diff --git a/.travis.yml b/.travis.yml index 9dd5f48164a3417940880f2b15bf7d9906453fb8..bf6a41d13c4eabc2d8543ab821ce0ff747a061df 100644 --- a/.travis.yml +++ b/.travis.yml @@ -56,7 +56,7 @@ script: export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh export DOCS_DIR=`pwd` cd .. - curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc/v2 + curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc/ notifications: email: on_success: change diff --git a/CMakeLists.txt b/CMakeLists.txt index c86889c05c8cf0d521dce9adbf3e918ba91729a1..0ec65bac84b0b0d89123473a8941f80c90f1b339 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -53,7 +53,7 @@ option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF) option(ON_TRAVIS "Exclude special unit test on Travis CI" OFF) option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF) # TODO: Only compile PaddlePaddle fluid version by WITH_FLUID option. -option(WITH_FLUID "Compile PaddlePaddle fluid only(TODO)" ON) +option(WITH_FLUID "Compile PaddlePaddle fluid only(TODO)" OFF) option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) option(GLIDE_INSTALL "Download and install go dependencies " ON) option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) diff --git a/Dockerfile b/Dockerfile index 60e76c7f2ede6beaca11659020d5991a75d5b741..fbec88c7966d6ea93495519843d6cda63f622661 100644 --- a/Dockerfile +++ b/Dockerfile @@ -53,10 +53,14 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8 # FIXME: due to temporary ipykernel dependency issue, specify ipykernel jupyter # version util jupyter fixes this issue. + +# specify sphinx version as 1.5.6 and remove -U option for [pip install -U +# sphinx-rtd-theme] since -U option will cause sphinx being updated to newest +# version(1.7.1 for now), which causes building documentation failed. RUN pip install --upgrade pip && \ pip install -U wheel && \ - pip install -U docopt PyYAML sphinx && \ - pip install -U sphinx-rtd-theme==0.1.9 recommonmark + pip install -U docopt PyYAML sphinx==1.5.6 && \ + pip install sphinx-rtd-theme==0.1.9 recommonmark RUN pip install pre-commit 'ipython==5.3.0' && \ pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index e2b7ef8d54748de46370e808e28f9b9fcc988a84..8af2765f58717408e3a1ef6b500bb01511bfd8d3 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -77,7 +77,8 @@ IF(NOT ${CBLAS_FOUND}) INSTALL_DIR ${CBLAS_INSTALL_DIR} BUILD_IN_SOURCE 1 BUILD_COMMAND ${CMAKE_MAKE_PROGRAM} ${COMMON_ARGS} ${OPTIONAL_ARGS} - INSTALL_COMMAND ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 NO_LAPACK=1 PREFIX= + INSTALL_COMMAND ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 NO_LAPACK=1 PREFIX= + && rm -r ${CBLAS_INSTALL_DIR}/lib/cmake ${CBLAS_INSTALL_DIR}/lib/pkgconfig UPDATE_COMMAND "" CONFIGURE_COMMAND "" ) @@ -100,11 +101,6 @@ IF(NOT ${CBLAS_FOUND}) \"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\" )" ) - INSTALL(CODE "execute_process( - COMMAND rm -r ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}/cmake - ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}/pkgconfig - )" - ) ENDIF() ENDIF(NOT ${CBLAS_FOUND}) diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 356da582d1f8b6a8858af90ccdf5af2100e5db87..471e3929069d0d28105404b4f0f6baa303faf0e0 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -186,7 +186,9 @@ function(cc_library TARGET_NAME) add_library(${TARGET_NAME} SHARED ${cc_library_SRCS}) else() add_library(${TARGET_NAME} STATIC ${cc_library_SRCS}) + find_fluid_modules(${TARGET_NAME}) endif() + if(cc_library_DEPS) # Don't need link libwarpctc.so if("${cc_library_DEPS};" MATCHES "warpctc;") @@ -242,11 +244,11 @@ function(cc_test TARGET_NAME) cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) add_executable(${TARGET_NAME} ${cc_test_SRCS}) # Support linking flags: --whole-archive (Linux) / -force_load (MacOS) - target_circle_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) + target_circle_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog) if("${cc_test_DEPS}" MATCHES "ARCHIVE_START") list(REMOVE_ITEM cc_test_DEPS ARCHIVE_START ARCHIVE_END) endif() - add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) + add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog) add_test(NAME ${TARGET_NAME} COMMAND ${TARGET_NAME} ${cc_test_ARGS} WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) @@ -263,7 +265,8 @@ function(nv_library TARGET_NAME) if (nv_library_SHARED OR nv_library_shared) # build *.so cuda_add_library(${TARGET_NAME} SHARED ${nv_library_SRCS}) else() - cuda_add_library(${TARGET_NAME} STATIC ${nv_library_SRCS}) + cuda_add_library(${TARGET_NAME} STATIC ${nv_library_SRCS}) + find_fluid_modules(${TARGET_NAME}) endif() if (nv_library_DEPS) add_dependencies(${TARGET_NAME} ${nv_library_DEPS}) @@ -308,8 +311,8 @@ function(nv_test TARGET_NAME) set(multiValueArgs SRCS DEPS) cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS}) - target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) - add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) + target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog) + add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main paddle_memory gtest gflags glog) add_test(${TARGET_NAME} ${TARGET_NAME}) endif() endfunction(nv_test) diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index 4471df36b0717171da4dff92ca0ec98b4f981028..6b2237b858380f384be0aa3c6ae24a4c83ad646d 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -1,9 +1,22 @@ +set_property(GLOBAL PROPERTY FLUID_MODULES "") +# find all fluid modules is used for paddle fluid static library +function(find_fluid_modules TARGET_NAME) + get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE) + string(FIND "${__target_path}" "fluid" pos) + if(pos GREATER 1) + get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES) + set(fluid_modules ${fluid_modules} ${TARGET_NAME}) + set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}") + endif() +endfunction(find_fluid_modules) + # make package for paddle fluid shared and static library function(copy TARGET) set(options "") set(oneValueArgs "") set(multiValueArgs SRCS DSTS DEPS) cmake_parse_arguments(copy_lib "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + set(inference_lib_dist_dep ${TARGET} ${inference_lib_dist_dep} PARENT_SCOPE) list(LENGTH copy_lib_SRCS copy_lib_SRCS_len) list(LENGTH copy_lib_DSTS copy_lib_DSTS_len) @@ -42,13 +55,21 @@ copy(glog_lib DSTS ${dst_dir} ${dst_dir}/lib ) -IF(NOT PROTOBUF_FOUND) +if(NOT PROTOBUF_FOUND) set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/protobuf") copy(protobuf_lib - SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LITE_LIBRARY} + SRCS ${PROTOBUF_INCLUDE_DIR} ${PROTOBUF_LIBRARY} DSTS ${dst_dir} ${dst_dir}/lib ) -ENDIF(NOT PROTOBUF_FOUND) +endif() + +if(NOT CBLAS_FOUND) + set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/openblas") + copy(openblas_lib + SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include + DSTS ${dst_dir} ${dst_dir} + ) +endif() # paddle fluid module set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid") @@ -66,8 +87,8 @@ copy(memory_lib ) set(module "inference") -copy(inference_lib DEPENDS paddle_fluid_shared - SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.so +copy(inference_lib DEPS paddle_fluid_shared paddle_fluid + SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.* DSTS ${dst_dir}/${module} ${dst_dir}/${module} ) @@ -83,6 +104,4 @@ copy(string_lib DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat ) -add_custom_target(inference_lib_dist DEPENDS - inference_lib framework_lib memory_lib platform_lib string_lib - gflags_lib glog_lib protobuf_lib eigen3_lib) +add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) diff --git a/doc/design/cpp_data_feeding.md b/doc/design/cpp_data_feeding.md deleted file mode 100644 index 40205350f99722f0b71bfa6f390fe9d01d831966..0000000000000000000000000000000000000000 --- a/doc/design/cpp_data_feeding.md +++ /dev/null @@ -1,79 +0,0 @@ -# C++ Data Feeding - -In training with Paddle V2 API, data feeding wholly dependents on Python code. To get rid of the Python environment and achieve the goal of "wrapping the whole training by a while loop op" in Paddle Fluid, a C++ data feeding mechanism is required. - -In this document we show the fundamental design of C++ data feeding process, which includes the data reading, shuffling and batching. - -## Reader - -A new concept named 'Reader' is introduced. `Reader` is a series of inherited classes which can be hold by our `Variable` and they are used to read or process file data. - - -### `ReaderBase` - -`ReaderBase` is the abstract base class of all readers. It defines the all readers' interfaces. - -```cpp -class ReaderBase { - public: - explicit ReaderBase(const std::vector& shapes) : shapes_(shapes) { - PADDLE_ENFORCE(!shapes_.empty()); - } - // Read the next batch of data. (A 'batch' can be only one instance) - virtual void ReadNext(std::vector* out) = 0; - // Show whether the next bacth exists. - virtual bool HasNext() const = 0; - - // Reinitialize the reader and read the file from the begin. - virtual void ReInit() = 0; - - // Get a certain read in data's shape. - DDim shape(size_t idx) const; - // Get shapes of all read in data. - std::vector shapes() const { return shapes_; } - // Set shapes of read in data. - void set_shapes(const std::vector& shapes) { shapes_ = shapes; } - - virtual ~ReaderBase() {} - - protected: - std::vector shapes_; -}; -``` - -### `FileReader` and `DecoratedReader` - -These two classes are derived from the `ReaderBase` and will further be derived by respective specific readers. That is to say, in our design, there are two kinds of readers: file readers and decorated readers. A file reader reads from a file of some specific format, and yield only one instance of data at a time. e.g. RecordIO reader, jpg reader, .... A decorated reader takes another reader(both file reader and decorated reader are OK) as its 'underlying reader'. It gets data from its underlying reader, does some process on them(shuffling, or batching), then yields processed data. The output data of a decorated reader can be a single instance or a batch. `ShuffleReader` and `BatchReader` are both decorated readers. - -All the readers share exactly the same interfaces defined in `ReaderBase`. So they can be decorated for more than one time: We can **shuffle** a reader's outputs and then **batch** the shuffle outputs. The interface consistency also allows related ops use readers without knowing what they are exactly. - - -### `ReaderHolder` - -Different readers belong to different class types. It leads to a problem: How can we drop them into `Variable`s and fetch them out by a unified method? For example, if a Variable holds a `BatchReader`, we can not get it by the following code: - -```cpp -var->Get("batch_reader"); -``` - -we have to write: - -```cpp -var->Get("batch_reader"); -``` - -This requires each time getting a reader from a variable we must know the reader's type exactly. It is nearly impossible. - -To solve this problem, we introduce `ReaderHolder` as a wrapper. It acts as an empty decorator of `ReaderBase`, which erases reader's type. With `ReaderHolder` we are able to fetch all types of readers by `var->Get("...")` and regard the obtained object as a reader. - -## Related Operators - -To create and invoke readers, some now ops are introduced: - -### `CreateReaderOp` - -Each reader has its creating op. File readers' creating ops have no input and yield the created file reader as its output. Decorated readers' creating ops take the underlying readers as inputs and then yield new decorated readers. - -### `ReadOp` - -A reader is only a Variable. It cannot trigger the reading process by itself. So we add the `ReadOp` to execute it. A `ReadOp` takes a reader Variable as its input. Each time it runs, it invokes the reader‘s `ReadNext()` function and gets a new batch of data(or only one instance of data, if we use file reader directly). The output data of a reader are in the form of `std::vector`, so the `ReadOp` also needs to split the vector and move LoDTensors to their respective output Variables. diff --git a/doc/design/distributed_lookup_table_design.md b/doc/design/distributed_lookup_table_design.md new file mode 100644 index 0000000000000000000000000000000000000000..a09f2818c888397b07fc7d09ecd20056f4176982 --- /dev/null +++ b/doc/design/distributed_lookup_table_design.md @@ -0,0 +1,128 @@ +## Design Doc: Distributed Lookup Table Operator + +A lookup table operator in PaddlePaddle where the table could be out +of the memory of a computer. + +## Background + +A lookup table operator is well-used in deep learning for learning the +representation, or the +[*embedding*](http://www.cs.toronto.edu/~fritz/absps/ieee-lre.pdf), of +symbols. + +### The Forward Algorithm + +The forward algorithm of the lookup table is a multiplication of the +input vector x and the lookup table matrix W: + +$$y = x * W$$ + +When x is a sparse vector of symbols, the above multiplication +simplifies into looking up rows in W that correspond to symbols in x, +denoted by W(x). Please be aware that W could be huge and out of the +memory, so we'd need a distributed storage service, which supports the +lookup of rows. + +The following figure illustrates the multiplication of x with two +non-zero elements, or say, two symbols, and a lookup table W: + +![lookup table](./lookup_table.png) + +### The Backward Algorithm + +The backward algorithm computes W'(x) using W(x). W'(x) has the same +scale of size as W(x) and is much smaller than W. + +To optimize W given W', we can do simple SGD update: + +$$W = f(W') = \lambda * W'$$ + +or some more sophisticated algorithms that rely on both W' and W: + +$$W = f(W, W')$$ + +The following figure illustrates the backward pass of the lookup +operator: ![lookup table training](./lookup_table_training.png) + +## Distributed Storage Service + +The forward algorithm requires a distributed storage service for W. +The backward algorithm prefers that the storage system can apply the +optimization algorithm on W. The following two sections describe two +solutions -- the former doesn't require that the storage service can +do optimization, the latter does. + +### Storage Service Doesn't Optimize + +In this design, we use highly-optimized distributed storage, e.g., +memcached, as the storage service, and we run the optimization +algorithm on parameter servers of PaddlePaddle. The following figure +illustrates the training process. + + + + + +Each trainer runs the forward and backward passes using their local +data: + +1. In the forward pass, when a trainer runs the forward algorithm of a + lookup operator, it retrieves W(x) from the storage service. +1. The trainer computes W'(x) in the backward pass using W(x). + +During the global update process: + +1. Each trainer uploads its W'(x) to parameter servers. +1. The parameter server runs the optimization algorithm, e.g., the + Adam optimization algorithm, which requires that + 1. The parameter server retrieves W(x) from memcached, and + 1. The parameter server pushes $\Delta W(x)=f(W(x), lambda \sum_j + W'(x))$ to memcached, where $f$ denotes the optimization + algorithm. + +### Storage Service Does Optimize + +This design is very similar to the above one, except that the +optimization algorithm $f$ runs on the storage service. + +- Pro: parameter servers do not retrieve W(x) from the storage + service, thus saves half network communication. +- Con: the storage service needs to be able to run the optimization + algorithm. + +## Conclusion + +Let us do the "storage service does not optimize" solution first, as a +baseline at least, because it is easier to use a well-optimized +distributed storage service like memcached. We can do the "storage +service does optimize" solution later or at the same time, which, if +implemented carefully, should have better performance than the former. diff --git a/doc/design/images/duplicate_op.graffle b/doc/design/images/duplicate_op.graffle deleted file mode 100644 index 5979f792e252f028a615729215529c2be42d9165..0000000000000000000000000000000000000000 Binary files a/doc/design/images/duplicate_op.graffle and /dev/null differ diff --git a/doc/design/images/duplicate_op.png b/doc/design/images/duplicate_op.png deleted file mode 100644 index f299c5d37f260a1bb0daec886f0a4ee1c1f31c92..0000000000000000000000000000000000000000 Binary files a/doc/design/images/duplicate_op.png and /dev/null differ diff --git a/doc/design/images/duplicate_op2.graffle b/doc/design/images/duplicate_op2.graffle