceinline__ void GetTopK(Pair topk[], const T* src, int idx, + int dim, const Pair& max, + int beam_size) { + while (idx < dim) { + if (topk[beam_size - 1] < src[idx]) { + Pair tmp(src[idx], idx); + if (tmp < max) { + AddTo(topk, tmp, beam_size); + } + } + idx += BlockSize; + } +} + +template +__device__ __forceinline__ void GetTopK(Pair topk[], const T* val, int* col, + int idx, int dim, int beam_size) { + while (idx < dim) { + if (topk[beam_size - 1] < val[idx]) { + Pair tmp(val[idx], col[idx]); + AddTo(topk, tmp, beam_size); + } + idx += BlockSize; + } +} + +template +__device__ __forceinline__ void GetTopK(Pair topk[], const T* val, int* col, + int idx, int dim, const Pair& max, + int beam_size) { + while (idx < dim) { + if (topk[beam_size - 1] < val[idx]) { + Pair tmp(val[idx], col[idx]); + if (tmp < max) { + AddTo(topk, tmp, beam_size); + } + } + idx += BlockSize; + } +} + +template +__device__ __forceinline__ void ThreadGetTopK(Pair topk[], int& beam, + int beam_size, const T* src, + bool& firstStep, bool& is_empty, + Pair& max, int dim, + const int tid) { + if (beam > 0) { + int length = beam < beam_size ? beam : beam_size; + if (firstStep) { + firstStep = false; + GetTopK(topk, src, tid, dim, length); + } else { + for (int k = 0; k < MaxLength; k++) { + if (k < MaxLength - beam) { + topk[k] = topk[k + beam]; + } else { + topk[k].set(-INFINITY, -1); + } + } + if (!is_empty) { + GetTopK(topk + MaxLength - beam, src, tid, dim, max, + length); + } + } + + max = topk[MaxLength - 1]; + if (max.v == -1) is_empty = true; + beam = 0; + } +} + +template +__device__ __forceinline__ void ThreadGetTopK(Pair topk[], int& beam, + int beam_size, const T* val, + int* col, bool& firstStep, + bool& is_empty, Pair& max, + int dim, const int tid) { + if (beam > 0) { + int length = beam < beam_size ? beam : beam_size; + if (firstStep) { + firstStep = false; + GetTopK(topk, val, col, tid, dim, length); + } else { + for (int k = 0; k < MaxLength; k++) { + if (k < MaxLength - beam) { + topk[k] = topk[k + beam]; + } else { + topk[k].set(-INFINITY, -1); + } + } + if (!is_empty) { + GetTopK(topk + MaxLength - beam, val, col, tid, dim, max, + length); + } + } + + max = topk[MaxLength - 1]; + if (max.v == -1) is_empty = true; + beam = 0; + } +} + +template +__device__ __forceinline__ void BlockReduce(Pair* sh_topk, int* maxid, + Pair topk[], T** topVal, + int** topIds, int& beam, int& k, + const int tid, const int warp) { + while (true) { + __syncthreads(); + if (tid < BlockSize / 2) { + if (sh_topk[tid] < sh_topk[tid + BlockSize / 2]) { + maxid[tid] = tid + BlockSize / 2; + } else { + maxid[tid] = tid; + } + } + __syncthreads(); + for (int stride = BlockSize / 4; stride > 0; stride = stride / 2) { + if (tid < stride) { + if (sh_topk[maxid[tid]] < sh_topk[maxid[tid + stride]]) { + maxid[tid] = maxid[tid + stride]; + } + } + __syncthreads(); + } + __syncthreads(); + + if (tid == 0) { + **topVal = sh_topk[maxid[0]].v; + **topIds = sh_topk[maxid[0]].id; + (*topVal)++; + (*topIds)++; + } + if (tid == maxid[0]) beam++; + if (--k == 0) break; + __syncthreads(); + + if (tid == maxid[0]) { + if (beam < MaxLength) { + sh_topk[tid] = topk[beam]; + } + } + if (maxid[0] / 32 == warp) { + if (__shfl(beam, (maxid[0]) % 32, 32) == MaxLength) break; + } + } +} + +/** + * Each block compute one sample. + * In a block: + * 1. every thread get top MaxLength value; + * 2. merge to sh_topk, block reduce and get max value; + * 3. go to the second setp, until one thread's topk value is null; + * 4. go to the first setp, until get the topk value. + */ +template +__global__ void KeMatrixTopK(T* output, int output_stride, int* indices, + const T* src, int lds, int dim, int k) { + __shared__ Pair sh_topk[BlockSize]; + __shared__ int maxid[BlockSize / 2]; + const int tid = threadIdx.x; + const int warp = threadIdx.x / 32; + output += blockIdx.x * output_stride; + indices += blockIdx.x * k; + + Pair topk[MaxLength]; + int beam = MaxLength; + Pair max; + bool is_empty = false; + bool firststep = true; + + for (int k = 0; k < MaxLength; k++) { + topk[k].set(-INFINITY, -1); + } + while (k) { + ThreadGetTopK(topk, beam, k, + src + blockIdx.x * lds, firststep, + is_empty, max, dim, tid); + + sh_topk[tid] = topk[0]; + BlockReduce(sh_topk, maxid, topk, &output, + &indices, beam, k, tid, warp); + } +} + +template +class TopkOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use GPUPlace."); + auto* input = ctx.Input("X"); + auto* output = ctx.Output("Out"); + auto* indices = ctx.Output("Indices"); + size_t k = static_cast(ctx.Attr("k")); + + const T* input_data = input->data(); + + T* output_data = output->mutable_data(ctx.GetPlace()); + // FIXME(typhoonzero): data is always converted to type T? + int* indices_data = indices->mutable_data(ctx.GetPlace()); + + size_t input_height = input->dims()[0]; + size_t input_width = input->dims()[1]; + if (k > input_width) k = input_width; + + // NOTE: pass lds and dim same to input width. + // NOTE: old matrix implementation of stride is different to eigen. + // TODO(typhoonzero): launch kernel on specified stream. + // TODO(typhoonzero): refine this kernel. + dim3 threads(256, 1); + dim3 grid(input_height, 1); + + KeMatrixTopK<<>>( + output_data, output->dims()[1], indices_data, input_data, input_width, + input_width, int(k)); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_GPU_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel); diff --git a/paddle/operators/top_k_op.h b/paddle/operators/top_k_op.h new file mode 100644 index 0000000000..ef66acc1d5 --- /dev/null +++ b/paddle/operators/top_k_op.h @@ -0,0 +1,76 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +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 +#include +#include +#include "paddle/framework/eigen.h" +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +using EigenMatrix = framework::EigenMatrix; + +template +class TopkKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + // Get the top k elements of each row of input tensor + // FIXME: only deal with matrix(2d tensor). + auto* input = ctx.Input("X"); + auto* output = ctx.Output("Out"); + auto* indices = ctx.Output("Indices"); + // k is determined by Attr + const size_t k = static_cast(ctx.Attr("k")); + + T* output_data = output->mutable_data(ctx.GetPlace()); + T* indices_data = indices->mutable_data(ctx.GetPlace()); + + auto eg_input = EigenMatrix::From(*input); + + // reshape input to a flattern matrix(like flat_inner_dims) + framework::DDim inputdims = input->dims(); + const size_t row = framework::product( + framework::slice_ddim(inputdims, 0, inputdims.size() - 1)); + const size_t col = inputdims[inputdims.size() - 1]; + Eigen::DSizes flat2dims(row, col); + // NOTE: eigen shape doesn't affect paddle tensor. + eg_input.reshape(flat2dims); + + for (size_t i = 0; i < row; i++) { + std::vector> vec; + for (size_t j = 0; j < col; j++) { + vec.push_back(std::pair(eg_input(i, j), j)); + } + + std::partial_sort( + vec.begin(), vec.begin() + k, vec.end(), + [](const std::pair& l, const std::pair& r) { + return l.first > r.first; + }); + for (size_t j = 0; j < k; j++) { + output_data[i * k + j] = vec[j].first; + indices_data[i * k + j] = vec[j].second; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc index 2d943c4508..f2aeef6c31 100644 --- a/paddle/operators/uniform_random_op.cc +++ b/paddle/operators/uniform_random_op.cc @@ -26,17 +26,17 @@ class CPUUniformRandomKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& context) const override { auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); - unsigned int seed = static_cast(context.GetAttr("seed")); + unsigned int seed = static_cast(context.Attr("seed")); std::minstd_rand engine; if (seed == 0) { seed = std::random_device()(); } engine.seed(seed); std::uniform_real_distribution dist( - static_cast(context.GetAttr("min")), - static_cast(context.GetAttr("max"))); - ssize_t size = framework::product(tensor->dims()); - for (ssize_t i = 0; i < size; ++i) { + static_cast(context.Attr("min")), + static_cast(context.Attr("max"))); + int64_t size = framework::product(tensor->dims()); + for (int64_t i = 0; i < size; ++i) { data[i] = dist(engine); } } @@ -48,11 +48,16 @@ class UniformRandomOp : public framework::OperatorWithKernel { protected: void InferShape(const framework::InferShapeContext& ctx) const override { - PADDLE_ENFORCE(GetAttr("min") < GetAttr("max"), + PADDLE_ENFORCE(Attr("min") < Attr("max"), "uniform_random's min must less then max"); auto* tensor = ctx.Output("Out"); - auto dims = GetAttr>("dims"); - tensor->Resize(framework::make_ddim(dims)); + auto dims = Attr>("dims"); + std::vector temp; + temp.reserve(dims.size()); + for (auto dim : dims) { + temp.push_back(static_cast(dim)); + } + tensor->Resize(framework::make_ddim(temp)); } }; diff --git a/paddle/operators/uniform_random_op.cu b/paddle/operators/uniform_random_op.cu index df993c0779..c2c041b144 100644 --- a/paddle/operators/uniform_random_op.cu +++ b/paddle/operators/uniform_random_op.cu @@ -45,13 +45,13 @@ class GPUUniformRandomKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& context) const override { auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); - unsigned int seed = static_cast(context.GetAttr("seed")); + unsigned int seed = static_cast(context.Attr("seed")); if (seed == 0) { std::random_device rd; seed = rd(); } - T min = static_cast(context.GetAttr("min")); - T max = static_cast(context.GetAttr("max")); + T min = static_cast(context.Attr("min")); + T max = static_cast(context.Attr("max")); thrust::counting_iterator index_sequence_begin(0); ssize_t N = framework::product(tensor->dims()); thrust::transform(index_sequence_begin, index_sequence_begin + N, diff --git a/paddle/platform/cudnn_helper.h b/paddle/platform/cudnn_helper.h index 24ddf3441c..2841d2a2db 100644 --- a/paddle/platform/cudnn_helper.h +++ b/paddle/platform/cudnn_helper.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "paddle/platform/dynload/cudnn.h" #include "paddle/platform/enforce.h" #include "paddle/platform/macros.h" diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc index e86f4dfe26..7d363b3108 100644 --- a/paddle/pybind/pybind.cc +++ b/paddle/pybind/pybind.cc @@ -30,7 +30,7 @@ limitations under the License. */ namespace py = pybind11; -USE_OP(add_two); +USE_OP(add); USE_OP(onehot_cross_entropy); USE_OP(sgd); USE_OP(mul); @@ -47,8 +47,11 @@ USE_OP(lookup_table); USE_OP(scale); USE_NO_KERNEL_OP(identity); USE_OP(minus); +USE_OP(cos_sim); USE_CPU_ONLY_OP(gather); USE_CPU_ONLY_OP(scatter); +USE_OP(top_k); +USE_OP(squared_l2_distance); namespace paddle { namespace framework { @@ -77,7 +80,7 @@ PYBIND11_PLUGIN(core) { .def("get_dims", [](const Tensor &self) { return vectorize(self.dims()); }) .def("set_dims", - [](Tensor &self, const std::vector &dim) { + [](Tensor &self, const std::vector &dim) { self.Resize(make_ddim(dim)); }) .def("alloc_float", diff --git a/paddle/pybind/tensor_py.h b/paddle/pybind/tensor_py.h index 39ba60b4dc..95171acf72 100644 --- a/paddle/pybind/tensor_py.h +++ b/paddle/pybind/tensor_py.h @@ -85,7 +85,7 @@ void PyCPUTensorSetFromArray( framework::Tensor &self, py::array_t array, paddle::platform::CPUPlace &place) { - std::vector dims; + std::vector dims; dims.reserve(array.ndim()); for (size_t i = 0; i < array.ndim(); ++i) { dims.push_back((int)array.shape()[i]); @@ -102,7 +102,7 @@ void PyCUDATensorSetFromArray( framework::Tensor &self, py::array_t array, paddle::platform::GPUPlace &place) { - std::vector dims; + std::vector dims; dims.reserve(array.ndim()); for (size_t i = 0; i < array.ndim(); ++i) { dims.push_back((int)array.shape()[i]); diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index 1798642022..e57f793ac4 100644 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -37,7 +37,7 @@ Configuring cmake in /paddle/build ... -DWITH_PYTHON=${WITH_PYTHON:-ON} -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} -DCUDNN_ROOT=/usr/ - -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF} + -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} -DWITH_TESTING=${WITH_TESTING:-ON} -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ======================================== diff --git a/paddle/scripts/docker/build_android.sh b/paddle/scripts/docker/build_android.sh index 5584e29e2a..aabd2da5e4 100644 --- a/paddle/scripts/docker/build_android.sh +++ b/paddle/scripts/docker/build_android.sh @@ -2,22 +2,58 @@ set -xe -mkdir -p /paddle/build_android -cd /paddle/build_android -rm -rf /paddle/install 2>/dev/null || true -cmake -DCMAKE_SYSTEM_NAME=Android \ - -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \ - -DANDROID_ABI=armeabi-v7a \ - -DANDROID_ARM_NEON=ON \ - -DANDROID_ARM_MODE=ON \ - -DHOST_C_COMPILER=/usr/bin/gcc \ - -DHOST_CXX_COMPILER=/usr/bin/g++ \ - -DCMAKE_INSTALL_PREFIX=/paddle/install \ - -DCMAKE_BUILD_TYPE=RelWithDebInfo \ - -DCMAKE_C_FLAGS_RELWITHDEBINFO="-O3" \ - -DCMAKE_CXX_FLAGS_RELWITHDEBINFO="-O3" \ - -DWITH_C_API=ON \ - -DWITH_SWIG_PY=OFF \ - .. +BUILD_ROOT=/paddle/build_android +DEST_ROOT=/paddle/install + +rm -rf $BUILD_ROOT 2>/dev/null || true +mkdir -p $BUILD_ROOT +cd $BUILD_ROOT + +if [ $ANDROID_ABI == "armeabi-v7a" ]; then + cmake -DCMAKE_SYSTEM_NAME=Android \ + -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \ + -DANDROID_ABI=$ANDROID_ABI \ + -DANDROID_ARM_NEON=ON \ + -DANDROID_ARM_MODE=ON \ + -DHOST_C_COMPILER=/usr/bin/gcc \ + -DHOST_CXX_COMPILER=/usr/bin/g++ \ + -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \ + -DCMAKE_BUILD_TYPE=Release \ + -DUSE_EIGEN_FOR_BLAS=ON \ + -DWITH_C_API=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_STYLE_CHECK=OFF \ + .. +elif [ $ANDROID_ABI == "arm64-v8a" ]; then + cmake -DCMAKE_SYSTEM_NAME=Android \ + -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM64_STANDALONE_TOOLCHAIN \ + -DANDROID_ABI=$ANDROID_ABI \ + -DANDROID_ARM_MODE=ON \ + -DHOST_C_COMPILER=/usr/bin/gcc \ + -DHOST_CXX_COMPILER=/usr/bin/g++ \ + -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \ + -DCMAKE_BUILD_TYPE=Release \ + -DUSE_EIGEN_FOR_BLAS=OFF \ + -DWITH_C_API=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_STYLE_CHECK=OFF \ + .. +elif [ $ANDROID_ABI == "armeabi" ]; then + cmake -DCMAKE_SYSTEM_NAME=Android \ + -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \ + -DANDROID_ABI=$ANDROID_ABI \ + -DANDROID_ARM_MODE=ON \ + -DHOST_C_COMPILER=/usr/bin/gcc \ + -DHOST_CXX_COMPILER=/usr/bin/g++ \ + -DCMAKE_INSTALL_PREFIX=/paddle/install \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_C_API=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_STYLE_CHECK=OFF \ + .. +else + echo "Invalid ANDROID_ABI: $ANDROID_ABI" +fi + make -j `nproc` make install -j `nproc` diff --git a/paddle/scripts/travis/build_android.sh b/paddle/scripts/travis/build_android.sh index 004067a8f5..9da71d1e8c 100755 --- a/paddle/scripts/travis/build_android.sh +++ b/paddle/scripts/travis/build_android.sh @@ -22,6 +22,7 @@ cmake -DCMAKE_SYSTEM_NAME=Android \ -DANDROID_ABI=armeabi-v7a \ -DANDROID_ARM_NEON=ON \ -DANDROID_ARM_MODE=ON \ + -DUSE_EIGEN_FOR_BLAS=ON \ -DWITH_C_API=ON \ -DWITH_SWIG_PY=OFF \ -DWITH_STYLE_CHECK=OFF \ diff --git a/paddle/utils/Util.cpp b/paddle/utils/Util.cpp index b18b73e06a..2755fdd9cd 100644 --- a/paddle/utils/Util.cpp +++ b/paddle/utils/Util.cpp @@ -320,6 +320,9 @@ void loadFileList(const std::string& fileListFileName, } double getMemoryUsage() { +#if defined(__ANDROID__) + return 0.0; +#else FILE* fp = fopen("/proc/meminfo", "r"); CHECK(fp) << "failed to fopen /proc/meminfo"; size_t bufsize = 256 * sizeof(char); @@ -357,6 +360,7 @@ double getMemoryUsage() { delete[] buf; double usedMem = 1.0 - 1.0 * (freeMem + bufMem + cacheMem) / totalMem; return usedMem; +#endif } SyncThreadPool* getGlobalSyncThreadPool() { diff --git a/paddle/utils/Util.h b/paddle/utils/Util.h index 613844669d..22ce2534d3 100644 --- a/paddle/utils/Util.h +++ b/paddle/utils/Util.h @@ -33,6 +33,13 @@ limitations under the License. */ #include "Flags.h" #include "hl_gpu.h" +#if defined(__ANDROID__) && (__ANDROID_API__ < 21) +inline int rand_r(unsigned int* seedp) { + (void)seedp; + return rand(); +} +#endif + /** * Loop over the elements in a container * TODO(yuyang18): It's this foreach useful? Why not use C++ 11 foreach, diff --git a/proto/ModelConfig.proto b/proto/ModelConfig.proto index 4ddf023780..7d7fc23a46 100644 --- a/proto/ModelConfig.proto +++ b/proto/ModelConfig.proto @@ -287,6 +287,11 @@ message PadConfig { repeated uint32 pad_w = 4; } +message ReshapeConfig { + repeated uint32 height_axis = 1; + repeated uint32 width_axis = 2; +} + message MultiBoxLossConfig { required uint32 num_classes = 1; required float overlap_threshold = 2; @@ -339,7 +344,6 @@ message LayerInputConfig { } message LayerConfig { - required string name = 1; required string type = 2; optional uint64 size = 3; @@ -516,6 +520,9 @@ message LayerConfig { optional double delta = 57 [ default = 1.0 ]; optional uint64 depth = 58 [ default = 1 ]; + + // for switch order layer + optional ReshapeConfig reshape_conf = 59; } message EvaluatorConfig { diff --git a/python/paddle/trainer/PyDataProvider2.py b/python/paddle/trainer/PyDataProvider2.py index 7e305e2cd9..248da4ae8d 100644 --- a/python/paddle/trainer/PyDataProvider2.py +++ b/python/paddle/trainer/PyDataProvider2.py @@ -27,6 +27,14 @@ class SequenceType(object): SEQUENCE = 1 SUB_SEQUENCE = 2 + @classmethod + def tostring(cls, value): + for k in cls.__dict__: + if not k.startswith('__'): + if getattr(cls, k) == value: + return cls.__name__ + '.' + k + return 'INVALID(' + str(value) + ')' + # TODO(yuyang18): Add string data type here. class DataType(object): @@ -35,6 +43,14 @@ class DataType(object): SparseValue = 2 Index = 3 + @classmethod + def tostring(cls, value): + for k in cls.__dict__: + if not k.startswith('__'): + if getattr(cls, k) == value: + return cls.__name__ + '.' + k + return 'INVALID(' + str(value) + ')' + class CacheType(object): NO_CACHE = 0 # No cache at all @@ -69,6 +85,26 @@ class InputType(object): self.seq_type = seq_type self.type = tp + def __repr__(self): + """ + Return a human readable representation like 'InputType(dim=25921, + seq_type=SequenceType.NO_SEQUENCE, type=DataType.Dense)' + """ + repr_str = type(self).__name__ + repr_str += '(' + serialize_func_map = { + 'dim': repr, + 'seq_type': SequenceType.tostring, + 'type': DataType.tostring + } + for idx, k in enumerate(self.