deleted file mode 100644 index 5cec3bc64dbd44dc99e348485969f29bd128ceb1..0000000000000000000000000000000000000000 Binary files a/doc/design/images/duplicate_op2.graffle and /dev/null differ diff --git a/doc/design/images/duplicate_op2.png b/doc/design/images/duplicate_op2.png deleted file mode 100644 index 21cdd5cabf1b5203e1435a75b57770d2f702fa92..0000000000000000000000000000000000000000 Binary files a/doc/design/images/duplicate_op2.png and /dev/null differ diff --git a/doc/design/images/replica.png b/doc/design/images/replica.png deleted file mode 100644 index ef59e56b01d792a059279e6bb9a29f3db6a59a41..0000000000000000000000000000000000000000 Binary files a/doc/design/images/replica.png and /dev/null differ diff --git a/doc/design/images/two_phase_commit.png b/doc/design/images/two_phase_commit.png deleted file mode 100644 index ef6f7317bd440cc7d9fe08fcbbf2b7a542f99049..0000000000000000000000000000000000000000 Binary files a/doc/design/images/two_phase_commit.png and /dev/null differ diff --git a/doc/design/lookup_table.png b/doc/design/lookup_table.png new file mode 100644 index 0000000000000000000000000000000000000000..72dfe3547f731d0d090338afb206b0549dff472e Binary files /dev/null and b/doc/design/lookup_table.png differ diff --git a/doc/design/lookup_table_training.png b/doc/design/lookup_table_training.png new file mode 100644 index 0000000000000000000000000000000000000000..cc7cc4aeb3b885850fe2f70f19fb84d5873bed1e Binary files /dev/null and b/doc/design/lookup_table_training.png differ diff --git a/doc/design/images/asgd.gif b/doc/fluid/design/algorithm/images/asgd.gif similarity index 100% rename from doc/design/images/asgd.gif rename to doc/fluid/design/algorithm/images/asgd.gif diff --git a/doc/design/images/theta_star.gif b/doc/fluid/design/algorithm/images/theta_star.gif similarity index 100% rename from doc/design/images/theta_star.gif rename to doc/fluid/design/algorithm/images/theta_star.gif diff --git a/doc/design/parameter_average.md b/doc/fluid/design/algorithm/parameter_average.md similarity index 100% rename from doc/design/parameter_average.md rename to doc/fluid/design/algorithm/parameter_average.md diff --git a/doc/design/build_system/README.md b/doc/fluid/design/concepts/README.md similarity index 100% rename from doc/design/build_system/README.md rename to doc/fluid/design/concepts/README.md diff --git a/doc/design/block.md b/doc/fluid/design/concepts/block.md similarity index 100% rename from doc/design/block.md rename to doc/fluid/design/concepts/block.md diff --git a/doc/fluid/design/concepts/cpp_data_feeding.md b/doc/fluid/design/concepts/cpp_data_feeding.md new file mode 100644 index 0000000000000000000000000000000000000000..8607b40ccbbe01db77afed72c1efa780b520744c --- /dev/null +++ b/doc/fluid/design/concepts/cpp_data_feeding.md @@ -0,0 +1,171 @@ +# C++ Data Feeding + +While using Paddle V2 API for training, data feeding completely depends on the Python code. To get rid of the Python environment and achieve the goal of "wrapping the whole training by a while loop op" in Paddle Fluid, a C++ data feeding mechanism is required. + +In this document, we show the fundamental design of a C++ data feeding process, which includes data reading, shuffling and batching. + +## Overview + +![](images/readers.png) + +## Reader + +In order to handle the above-mentioned problem, a new concept called 'Reader' is introduced. `Reader` is a series of inherited classes which can be held by our `Variable` and they are used to read or process file data. + + +### ReaderBase + +`ReaderBase` is the abstract base class for all readers. It defines the interface for all readers. + +```cpp +class ReaderBase { + public: + // Reads the next batch of data. (A 'batch' can be only one instance) + // If the next batch doesn't exist, it throws an exception + virtual void ReadNext(std::vector* out) = 0; + + // Checks whether the next instance exists. + virtual bool HasNext() = 0; + + // Reinitializes the reader and read the file from the beginning. + virtual void ReInit() = 0; + + virtual ~ReaderBase(); +}; +``` + +### FileReader + +`FileReader` is derived from the `ReaderBase`. It is still an abstract class and will further be derived by Readers of respective specific format. + +```cpp +class FileReader : public ReaderBase { + public: + explicit FileReader(const std::vector& dims); + + void ReadNext(std::vector* out) override; + + protected: + virtual void ReadNextImpl(std::vector* out) = 0; + + private: + std::vector dims_; +}; +``` + +A file reader binds with a single file and reads one data instance at a time. Each type of file reader shall implement its own `ReadNextImpl()`, `HasNext()` and `ReInit()`. + +The `ReadNextImpl()` is invoked by `ReadNext()`. Besides invoking `ReadNextImpl()`, `ReadNext()` is also responsible for checking the output, making sure that each shape of `LoDTensor` in `*out` is consistent with the one in `dims_`. + +### DecoratedReader + +A decorated reader takes another reader(both file reader and decorated reader are OK) as its 'underlying reader'. It gets data from its underlying reader, does some processing on them(shuffling, batching or something else), then yields processed data. The output data of a decorated reader can be a single instance or a batch. `ShuffleReader` and `BatchReader` are both decorated readers. + +```cpp +class DecoratedReader : public ReaderBase { + public: + explicit DecoratedReader(ReaderBase* reader) : ReaderBase(), reader_(reader) { + PADDLE_ENFORCE_NOT_NULL(reader_); + } + + void ReInit() override { reader_->ReInit(); } + + bool HasNext() const override { return reader_->HasNext(); } + + protected: + ReaderBase* reader_; +}; +``` + +Both the `FileReader` and `DecoratedReader` share exactly the same interface as defined in `ReaderBase`. So they can be decorated for multiple times: We can **shuffle** a reader's outputs and then **batch** the shuffled outputs. The interface consistency also allows related ops use readers without knowing their underlying type. + +### MultipleReader + +All `FileReader` binds with a single file and are single-threaded. However, sometimes we need to read data from more than one file. In this case, it's not enough to only have `FileReader` and `DecoratedReader`. + +So `MultipleReader` is introduced. It is also derived from `ReaderBase`. A `MultipleReader` holds several prefetching `FileReaders` and these readers run concurrently. Another pivotal part of a `MultipleReader` is a buffer channel. The channel collects data yield by all prefetching readers and makes subsequent OPs or decorated readers be able to fetch data without concerning about multiple readers scheduling. + +![](images/multiple_reader.png) + +This graph shows how a `MultipleReader` works with three prefetching file readers and two GPUs. There is a queue of files which are going to be read. Each time when a prefetching file reader is free(complete reading from one file), it fetches a new file from the queue. Each prefetching file reader runs in a separated prefetch thread and dumps their outputs to the same channel. + +To the subsequent two decorated readers, the `MultipleReader` is **a single reader**. They don't need to concern about how prefetch readers are scheduled. They only need to invoke `MultipleReader::ReadNext()` to get the next data from the buffer channel. + +### ReaderHolder + +Different readers belong to different class types. This leads to a problem: How can we drop them into `Variable`s and fetch them out by a unified method? For example, if a Variable holds a `BatchReader`, we can not get it by the following code: + +```cpp +var->Get("batch_reader"); +``` + +We would have to write: + +```cpp +var->Get("batch_reader"); +``` + +This requires that in order to get a reader from a variable, every time, we must know the reader's type exactly. This is nearly impossible. + +To solve this problem, we introduce `ReaderHolder` as a wrapper. It acts as an empty decorator of `ReaderBase`, which hides reader's type. With `ReaderHolder` we are able to fetch all types of readers by `var->Get("...")` and regard the obtained object as a reader. + +## Related Operators + +To create and invoke readers, some new ops are introduced: + +### CreateReaderOp + +Each reader has its creation op. File readers' creation ops have no input and yield the created file reader as its output. Decorated readers' creation ops take the underlying readers as inputs and then yield new decorated readers. + +However, direct usage of file readers' creation ops is not recommended because a file reader can only read one file via a single thread. Using `OpenFilesOp` is a better choice. + +### OpenFilesOp + +The `OpenFilesOp` is the creation op of `MultipleReader`. It takes no input but requires a list of file names as one of its attributes. The newly created `MultipleReader` then creates its own prefetching readers according to given file names. + +To make sure that created prefetching readers match file formats, we need a name prefix rule to append file format tags to file names, as well as a file reader registry mechanism to map file format tags to their corresponding file readers' constructors. + +### HasNextOp + +`HasNextOp` is used to check whether the next data batch exists via the reader's `HasNext()` interface. + +### ResetOp + +`ResetOp` is used to reset a reader via its `ReInit()` interface. + +### ReadOp + +A reader is only a Variable. It cannot trigger the reading process by itself. So we add the `ReadOp` to execute it. A `ReadOp` takes a reader Variable as its input. Each time it runs, it invokes the reader‘s `ReadNext()` function and gets a new batch of data(or only one instance of data, if we use file reader directly). The output data of a reader are in the form of `std::vector`, so the `ReadOp` also needs to split the vector and move LoDTensors to their respective output Variables. + +## Program with Readers + +A `Program` holds readers as its persistable variables. These variables are created by `CreateReaderOp` or `OpenFilesOp`. These ops shall run only once. So they shall be settled in the `startup_program`. `HasNextOp`, `ResetOp` and `ReadOp` are required by training loop, so they shall be in the `main_program`. + +The ops of a `startup_program` with readers would be like this: + +``` +multiple_reader = open_files_op(...) +batch_reader = create_batch_reader_op(multiple_reader) +double_buffer_reader = create_double_buffer_op(batch_reader) +... (other initializers) +``` + +The forwarding ops of the corresponding `main_program` would be like this: + +``` +while_op { + has_next = has_next_op(double_buffer_reader) + if_else_op(has_next) { + batch_data = read_op(double_buffer_reader) + ... (subsequent training ops) + } else { + reset_op(double_buffer_reader) + } +} +``` + +Two important considerations for these programs are as follows: + +1. The multiple\_reader is the batch\_reader's underlying reader, and the batch\_reader is the double\_buffer\_reader's underlying reader. `read_op`, `has_next_op` and other reader related ops will only invoke the top-most reader. In this case, it's the double\_buffer\_reader. + +2. All readers exist in both `startup_program` and `main_program`. And they are persistable. diff --git a/doc/design/executor.md b/doc/fluid/design/concepts/executor.md similarity index 100% rename from doc/design/executor.md rename to doc/fluid/design/concepts/executor.md diff --git a/doc/design/functions_operators_layers.md b/doc/fluid/design/concepts/functions_operators_layers.md similarity index 100% rename from doc/design/functions_operators_layers.md rename to doc/fluid/design/concepts/functions_operators_layers.md diff --git a/doc/fluid/design/concepts/images/multiple_reader.png b/doc/fluid/design/concepts/images/multiple_reader.png new file mode 100644 index 0000000000000000000000000000000000000000..b22126b31db4982c13fc3a0827805e6aaf955046 Binary files /dev/null and b/doc/fluid/design/concepts/images/multiple_reader.png differ diff --git a/doc/fluid/design/concepts/images/readers.png b/doc/fluid/design/concepts/images/readers.png new file mode 100644 index 0000000000000000000000000000000000000000..fd59168ce16c9e2a0ef45303c28c997cfd7740be Binary files /dev/null and b/doc/fluid/design/concepts/images/readers.png differ diff --git a/paddle/fluid/framework/lod_tensor.md b/doc/fluid/design/concepts/lod_tensor.md similarity index 100% rename from paddle/fluid/framework/lod_tensor.md rename to doc/fluid/design/concepts/lod_tensor.md diff --git a/doc/design/program.md b/doc/fluid/design/concepts/program.md similarity index 100% rename from doc/design/program.md rename to doc/fluid/design/concepts/program.md diff --git a/doc/design/scope.md b/doc/fluid/design/concepts/scope.md similarity index 100% rename from doc/design/scope.md rename to doc/fluid/design/concepts/scope.md diff --git a/paddle/fluid/framework/tensor.md b/doc/fluid/design/concepts/tensor.md similarity index 100% rename from paddle/fluid/framework/tensor.md rename to doc/fluid/design/concepts/tensor.md diff --git a/doc/design/tensor_array.md b/doc/fluid/design/concepts/tensor_array.md similarity index 100% rename from doc/design/tensor_array.md rename to doc/fluid/design/concepts/tensor_array.md diff --git a/doc/design/var_desc.md b/doc/fluid/design/concepts/var_desc.md similarity index 100% rename from doc/design/var_desc.md rename to doc/fluid/design/concepts/var_desc.md diff --git a/paddle/fluid/framework/variable.md b/doc/fluid/design/concepts/variable.md similarity index 100% rename from paddle/fluid/framework/variable.md rename to doc/fluid/design/concepts/variable.md diff --git a/doc/design/concurrent_programming.md b/doc/fluid/design/concurrent/concurrent_programming.md similarity index 100% rename from doc/design/concurrent_programming.md rename to doc/fluid/design/concurrent/concurrent_programming.md diff --git a/doc/design/csp.md b/doc/fluid/design/concurrent/csp.md similarity index 100% rename from doc/design/csp.md rename to doc/fluid/design/concurrent/csp.md diff --git a/doc/design/parallel_do.md b/doc/fluid/design/concurrent/parallel_do.md similarity index 100% rename from doc/design/parallel_do.md rename to doc/fluid/design/concurrent/parallel_do.md diff --git a/doc/design/float16.md b/doc/fluid/design/data_type/float16.md similarity index 100% rename from doc/design/float16.md rename to doc/fluid/design/data_type/float16.md diff --git a/doc/design/dist_refactor/distributed_architecture.md b/doc/fluid/design/dist_train/distributed_architecture.md similarity index 98% rename from doc/design/dist_refactor/distributed_architecture.md rename to doc/fluid/design/dist_train/distributed_architecture.md index 9368c5780dc922953f38bf0f86d9f797a4a8a6fe..a405cb6aaf80b9d2e8a1a9c774ca85cc7e62bbab 100644 --- a/doc/design/dist_refactor/distributed_architecture.md +++ b/doc/fluid/design/dist_train/distributed_architecture.md @@ -1,4 +1,4 @@ -# Design Doc: Distributed Training Architecture +# Design Doc: Fluid Distributed Training Architecture ## Abstract @@ -155,7 +155,7 @@ Cluster environment. `RemoteExecutor.run` sends the `ProgramDesc` and -[TrainingJob](https://github.com/PaddlePaddle/cloud/blob/develop/doc/autoscale/README.md#training-job-resource) +[TrainingJob](https://github.com/PaddlePaddle/cloud/blob/unreleased-tpr/doc/autoscale/README.md#training-job-resource) to a server in the cluster which executes `RemoteExecutor.listen`. This server is responsible to start the final Kubernetes Jobs to run the different role of `ProgramDesc` from `ConfigMap`. diff --git a/doc/design/dist_refactor/multi_cpu.md b/doc/fluid/design/dist_train/multi_cpu.md similarity index 100% rename from doc/design/dist_refactor/multi_cpu.md rename to doc/fluid/design/dist_train/multi_cpu.md diff --git a/doc/design/dist_refactor/parameter_server.md b/doc/fluid/design/dist_train/parameter_server.md similarity index 86% rename from doc/design/dist_refactor/parameter_server.md rename to doc/fluid/design/dist_train/parameter_server.md index 805dd13048d41b995d2a01cda52b2ea33e4bbe1d..6ce48dfbfce8b094684b412ebfda7e505ddc30ae 100644 --- a/doc/design/dist_refactor/parameter_server.md +++ b/doc/fluid/design/dist_train/parameter_server.md @@ -59,6 +59,17 @@ After converting: queue. It will block until the queue has the required number of tensors. +### Sparse Update + +For embedding layers, the gradient may have many rows containing only 0 when training, +if the gradient uses a dense tensor to do parameter optimization, +it could spend unnecessary memory, slow down the calculations and waste +the bandwidth while doing distributed training. +In Fluid, we introduce [SelectedRows](../selected_rows.md) to represent a list of rows containing +non-zero gradient data. So when we do parameter optimization both locally and remotely, +we only need to send those non-zero rows to the optimizer operators: + + ### Benefits @@ -91,6 +102,6 @@ After converting: `min_count` attribute), does our current design support it? (similar question for the *Add* OP) +### References -### References: [1] [TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45166.pdf) diff --git a/doc/design/dist_refactor/src/compiler.graffle b/doc/fluid/design/dist_train/src/compiler.graffle similarity index 100% rename from doc/design/dist_refactor/src/compiler.graffle rename to doc/fluid/design/dist_train/src/compiler.graffle diff --git a/doc/design/dist_refactor/src/compiler.png b/doc/fluid/design/dist_train/src/compiler.png similarity index 100% rename from doc/design/dist_refactor/src/compiler.png rename to doc/fluid/design/dist_train/src/compiler.png diff --git a/doc/design/dist_refactor/src/dist-graph.graffle b/doc/fluid/design/dist_train/src/dist-graph.graffle similarity index 100% rename from doc/design/dist_refactor/src/dist-graph.graffle rename to doc/fluid/design/dist_train/src/dist-graph.graffle diff --git a/doc/design/dist_refactor/src/dist-graph.png b/doc/fluid/design/dist_train/src/dist-graph.png similarity index 100% rename from doc/design/dist_refactor/src/dist-graph.png rename to doc/fluid/design/dist_train/src/dist-graph.png diff --git a/doc/design/dist_refactor/src/distributed_architecture.graffle b/doc/fluid/design/dist_train/src/distributed_architecture.graffle similarity index 100% rename from doc/design/dist_refactor/src/distributed_architecture.graffle rename to doc/fluid/design/dist_train/src/distributed_architecture.graffle diff --git a/doc/design/dist_refactor/src/distributed_architecture.png b/doc/fluid/design/dist_train/src/distributed_architecture.png similarity index 100% rename from doc/design/dist_refactor/src/distributed_architecture.png rename to doc/fluid/design/dist_train/src/distributed_architecture.png diff --git a/doc/design/dist_refactor/src/local-graph.graffle b/doc/fluid/design/dist_train/src/local-graph.graffle similarity index 100% rename from doc/design/dist_refactor/src/local-graph.graffle rename to doc/fluid/design/dist_train/src/local-graph.graffle diff --git a/doc/design/dist_refactor/src/local-graph.png b/doc/fluid/design/dist_train/src/local-graph.png similarity index 100% rename from doc/design/dist_refactor/src/local-graph.png rename to doc/fluid/design/dist_train/src/local-graph.png diff --git a/doc/design/dist_refactor/src/local_architecture.graffle b/doc/fluid/design/dist_train/src/local_architecture.graffle similarity index 100% rename from doc/design/dist_refactor/src/local_architecture.graffle rename to doc/fluid/design/dist_train/src/local_architecture.graffle diff --git a/doc/design/dist_refactor/src/local_architecture.png b/doc/fluid/design/dist_train/src/local_architecture.png similarity index 100% rename from doc/design/dist_refactor/src/local_architecture.png rename to doc/fluid/design/dist_train/src/local_architecture.png diff --git a/doc/design/dist_refactor/src/multi-threads.graffle b/doc/fluid/design/dist_train/src/multi-threads.graffle similarity index 100% rename from doc/design/dist_refactor/src/multi-threads.graffle rename to doc/fluid/design/dist_train/src/multi-threads.graffle diff --git a/doc/design/dist_refactor/src/multi-threads/multi-threads@3x.png b/doc/fluid/design/dist_train/src/multi-threads/multi-threads@3x.png similarity index 100% rename from doc/design/dist_refactor/src/multi-threads/multi-threads@3x.png rename to doc/fluid/design/dist_train/src/multi-threads/multi-threads@3x.png diff --git a/doc/design/dist_refactor/src/multi-threads/single-thread@3x.png b/doc/fluid/design/dist_train/src/multi-threads/single-thread@3x.png similarity index 100% rename from doc/design/dist_refactor/src/multi-threads/single-thread@3x.png rename to doc/fluid/design/dist_train/src/multi-threads/single-thread@3x.png diff --git a/doc/design/dist_refactor/src/paddle-compile.graffle b/doc/fluid/design/dist_train/src/paddle-compile.graffle similarity index 100% rename from doc/design/dist_refactor/src/paddle-compile.graffle rename to doc/fluid/design/dist_train/src/paddle-compile.graffle diff --git a/doc/design/dist_refactor/src/paddle-compile.png b/doc/fluid/design/dist_train/src/paddle-compile.png similarity index 100% rename from doc/design/dist_refactor/src/paddle-compile.png rename to doc/fluid/design/dist_train/src/paddle-compile.png diff --git a/doc/design/dist_refactor/src/remote_executor.graffle b/doc/fluid/design/dist_train/src/remote_executor.graffle similarity index 100% rename from doc/design/dist_refactor/src/remote_executor.graffle rename to doc/fluid/design/dist_train/src/remote_executor.graffle diff --git a/doc/design/dist_refactor/src/remote_executor.png b/doc/fluid/design/dist_train/src/remote_executor.png similarity index 100% rename from doc/design/dist_refactor/src/remote_executor.png rename to doc/fluid/design/dist_train/src/remote_executor.png diff --git a/doc/fluid/design/dist_train/src/sparse_update.graffle b/doc/fluid/design/dist_train/src/sparse_update.graffle new file mode 100644 index 0000000000000000000000000000000000000000..08d689a58f83698d8c1158ee3990ed8abf3a7a9a Binary files /dev/null and b/doc/fluid/design/dist_train/src/sparse_update.graffle differ diff --git a/doc/fluid/design/dist_train/src/sparse_update.png b/doc/fluid/design/dist_train/src/sparse_update.png new file mode 100644 index 0000000000000000000000000000000000000000..8c872e6ac479f7d1b818a4a207956c43155d0ad7 Binary files /dev/null and b/doc/fluid/design/dist_train/src/sparse_update.png differ diff --git a/doc/design/ops/images/2_level_rnn.dot b/doc/fluid/design/dynamic_rnn/2_level_rnn.dot similarity index 100% rename from doc/design/ops/images/2_level_rnn.dot rename to doc/fluid/design/dynamic_rnn/2_level_rnn.dot diff --git a/doc/design/ops/images/2_level_rnn.png b/doc/fluid/design/dynamic_rnn/2_level_rnn.png similarity index 100% rename from doc/design/ops/images/2_level_rnn.png rename to doc/fluid/design/dynamic_rnn/2_level_rnn.png diff --git a/doc/design/ops/images/rnn.dot b/doc/fluid/design/dynamic_rnn/rnn.dot similarity index 100% rename from doc/design/ops/images/rnn.dot rename to doc/fluid/design/dynamic_rnn/rnn.dot diff --git a/doc/design/ops/images/rnn.jpg b/doc/fluid/design/dynamic_rnn/rnn.jpg similarity index 100% rename from doc/design/ops/images/rnn.jpg rename to doc/fluid/design/dynamic_rnn/rnn.jpg diff --git a/doc/design/ops/rnn.md b/doc/fluid/design/dynamic_rnn/rnn.md similarity index 100% rename from doc/design/ops/rnn.md rename to doc/fluid/design/dynamic_rnn/rnn.md diff --git a/doc/design/ops/images/rnn.png b/doc/fluid/design/dynamic_rnn/rnn.png similarity index 100% rename from doc/design/ops/images/rnn.png rename to doc/fluid/design/dynamic_rnn/rnn.png diff --git a/doc/design/ops/images/rnn_2level_data.dot b/doc/fluid/design/dynamic_rnn/rnn_2level_data.dot similarity index 100% rename from doc/design/ops/images/rnn_2level_data.dot rename to doc/fluid/design/dynamic_rnn/rnn_2level_data.dot diff --git a/doc/design/ops/images/rnn_2level_data.png b/doc/fluid/design/dynamic_rnn/rnn_2level_data.png similarity index 100% rename from doc/design/ops/images/rnn_2level_data.png rename to doc/fluid/design/dynamic_rnn/rnn_2level_data.png diff --git a/paddle/fluid/operators/op_documentation/rnn_design.md b/doc/fluid/design/dynamic_rnn/rnn_design.md similarity index 100% rename from paddle/fluid/operators/op_documentation/rnn_design.md rename to doc/fluid/design/dynamic_rnn/rnn_design.md diff --git a/doc/design/if_else_op.md b/doc/fluid/design/execution/if_else_op.md similarity index 100% rename from doc/design/if_else_op.md rename to doc/fluid/design/execution/if_else_op.md diff --git a/doc/design/switch.md b/doc/fluid/design/execution/switch.md similarity index 100% rename from doc/design/switch.md rename to doc/fluid/design/execution/switch.md diff --git a/doc/design/multi_language_interface/00.why_plain_c.md b/doc/fluid/design/interface/00.why_plain_c.md similarity index 100% rename from doc/design/multi_language_interface/00.why_plain_c.md rename to doc/fluid/design/interface/00.why_plain_c.md diff --git a/doc/design/multi_language_interface/01.inference_implementation.md b/doc/fluid/design/interface/01.inference_implementation.md similarity index 100% rename from doc/design/multi_language_interface/01.inference_implementation.md rename to doc/fluid/design/interface/01.inference_implementation.md diff --git a/paddle/fluid/memory/README.md b/doc/fluid/design/memory/README.md similarity index 100% rename from paddle/fluid/memory/README.md rename to doc/fluid/design/memory/README.md diff --git a/doc/design/images/control_flow_graph.png b/doc/fluid/design/memory/images/control_flow_graph.png similarity index 100% rename from doc/design/images/control_flow_graph.png rename to doc/fluid/design/memory/images/control_flow_graph.png diff --git a/doc/design/images/dataflow_equations.png b/doc/fluid/design/memory/images/dataflow_equations.png similarity index 100% rename from doc/design/images/dataflow_equations.png rename to doc/fluid/design/memory/images/dataflow_equations.png diff --git a/doc/design/images/deep_learning.png b/doc/fluid/design/memory/images/deep_learning.png similarity index 100% rename from doc/design/images/deep_learning.png rename to doc/fluid/design/memory/images/deep_learning.png diff --git a/doc/design/memory_optimization.md b/doc/fluid/design/memory/memory_optimization.md similarity index 100% rename from doc/design/memory_optimization.md rename to doc/fluid/design/memory/memory_optimization.md diff --git a/doc/design/backward.md b/doc/fluid/design/modules/backward.md similarity index 100% rename from doc/design/backward.md rename to doc/fluid/design/modules/backward.md diff --git a/paddle/fluid/operators/op_documentation/batch_norm_op.md b/doc/fluid/design/modules/batch_norm_op.md similarity index 100% rename from paddle/fluid/operators/op_documentation/batch_norm_op.md rename to doc/fluid/design/modules/batch_norm_op.md diff --git a/doc/design/evaluator.md b/doc/fluid/design/modules/evaluator.md similarity index 100% rename from doc/design/evaluator.md rename to doc/fluid/design/modules/evaluator.md diff --git a/paddle/fluid/operators/images/batch_norm_fork.dot b/doc/fluid/design/modules/images/batch_norm_fork.dot similarity index 100% rename from paddle/fluid/operators/images/batch_norm_fork.dot rename to doc/fluid/design/modules/images/batch_norm_fork.dot diff --git a/paddle/fluid/operators/images/batch_norm_fork.png b/doc/fluid/design/modules/images/batch_norm_fork.png similarity index 100% rename from paddle/fluid/operators/images/batch_norm_fork.png rename to doc/fluid/design/modules/images/batch_norm_fork.png diff --git a/paddle/fluid/operators/images/batch_norm_op_kernel.png b/doc/fluid/design/modules/images/batch_norm_op_kernel.png similarity index 100% rename from paddle/fluid/operators/images/batch_norm_op_kernel.png rename to doc/fluid/design/modules/images/batch_norm_op_kernel.png diff --git a/doc/design/images/feed_forward.png b/doc/fluid/design/modules/images/feed_forward.png similarity index 100% rename from doc/design/images/feed_forward.png rename to doc/fluid/design/modules/images/feed_forward.png diff --git a/doc/design/images/feed_forward_regularized.png b/doc/fluid/design/modules/images/feed_forward_regularized.png similarity index 100% rename from doc/design/images/feed_forward_regularized.png rename to doc/fluid/design/modules/images/feed_forward_regularized.png diff --git a/doc/design/images/l1_regularization.png b/doc/fluid/design/modules/images/l1_regularization.png similarity index 100% rename from doc/design/images/l1_regularization.png rename to doc/fluid/design/modules/images/l1_regularization.png diff --git a/doc/design/images/l2_regularization.png b/doc/fluid/design/modules/images/l2_regularization.png similarity index 100% rename from doc/design/images/l2_regularization.png rename to doc/fluid/design/modules/images/l2_regularization.png diff --git a/doc/design/images/loss_equation.png b/doc/fluid/design/modules/images/loss_equation.png similarity index 100% rename from doc/design/images/loss_equation.png rename to doc/fluid/design/modules/images/loss_equation.png diff --git a/doc/design/infer_var_type.md b/doc/fluid/design/modules/infer_var_type.md similarity index 100% rename from doc/design/infer_var_type.md rename to doc/fluid/design/modules/infer_var_type.md diff --git a/paddle/fluid/operators/op_documentation/net_op_design.md b/doc/fluid/design/modules/net_op_design.md similarity index 100% rename from paddle/fluid/operators/op_documentation/net_op_design.md rename to doc/fluid/design/modules/net_op_design.md diff --git a/doc/design/optimizer.md b/doc/fluid/design/modules/optimizer.md similarity index 100% rename from doc/design/optimizer.md rename to doc/fluid/design/modules/optimizer.md diff --git a/doc/design/prune.md b/doc/fluid/design/modules/prune.md similarity index 100% rename from doc/design/prune.md rename to doc/fluid/design/modules/prune.md diff --git a/doc/design/python_api.md b/doc/fluid/design/modules/python_api.md similarity index 100% rename from doc/design/python_api.md rename to doc/fluid/design/modules/python_api.md diff --git a/doc/design/register_grad_op.md b/doc/fluid/design/modules/register_grad_op.md similarity index 100% rename from doc/design/register_grad_op.md rename to doc/fluid/design/modules/register_grad_op.md diff --git a/doc/design/regularization.md b/doc/fluid/design/modules/regularization.md similarity index 100% rename from doc/design/regularization.md rename to doc/fluid/design/modules/regularization.md diff --git a/doc/design/selected_rows.md b/doc/fluid/design/modules/selected_rows.md similarity index 100% rename from doc/design/selected_rows.md rename to doc/fluid/design/modules/selected_rows.md diff --git a/doc/design/api.md b/doc/fluid/design/motivation/api.md similarity index 100% rename from doc/design/api.md rename to doc/fluid/design/motivation/api.md diff --git a/doc/design/fluid-compiler.graffle b/doc/fluid/design/motivation/fluid-compiler.graffle similarity index 100% rename from doc/design/fluid-compiler.graffle rename to doc/fluid/design/motivation/fluid-compiler.graffle diff --git a/doc/design/fluid-compiler.png b/doc/fluid/design/motivation/fluid-compiler.png similarity index 100% rename from doc/design/fluid-compiler.png rename to doc/fluid/design/motivation/fluid-compiler.png diff --git a/doc/design/fluid.md b/doc/fluid/design/motivation/fluid.md similarity index 100% rename from doc/design/fluid.md rename to doc/fluid/design/motivation/fluid.md diff --git a/doc/design/fluid_compiler.md b/doc/fluid/design/motivation/fluid_compiler.md similarity index 100% rename from doc/design/fluid_compiler.md rename to doc/fluid/design/motivation/fluid_compiler.md diff --git a/doc/design/refactorization.md b/doc/fluid/design/motivation/refactorization.md similarity index 100% rename from doc/design/refactorization.md rename to doc/fluid/design/motivation/refactorization.md diff --git a/doc/design/kernel_hint_design.md b/doc/fluid/design/muti_devices/kernel_hint_design.md similarity index 100% rename from doc/design/kernel_hint_design.md rename to doc/fluid/design/muti_devices/kernel_hint_design.md diff --git a/doc/design/kernel_selection.md b/doc/fluid/design/muti_devices/kernel_selection.md similarity index 100% rename from doc/design/kernel_selection.md rename to doc/fluid/design/muti_devices/kernel_selection.md diff --git a/doc/design/operator_kernel_type.md b/doc/fluid/design/muti_devices/operator_kernel_type.md similarity index 100% rename from doc/design/operator_kernel_type.md rename to doc/fluid/design/muti_devices/operator_kernel_type.md diff --git a/doc/design/speech/deep_speech_2.md b/doc/fluid/design/network/deep_speech_2.md similarity index 98% rename from doc/design/speech/deep_speech_2.md rename to doc/fluid/design/network/deep_speech_2.md index cfdc4d6df04344c70d3334626bd38eca997c31ff..af0c6ef36feba9e0239e7a5f81a8dc9108b2471a 100644 --- a/doc/design/speech/deep_speech_2.md +++ b/doc/fluid/design/network/deep_speech_2.md @@ -94,7 +94,7 @@ The classical DS2 network contains 15 layers (from bottom to top): - **One** CTC-loss layer
-
+
Figure 1. Archetecture of Deep Speech 2 Network.