__slots__): + if idx != 0: + repr_str += ', ' + repr_str += ( + k + '=' + serialize_func_map.get(k, repr)(getattr(self, k))) + repr_str += ')' + return repr_str + def dense_slot(dim, seq_type=SequenceType.NO_SEQUENCE): """ diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py index 152a56190c..11dc84ae20 100644 --- a/python/paddle/trainer/config_parser.py +++ b/python/paddle/trainer/config_parser.py @@ -3670,6 +3670,15 @@ class RecurrentLayerGroup(LayerBase): name, 'recurrent_layer_group', 0, inputs=[], device=device) +@config_layer('switch_order') +class SwitchOrderLayer(LayerBase): + def __init__(self, name, inputs, reshape, **xargs): + super(SwitchOrderLayer, self).__init__( + name, 'switch_order', 0, inputs=inputs, **xargs) + self.config.reshape_conf.heightAxis.extend(reshape['height']) + self.config.reshape_conf.widthAxis.extend(reshape['width']) + + # Deprecated, use a new layer specific class instead @config_func def Layer(name, type, **xargs): diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index 47ac601e67..cba45bd3af 100644 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -131,6 +131,7 @@ __all__ = [ 'row_conv_layer', 'dropout_layer', 'prelu_layer', + 'switch_order_layer', 'gated_unit_layer', 'crop_layer', 'sub_nested_seq_layer', @@ -239,6 +240,7 @@ class LayerType(object): SMOOTH_L1 = 'smooth_l1' PRELU = 'prelu' + SWITCH_ORDER_LAYER = 'switch_order' CROP_LAYER = 'crop' SUB_NESTED_SEQ = 'sub_nested_seq' CLIP_LAYER = 'clip' @@ -6404,6 +6406,48 @@ def gated_unit_layer(input, layer_attr=layer_attr) +@layer_support() +@wrap_name_default('switch_order') +def switch_order_layer(input, + name=None, + reshape=None, + act=None, + layer_attr=None): + """ + This layer switch dimension order of image input. + From order "batchSize, channels, height, width" + to order "batchSize, height, width, channels". + + The example usage is: + + .. code-block:: python + reshape = {'height':[ 0, 1, 2], 'width':[3]} + switch = switch_order(input=layer, name='switch', reshape=reshape) + + :param input: The input layer. + :type input: LayerOutput + :param name: Name of this layer. + :type name: basestring + :param reshape: reshape matrix by axises. + :type reshape: Dict + :return: LayerOutput object. + :rtype: LayerOutput + """ + assert isinstance(input, LayerOutput) + l = Layer( + name=name, + inputs=input.name, + reshape=reshape, + type=LayerType.SWITCH_ORDER_LAYER, + active_type=act.name, + **ExtraLayerAttribute.to_kwargs(layer_attr)) + return LayerOutput( + name=name, + layer_type=LayerType.SWITCH_ORDER_LAYER, + parents=input, + size=l.config.size) + + @wrap_name_default() @layer_support() def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None): diff --git a/python/paddle/v2/framework/op.py b/python/paddle/v2/framework/op.py index e7e932f6fe..c1585bcffc 100644 --- a/python/paddle/v2/framework/op.py +++ b/python/paddle/v2/framework/op.py @@ -4,8 +4,8 @@ import paddle.v2.framework.proto.framework_pb2 as framework_pb2 def get_all_op_protos(): """ - Get all registered op proto from Paddle C++ - :return: list of OpProto + Get all registered op proto from PaddlePaddle C++ end. + :return: A list of registered OpProto. """ protostrs = core.get_all_op_protos() ret_values = [] @@ -21,8 +21,8 @@ def is_str(s): class OpDescCreationMethod(object): """ - A Functor object to convert user input(use key word args) to OpDesc based on - OpProto. + Convert the user's input(only keyword arguments are supported) to OpDesc + based on the OpProto. :param op_proto: The OpProto object. :type op_proto: op_proto_pb2.OpProto @@ -30,17 +30,18 @@ class OpDescCreationMethod(object): def __init__(self, op_proto): if not isinstance(op_proto, framework_pb2.OpProto): - raise TypeError("Argument should be OpProto") + raise TypeError( + "Type of op_proto should be OpProto in PaddlePaddle.") self.__op_proto__ = op_proto def __call__(self, *args, **kwargs): """ - Convert user input to OpDesc. Only key-word args are supported. - :return: OpDesc based on user input + Convert user's input to OpDesc. Only keyword arguments are supported. + :return: The OpDesc based on user input. :rtype: op_desc_pb2.OpDesc """ if len(args) != 0: - raise ValueError("Only keyword arguments is supported by Paddle") + raise ValueError("Only keyword arguments are supported.") op_desc = framework_pb2.OpDesc() for input_parameter in self.__op_proto__.inputs: @@ -49,8 +50,9 @@ class OpDescCreationMethod(object): input_arguments = [input_arguments] if not input_parameter.duplicable and len(input_arguments) > 1: - raise ValueError("Input %s only accepts one input, but give %d" - % (input_parameter.name, len(input_arguments))) + raise ValueError( + "Input %s expects only one input, but %d are given." % + (input_parameter.name, len(input_arguments))) ipt = op_desc.inputs.add() ipt.parameter = input_parameter.name @@ -63,7 +65,7 @@ class OpDescCreationMethod(object): if not output_parameter.duplicable and len(output_arguments) > 1: raise ValueError( - "Output %s only accepts one output, but give %d" % + "Output %s expects only one output, but %d are given." % (output_parameter.name, len(output_arguments))) out = op_desc.outputs.add() @@ -94,16 +96,23 @@ class OpDescCreationMethod(object): new_attr.floats.extend(user_defined_attr) elif attr.type == framework_pb2.STRINGS: new_attr.strings.extend(user_defined_attr) + elif attr.type == framework_pb2.INT_PAIRS: + for p in user_defined_attr: + pair = new_attr.pairs.add() + pair.first = p[0] + pair.second = p[1] else: - raise NotImplementedError("Not support attribute type " + - attr.type) + raise NotImplementedError( + "A not supported attribute type: %s." % ( + str(attr.type))) return op_desc @staticmethod def any_is_true(generator): """ - Reduce a bool array to one. If any of them is True, then return True. + Reduce a boolean array to a single boolean parameter. If any element in + the array is True, this function will return True, otherwise False. """ for flag in generator: if flag: @@ -122,7 +131,7 @@ class OpInfo(object): def create_op_creation_method(op_proto): """ - Generate op creation method for an OpProto + Generate op creation method for an OpProto. """ method = OpDescCreationMethod(op_proto) @@ -141,20 +150,23 @@ def create_op_creation_method(op_proto): class OperatorFactory(object): def __init__(self): self.op_methods = dict() + for op_proto in get_all_op_protos(): method = create_op_creation_method(op_proto) self.op_methods[method.name] = method def __call__(self, *args, **kwargs): - if 'type' in