@@ -141,7 +141,7 @@ TODO by Assignees ### Beam Search with CTC and LM
-
+
Figure 2. Algorithm for CTC Beam Search Decoder.
diff --git a/doc/design/ops/images/LOD-and-shape-changes-during-decoding.jpg b/doc/fluid/design/network/images/LOD-and-shape-changes-during-decoding.jpg similarity index 100% rename from doc/design/ops/images/LOD-and-shape-changes-during-decoding.jpg rename to doc/fluid/design/network/images/LOD-and-shape-changes-during-decoding.jpg diff --git a/doc/design/speech/image/beam_search.png b/doc/fluid/design/network/images/beam_search.png similarity index 100% rename from doc/design/speech/image/beam_search.png rename to doc/fluid/design/network/images/beam_search.png diff --git a/doc/design/speech/image/ds2_network.png b/doc/fluid/design/network/images/ds2_network.png similarity index 100% rename from doc/design/speech/image/ds2_network.png rename to doc/fluid/design/network/images/ds2_network.png diff --git a/doc/design/ops/sequence_decoder.md b/doc/fluid/design/network/sequence_decoder.md similarity index 100% rename from doc/design/ops/sequence_decoder.md rename to doc/fluid/design/network/sequence_decoder.md diff --git a/doc/design/auto_gradient_check.md b/doc/fluid/design/others/auto_gradient_check.md similarity index 100% rename from doc/design/auto_gradient_check.md rename to doc/fluid/design/others/auto_gradient_check.md diff --git a/doc/design/dcgan.png b/doc/fluid/design/others/dcgan.png similarity index 100% rename from doc/design/dcgan.png rename to doc/fluid/design/others/dcgan.png diff --git a/doc/design/gan_api.md b/doc/fluid/design/others/gan_api.md similarity index 100% rename from doc/design/gan_api.md rename to doc/fluid/design/others/gan_api.md diff --git a/doc/design/graph.md b/doc/fluid/design/others/graph.md similarity index 100% rename from doc/design/graph.md rename to doc/fluid/design/others/graph.md diff --git a/doc/design/graph_survey.md b/doc/fluid/design/others/graph_survey.md similarity index 100% rename from doc/design/graph_survey.md rename to doc/fluid/design/others/graph_survey.md diff --git a/doc/design/images/graph_construction_example.bash b/doc/fluid/design/others/images/graph_construction_example.bash similarity index 100% rename from doc/design/images/graph_construction_example.bash rename to doc/fluid/design/others/images/graph_construction_example.bash diff --git a/doc/design/images/graph_construction_example.dot b/doc/fluid/design/others/images/graph_construction_example.dot similarity index 100% rename from doc/design/images/graph_construction_example.dot rename to doc/fluid/design/others/images/graph_construction_example.dot diff --git a/doc/design/images/graph_construction_example_all.png b/doc/fluid/design/others/images/graph_construction_example_all.png similarity index 100% rename from doc/design/images/graph_construction_example_all.png rename to doc/fluid/design/others/images/graph_construction_example_all.png diff --git a/doc/design/images/graph_construction_example_forward_backward.png b/doc/fluid/design/others/images/graph_construction_example_forward_backward.png similarity index 100% rename from doc/design/images/graph_construction_example_forward_backward.png rename to doc/fluid/design/others/images/graph_construction_example_forward_backward.png diff --git a/doc/design/images/graph_construction_example_forward_only.png b/doc/fluid/design/others/images/graph_construction_example_forward_only.png similarity index 100% rename from doc/design/images/graph_construction_example_forward_only.png rename to doc/fluid/design/others/images/graph_construction_example_forward_only.png diff --git a/doc/design/parameters_in_cpp.md b/doc/fluid/design/others/parameters_in_cpp.md similarity index 100% rename from doc/design/parameters_in_cpp.md rename to doc/fluid/design/others/parameters_in_cpp.md diff --git a/doc/design/simple_op_design.md b/doc/fluid/design/others/simple_op_design.md similarity index 100% rename from doc/design/simple_op_design.md rename to doc/fluid/design/others/simple_op_design.md diff --git a/doc/design/test.dot b/doc/fluid/design/others/test.dot similarity index 100% rename from doc/design/test.dot rename to doc/fluid/design/others/test.dot diff --git a/doc/design/test.dot.png b/doc/fluid/design/others/test.dot.png similarity index 100% rename from doc/design/test.dot.png rename to doc/fluid/design/others/test.dot.png diff --git a/doc/fluid/dev/api_doc_std_cn.md b/doc/fluid/dev/api_doc_std_cn.md new file mode 100644 index 0000000000000000000000000000000000000000..5596b2653ae6ed9917f77dad08f926bcb1fb3419 --- /dev/null +++ b/doc/fluid/dev/api_doc_std_cn.md @@ -0,0 +1,220 @@ +# API注释撰写标准 + +- [API注释模块](#API注释模块) +- [格式及示例](#格式及示例) +- [完整示例](#完整示例) + + +## API注释模块 + +API文档须包含以下几个模块（排列顺序为文档撰写顺序）： + +- Python API Definition + + API的代码定义。 + +- Function Description + + API的功能描述。描述该API的含义、作用或对输入所做的操作，及参考文献和对应链接（如果有），必要时给出公式，并解释公式中关键变量的含义。 + +- Args Description + + API参数介绍。按代码定义中的参数顺序逐个介绍，介绍内容包含数据类型、默认值（如果有）、含义等。 + +- Returns + + API返回值介绍。介绍返回值含义，必要时给出对应的形状。若返回值为包含多个参数的tuple，则按顺序逐个介绍各参数。 + +- Raises（如果有） + + 可能抛出的异常或错误及可能的产生原因，当可能抛出多种异常或错误时应分条列出。 + +- Note（如果有） + + 注意事项。当有多条注意事项时，应分条列出。 + +- Examples + + API的使用示例。 + + +## 格式及示例 + +API文档须使用reStructuredText格式撰写，该格式详情请参考[链接](http://sphinx-doc-zh.readthedocs.io/en/latest/rest.html)。API文档各模块的内容格式及示例如下（以下以fc为例进行说明）： + +- Python API Definition + + - 格式： + + [Python API Definition] + + - 示例 + + ``` + fc(input, + size, + num_flatten_dims=1, + param_attr=None, + bias_attr=None, + act=None, + name=None, + main_program=None, + startup_program=None) + ``` + +- Function Description + + - 格式 + + 本模块应包含以下内容（排列顺序为文档撰写顺序）： + + [Function Description] + + [Formula] + + [Symbols' Descriptions if necessary] + + [References if necessary] + + - 示例 + + [Function Description] + + ``` + **Fully Connected Layer** + + The fully connected layer can take multiple tensors as its inputs. It + creates a variable called weights for each input tensor, which represents + a fully connected weight matrix from each input unit to each output unit. + The fully connected layer multiplies each input tensor with its coresponding + weight to produce an output Tensor. If multiple input tensors are given, + the results of multiple multiplications will be sumed up. If bias_attr is + not None, a bias variable will be created and added to the output. Finally, + if activation is not None, it will be applied to the output as well. + ``` + + [Formula] + + ``` + This process can be formulated as follows: + + .. math:: + + Out = Act({\sum_{i=0}^{N-1}X_iW_i + b}) + ``` + + [Symbols' Descriptions if necessary] + + ``` + In the above equation: + + * :math:`N`: Number of the input. + * :math:`X_i`: The input tensor. + * :math:`W`: The weights created by this layer. + * :math:`b`: The bias parameter created by this layer (if needed). + * :math:`Act`: The activation function. + * :math:`Out`: The output tensor. + ``` + + [References if necessary] + + 因fc没有必要列出的参考文献，故该内容省略。其他情况下需明确给出对应的参考文献和对应连接，以 layer_norm 为例： + + ``` + Refer to `Layer Normalization `_ for more details. + ``` + + +- Args Description + + - 格式 + + \[Arg's Name\][(Data Type, Default Value)][Description] + + - 示例 + + fc的部分参数注释如下： + + ``` + Args: + input (Variable|list of Variable): The input tensor(s) of this layer, and the dimension of + the input tensor(s) is at least 2. + param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable + parameters/weights of this layer. + name (str, default None): The name of this layer. + ``` + +- Returns + + - 格式 + + [Name][Shape] + + - 示例 + + ``` + Returns: + A tensor variable storing the transformation result. + ``` + + 当返回值为包含多个参数的tuple时，应按顺序逐个介绍各参数，以dynamic_lstm为例： + + ``` + Returns: + A tuple containing: + The hidden state of LSTM whose shape is (T X D). + The cell state of LSTM whose shape is (T X D). + ``` + +- Raises + + - 格式 + + [Exception Type][Condition] + + - 示例 + + ``` + Raises: + ValueError: If the rank of the input is less than 2. + ``` + +- Note + + - 格式 + + [Note] + + - 示例 + + fc没有注意事项，故该模块省略不写。如有注意事项应明确给出，当有多条注意事项，须分条列出，以scaled\_dot\_product\_attention为例： + + ``` + Note: + 1. When num_heads > 1, three linear projections are learned respectively + to map input queries, keys and values into queries', keys' and values'. + queries', keys' and values' have the same shapes with queries, keys + and values. + 2. When num_heads == 1, scaled_dot_product_attention has no learnable + parameters. + ``` + +- Examples + + - 格式 + + \[Python Code Snipper] + + - 示例 + + ``` + Examples: + .. code-block:: python + + data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") + fc = fluid.layers.fc(input=data, size=1000, act="tanh") + ``` + +## 完整示例 + +fc 的完整注释见[示例](src/fc.py)。 diff --git a/doc/design/ci_build_whl.png b/doc/fluid/dev/ci_build_whl.png similarity index 100% rename from doc/design/ci_build_whl.png rename to doc/fluid/dev/ci_build_whl.png diff --git a/paddle/fluid/operators/op_documentation/name_convention.md b/doc/fluid/dev/name_convention.md similarity index 100% rename from paddle/fluid/operators/op_documentation/name_convention.md rename to doc/fluid/dev/name_convention.md diff --git a/paddle/fluid/operators/op_documentation/op_markdown_format.md b/doc/fluid/dev/op_markdown_format.md similarity index 100% rename from paddle/fluid/operators/op_documentation/op_markdown_format.md rename to doc/fluid/dev/op_markdown_format.md diff --git a/doc/design/releasing_process.md b/doc/fluid/dev/releasing_process.md similarity index 100% rename from doc/design/releasing_process.md rename to doc/fluid/dev/releasing_process.md diff --git a/doc/fluid/dev/src/fc.py b/doc/fluid/dev/src/fc.py new file mode 100644 index 0000000000000000000000000000000000000000..3b074821cc2276a29b2a8639e82199fcf4d72020 --- /dev/null +++ b/doc/fluid/dev/src/fc.py @@ -0,0 +1,81 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# 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. + + +def fc(input, + size, + num_flatten_dims=1, + param_attr=None, + bias_attr=None, + act=None, + name=None): + """ + **Fully Connected Layer** + + The fully connected layer can take multiple tensors as its inputs. It + creates a variable called weights for each input tensor, which represents + a fully connected weight matrix from each input unit to each output unit. + The fully connected layer multiplies each input tensor with its coresponding + weight to produce an output Tensor. If multiple input tensors are given, + the results of multiple multiplications will be sumed up. If bias_attr is + not None, a bias variable will be created and added to the output. Finally, + if activation is not None, it will be applied to the output as well. + + This process can be formulated as follows: + + .. math:: + + Out = Act({\sum_{i=0}^{N-1}X_iW_i + b}) + + In the above equation: + + * :math:`N`: Number of the input. + * :math:`X_i`: The input tensor. + * :math:`W`: The weights created by this layer. + * :math:`b`: The bias parameter created by this layer (if needed). + * :math:`Act`: The activation function. + * :math:`Out`: The output tensor. + + Args: + input (Variable|list of Variable): The input tensor(s) of this layer, and the dimension of + the input tensor(s) is at least 2. + size(int): The number of output units in this layer. + num_flatten_dims (int, default 1): The fc layer can accept an input tensor with more than + two dimensions. If this happens, the multidimensional tensor will first be flattened + into a 2-dimensional matrix. The parameter `num_flatten_dims` determines how the input + tensor is flattened: the first `num_flatten_dims` (inclusive, index starts from 1) + dimensions will be flatten to form the first dimension of the final matrix (height of + the matrix), and the rest `rank(X) - num_flatten_dims` dimensions are flattened to + form the second dimension of the final matrix (width of the matrix). For example, suppose + `X` is a 6-dimensional tensor with a shape [2, 3, 4, 5, 6], and `num_flatten_dims` = 3. + Then, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30]. + param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable + parameters/weights of this layer. + bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias + of this layer. If it is set to None, no bias will be added to the output units. + act (str, default None): Activation to be applied to the output of this layer. + name (str, default None): The name of this layer. + + Returns: + A tensor variable storing the transformation result. + + Raises: + ValueError: If rank of the input tensor is less than 2. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") + fc = fluid.layers.fc(input=data, size=1000, act="tanh") + """ diff --git a/doc/design/support_new_device.md b/doc/fluid/dev/support_new_device.md similarity index 100% rename from doc/design/support_new_device.md rename to doc/fluid/dev/support_new_device.md diff --git a/doc/fluid/dev/use_eigen_cn.md b/doc/fluid/dev/use_eigen_cn.md index 1367323b71277984834d9d4f0d9bea0f69478479..f36843b4408c21bdca1fa83853e5b0a40116791c 100644 --- a/doc/fluid/dev/use_eigen_cn.md +++ b/doc/fluid/dev/use_eigen_cn.md @@ -107,7 +107,7 @@ void Compute(const framework::ExecutionContext& context) const override { ### 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。 +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/fluid/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 @@ -125,7 +125,7 @@ From是EigenTensor模板提供的一个接口，可以实现从paddle::framework 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)。 +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/fluid/framework/eigen_test.cc)。 diff --git a/doc/fluid/dev/use_eigen_en.md b/doc/fluid/dev/use_eigen_en.md index e169106e12f5d62696f1f0e7163562793b32c18c..3a466f73d1f9b94a29b171015279c782ca50bd02 100644 --- a/doc/fluid/dev/use_eigen_en.md +++ b/doc/fluid/dev/use_eigen_en.md @@ -107,7 +107,7 @@ void Compute(const framework::ExecutionContext& context) const override { ### paddle::framework::Tensor到EigenTensor的转换 -As shown above, in actual computation, we need to transform the input and output `Tensor`s into formats Eigen supports. We show some functions in [eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen.h) to implement the transformation from `paddle::framework::Tensor`to `EigenTensor/EigenMatrix/EigenVector/EigenScalar`. +As shown above, in actual computation, we need to transform the input and output `Tensor`s into formats Eigen supports. We show some functions in [eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/eigen.h) to implement the transformation from `paddle::framework::Tensor`to `EigenTensor/EigenMatrix/EigenVector/EigenScalar`. Using EigenTensor as an example: @@ -125,7 +125,7 @@ EigenTensor::Type et = EigenTensor::From(t); In Eigen, tensors with different ranks are different types, with `Vector` bring a rank-1 instance. Note that `EigenVector::From` uses a transformation from an 1-dimensional Paddle tensor to a 1-dimensional Eigen tensor while `EigenVector::Flatten` reshapes a paddle tensor and flattens it into a 1-dimensional Eigen tensor. Both resulting tensors are still typed EigenVector. -For more transformations, see the [unit tests](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen_test.cc) in the `eigen_test.cc` file. +For more transformations, see the [unit tests](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/eigen_test.cc) in the `eigen_test.cc` file. diff --git a/doc/design/reader/README.md b/doc/fluid/getstarted/concepts/reader/README.md similarity index 100% rename from doc/design/reader/README.md rename to doc/fluid/getstarted/concepts/reader/README.md diff --git a/doc/design/model_format.md b/doc/fluid/getstarted/concepts/save_model/model_format.md similarity index 100% rename from doc/design/model_format.md rename to doc/fluid/getstarted/concepts/save_model/model_format.md diff --git a/doc/v2/howto/optimization/pprof_1.png b/doc/fluid/howto/optimization/pprof_1.png similarity index 100% rename from doc/v2/howto/optimization/pprof_1.png rename to doc/fluid/howto/optimization/pprof_1.png diff --git a/doc/v2/howto/optimization/pprof_2.png b/doc/fluid/howto/optimization/pprof_2.png similarity index 100% rename from doc/v2/howto/optimization/pprof_2.png rename to doc/fluid/howto/optimization/pprof_2.png diff --git a/doc/fluid/howto/optimization/timeline.jpeg b/doc/fluid/howto/optimization/timeline.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..38ec3f80c982857531f30a8bb0fa26ea5bf05385 Binary files /dev/null and b/doc/fluid/howto/optimization/timeline.jpeg differ diff --git a/doc/fluid/howto/optimization/timeline.md b/doc/fluid/howto/optimization/timeline.md new file mode 100644 index 0000000000000000000000000000000000000000..9d9565a3e698a83ca465c5da83ff892360c33b8f --- /dev/null +++ b/doc/fluid/howto/optimization/timeline.md @@ -0,0 +1,27 @@ +## how to use timeline tool to do profile + +1. Add `with profiler.profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number. + + ```python + with profiler.profiler('All', 'total', '/tmp/profile') as prof: + for pass_id in range(pass_num): + for batch_id, data in enumerate(train_reader()): + exe.run(fluid.default_main_program(), + feed=feeder.feed(data), + fetch_list=[], + use_program_cache=True) + ... + ``` + +1. Run `python paddle/tools/timeline.py` to process `/tmp/profile`, it will generate another +file `/tmp/timeline` by default. You can change the path by cmd parameter, please take a look at +[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py) for details. + +1. Open chrome and visit , use `load` button to load the generated `timeline` file. + + ![chrome tracing](./tracing.jpeg) + +1. The resulting timeline should be like: + + + ![chrome timeline](./timeline.jpeg) diff --git a/doc/fluid/howto/optimization/tracing.jpeg b/doc/fluid/howto/optimization/tracing.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..3a49fc4f8a401a9463b0157e2f38c164ca02dcc5 Binary files /dev/null and b/doc/fluid/howto/optimization/tracing.jpeg differ diff --git a/doc/design/error_clip.md b/doc/fluid/howto/performance/error_clip.md similarity index 100% rename from doc/design/error_clip.md rename to doc/fluid/howto/performance/error_clip.md diff --git a/doc/design/images/profiler.png b/doc/fluid/howto/performance/images/profiler.png similarity index 100% rename from doc/design/images/profiler.png rename to doc/fluid/howto/performance/images/profiler.png diff --git a/doc/design/profiler.md b/doc/fluid/howto/performance/profiler.md similarity index 100% rename from doc/design/profiler.md rename to doc/fluid/howto/performance/profiler.md diff --git a/doc/design/images/multigpu_allreduce.graffle b/doc/fluid/howto/third_party/images/multigpu_allreduce.graffle similarity index 100% rename from doc/design/images/multigpu_allreduce.graffle rename to doc/fluid/howto/third_party/images/multigpu_allreduce.graffle diff --git a/doc/design/images/multigpu_allreduce.png b/doc/fluid/howto/third_party/images/multigpu_allreduce.png similarity index 100% rename from doc/design/images/multigpu_allreduce.png rename to doc/fluid/howto/third_party/images/multigpu_allreduce.png diff --git a/doc/design/images/multigpu_before_convert.graffle b/doc/fluid/howto/third_party/images/multigpu_before_convert.graffle similarity index 100% rename from doc/design/images/multigpu_before_convert.graffle rename to doc/fluid/howto/third_party/images/multigpu_before_convert.graffle diff --git a/doc/design/images/multigpu_before_convert.png b/doc/fluid/howto/third_party/images/multigpu_before_convert.png similarity index 100% rename from doc/design/images/multigpu_before_convert.png rename to doc/fluid/howto/third_party/images/multigpu_before_convert.png diff --git a/doc/design/mkl/mkldnn_fluid.md b/doc/fluid/howto/third_party/mkldnn_fluid.md similarity index 100% rename from doc/design/mkl/mkldnn_fluid.md rename to doc/fluid/howto/third_party/mkldnn_fluid.md diff --git a/doc/design/paddle_nccl.md b/doc/fluid/howto/third_party/paddle_nccl.md similarity index 100% rename from doc/design/paddle_nccl.md rename to doc/fluid/howto/third_party/paddle_nccl.md diff --git a/doc/fluid/read_source.md b/doc/fluid/read_source.md index edf46aff8c6cc9fc01d26c6453b3a8123238ef91..bb6d4563f5617fb98af055bca2f6f0479bdb4393 100644 --- a/doc/fluid/read_source.md +++ b/doc/fluid/read_source.md @@ -2,17 +2,17 @@ Examples: https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/fluid/tests/book -Core: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework +Core: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/framework -Operator: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators +Operator: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators -Memory: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory +Memory: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/memory -Platform: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform +Platform: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/platform # Compile Time -The following **defines** the NN. The definition goes into this [protocol buffer](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto). +The following **defines** the NN. The definition goes into this [protocol buffer](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto). ```python x = fluid.layers.data(name='x', shape=[13], dtype='float32') @@ -29,10 +29,10 @@ sgd_optimizer.minimize(avg_cost) - Variables: `x`, `y`, `y_predict`, `cost` and `avg_cost`. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/framework.py#) - Layers: `fluid.layers.data`, `fluid.layers.fc` and `fluid.layers.mean` are layers. [Python](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/fluid/layers) - Every Layer has one or more operators and variables/parameters - - All the operators are defined at [`paddle/operators/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators). Other worth-looking files: - - Base class: [`paddle/framework/operator.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h) - - Operator Registration: [`paddle/framework/op_registry.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h) - - Operator Lookup: [`paddle/framework/op_info.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_info.h) + - All the operators are defined at [`paddle/fluid/operators/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators). Other worth-looking files: + - Base class: [`paddle/fluid/framework/operator.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/operator.h) + - Operator Registration: [`paddle/fluid/framework/op_registry.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/op_registry.h) + - Operator Lookup: [`paddle/fluid/framework/op_info.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/op_info.h) - Optimizer: `fluid.optimizer.SGD`. It does the following - Add backward operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/backward.py)] - Add optimizer operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/optimizer.py)] @@ -55,13 +55,13 @@ exe.run(fluid.default_main_program(), fetch_list=[avg_cost]) ``` -- Place: `place`. one of CPU, GPU or FPGA. [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h) - - The device handle are at [paddle/platform/device_context.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h) -- Executor: `fluid.Executor(place)`. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/executor.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.cc)] +- Place: `place`. one of CPU, GPU or FPGA. [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/place.h) + - The device handle are at [paddle/fluid/platform/device_context.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/device_context.h) +- Executor: `fluid.Executor(place)`. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/executor.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/executor.cc)] - Feeds the data: `feed=feeder.feed(data)` - Evaluates all the operators - Fetches the result: `fetch_list=[avg_cost]` - Other worth looking files: - - Scope: [paddle/framework/scope.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/scope.h). Where all the variables live - - Variable: [paddle/framework/variable.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h). Where all the data (most likely tensors) live - - Tensor: [paddle/framework/tensor.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h). Where we allocate memory through [`paddle/memory/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory) + - Scope: [paddle/fluid/framework/scope.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/scope.h). Where all the variables live + - Variable: [paddle/fluid/framework/variable.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/variable.h). Where all the data (most likely tensors) live + - Tensor: [paddle/fluid/framework/tensor.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/tensor.h). Where we allocate memory through [`paddle/fluid/memory/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/memory) diff --git a/doc/v2/build_and_install/pip_install_cn.rst b/doc/v2/build_and_install/pip_install_cn.rst index 8e4165da6b8135d083766c650f1092158f9d01c2..b3d882743785e8ee301b71b696230531d2b7ba58 100644 --- a/doc/v2/build_and_install/pip_install_cn.rst +++ b/doc/v2/build_and_install/pip_install_cn.rst @@ -34,15 +34,15 @@ PaddlePaddle可以使用常用的Python包管理工具 :align: center .. csv-table:: 各个版本最新的whl包 - :header: "版本说明", "cp27-cp27mu", "cp27-cp27m", "C-API" - :widths: 1, 3, 3, 3 - - "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "`paddle.tgz `_" - "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "暂无" - "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "暂无" - "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "`paddle.tgz `_" - "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "`paddle.tgz `_" - "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "`paddle.tgz `_" + :header: "版本说明", "cp27-cp27mu", "cp27-cp27m" + :widths: 1, 3, 3 + + "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" .. _pip_dependency: diff --git a/doc/v2/build_and_install/pip_install_en.rst b/doc/v2/build_and_install/pip_install_en.rst index 0d4c925b6e2731bdfd76582309aa7e8adbafa6ae..1e409d86b9775094998f72f92954f4bbc1013ea1 100644 --- a/doc/v2/build_and_install/pip_install_en.rst +++ b/doc/v2/build_and_install/pip_install_en.rst @@ -37,15 +37,15 @@ If the links below shows up the login form, just click "Log in as guest" to star :align: center .. csv-table:: whl package of each version - :header: "version", "cp27-cp27mu", "cp27-cp27m", "C-API" - :widths: 1, 3, 3, 3 - - "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "`paddle.tgz `_" - "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "Not Available" - "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "Not Available" - "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "`paddle.tgz `_" - "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "`paddle.tgz `_" - "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_", "`paddle.tgz `_" + :header: "version", "cp27-cp27mu", "cp27-cp27m" + :widths: 1, 3, 3 + + "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cpu_noavx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" + "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl `_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl `_" .. _pip_dependency: diff --git a/doc/design/cluster_train/README.md b/doc/v2/design/cluster_train/README.md similarity index 100% rename from doc/design/cluster_train/README.md rename to doc/v2/design/cluster_train/README.md diff --git a/doc/design/cluster_train/checkpointing.md b/doc/v2/design/cluster_train/checkpointing.md similarity index 100% rename from doc/design/cluster_train/checkpointing.md rename to doc/v2/design/cluster_train/checkpointing.md diff --git a/doc/design/cluster_train/data_dispatch.md b/doc/v2/design/cluster_train/data_dispatch.md similarity index 100% rename from doc/design/cluster_train/data_dispatch.md rename to doc/v2/design/cluster_train/data_dispatch.md diff --git a/doc/design/cluster_train/large_model_dist_train.md b/doc/v2/design/cluster_train/large_model_dist_train.md similarity index 100% rename from doc/design/cluster_train/large_model_dist_train.md rename to doc/v2/design/cluster_train/large_model_dist_train.md diff --git a/doc/design/cluster_train/master_server.md b/doc/v2/design/cluster_train/master_server.md similarity index 100% rename from doc/design/cluster_train/master_server.md rename to doc/v2/design/cluster_train/master_server.md diff --git a/doc/design/cluster_train/pserver_client.md b/doc/v2/design/cluster_train/pserver_client.md similarity index 100% rename from doc/design/cluster_train/pserver_client.md rename to doc/v2/design/cluster_train/pserver_client.md diff --git a/doc/design/cluster_train/remote_parameter_updater.md b/doc/v2/design/cluster_train/remote_parameter_updater.md similarity index 100% rename from doc/design/cluster_train/remote_parameter_updater.md rename to doc/v2/design/cluster_train/remote_parameter_updater.md diff --git a/doc/design/cluster_train/save_model.md b/doc/v2/design/cluster_train/save_model.md similarity index 100% rename from doc/design/cluster_train/save_model.md rename to doc/v2/design/cluster_train/save_model.md diff --git a/doc/design/cluster_train/src/checkpointing.png b/doc/v2/design/cluster_train/src/checkpointing.png similarity index 100% rename from doc/design/cluster_train/src/checkpointing.png rename to doc/v2/design/cluster_train/src/checkpointing.png diff --git a/doc/design/cluster_train/src/data_dispatch.png b/doc/v2/design/cluster_train/src/data_dispatch.png similarity index 100% rename from doc/design/cluster_train/src/data_dispatch.png rename to doc/v2/design/cluster_train/src/data_dispatch.png diff --git a/doc/design/cluster_train/src/dataset.graffle b/doc/v2/design/cluster_train/src/dataset.graffle similarity index 100% rename from doc/design/cluster_train/src/dataset.graffle rename to doc/v2/design/cluster_train/src/dataset.graffle diff --git a/doc/design/cluster_train/src/dataset.png b/doc/v2/design/cluster_train/src/dataset.png similarity index 100% rename from doc/design/cluster_train/src/dataset.png rename to doc/v2/design/cluster_train/src/dataset.png diff --git a/doc/design/cluster_train/src/file_storage.graffle b/doc/v2/design/cluster_train/src/file_storage.graffle