kwargs: + if "type" in kwargs: if len(args) != 0: - raise ValueError("All Paddle argument should be key-word " - "argument except type") - t = kwargs.pop('type') + raise ValueError( + "Except the argument \"type\"," + "all of the other arguments should be keyword arguments.") + t = kwargs.pop("type") else: if len(args) != 1: - raise ValueError("All Paddle argument should be key-word " - "argument except type") + raise ValueError( + "Except the argument \"type\"," + "all of the other arguments should be keyword arguments.") t = args[0] return self.get_op_info(t).method(**kwargs) @@ -164,7 +176,7 @@ class OperatorFactory(object): def get_op_info(self, t): if t not in self.op_methods: - raise ValueError("operator %s is not registered", t) + raise ValueError("The operator: %s is not registered." % t) return self.op_methods.get(t) def get_op_input_names(self, type): @@ -179,7 +191,7 @@ class OperatorFactory(object): class __RecurrentOp__(object): __proto__ = None - type = 'recurrent' + type = "recurrent" def __init__(self): # cache recurrent_op's proto @@ -189,8 +201,8 @@ class __RecurrentOp__(object): self.__proto__ = op_proto def __call__(self, *args, **kwargs): - if self.type not in args and 'type' not in kwargs: - kwargs['type'] = self.type + if self.type not in args and "type" not in kwargs: + kwargs["type"] = self.type # create proto create_method = OpDescCreationMethod(self.__proto__) proto = create_method(*args, **kwargs) @@ -198,5 +210,5 @@ class __RecurrentOp__(object): return core.RecurrentOp.create(proto.SerializeToString()) -Operator = OperatorFactory() # Default global factory +Operator = OperatorFactory() # The default global factory RecurrentOp = __RecurrentOp__() diff --git a/python/paddle/v2/framework/tests/CMakeLists.txt b/python/paddle/v2/framework/tests/CMakeLists.txt index 661ebd8964..ef910f939b 100644 --- a/python/paddle/v2/framework/tests/CMakeLists.txt +++ b/python/paddle/v2/framework/tests/CMakeLists.txt @@ -4,6 +4,7 @@ py_test(test_scope SRCS test_scope.py) py_test(test_tensor SRCS test_tensor.py) py_test(test_mul_op SRCS test_mul_op.py) +py_test(test_cos_sim_op SRCS test_cos_sim_op.py) py_test(test_mean_op SRCS test_mean_op.py) @@ -16,6 +17,7 @@ py_test(test_cross_entropy_op SRCS test_cross_entropy_op.py) py_test(test_gather_op SRCS test_gather_op.py) py_test(test_scatter_op SRCS test_scatter_op.py) py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py) +py_test(test_top_k_op SRCS test_top_k_op.py) py_test(gradient_checker SRCS gradient_checker.py) @@ -32,3 +34,4 @@ py_test(test_gradient_checker SRCS test_gradient_checker.py) py_test(test_lookup_table SRCS test_lookup_table.py) py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py) py_test(mnist SRCS mnist.py) +py_test(test_squared_l2_distance_op SRCS test_squared_l2_distance_op.py) diff --git a/python/paddle/v2/framework/tests/gradient_checker.py b/python/paddle/v2/framework/tests/gradient_checker.py index 518f828bac..fdb06b7988 100644 --- a/python/paddle/v2/framework/tests/gradient_checker.py +++ b/python/paddle/v2/framework/tests/gradient_checker.py @@ -36,13 +36,13 @@ def get_numeric_gradient(op, in_place=False): """ Get Numeric Gradient for an operator's input. - - :param op: C++ operator instance, could be an network - :param input_values: The input variables. Should be an dictionary, key is + + :param op: C++ operator instance, could be an network + :param input_values: The input variables. Should be an dictionary, key is variable name. Value is numpy array. - :param output_name: The final output variable name. + :param output_name: The final output variable name. :param input_to_check: The input variable need to get gradient. - :param delta: The perturbation value for numeric gradient method. The + :param delta: The perturbation value for numeric gradient method. The smaller delta is, the more accurate result will get. But if that delta is too small, it could occur numerical stability problem. :param local_scope: The local scope used for get_numeric_gradient. @@ -229,9 +229,9 @@ class GradientChecker(unittest.TestCase): """Use relative error for the comparison. :param numeric_grads: the numerical graidents. - :type numeric_grads: a list of numpy.array + :type numeric_grads: a list of numpy.array :param analytic_grads: the analytical graidents. - :type analytic_grads: a list of numpy.array + :type analytic_grads: a list of numpy.array :param name: the names of gradients, used to print for debug. :type names: a list of string :param msg_prefix: string info, used to print for debug. @@ -286,6 +286,9 @@ class GradientChecker(unittest.TestCase): for no_grad in no_grad_set: if no_grad not in in_names: raise ValueError("no_grad should be in in_names") + if no_grad in inputs_to_check: + raise ValueError("no_grad should not be in inputs_to_check") + backward_op = core.Operator.backward(forward_op, no_grad_set) places = [core.CPUPlace()] @@ -301,7 +304,6 @@ class GradientChecker(unittest.TestCase): check_names = [grad_var_name(name) for name in inputs_to_check] for place in places: - # get analytical gradients according to different device analytic_grads = self.__get_gradient(forward_op, backward_op, input_vars, check_names, place) self.__assert_is_close(numeric_grads, analytic_grads, check_names, diff --git a/python/paddle/v2/framework/tests/mnist.py b/python/paddle/v2/framework/tests/mnist.py index a68f302f9c..f6f8f49b79 100644 --- a/python/paddle/v2/framework/tests/mnist.py +++ b/python/paddle/v2/framework/tests/mnist.py @@ -38,9 +38,9 @@ def feed_data(name, data): assert isinstance(data, numpy.ndarray) tensor = scope.find_var(name).get_tensor() tensor.set_dims(data.shape) - if data.dtype == numpy.dtype('int32'): + if data.dtype == numpy.dtype("int32"): tensor.alloc_int(place) - elif data.dtype == numpy.dtype('float32'): + elif data.dtype == numpy.dtype("float32"): tensor.alloc_float(place) else: raise ValueError("data type not supported") @@ -74,22 +74,25 @@ def init_param(net, param_name, dims): # fc_layer def fc_layer(net, input, size, act="softmax", bias=True, param=None, name=None): """ - Add a fc layer to net + The fully connected layer. - :param input: input variable name. + :param input: The name of input variable. :type input: str - :param size: fully connected layer size. - :param act: activation name - :param param: parameter attribute, used for initialize parameters. - :param bias: bias attribute. False will not have a bias. - :param name: the name of fc layer. If not set, model will generate a - readable name - :return: output variable name. + :param size: The size of fully connected layer. + :param act: The name of activation. + :param param: The attribute of learnable parameter which can be used to + modify initialization mean and std of the parameter. + :param bias: The attribute of bias. If set False, this layer does not have + a bias. + :param name: The name of this layer. If it is not set explictly, a name + will be generated automatically. + :return: The name of the output variable. """ + if name is None: - name = 'fc_%d' % uniq_id() + name = "fc_%d" % uniq_id() if not isinstance(name, str): - raise ValueError("name should be string") + raise ValueError("The name of a layer should be a string.") input_dims = scope.find_var(input).get_tensor().get_dims() @@ -123,7 +126,7 @@ def fc_layer(net, input, size, act="softmax", bias=True, param=None, name=None): def cross_entropy_layer(net, input, label): - cost_name = 'cross_entropy_%d' % uniq_id() + cost_name = "cross_entropy_%d" % uniq_id() cross_entropy_op = Operator( "onehot_cross_entropy", X=input, label=label, Y=cost_name) net.append_op(cross_entropy_op) @@ -177,8 +180,8 @@ def error_rate(predict, label): return error_num / float(len(label)) -images = data_layer(name='pixel', dims=[BATCH_SIZE, 784]) -labels = data_layer(name='label', dims=[BATCH_SIZE]) +images = data_layer(name="pixel", dims=[BATCH_SIZE, 784]) +labels = data_layer(name="label", dims=[BATCH_SIZE]) fc1 = fc_layer(net=forward_net, input=images, size=100, act="sigmoid") fc2 = fc_layer(net=forward_net, input=fc1, size=100, act="sigmoid") predict = fc_layer(net=forward_net, input=fc2, size=10, act="softmax") diff --git a/python/paddle/v2/framework/tests/op_test_util.py b/python/paddle/v2/framework/tests/op_test_util.py index 3bc05a0fec..370f27eaf6 100644 --- a/python/paddle/v2/framework/tests/op_test_util.py +++ b/python/paddle/v2/framework/tests/op_test_util.py @@ -6,13 +6,13 @@ from paddle.v2.framework.op import Operator class OpTestMeta(type): """ Operator Test ClassMeta. - - It injects `test_all` method into user's OperatorTest class, to make Python + + It injects `test_all` method into user's OperatorTest class, to make Python unittest module run that method. - + The `test_all` read what value is stored in `self`. It use self's values to create and run a operator, and check whether that op is OK or not. - + See `test_add_two_op` for example usage. """ @@ -66,7 +66,7 @@ class OpTestMeta(type): self.assertTrue( numpy.allclose( actual, expect, atol=1e-05), - "output name: " + out_name + "has diff") + "output name: " + out_name + " has diff") obj.test_all = test_all return obj diff --git a/python/paddle/v2/framework/tests/test_add_two_op.py b/python/paddle/v2/framework/tests/test_add_two_op.py index 0def484edd..a578e74eca 100644 --- a/python/paddle/v2/framework/tests/test_add_two_op.py +++ b/python/paddle/v2/framework/tests/test_add_two_op.py @@ -11,7 +11,7 @@ class TestAddOp(unittest.TestCase): __metaclass__ = OpTestMeta def setUp(self): - self.type = "add_two" + self.type = "add" self.inputs = { 'X': numpy.random.random((102, 105)).astype("float32"), 'Y': numpy.random.random((102, 105)).astype("float32") diff --git a/python/paddle/v2/framework/tests/test_cos_sim_op.py b/python/paddle/v2/framework/tests/test_cos_sim_op.py new file mode 100644 index 0000000000..32013a7999 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_cos_sim_op.py @@ -0,0 +1,60 @@ +import unittest +import numpy as np +from gradient_checker import GradientChecker, create_op +from op_test_util import OpTestMeta + + +class TestCosSimOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "cos_sim" + self.inputs = { + 'X': np.random.random((32, 64)).astype("float32"), + 'Y': np.random.random((32, 64)).astype("float32") + } + expect_x_norm = np.linalg.norm(self.inputs['X'], axis=1) + expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=1) + expect_out = (self.inputs['X'] * self.inputs['Y']).sum(axis=1) / \ + expect_x_norm / expect_y_norm + self.outputs = { + 'XNorm': np.expand_dims(expect_x_norm, 1), + 'YNorm': np.expand_dims(expect_y_norm, 1), + 'Out': np.expand_dims(expect_out, 1) + } + + +class TestCosSimGradOp(GradientChecker): + def setUp(self): + self.op = create_op("cos_sim") + self.inputs = { + 'X': np.random.random((10, 5)).astype("float32"), + 'Y': np.random.random((10, 5)).astype("float32") + } + + def test_cpu_gpu_compare(self): + self.compare_grad(self.op, self.inputs) + + def test_normal(self): + self.check_grad( + self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.05) + + def test_ignore_x(self): + self.check_grad( + self.op, + self.inputs, ["Y"], + "Out", + max_relative_error=0.05, + no_grad_set={"X"}) + + def test_ignore_y(self): + self.check_grad( + self.op, + self.inputs, ["X"], + "Out", + max_relative_error=0.05, + no_grad_set={"Y"}) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_gradient_checker.py b/python/paddle/v2/framework/tests/test_gradient_checker.py index e0b3151208..e8a7f848df 100644 --- a/python/paddle/v2/framework/tests/test_gradient_checker.py +++ b/python/paddle/v2/framework/tests/test_gradient_checker.py @@ -7,11 +7,11 @@ from gradient_checker import get_numeric_gradient class GetNumericGradientTest(unittest.TestCase): def test_add_op(self): - add_op = Operator('add_two', X="X", Y="Y", Out="Z") + add_op = Operator("add", X="X", Y="Y", Out="Z") x = numpy.random.random((10, 1)).astype("float32") y = numpy.random.random((10, 1)).astype("float32") - arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X') + arr = get_numeric_gradient(add_op, {"X": x, "Y": y}, "Z", "X") self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-4) def test_softmax_op(self): @@ -35,9 +35,9 @@ class GetNumericGradientTest(unittest.TestCase): dY = numpy.ones(Y.shape) dX = label_softmax_grad(Y, dY) - arr = get_numeric_gradient(softmax_op, {"X": X}, 'Y', 'X') + arr = get_numeric_gradient(softmax_op, {"X": X}, "Y", "X") numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2) -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_mul_op.py b/python/paddle/v2/framework/tests/test_mul_op.py index ee0d81a64e..8c827e242e 100644 --- a/python/paddle/v2/framework/tests/test_mul_op.py +++ b/python/paddle/v2/framework/tests/test_mul_op.py @@ -2,6 +2,7 @@ import unittest import numpy as np from gradient_checker import GradientChecker, create_op from op_test_util import OpTestMeta +from paddle.v2.framework.op import Operator class TestMulOp(unittest.TestCase): @@ -16,19 +17,87 @@ class TestMulOp(unittest.TestCase): self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} -class MulGradOpTest(GradientChecker): - def test_mul(self): - op = create_op("mul") - inputs = { +class TestMulOp2(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "mul" + self.inputs = { + 'X': np.random.random((15, 4, 12, 10)).astype("float32"), + 'Y': np.random.random((4, 30, 8, 2, 9)).astype("float32") + } + self.attrs = {'x_num_col_dims': 2, 'y_num_col_dims': 2} + self.outputs = { + 'Out': np.dot(self.inputs['X'].reshape(15 * 4, 12 * 10), + self.inputs['Y'].reshape(4 * 30, 8 * 2 * 9)) + } + + +class TestMulGradOp(GradientChecker): + def setUp(self): + self.op = create_op("mul") + self.inputs = { 'X': np.random.random((32, 84)).astype("float32"), 'Y': np.random.random((84, 100)).astype("float32") } + + def test_cpu_gpu_compare(self): + self.compare_grad(self.op, self.inputs) + + def test_normal(self): # mul op will enlarge the relative error self.check_grad( - op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5) + self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5) + + def test_ignore_x(self): + self.check_grad( + self.op, + self.inputs, ["Y"], + "Out", + max_relative_error=0.5, + no_grad_set={"X"}) + + def test_ignore_y(self): + self.check_grad( + self.op, + self.inputs, ["X"], + "Out", + max_relative_error=0.5, + no_grad_set={"Y"}) + + +class TestMulGradTest2(GradientChecker): + def setUp(self): + self.op = Operator( + "mul", X="X", Y="Y", Out="Out", x_num_col_dims=2, y_num_col_dims=2) + self.inputs = { + "X": np.random.random((15, 4, 12, 10)).astype("float32"), + "Y": np.random.random((4, 30, 8, 2, 9)).astype("float32") + } + + def test_cpu_gpu_compare(self): + self.compare_grad(self.op, self.inputs) + + def test_normal(self): + self.check_grad( + self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5) + def test_ignore_x(self): + self.check_grad( + self.op, + self.inputs, ["Y"], + "Out", + max_relative_error=0.5, + no_grad_set={"X"}) + + def test_ignore_y(self): + self.check_grad( + self.op, + self.inputs, ["X"], + "Out", + max_relative_error=0.5, + no_grad_set={"Y"}) -# TODO(dzh,qijun) : mulgrad test case need transpose feature of blas library if __name__ == '__main__': unittest.main() diff --git a/python/paddle/v2/framework/tests/test_net.py b/python/paddle/v2/framework/tests/test_net.py index 9339cf28da..e4b7cd480c 100644 --- a/python/paddle/v2/framework/tests/test_net.py +++ b/python/paddle/v2/framework/tests/test_net.py @@ -15,7 +15,7 @@ def fc(X, W, Y): class TestNet(unittest.TestCase): def test_net_all(self): net = core.Net.create() - op1 = Operator("add_two", X="X", Y="Y", Out="Out") + op1 = Operator("add", X="X", Y="Y", Out="Out") net.append_op(op1) net2 = core.Net.create() @@ -26,7 +26,7 @@ class TestNet(unittest.TestCase): expected = ''' Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]}. - Op(add_two), inputs:{X[X], Y[Y]}, outputs:{Out[Out]}. + Op(add), inputs:{X[X], Y[Y]}, outputs:{Out[Out]}. Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}. Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}. Op(mul), inputs:{X[X], Y[W]}, outputs:{Out[pre_activation]}. diff --git a/python/paddle/v2/framework/tests/test_operator.py b/python/paddle/v2/framework/tests/test_operator.py index 1abc4eeb57..040556322d 100644 --- a/python/paddle/v2/framework/tests/test_operator.py +++ b/python/paddle/v2/framework/tests/test_operator.py @@ -193,10 +193,10 @@ class TestOpDescCreationMethod(unittest.TestCase): class TestOpCreations(unittest.TestCase): def test_all(self): - add_op = op.Operator("add_two", X="a", Y="b", Out="z") + add_op = op.Operator("add", X="a", Y="b", Out="z") self.assertIsNotNone(add_op) # Invoke C++ DebugString() - self.assertEqual('Op(add_two), inputs:{X[a], Y[b]}, outputs:{Out[z]}.', + self.assertEqual('Op(add), inputs:{X[a], Y[b]}, outputs:{Out[z]}.', str(add_op)) diff --git a/python/paddle/v2/framework/tests/test_recurrent_op.py b/python/paddle/v2/framework/tests/test_recurrent_op.py index d6000ab9f9..22e680fd78 100644 --- a/python/paddle/v2/framework/tests/test_recurrent_op.py +++ b/python/paddle/v2/framework/tests/test_recurrent_op.py @@ -146,7 +146,7 @@ class TestRecurrentOp(unittest.TestCase): stepnet = core.Net.create() x_fc_op = Operator("mul", X="x@alias", Y="W", Out="Wx") h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh") - sum_op = Operator("add_two", X="Wx", Y="Uh", Out="sum") + sum_op = Operator("add", X="Wx", Y="Uh", Out="sum") sig_op = Operator("sigmoid", X="sum", Y="h@alias") for op in [x_fc_op, h_fc_op, sum_op, sig_op]: diff --git a/python/paddle/v2/framework/tests/test_rowwise_add_op.py b/python/paddle/v2/framework/tests/test_rowwise_add_op.py index 45d569da29..8378c1cd21 100644 --- a/python/paddle/v2/framework/tests/test_rowwise_add_op.py +++ b/python/paddle/v2/framework/tests/test_rowwise_add_op.py @@ -16,14 +16,52 @@ class TestRowwiseAddOp(unittest.TestCase): self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])} -class RowwiseAddGradOpTest(GradientChecker): - def test_rowwise_add(self): - op = create_op("rowwise_add") - inputs = { +class TestRowwiseAddOp2(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "rowwise_add" + self.inputs = { + 'X': np.random.random((13, 6, 7, 8)).astype("float32"), + 'b': np.random.random((7, 8)).astype("float32") + } + self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])} + + +class TestRowwiseAddGradOp(GradientChecker): + def setUp(self): + self.op = create_op("rowwise_add") + self.inputs = { "X": np.random.uniform(0.1, 1, [5, 10]).astype("float32"), "b": np.random.uniform(0.1, 1, [10]).astype("float32") } - self.check_grad(op, inputs, set(["X", "b"]), "Out") + + def