similarity index 100% rename from doc/design/cluster_train/src/file_storage.graffle rename to doc/v2/design/cluster_train/src/file_storage.graffle diff --git a/doc/design/cluster_train/src/file_storage.png b/doc/v2/design/cluster_train/src/file_storage.png similarity index 100% rename from doc/design/cluster_train/src/file_storage.png rename to doc/v2/design/cluster_train/src/file_storage.png diff --git a/doc/design/cluster_train/src/init_lock.graffle b/doc/v2/design/cluster_train/src/init_lock.graffle similarity index 100% rename from doc/design/cluster_train/src/init_lock.graffle rename to doc/v2/design/cluster_train/src/init_lock.graffle diff --git a/doc/design/cluster_train/src/init_lock.png b/doc/v2/design/cluster_train/src/init_lock.png similarity index 100% rename from doc/design/cluster_train/src/init_lock.png rename to doc/v2/design/cluster_train/src/init_lock.png diff --git a/doc/design/cluster_train/src/paddle-cloud-in-data-center.png b/doc/v2/design/cluster_train/src/paddle-cloud-in-data-center.png similarity index 100% rename from doc/design/cluster_train/src/paddle-cloud-in-data-center.png rename to doc/v2/design/cluster_train/src/paddle-cloud-in-data-center.png diff --git a/doc/design/cluster_train/src/paddle-etcd.graffle b/doc/v2/design/cluster_train/src/paddle-etcd.graffle similarity index 100% rename from doc/design/cluster_train/src/paddle-etcd.graffle rename to doc/v2/design/cluster_train/src/paddle-etcd.graffle diff --git a/doc/design/cluster_train/src/paddle-etcd.png b/doc/v2/design/cluster_train/src/paddle-etcd.png similarity index 100% rename from doc/design/cluster_train/src/paddle-etcd.png rename to doc/v2/design/cluster_train/src/paddle-etcd.png diff --git a/doc/design/cluster_train/src/paddle-model-sharding.graffle b/doc/v2/design/cluster_train/src/paddle-model-sharding.graffle similarity index 100% rename from doc/design/cluster_train/src/paddle-model-sharding.graffle rename to doc/v2/design/cluster_train/src/paddle-model-sharding.graffle diff --git a/doc/design/cluster_train/src/paddle-model-sharding.png b/doc/v2/design/cluster_train/src/paddle-model-sharding.png similarity index 100% rename from doc/design/cluster_train/src/paddle-model-sharding.png rename to doc/v2/design/cluster_train/src/paddle-model-sharding.png diff --git a/doc/design/cluster_train/src/paddle-ps-0.png b/doc/v2/design/cluster_train/src/paddle-ps-0.png similarity index 100% rename from doc/design/cluster_train/src/paddle-ps-0.png rename to doc/v2/design/cluster_train/src/paddle-ps-0.png diff --git a/doc/design/cluster_train/src/paddle-ps-1.png b/doc/v2/design/cluster_train/src/paddle-ps-1.png similarity index 100% rename from doc/design/cluster_train/src/paddle-ps-1.png rename to doc/v2/design/cluster_train/src/paddle-ps-1.png diff --git a/doc/design/cluster_train/src/paddle-ps.graffle b/doc/v2/design/cluster_train/src/paddle-ps.graffle similarity index 100% rename from doc/design/cluster_train/src/paddle-ps.graffle rename to doc/v2/design/cluster_train/src/paddle-ps.graffle diff --git a/doc/design/cluster_train/src/paddle-task-queues.graffle b/doc/v2/design/cluster_train/src/paddle-task-queues.graffle similarity index 100% rename from doc/design/cluster_train/src/paddle-task-queues.graffle rename to doc/v2/design/cluster_train/src/paddle-task-queues.graffle diff --git a/doc/design/cluster_train/src/paddle-task-queues.png b/doc/v2/design/cluster_train/src/paddle-task-queues.png similarity index 100% rename from doc/design/cluster_train/src/paddle-task-queues.png rename to doc/v2/design/cluster_train/src/paddle-task-queues.png diff --git a/doc/design/cluster_train/src/paddle-task-states.graffle b/doc/v2/design/cluster_train/src/paddle-task-states.graffle similarity index 100% rename from doc/design/cluster_train/src/paddle-task-states.graffle rename to doc/v2/design/cluster_train/src/paddle-task-states.graffle diff --git a/doc/design/cluster_train/src/paddle-task-states.png b/doc/v2/design/cluster_train/src/paddle-task-states.png similarity index 100% rename from doc/design/cluster_train/src/paddle-task-states.png rename to doc/v2/design/cluster_train/src/paddle-task-states.png diff --git a/doc/design/cluster_train/src/pserver_init.graffle b/doc/v2/design/cluster_train/src/pserver_init.graffle similarity index 100% rename from doc/design/cluster_train/src/pserver_init.graffle rename to doc/v2/design/cluster_train/src/pserver_init.graffle diff --git a/doc/design/cluster_train/src/pserver_init.png b/doc/v2/design/cluster_train/src/pserver_init.png similarity index 100% rename from doc/design/cluster_train/src/pserver_init.png rename to doc/v2/design/cluster_train/src/pserver_init.png diff --git a/doc/design/cluster_train/src/submit-job.graffle b/doc/v2/design/cluster_train/src/submit-job.graffle similarity index 100% rename from doc/design/cluster_train/src/submit-job.graffle rename to doc/v2/design/cluster_train/src/submit-job.graffle diff --git a/doc/design/cluster_train/src/submit-job.png b/doc/v2/design/cluster_train/src/submit-job.png similarity index 100% rename from doc/design/cluster_train/src/submit-job.png rename to doc/v2/design/cluster_train/src/submit-job.png diff --git a/doc/design/cluster_train/src/trainer.graffle b/doc/v2/design/cluster_train/src/trainer.graffle similarity index 100% rename from doc/design/cluster_train/src/trainer.graffle rename to doc/v2/design/cluster_train/src/trainer.graffle diff --git a/doc/design/cluster_train/src/trainer.png b/doc/v2/design/cluster_train/src/trainer.png similarity index 100% rename from doc/design/cluster_train/src/trainer.png rename to doc/v2/design/cluster_train/src/trainer.png diff --git a/doc/design/cluster_train/submit-job.md b/doc/v2/design/cluster_train/submit-job.md similarity index 100% rename from doc/design/cluster_train/submit-job.md rename to doc/v2/design/cluster_train/submit-job.md diff --git a/doc/design/mkl/image/engine.png b/doc/v2/design/mkl/image/engine.png similarity index 100% rename from doc/design/mkl/image/engine.png rename to doc/v2/design/mkl/image/engine.png diff --git a/doc/design/mkl/image/gradients.png b/doc/v2/design/mkl/image/gradients.png similarity index 100% rename from doc/design/mkl/image/gradients.png rename to doc/v2/design/mkl/image/gradients.png diff --git a/doc/design/mkl/image/layers.png b/doc/v2/design/mkl/image/layers.png similarity index 100% rename from doc/design/mkl/image/layers.png rename to doc/v2/design/mkl/image/layers.png diff --git a/doc/design/mkl/image/matrix.png b/doc/v2/design/mkl/image/matrix.png similarity index 100% rename from doc/design/mkl/image/matrix.png rename to doc/v2/design/mkl/image/matrix.png diff --git a/doc/design/mkl/image/overview.png b/doc/v2/design/mkl/image/overview.png similarity index 100% rename from doc/design/mkl/image/overview.png rename to doc/v2/design/mkl/image/overview.png diff --git a/doc/design/mkl/mkl_packed.md b/doc/v2/design/mkl/mkl_packed.md similarity index 100% rename from doc/design/mkl/mkl_packed.md rename to doc/v2/design/mkl/mkl_packed.md diff --git a/doc/design/mkl/mkldnn.md b/doc/v2/design/mkl/mkldnn.md similarity index 100% rename from doc/design/mkl/mkldnn.md rename to doc/v2/design/mkl/mkldnn.md diff --git a/doc/v2/dev/index_cn.rst b/doc/v2/dev/index_cn.rst index c488191b8174531905e44cb9443ee539d4cb1ed3..aee3c68de05de26df3cd79170fa7f4ecad4bf386 100644 --- a/doc/v2/dev/index_cn.rst +++ b/doc/v2/dev/index_cn.rst @@ -1,9 +1,24 @@ 开发标准 ======== +PaddlePaddle遵守如下三个部分的代码和文档规范。 + +PaddlePaddle使用git做版本管理，docker作为构建和测试环境。代码中包含了Cuda, C++, Python, Shell等多种编程语言。语言规范遵守Google C++ Style, Pep-8, 代码库中包含自动化检查工具做风格检查。代码注释需要遵守Doxygen规范，不满足风格要求的代码会编译失败。关于如何使用git, 构建测试及代码开发, 我们提供了如下指南。 .. toctree:: :maxdepth: 1 contribute_to_paddle_cn.md + +PaddlePaddle面向国内外用户，包含了中文和英文两部分的文档。设计文档和issue问题描述都推荐使用英文。对于设计文档，重在问题描述，背景阐述，然后才是解决方案。文档由Sphinx生成，因此代码注释也需要符合Sphinx文档标准。推荐本地使用paddlepaddle.org工具编译生成和预览文档，请参阅如下文档。 + +.. toctree:: + :maxdepth: 1 + write_docs_cn.rst + +PaddlePaddle V2 使用新增Layer方式定义新的操作。组合基础API可以实现多种复杂Layer, 满足绝大多数应用。如需要定制Layer，请参阅如下文档，欢迎提交patch。 + +.. toctree:: + :maxdepth: 1 + new_layer_cn.rst diff --git a/doc/v2/dev/new_layer_cn.rst b/doc/v2/dev/new_layer_cn.rst index 0ded1c262adad44f4df000ef2933c7b68050f2fc..3115654b2bd87995fa63bb7828fd1b3039aea8cc 100644 --- a/doc/v2/dev/new_layer_cn.rst +++ b/doc/v2/dev/new_layer_cn.rst @@ -16,7 +16,7 @@ 下图是一个全连接层的示意图。在全连接层中，每个输出节点都连接到所有的输入节点上。 -.. image:: FullyConnected.jpg +.. image:: src/FullyConnected.jpg :align: center :scale: 60 % diff --git a/doc/v2/dev/new_layer_en.rst b/doc/v2/dev/new_layer_en.rst index 110a9fb38f890a766bb4480e91feb22d3b0838a5..b05bb45f11eb253dfb87d6283c29ec6689394d22 100644 --- a/doc/v2/dev/new_layer_en.rst +++ b/doc/v2/dev/new_layer_en.rst @@ -16,7 +16,7 @@ First we need to derive equations of the *forward* and *backward* part of the la The illustration of a fully connected layer is shown in the following figure. In a fully connected layer, all output nodes are connected to all the input nodes. -.. image:: FullyConnected.jpg +.. image:: src/FullyConnected.jpg :align: center :scale: 60 % diff --git a/doc/v2/dev/FullyConnected.jpg b/doc/v2/dev/src/FullyConnected.jpg similarity index 100% rename from doc/v2/dev/FullyConnected.jpg rename to doc/v2/dev/src/FullyConnected.jpg diff --git a/doc/v2/dev/src/doc_en.png b/doc/v2/dev/src/doc_en.png new file mode 100644 index 0000000000000000000000000000000000000000..ed6b9178fba91a3bdf45ae797a9924f84146fbc8 Binary files /dev/null and b/doc/v2/dev/src/doc_en.png differ diff --git a/doc/v2/dev/write_docs_cn.rst b/doc/v2/dev/write_docs_cn.rst index f79769b810b91c6984016d95f40b89186bfb61b0..a055bb04c0c093c9159290067e5ccbd2525cd519 100644 --- a/doc/v2/dev/write_docs_cn.rst +++ b/doc/v2/dev/write_docs_cn.rst @@ -2,20 +2,19 @@ 如何贡献文档 ############# -PaddlePaddle的文档包括英文文档 ``doc`` 和中文文档 ``doc_cn`` 两个部分。文档都是通过 `cmake`_ 驱动 `sphinx`_ 编译生成，生成后的文档分别存储在编译目录的 ``doc`` 和 ``doc_cn`` 两个子目录下。 -也可以利用PaddlePaddle 工具来编译文档，这个情况下所有的文件会存在整理过的的文件目录 .ppo_workspace/content 下 +PaddlePaddle的文档包括中英文两个部分。文档都是通过 ``cmake`` 驱动 ``sphinx`` 编译生成，也可以利用paddlepaddle.org工具来编译和预览文档。 如何构建文档 ============ -PaddlePaddle的文档构建有三种方式。 +PaddlePaddle的文档构建有两种方式，分别为使用paddlepaddle.org工具和不使用paddlepaddle.org工具，两种方式都有各自的优点，前者方便预览，后者方便开发者进行调试。这两种方式中又分别有使用docker和不使用docker的两种构建方法。 使用PaddlePaddle.org工具 --------------- -这个是目前推荐的使用方法。除了可以自动编译文档，也可以直接在网页预览文档。 +------------------------ +这个是目前推荐的使用方法。除了可以自动编译文档，还可以直接在网页中预览文档，需要注意的是，采用后续说明的其它方式虽然也可以预览文档，但是文档的样式与官网文档是不一致的，使用PaddlePaddle.org工具进行编译才能产生与官网文档样式一致的预览效果。 -文件工具是使用Docker，需要在系统里先安装好Docker工具包。Docker安装请参考Docker的官网。安装好Docker之后及可用以下命令启动工具 +PaddlePaddle.org工具可以配合Docker使用，需要在系统里先安装好Docker工具包。Docker安装请参考 `Docker的官网 `_ 。安装好Docker之后即可用以下命令启动工具 .. code-block:: bash @@ -35,7 +34,7 @@ PaddlePaddle的文档构建有三种方式。 之后再用网页连到http://localhost:8000就可以在网页上生成需要的文档 编译后的文件将被存储在工作目录 /.ppo_workspace/content。 -如果不想使用 Docker，你还可以通过运行Django框架直接激活工具的服务器。使用下面的命令来运行它。 +如果不想使用Docker，你还可以通过运行Django框架直接激活工具的服务器。使用下面的命令来运行它。 .. code-block:: bash @@ -62,37 +61,46 @@ PaddlePaddle的文档构建有三种方式。 想了解更多PaddlePaddle.org工具的详细信息，可以 `点击这里 `_ 。 -使用Docker构建 --------------- +不使用PaddlePaddle.org工具 +-------------------------- 使用Docker构建PaddlePaddle的文档，需要在系统里先安装好Docker工具包。Docker安装请参考 `Docker的官网 `_ 。安装好Docker之后可以使用源码目录下的脚本构建文档，即 -.. code-block:: bash +[TBD] - cd TO_YOUR_PADDLE_CLONE_PATH - cd paddle/scripts/tools/build_docs - sh build_docs.sh +如果不想使用Docker，也可以使用以下命令直接构建PaddlePaddle文档，即 -编译完成之后，会在当前目录生成两个子目录\: doc(英文文档目录)和 doc_cn(中文文档目录)。 -打开浏览器访问对应目录下的index.html即可访问本地文档。 +.. code-block:: bash -直接构建 --------- + mkdir paddle + cd paddle + git clone https://github.com/PaddlePaddle/Paddle.git + mkdir -p build + cd build + cmake .. -DCMAKE_BUILD_TYPE=Release -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON -如果提示正确，可以执行以下命令编译生成文档，即 + # 如果只需要构建使用文档，则执行以下命令 + make -j $processors gen_proto_py + make -j $processors paddle_docs paddle_docs_cn -.. code-block:: bash + # 如果只需要构建API，则执行以下命令 + make -j $processors gen_proto_py framework_py_proto + make -j $processors copy_paddle_pybind + make -j $processors paddle_api_docs + +其中$processors代表启动和CPU核一样多的进程来并行编译，可以根据本机的CPU核数设置相应的值。 + +编译完成后，进入 ``doc/v2`` 目录，如果选择构建文档则会在该目录下生成 ``cn/html/`` 、 ``en/html`` 两个子目录，选择构建API则会生成 ``api/en/html`` 目录，分别进入这些目录下，执行以下命令： + +.. code-block:: bash - cd TO_YOUR_PADDLE_CLONE_PATH - mkdir -p build - cd build - cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON - make gen_proto_py - make paddle_docs paddle_docs_cn + python -m SimpleHTTPServer 8088 -编译完成之后，会在当前目录生成两个子目录\: doc(英文文档目录)和 doc_cn(中文文档目录)。 -打开浏览器访问对应目录下的index.html即可访问本地文档。 +在浏览器中输入http://localhost:8088就可以看到编译生成的中/英文的文档页面和英文的API页面,下图为生成的英文文档首页示例。注意，示例中由于使用了sphinx的原始主题，所以页面的风格与官网并不一致，但这并不影响开发者进行调试。 +.. image:: src/doc_en.png + :align: center + :scale: 60 % 如何书写文档 ============ @@ -102,7 +110,7 @@ PaddlePaddle文档使用 `sphinx`_ 自动生成，用户可以参考sphinx教程 如何更新www.paddlepaddle.org ============================ -更新的文档以PR的形式提交到github中，提交方式参见 `贡献文档 `_ 。 +更新的文档以PR的形式提交到github中，提交方式参见 `如何贡献文档 `_ 。 目前PaddlePaddle的develop分支的文档是自动触发更新的，用户可以分别查看最新的 `中文文档 `_ 和 `英文文档 `_ 。 diff --git a/doc/v2/getstarted/index_cn.rst b/doc/v2/getstarted/index_cn.rst index 1dc141396b95bda776aeff87ac30fad6baf37bd2..75af7354be93a6eeabfa9ccf86903505402a7ca6 100644 --- a/doc/v2/getstarted/index_cn.rst +++ b/doc/v2/getstarted/index_cn.rst @@ -1,8 +1,19 @@ 新手入门 ============ + +如果需要快速了解PaddlePaddle的使用，可以参考以下指南。 + .. toctree:: :maxdepth: 1 quickstart_cn.rst + + +在使用PaddlePaddle构建应用时，需要了解一些基本概念。 +这里以一个线性回归为例子，详细介绍了PaddlePaddle的使用流程，包括数据格式，模型配置与训练等。 + +.. toctree:: + :maxdepth: 1 + concepts/use_concepts_cn.rst diff --git a/doc/v2/howto/capi/compile_paddle_lib_cn.md b/doc/v2/howto/capi/compile_paddle_lib_cn.md index fd8dec8164580b9dcb716e69f3cc5357639f17d3..e223fd33a8420abcdfdad53d1cfc5ed160a1b37e 100644 --- a/doc/v2/howto/capi/compile_paddle_lib_cn.md +++ b/doc/v2/howto/capi/compile_paddle_lib_cn.md @@ -1,22 +1,80 @@ -## 安装与编译C-API预测库 - -### 概述 - -使用 C-API 进行预测依赖于将 PaddlePaddle 核心代码编译成链接库，只需在编译时需配制下面这些编译选项： - -必须配置选项： -- `WITH_C_API`，必须配置为`ON`。 - -推荐配置选项： -- `WITH_PYTHON`，推荐配置为`OFF` -- `WITH_SWIG_PY`，推荐配置为`OFF` -- `WITH_GOLANG`，推荐设置为`OFF` - -可选配置选项： -- `WITH_GPU`，可配置为`ON/OFF` -- `WITH_MKL`，可配置为`ON/OFF` - -对推荐配置中的选项建议按照设置，以避免链接不必要的库。其它可选编译选项按需进行设定。 +## 安装、编译与链接C-API预测库 + +### 直接下载安装 + +从CI系统中下载最新的C-API开发包进行安装，用户可以从下面的表格中找到需要的版本： + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
版本说明C-API
cpu_avx_mklpaddle.tgz
cpu_avx_openblas暂无
cpu_noavx_openblaspaddle.tgz
cuda7.5_cudnn5_avx_mklpaddle.tgz
cuda8.0_cudnn5_avx_mklpaddle.tgz
cuda8.0_cudnn7_avx_mklpaddle.tgz
+ +### 从源码编译 + +用户也可以从 PaddlePaddle 核心代码编译C-API链接库，只需在编译时配制下面这些编译选项： + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
选项值
WITH_C_APION
WITH_PYTHONOFF（推荐）
WITH_SWIG_PYOFF（推荐）
WITH_GOLANGOFF（推荐）
WITH_GPUON/OFF
WITH_MKLON/OFF
+ +建议按照推荐值设置，以避免链接不必要的库。其它可选编译选项按需进行设定。 下面的代码片段从github拉取最新代码，配制编译选项（需要将PADDLE_ROOT替换为PaddlePaddle预测库的安装路径）： @@ -100,23 +158,19 @@ cmake -DCMAKE_INSTALL_PREFIX=$PADDLE_ROOT \ 目前提供三种链接方式： -1. 链接`libpaddle_capi_shared.so` 动态库 - - 使用 PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_shared.so`时，需注意： - 1. 如果编译时指定编译CPU版本，且使用`OpenBLAS`数学库，在使用C-API开发预测程序时，只需要链接`libpaddle_capi_shared.so`这一个库。 - 1. 如果是用编译时指定CPU版本，且使用`MKL`数学库，由于`MKL`库有自己独立的动态库文件，在使用PaddlePaddle C-API开发预测程序时，需要自己链接MKL链接库。 - 1. 如果编译时指定编译GPU版本，CUDA相关库会在预测程序运行时动态装载，需要将CUDA相关的库设置到`LD_LIBRARY_PATH`环境变量中。 - - 这种方式最为简便，链接相对容易，**在无特殊需求情况下，推荐使用此方式**。 - -2. 链接静态库 `libpaddle_capi_whole.a` - - 使用PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_whole.a`时，需注意： - 1. 需要指定`-Wl,--whole-archive`链接选项。 - 1. 需要显式地链接 `gflags`、`glog`、`libz`、`protobuf` 等第三方库，可在`PADDLE_ROOT/third_party`下找到。 - 1. 如果在编译 C-API 时使用OpenBLAS数学库，需要显示地链接`libopenblas.a`。 - 1. 如果在编译 C-API 是使用MKL数学库，需要显示地链接MKL的动态库。 - -3. 链接静态库 `libpaddle_capi_layers.a`和`libpaddle_capi_engine.a` - - 使用PaddlePaddle C-API 开发预测程序链接`libpaddle_capi_whole.a`时，需注意： - 1. 这种链接方式主要用于移动端预测。 - 1. 为了减少生成链接库的大小把`libpaddle_capi_whole.a`拆成以上两个静态链接库。 - 1. 需指定`-Wl,--whole-archive -lpaddle_capi_layers` 和 `-Wl,--no-whole-archive -lpaddle_capi_engine` 进行链接。 - 1. 第三方依赖库需要按照与方式2同样方法显示地进行链接。 +1. 链接`libpaddle_capi_shared.so` 动态库（这种方式最为简便，链接相对容易，**在无特殊需求情况下，推荐使用此方式**），需注意： + 1. 如果编译时指定编译CPU版本，且使用`OpenBLAS`数学库，在使用C-API开发预测程序时，只需要链接`libpaddle_capi_shared.so`这一个库。 + 1. 如果是用编译时指定CPU版本，且使用`MKL`数学库，由于`MKL`库有自己独立的动态库文件，在使用PaddlePaddle C-API开发预测程序时，需要自己链接MKL链接库。 + 1. 如果编译时指定编译GPU版本，CUDA相关库会在预测程序运行时动态装载，需要将CUDA相关的库设置到`LD_LIBRARY_PATH`环境变量中。 + +2. 链接静态库 `libpaddle_capi_whole.a`，需注意： + 1. 需要指定`-Wl,--whole-archive`链接选项。 + 1. 需要显式地链接 `gflags`、`glog`、`libz`、`protobuf` 等第三方库，可在`PADDLE_ROOT/third_party`下找到。 + 1. 如果在编译 C-API 时使用OpenBLAS数学库，需要显示地链接`libopenblas.a`。 + 1. 如果在编译 C-API 是使用MKL数学库，需要显示地链接MKL的动态库。 + +3. 链接静态库 `libpaddle_capi_layers.a`和`libpaddle_capi_engine.a`，需注意： + 1. 这种链接方式主要用于移动端预测。 + 1. 为了减少生成链接库的大小把`libpaddle_capi_whole.a`拆成以上两个静态链接库。 + 1. 需指定`-Wl,--whole-archive -lpaddle_capi_layers` 和 `-Wl,--no-whole-archive -lpaddle_capi_engine` 进行链接。 + 1. 第三方依赖库需要按照与方式2同样方法显示地进行链接。 diff --git a/doc/v2/howto/cluster/cmd_argument_cn.md b/doc/v2/howto/cluster/cmd_argument_cn.md index 40e1dde4858b802c2e703bcca4b71730facde5ef..c0ba093cbf2eac5c3b60a0b071b31776a11998f3 100644 --- a/doc/v2/howto/cluster/cmd_argument_cn.md +++ b/doc/v2/howto/cluster/cmd_argument_cn.md @@ -71,6 +71,13 @@ paddle.init( - trainer_id：**必选，默认0**，每个trainer的唯一ID，从0开始的整数 - pservers：**必选，默认127.0.0.1**，当前训练任务启动的pserver的IP列表，多个IP使用“,”隔开 +```python +trainer = paddle.trainer.SGD(..., is_local=False) +``` + +参数说明 + +- is_local: **必选, 默认True**, 是否使用PServer更新参数 ## 准备数据集 diff --git a/doc/v2/howto/cluster/cmd_argument_en.md b/doc/v2/howto/cluster/cmd_argument_en.md index 40179c28f83800c1c74a6045f8fac6841bdafeaa..df1381a00fa0fa129eecffe002164c489a4183aa 100644 --- a/doc/v2/howto/cluster/cmd_argument_en.md +++ b/doc/v2/howto/cluster/cmd_argument_en.md @@ -73,6 +73,14 @@ Parameter Description - trainer_id: **required, default 0**, ID for every trainer, start from 0. - pservers: **required, default 127.0.0.1**, list of IPs of parameter servers, separated by ",". +```python +trainer = paddle.trainer.SGD(..., is_local=False) +``` + +Parameter Description + +- is_local: **required, default True**, whether update parameters by PServer. + ## Prepare Training Dataset Here's some example code [prepare.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/prepare.py), it will download public `imikolov` dataset and split it into multiple files according to job parallelism(trainers count). Modify `SPLIT_COUNT` at the begining of `prepare.py` to change the count of output files. diff --git a/doc/v2/howto/cluster/index_en.rst b/doc/v2/howto/cluster/index_en.rst index 2640a09dcc904619bc97c9bd3f3d81a9dc307663..c965d30d54e71339cf10d4b05f25e740c81adbf9 100644 --- a/doc/v2/howto/cluster/index_en.rst +++ b/doc/v2/howto/cluster/index_en.rst @@ -1,8 +1,7 @@ Distributed Training ==================== -In this section, we'll explain how to run distributed training jobs with PaddlePaddle on different types of clusters. The diagram below shows the main architecture of a distributed trainning job: - +The effectiveness of the deep learning model is often directly related to the scale of the data: it can generally achieve better results after increasing the size of the dataset on the same model. However, it can not fit in one single computer when the amount of data increases to a certain extent. At this point, using multiple computers for distributed training is a natural solution. In distributed training, the training data is divided into multiple copies (sharding), and multiple machines participating in the training read their own data for training and collaboratively update the parameters of the overall model. .. image:: src/ps_en.png :width: 500 @@ -10,13 +9,27 @@ In this section, we'll explain how to run distributed training jobs with PaddleP - Trainer: each trainer reads the data shard, and train the neural network. Then the trainer will upload calculated "gradients" to parameter servers, and wait for parameters to be optimized on the parameter server side. When that finishes, the trainer download optimized parameters and continues its training. - Parameter server: every parameter server stores part of the whole neural network model data. They will do optimization calculations when gradients are uploaded from trainers, and then send updated parameters to trainers. -PaddlePaddle can support both synchronize stochastic gradient descent (SGD) and asynchronous SGD. +The training of synchronous random gradient descent for neural network can be achieved by cooperation of trainers and parameter servers. + +PaddlePaddle supports both synchronize stochastic gradient descent (SGD) and asynchronous SGD. -When training with synchronize SGD, PaddlePaddle uses an internal "synchronize barrier" which makes gradients update and parameter download in strict order. On the other hand, asynchronous SGD won't wait for all trainers to finish upload at a single step, this will increase the parallelism of distributed training: parameter servers do not depend on each other, they'll do parameter optimization concurrently. Parameter servers will not wait for trainers, so trainers will also do their work concurrently. But asynchronous SGD will introduce more randomness and noises in the gradient. +Before starting the cluster training, you need to prepare the cluster configuration, PaddlePaddle installation, and other preparations. To understand how to configure the basic environment for distributed training, check the link below: .. toctree:: :maxdepth: 1 preparations_en.md + +Cluster training has a large number of configurable parameters, such as the number of machines used, communication ports, etc. To learn how to configure the distributed training process by setting startup these parameters, check the link below: + +.. toctree:: + :maxdepth: 1 + cmd_argument_en.md + +PaddlePaddle is compatible with a variety of different clusters. Each cluster has its own advantages, To learn how to run PaddlePaddle in different types of them, check the link below: + +.. toctree:: + :maxdepth: 1 + multi_cluster/index_en.rst diff --git a/doc/v2/howto/index_cn.rst b/doc/v2/howto/index_cn.rst index 0c534f107b6e047035c424ed2ea59f3982799b63..b0268907bceb11cd53a4630c3f8b8e0424abe247 100644 --- a/doc/v2/howto/index_cn.rst +++ b/doc/v2/howto/index_cn.rst @@ -1,11 +1,37 @@ 进阶使用 ======== +PaddlePaddle支持用户灵活地设置各种命令行参数，以实现对模型训练或预测流程的控制。使用方式请参考： + .. toctree:: :maxdepth: 1 cmd_parameter/index_cn.rst + +PaddlePaddle支持在fabric集群、MPI集群、kubernetes集群上分布式训练任务，具体环境配置和使用说明请参考： + +.. toctree:: + :maxdepth: 1 + cluster/index_cn.rst + +PaddlePaddle提供了用于预测的C-API，关于C-API的使用，我们提供了如下指南: + +.. toctree:: + :maxdepth: 1 + capi/index_cn.rst + +PaddlePaddle支持多种灵活和高效的循环神经网络，具体配置使用方式请参考： + +.. toctree:: + :maxdepth: 1 + rnn/index_cn.rst + +关于如何使用内置的定时工具、nvprof 或 nvvp 来运行性能分析和调优，请参考： + +.. toctree:: + :maxdepth: 1 + optimization/gpu_profiling_cn.rst diff --git a/doc/v2/howto/rnn/hierarchical_layer_cn.rst b/doc/v2/howto/rnn/hierarchical_layer_cn.rst index e05173c2006ff47ecb6ca5a4fe1502de750acc59..2f8f408b40299890da694862a7b9418cf9ff07f2 100644 --- a/doc/v2/howto/rnn/hierarchical_layer_cn.rst +++ b/doc/v2/howto/rnn/hierarchical_layer_cn.rst @@ -22,7 +22,7 @@ pooling ======== -pooling 的使用示例如下，详细见 :ref:`api_v2.layer_pooling` 配置API。 +pooling 的使用示例如下。 .. code-block:: bash @@ -47,7 +47,7 @@ pooling 的使用示例如下，详细见 :ref:`api_v2.layer_pooling` 配置API last_seq 和 first_seq ===================== -last_seq 的使用示例如下（ :ref:`api_v2.layer_first_seq` 类似），详细见 :ref:`api_v2.layer_last_seq` 配置API。 +last_seq 的使用示例如下（first_seq 类似）。 .. code-block:: bash @@ -68,7 +68,7 @@ last_seq 的使用示例如下（ :ref:`api_v2.layer_first_seq` 类似），详 expand ====== -expand 的使用示例如下，详细见 :ref:`api_v2.layer_expand` 配置API。 +expand 的使用示例如下。 .. code-block:: bash diff --git a/doc/v2/howto/rnn/hrnn_rnn_api_compare_cn.rst b/doc/v2/howto/rnn/hrnn_rnn_api_compare_cn.rst index efdc44455ea4dc81a87b4d4fc8a81e78b15cb06a..b05b66415fbb829f471b1491b9881f65137bfe17 100644 --- a/doc/v2/howto/rnn/hrnn_rnn_api_compare_cn.rst +++ b/doc/v2/howto/rnn/hrnn_rnn_api_compare_cn.rst @@ -4,7 +4,7 @@ 单双层RNN API对比介绍 ##################### -本文以PaddlePaddle的双层RNN单元测试为示例，用多对效果完全相同的、分别使用单双层RNN作为网络配置的模型，来讲解如何使用双层RNN。本文中所有的例子，都只是介绍双层RNN的API接口，并不是使用双层RNN解决实际的问题。如果想要了解双层RNN在具体问题中的使用，请参考\ :ref:`algo_hrnn_demo`\ 。本文中示例所使用的单元测试文件是\ `test_RecurrentGradientMachine.cpp `_\ 。 +本文以PaddlePaddle的双层RNN单元测试为示例，用多对效果完全相同的、分别使用单双层RNN作为网络配置的模型，来讲解如何使用双层RNN。本文中所有的例子，都只是介绍双层RNN的API接口，并不是使用双层RNN解决实际的问题。如果想要了解双层RNN在具体问题中的使用，请参考\ :ref:`algo_hrnn_demo`\ 。本文中示例所使用的单元测试文件是\ `test_RecurrentGradientMachine.cpp `_\ 。 示例1：双层RNN，子序列间无Memory ================================ @@ -166,11 +166,6 @@ 在上面代码中，单层和双层序列的使用和示例2中的示例类似，区别是同时处理了两个输入。而对于双层序列，两个输入的子序列长度也并不相同。但是，我们使用了\ :code:`targetInlink`\ 参数设置了外层\ :code:`recurrent_group`\ 的输出格式。所以外层输出的序列形状，和\ :code:`emb2`\ 的序列形状一致。 -示例4：beam_search的生成 -======================== - -TBD - 词汇表 ====== diff --git a/doc/v2/howto/rnn/index_cn.rst b/doc/v2/howto/rnn/index_cn.rst index bcc8c2f46eb662ec3650e829a77992224dbbb8e7..2032fb9e296ab024c68da1348064580c8c88d5be 100644 --- a/doc/v2/howto/rnn/index_cn.rst +++ b/doc/v2/howto/rnn/index_cn.rst @@ -1,10 +1,34 @@ RNN模型 =========== +循环神经网络（RNN）是对序列数据建模的重要工具。PaddlePaddle提供了灵活的接口以支持复杂循环神经网络的构建。 +这里将分为以下四个部分详细介绍如何使用PaddlePaddle搭建循环神经网络。 + +第一部分由浅入深的展示了使用PaddlePaddle搭建循环神经网络的全貌：首先以简单的循环神经网络（vanilla RNN）为例， +说明如何封装配置循环神经网络组件；然后更进一步的通过序列到序列（sequence to sequence）模型，逐步讲解如何构建完整而复杂的循环神经网络模型。 .. toctree:: :maxdepth: 1 rnn_config_cn.rst + +Recurrent Group是PaddlePaddle中实现复杂循环神经网络的关键，第二部分阐述了PaddlePaddle中Recurrent Group的相关概念和原理， +对Recurrent Group接口进行了详细说明。另外，对双层RNN（对应的输入为双层序列）及Recurrent Group在其中的使用进行了介绍。 + +.. toctree:: + :maxdepth: 1 + recurrent_group_cn.md + +第三部分对双层序列进行了解释说明，列出了PaddlePaddle中支持双层序列作为输入的Layer，并对其使用进行了逐一介绍。 + +.. toctree:: + :maxdepth: 1 + hierarchical_layer_cn.rst + +第四部分以PaddlePaddle的双层RNN单元测试中的网络配置为示例，辅以效果相同的单层RNN网络配置作为对比，讲解了多种情况下双层RNN的使用。 + +.. toctree:: + :maxdepth: 1 + hrnn_rnn_api_compare_cn.rst diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index a7b249d43bf3ad9924749d5e66618750f19d8bf7..d2a4b1335464f553a361728e64ed5ca177ca53da 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -1,27 +1,29 @@ -add_subdirectory(cuda) -add_subdirectory(function) -add_subdirectory(utils) -add_subdirectory(math) -add_subdirectory(gserver) -add_subdirectory(parameter) -add_subdirectory(testing) - -if(MOBILE_INFERENCE) - add_subdirectory(capi) -else() - add_subdirectory(pserver) - add_subdirectory(trainer) - add_subdirectory(scripts) +if(NOT WITH_FLUID) + add_subdirectory(cuda) + add_subdirectory(function) + add_subdirectory(utils) + add_subdirectory(math) + add_subdirectory(gserver) + add_subdirectory(parameter) - if(WITH_C_API) + if(MOBILE_INFERENCE) add_subdirectory(capi) - endif() + else() + add_subdirectory(pserver) + add_subdirectory(trainer) + add_subdirectory(scripts) - if(NOT ANDROID AND NOT IOS) - add_subdirectory(fluid) - endif() + if(WITH_C_API) + add_subdirectory(capi) + endif() - if(WITH_SWIG_PY) - add_subdirectory(api) + if(WITH_SWIG_PY) + add_subdirectory(api) + endif() endif() endif() + +add_subdirectory(testing) +if(NOT MOBILE_INFERENCE AND NOT ANDROID AND NOT IOS) + add_subdirectory(fluid) +endif() diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt index ebb083c5a477d5be91ef14be74dd9de349d07931..e06e9a2b363d1ffc6876b98bcb7304b0a54dbcaa 100644 --- a/paddle/capi/CMakeLists.txt +++ b/paddle/capi/CMakeLists.txt @@ -36,7 +36,7 @@ target_include_directories(paddle_capi PUBLIC ${CMAKE_CURRENT_BINARY_DIR}) add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER} ${CAPI_PRIVATE_HEADER}) -add_dependencies(paddle_capi paddle_proto) +add_dependencies(paddle_capi paddle_proto paddle_gserver) # TODO: paddle_capi_whole will be removed. set(PADDLE_CAPI_LAYERS_LIBS diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 48713f2c2ac62a37b7b7a4602f7f6a325aecb0b8..a4ea74a6d2fbc29dc33a6b57ee453f49ed36c7fa 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -21,7 +21,7 @@ endif() cc_test(eigen_test SRCS eigen_test.cc DEPS tensor) nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place paddle_memory device_context init) -cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto) +cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio) cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor paddle_memory) nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor init) @@ -103,4 +103,5 @@ cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc) cc_test(channel_test SRCS channel_test.cc) cc_test(tuple_test SRCS tuple_test.cc ) cc_test(concurrency_test SRCS concurrency_test.cc DEPS go_op channel_close_op channel_create_op - channel_send_op channel_recv_op sum_op elementwise_add_op executor proto_desc) + channel_send_op channel_recv_op sum_op select_op elementwise_add_op compare_op + conditional_block_op while_op assign_op print_op executor proto_desc) diff --git a/paddle/fluid/framework/block_desc.cc b/paddle/fluid/framework/block_desc.cc index d72b64700f7cf680501fd3e355d20e694f1f097d..3693bc25d81a8309df1a6ddf3d9b08d484596ea9 100644 --- a/paddle/fluid/framework/block_desc.cc +++ b/paddle/fluid/framework/block_desc.cc @@ -135,6 +135,14 @@ OpDesc *BlockDesc::PrependOp() { return ops_.front().get(); } +OpDesc *BlockDesc::InsertOp(size_t index) { + need_update_ = true; + auto it = ops_.begin() + index; + std::unique_ptr new_op(new OpDesc(this)); + it = ops_.insert(it, std::move(new_op)); + return (*it).get(); +} + void BlockDesc::RemoveOp(size_t s, size_t e) { if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) { return; diff --git a/paddle/fluid/framework/block_desc.h b/paddle/fluid/framework/block_desc.h index 3bd90f38907c0a45ae0c9bb00706e5c127f08417..185f018ac1b5863e0ee86fdaa17df1ccbc6e030e 100644 --- a/paddle/fluid/framework/block_desc.h +++ b/paddle/fluid/framework/block_desc.h @@ -87,6 +87,8 @@ class BlockDesc { OpDesc *PrependOp(); + OpDesc *InsertOp(size_t index); + void RemoveOp(size_t s, size_t e); std::vector AllOps() const; diff --git a/paddle/fluid/framework/channel.h b/paddle/fluid/framework/channel.h index 9f8fb12098d622058a86f83c1c42a1feb1cfb2e2..adfaba26ace78f547161ad4029a741f3ca8a6764 100644 --- a/paddle/fluid/framework/channel.h +++ b/paddle/fluid/framework/channel.h @@ -15,23 +15,43 @@ limitations under the License. */ #pragma once #include // for size_t +#include #include #include "paddle/fluid/platform/enforce.h" namespace