test_normal(self): + self.check_grad(self.op, self.inputs, ["X", "b"], "Out") + + def test_ignore_b(self): + self.check_grad(self.op, self.inputs, ["X"], "Out", no_grad_set={"b"}) + + def test_ignore_x(self): + self.check_grad(self.op, self.inputs, ["b"], "Out", no_grad_set={"X"}) + + +class TestRowwiseAddGradOp2(GradientChecker): + def setUp(self): + self.op = create_op("rowwise_add") + self.inputs = { + "X": np.random.uniform(0.1, 1, [2, 3, 2, 5]).astype("float32"), + "b": np.random.uniform(0.1, 1, [2, 5]).astype("float32") + } + + def test_normal(self): + self.check_grad(self.op, self.inputs, ["X", "b"], "Out") + + def test_ignore_b(self): + self.check_grad(self.op, self.inputs, ["X"], "Out", no_grad_set={"b"}) + + def test_ignore_x(self): + self.check_grad(self.op, self.inputs, ["b"], "Out", no_grad_set={"X"}) if __name__ == '__main__': diff --git a/python/paddle/v2/framework/tests/test_softmax_op.py b/python/paddle/v2/framework/tests/test_softmax_op.py index e670d93653..0d590fa706 100644 --- a/python/paddle/v2/framework/tests/test_softmax_op.py +++ b/python/paddle/v2/framework/tests/test_softmax_op.py @@ -18,18 +18,22 @@ class TestSoftmaxOp(unittest.TestCase): def setUp(self): self.type = "softmax" - self.inputs = {'X': np.random.random((32, 100)).astype("float32")} + self.inputs = {"X": np.random.random((10, 10)).astype("float32")} self.outputs = { - 'Y': np.apply_along_axis(stable_softmax, 1, self.inputs['X']) + "Y": np.apply_along_axis(stable_softmax, 1, self.inputs["X"]) } -class SoftmaxGradOpTest(GradientChecker): - def test_softmax(self): - op = create_op("softmax") - inputs = {"X": np.random.uniform(0.1, 1, [10, 10]).astype("float32")} - self.check_grad(op, inputs, set("X"), "Y") +class TestSoftmaxGradOp(GradientChecker): + def setUp(self): + self.op = create_op("softmax") + self.inputs = { + "X": np.random.uniform(0.1, 1, [10, 10]).astype("float32") + } + + def test_softmax_grad(self): + self.check_grad(self.op, self.inputs, ["X"], "Y") -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py b/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py new file mode 100644 index 0000000000..2bcdf37df4 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_squared_l2_distance_op.py @@ -0,0 +1,89 @@ +import unittest +from op_test_util import OpTestMeta +from gradient_checker import GradientChecker, create_op +import numpy as np + + +class TestSquaredL2DistanceOp_f0(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = 'squared_l2_distance' + self.inputs = { + 'X': np.random.uniform(0.1, 1., (32, 64)).astype('float32'), + 'Y': np.random.uniform(0.1, 1., (32, 64)).astype('float32') + } + sub_res = self.inputs['X'] - self.inputs['Y'] + output = sub_res * sub_res + self.outputs = { + 'sub_result': sub_res, + 'Out': np.expand_dims(output.sum(1), 1) + } + + +class TestSquaredL2DistanceOp_f1(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = 'squared_l2_distance' + self.inputs = { + 'X': np.random.uniform(0.1, 1., (32, 64)).astype('float32'), + 'Y': np.random.uniform(0.1, 1., (1, 64)).astype('float32') + } + sub_res = self.inputs['X'] - self.inputs['Y'] + output = sub_res * sub_res + self.outputs = { + 'sub_result': sub_res, + 'Out': np.expand_dims(output.sum(1), 1) + } + + +class TestSquaredL2DistanceOp_f2(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = 'squared_l2_distance' + self.inputs = { + 'X': np.random.uniform(0.1, 1., (32, 64, 128)).astype('float32'), + 'Y': np.random.uniform(0.1, 1., (1, 64, 128)).astype('float32') + } + sub_res = self.inputs['X'] - self.inputs['Y'] + sub_res = sub_res.reshape((32, 64 * 128)) + output = sub_res * sub_res + self.outputs = { + 'sub_result': sub_res, + 'Out': np.expand_dims(output.sum(1), 1) + } + + +class TestSquaredL2DistanceGradOp(GradientChecker): + def test_squared_l2_distance_b0(self): + op = create_op("squared_l2_distance") + inputs = { + 'X': np.random.uniform(0.1, .6, (2, 3)).astype('float32'), + 'Y': np.random.uniform(0.1, .6, (2, 3)).astype('float32') + } + self.compare_grad(op, inputs) + self.check_grad(op, inputs, set(["X", "Y"]), "Out") + + def test_squared_l2_distance_b1(self): + op = create_op("squared_l2_distance") + inputs = { + 'X': np.random.uniform(0.1, .6, (2, 3)).astype('float32'), + 'Y': np.random.uniform(0.1, .6, (1, 3)).astype('float32') + } + self.compare_grad(op, inputs) + self.check_grad(op, inputs, set(["X", "Y"]), "Out") + + def test_squared_l2_distance_b2(self): + op = create_op("squared_l2_distance") + inputs = { + 'X': np.random.uniform(0.1, .6, (2, 3, 4)).astype('float32'), + 'Y': np.random.uniform(0.1, .6, (1, 3, 4)).astype('float32') + } + self.compare_grad(op, inputs) + self.check_grad(op, inputs, set(["X", "Y"]), "Out") + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/framework/tests/test_top_k_op.py b/python/paddle/v2/framework/tests/test_top_k_op.py new file mode 100644 index 0000000000..e841d96d26 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_top_k_op.py @@ -0,0 +1,52 @@ +import unittest +import numpy as np +from gradient_checker import GradientChecker, create_op +from op_test_util import OpTestMeta + + +class TestTopkOp(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "top_k" + k = 1 + input = np.random.random((32, 84)).astype("float32") + output = np.ndarray((32, k)) + indices = np.ndarray((32, k)) + + self.inputs = {'X': input} + self.attrs = {'k': k} + + for rowid in xrange(32): + row = input[rowid] + output[rowid] = np.sort(row)[-k:] + indices[rowid] = row.argsort()[-k:] + + self.outputs = {'Out': output, 'Indices': indices} + + +class TestTopkOp3d(unittest.TestCase): + __metaclass__ = OpTestMeta + + def setUp(self): + self.type = "top_k" + k = 1 + input = np.random.random((32, 2, 84)).astype("float32") + input_flat_2d = input.reshape(64, 84) + output = np.ndarray((64, k)) + indices = np.ndarray((64, k)).astype("int") + + # FIXME: should use 'X': input for a 3d input + self.inputs = {'X': input_flat_2d} + self.attrs = {'k': k} + + for rowid in xrange(64): + row = input_flat_2d[rowid] + output[rowid] = np.sort(row)[-k:] + indices[rowid] = row.argsort()[-k:] + + self.outputs = {'Out': output, 'Indices': indices} + + +if __name__ == '__main__': + unittest.main()