paddle { namespace framework { +enum class ChannelAction { + SEND = 0, + RECEIVE = 1, + CLOSE = 2, +}; + // Channel is the abstract class of buffered and un-buffered channels. template class Channel { public: + virtual bool CanSend() = 0; + virtual bool CanReceive() = 0; virtual bool Send(T*) = 0; virtual bool Receive(T*) = 0; virtual size_t Cap() = 0; virtual void Lock() = 0; + virtual void Unlock() = 0; + virtual bool IsClosed() = 0; virtual void Close() = 0; virtual ~Channel() {} + + virtual void AddToSendQ(const void* referrer, T* data, + std::shared_ptr cond, + std::function cb) = 0; + virtual void AddToReceiveQ(const void* referrer, T* data, + std::shared_ptr cond, + std::function cb) = 0; + virtual void RemoveFromSendQ(const void* referrer) = 0; + virtual void RemoveFromReceiveQ(const void* referrer) = 0; }; // Forward declaration of channel implementations. @@ -80,6 +100,27 @@ class ChannelHolder { return channel != nullptr ? channel->Receive(data) : false; } + bool IsClosed() { + if (IsInitialized()) { + return holder_->IsClosed(); + } + return false; + } + + bool CanSend() { + if (IsInitialized()) { + return holder_->CanSend(); + } + return false; + } + + bool CanReceive() { + if (IsInitialized()) { + return holder_->CanReceive(); + } + return false; + } + void close() { if (IsInitialized()) holder_->Close(); } @@ -97,6 +138,38 @@ class ChannelHolder { if (IsInitialized()) holder_->Unlock(); } + template + void AddToSendQ(const void* referrer, T* data, + std::shared_ptr cond, + std::function cb) { + if (IsInitialized()) { + Channel* channel = static_cast*>(holder_->Ptr()); + if (channel != nullptr) { + channel->AddToSendQ(referrer, data, cond, cb); + } + } + } + + template + void AddToReceiveQ(const void* referrer, T* data, + std::shared_ptr cond, + std::function cb) { + if (IsInitialized()) { + Channel* channel = static_cast*>(holder_->Ptr()); + if (channel != nullptr) { + channel->AddToReceiveQ(referrer, data, cond, cb); + } + } + } + + void RemoveFromSendQ(const void* referrer) { + if (IsInitialized()) holder_->RemoveFromSendQ(referrer); + } + + void RemoveFromReceiveQ(const void* referrer) { + if (IsInitialized()) holder_->RemoveFromReceiveQ(referrer); + } + inline bool IsInitialized() const { return holder_ != nullptr; } inline const std::type_index Type() { @@ -113,6 +186,11 @@ class ChannelHolder { virtual ~Placeholder() {} virtual const std::type_index Type() const = 0; virtual void* Ptr() const = 0; + virtual bool IsClosed() = 0; + virtual bool CanSend() = 0; + virtual bool CanReceive() = 0; + virtual void RemoveFromSendQ(const void* referrer) = 0; + virtual void RemoveFromReceiveQ(const void* referrer) = 0; virtual void Close() = 0; virtual void Lock() = 0; virtual void Unlock() = 0; @@ -129,6 +207,39 @@ class ChannelHolder { virtual void* Ptr() const { return static_cast(channel_.get()); } + virtual bool IsClosed() { + if (channel_) { + return channel_->IsClosed(); + } + return false; + } + + virtual bool CanSend() { + if (channel_) { + return channel_->CanSend(); + } + return false; + } + + virtual bool CanReceive() { + if (channel_) { + return channel_->CanReceive(); + } + return false; + } + + virtual void RemoveFromSendQ(const void* referrer) { + if (channel_) { + channel_->RemoveFromSendQ(referrer); + } + } + + virtual void RemoveFromReceiveQ(const void* referrer) { + if (channel_) { + channel_->RemoveFromReceiveQ(referrer); + } + } + virtual void Close() { if (channel_) channel_->Close(); } diff --git a/paddle/fluid/framework/channel_impl.h b/paddle/fluid/framework/channel_impl.h index a4561031fd8c49613269e7008ce558f25f9765e4..457abbf373d4549229e8fd8bd6b2087cc6b8f5c8 100644 --- a/paddle/fluid/framework/channel_impl.h +++ b/paddle/fluid/framework/channel_impl.h @@ -29,32 +29,50 @@ class ChannelImpl : public paddle::framework::Channel { friend void paddle::framework::CloseChannel(Channel *); public: + virtual bool CanSend(); + virtual bool CanReceive(); virtual bool Send(T *); virtual bool Receive(T *); virtual size_t Cap() { return cap_; } virtual void Lock(); virtual void Unlock(); + virtual bool IsClosed(); virtual void Close(); - ChannelImpl(size_t); virtual ~ChannelImpl(); + virtual void AddToSendQ(const void *referrer, T *data, + std::shared_ptr cond, + std::function cb); + virtual void AddToReceiveQ(const void *referrer, T *data, + std::shared_ptr cond, + std::function cb); + + virtual void RemoveFromSendQ(const void *referrer); + virtual void RemoveFromReceiveQ(const void *referrer); + private: struct QueueMessage { T *data; - std::condition_variable_any cond; + std::shared_ptr cond; bool chan_closed = false; bool completed = false; + const void *referrer; // TODO(thuan): figure out better way to do this + std::function callback; - QueueMessage(T *item) : data(item) {} + QueueMessage(T *item) + : data(item), cond(std::make_shared()) {} + + QueueMessage(T *item, std::shared_ptr cond) + : data(item), cond(cond) {} void Wait(std::unique_lock &lock) { - cond.wait(lock, [this]() { return completed; }); + cond->wait(lock, [this]() { return completed; }); } void Notify() { completed = true; - cond.notify_all(); + cond->notify_all(); } }; @@ -87,6 +105,18 @@ ChannelImpl::ChannelImpl(size_t capacity) PADDLE_ENFORCE_GE(capacity, 0); } +template +bool ChannelImpl::CanSend() { + std::lock_guard lock{mu_}; + return !closed_ && (!recvq.empty() || buf_.size() < cap_); +} + +template +bool ChannelImpl::CanReceive() { + std::lock_guard lock{mu_}; + return !(closed_ && buf_.empty()) && (!sendq.empty() || buf_.size() > 0); +} + template bool ChannelImpl::Send(T *item) { send_ctr++; @@ -105,7 +135,24 @@ bool ChannelImpl::Send(T *item) { std::shared_ptr m = recvq.front(); recvq.pop_front(); // Do the data transfer - *(m->data) = std::move(*item); + // We will do this data transfer if either of the following + // cases are true + // 1. callback == nullptr // This means it was a regular channel send + // 2. callback returns true + bool do_send = true; + if (m->callback != nullptr) do_send = m->callback(ChannelAction::SEND); + if (do_send) + *(m->data) = std::move(*item); + else + // We cannot do the data transfer because + // this QueueMessage was added by Select + // and some other case was executed. + // So call the Send function again. + // We do not care about notifying other + // because they would have been notified + // by the executed select case. + return send_return(Send(item)); + // Wake up the blocked process and unlock m->Notify(); lock.unlock(); @@ -150,7 +197,25 @@ bool ChannelImpl::Receive(T *item) { std::shared_ptr m = sendq.front(); sendq.pop_front(); // Do the data transfer - *item = std::move(*(m->data)); + // We will do this data transfer if either of the following + // cases are true + // 1. callback == nullptr // This means it was a regular channel send + // 2. callback returns true + bool do_receive = true; + if (m->callback != nullptr) + do_receive = m->callback(ChannelAction::RECEIVE); + if (do_receive) + *item = std::move(*(m->data)); + else + // We cannot do the data transfer because + // this QueueMessage was added by Select + // and some other case was executed. + // So call the Receive function again. + // We do not care about notifying other + // because they would have been notified + // by the executed select case. + return recv_return(Receive(item)); + // Wake up the blocked process and unlock m->Notify(); lock.unlock(); @@ -186,6 +251,12 @@ void ChannelImpl::Unlock() { mu_.unlock(); } +template +bool ChannelImpl::IsClosed() { + std::lock_guard lock{mu_}; + return closed_; +} + template void ChannelImpl::Close() { std::unique_lock lock{mu_}; @@ -203,6 +274,12 @@ void ChannelImpl::Close() { std::shared_ptr m = recvq.front(); recvq.pop_front(); m->chan_closed = true; + + // Execute callback function (if any) + if (m->callback != nullptr) { + m->callback(ChannelAction::CLOSE); + } + m->Notify(); } @@ -211,10 +288,70 @@ void ChannelImpl::Close() { std::shared_ptr m = sendq.front(); sendq.pop_front(); m->chan_closed = true; + + // Execute callback function (if any) + if (m->callback != nullptr) { + m->callback(ChannelAction::CLOSE); + } + m->Notify(); } } +template +void ChannelImpl::AddToSendQ( + const void *referrer, T *data, + std::shared_ptr cond, + std::function cb) { + std::lock_guard lock{mu_}; + auto m = std::make_shared(data, cond); + m->referrer = referrer; + m->callback = cb; + sendq.push_back(m); +} + +template +void ChannelImpl::AddToReceiveQ( + const void *referrer, T *data, + std::shared_ptr cond, + std::function cb) { + std::lock_guard lock{mu_}; + auto m = std::make_shared(data, cond); + m->referrer = referrer; + m->callback = cb; + recvq.push_back(m); +} + +template +void ChannelImpl::RemoveFromSendQ(const void *referrer) { + std::lock_guard lock{mu_}; + + for (auto it = sendq.begin(); it != sendq.end();) { + std::shared_ptr sendMsg = (std::shared_ptr)*it; + + if (sendMsg->referrer == referrer) { + it = sendq.erase(it); + } else { + ++it; + } + } +} + +template +void ChannelImpl::RemoveFromReceiveQ(const void *referrer) { + std::lock_guard lock{mu_}; + + for (auto it = recvq.begin(); it != recvq.end();) { + std::shared_ptr recvMsg = (std::shared_ptr)*it; + + if (recvMsg->referrer == referrer) { + it = recvq.erase(it); + } else { + ++it; + } + } +} + template ChannelImpl::~ChannelImpl() { Close(); diff --git a/paddle/fluid/framework/concurrency_test.cc b/paddle/fluid/framework/concurrency_test.cc index 5770b0a5a18659e615e80a7c48113d8b543b69ec..25152054eb8452a9667bd65b4441665476c1d46d 100644 --- a/paddle/fluid/framework/concurrency_test.cc +++ b/paddle/fluid/framework/concurrency_test.cc @@ -19,7 +19,6 @@ limitations under the License. */ #include "paddle/fluid/framework/channel.h" #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/program_desc.h" USE_NO_KERNEL_OP(go); USE_NO_KERNEL_OP(channel_close); @@ -27,6 +26,12 @@ USE_NO_KERNEL_OP(channel_create); USE_NO_KERNEL_OP(channel_recv); USE_NO_KERNEL_OP(channel_send); USE_NO_KERNEL_OP(elementwise_add); +USE_NO_KERNEL_OP(select); +USE_NO_KERNEL_OP(conditional_block); +USE_NO_KERNEL_OP(equal); +USE_NO_KERNEL_OP(assign); +USE_NO_KERNEL_OP(while); +USE_NO_KERNEL_OP(print); namespace f = paddle::framework; namespace p = paddle::platform; @@ -35,27 +40,15 @@ namespace paddle { namespace framework { template -void CreateIntVariable(Scope &scope, p::CPUPlace &place, std::string name, - T value) { - // Create LoDTensor of dim [1,1] +LoDTensor *CreateVariable(Scope &scope, p::CPUPlace &place, std::string name, + T value) { + // Create LoDTensor of dim [1] auto var = scope.Var(name); auto tensor = var->GetMutable(); - tensor->Resize({1, 1}); + tensor->Resize({1}); T *expect = tensor->mutable_data(place); expect[0] = value; -} - -void InitTensorsInScope(Scope &scope, p::CPUPlace &place) { - p::CPUDeviceContext ctx(place); - - // Create channel variable - scope.Var("Channel"); - - // Create Variables, x0 will be put into channel, - // result will be pulled from channel - CreateIntVariable(scope, place, "Status", false); - CreateIntVariable(scope, place, "x0", 99); - CreateIntVariable(scope, place, "result", 0); + return tensor; } void AddOp(const std::string &type, const VariableNameMap &inputs, @@ -73,12 +66,116 @@ void AddOp(const std::string &type, const VariableNameMap &inputs, op->SetAttrMap(attrs); } +void AddCase(ProgramDesc *program, Scope *scope, p::CPUPlace *place, + BlockDesc *casesBlock, int caseId, int caseType, + std::string caseChannel, std::string caseVarName, + std::function func) { + std::string caseCondName = std::string("caseCond") + std::to_string(caseId); + std::string caseCondXVarName = + std::string("caseCondX") + std::to_string(caseId); + + BlockDesc *caseBlock = program->AppendBlock(*casesBlock); + func(caseBlock, scope); + + CreateVariable(*scope, *place, caseCondName, false); + CreateVariable(*scope, *place, caseCondXVarName, caseId); + CreateVariable(*scope, *place, caseVarName, caseId); + + scope->Var("step_scope"); + + AddOp("equal", {{"X", {caseCondXVarName}}, {"Y", {"caseToExecute"}}}, + {{"Out", {caseCondName}}}, {}, casesBlock); + + AddOp("conditional_block", {{"X", {caseCondName}}, {"Params", {}}}, + {{"Out", {}}, {"Scope", {"step_scope"}}}, + {{"sub_block", caseBlock}, {"is_scalar_condition", true}}, casesBlock); +} + +void AddFibonacciSelect(Scope *scope, p::CPUPlace *place, ProgramDesc *program, + BlockDesc *parentBlock, std::string dataChanName, + std::string quitChanName) { + BlockDesc *whileBlock = program->AppendBlock(*parentBlock); + + CreateVariable(*scope, *place, "whileExitCond", true); + CreateVariable(*scope, *place, "caseToExecute", -1); + CreateVariable(*scope, *place, "case1var", 0); + + CreateVariable(*scope, *place, "xtemp", 0); + + // TODO(thuan): Need to create fibXToSend, since channel send moves the actual + // data, + // which causes the data to be no longer accessible to do the fib calculation + // TODO(abhinav): Change channel send to do a copy instead of a move! + CreateVariable(*scope, *place, "fibXToSend", 0); + + CreateVariable(*scope, *place, "fibX", 0); + CreateVariable(*scope, *place, "fibY", 1); + CreateVariable(*scope, *place, "quitVar", 0); + + BlockDesc *casesBlock = program->AppendBlock(*whileBlock); + std::function f = [](BlockDesc *caseBlock) {}; + + // TODO(thuan): Remove this once we change channel send to do a copy instead + // of move + AddOp("assign", {{"X", {"fibX"}}}, {{"Out", {"fibXToSend"}}}, {}, whileBlock); + + // Case 0: Send to dataChanName + std::function case0Func = [&]( + BlockDesc *caseBlock, Scope *scope) { + AddOp("assign", {{"X", {"fibX"}}}, {{"Out", {"xtemp"}}}, {}, caseBlock); + AddOp("assign", {{"X", {"fibY"}}}, {{"Out", {"fibX"}}}, {}, caseBlock); + AddOp("elementwise_add", {{"X", {"xtemp"}}, {"Y", {"fibY"}}}, + {{"Out", {"fibY"}}}, {}, caseBlock); + }; + AddCase(program, scope, place, casesBlock, 0, 1, dataChanName, "fibXToSend", + case0Func); + std::string case0Config = + std::string("0,1,") + dataChanName + std::string(",fibXToSend"); + + // Case 1: Receive from quitChanName + std::function case2Func = [&]( + BlockDesc *caseBlock, Scope *scope) { + // Exit the while loop after we receive from quit channel. + // We assign a false to "whileExitCond" variable, which will + // break out of while_op loop + CreateVariable(*scope, *place, "whileFalse", false); + AddOp("assign", {{"X", {"whileFalse"}}}, {{"Out", {"whileExitCond"}}}, {}, + caseBlock); + }; + AddCase(program, scope, place, casesBlock, 1, 2, quitChanName, "quitVar", + case2Func); + std::string case1Config = + std::string("1,2,") + quitChanName + std::string(",quitVar"); + + // Select block + AddOp("select", {{"X", {dataChanName, quitChanName}}, + {"case_to_execute", {"caseToExecute"}}}, + {}, {{"sub_block", casesBlock}, + {"cases", std::vector{case0Config, case1Config}}}, + whileBlock); + + scope->Var("stepScopes"); + AddOp("while", + {{"X", {dataChanName, quitChanName}}, {"Condition", {"whileExitCond"}}}, + {{"Out", {}}, {"StepScopes", {"stepScopes"}}}, + {{"sub_block", whileBlock}}, parentBlock); +} + TEST(Concurrency, Go_Op) { Scope scope; p::CPUPlace place; // Initialize scope variables - InitTensorsInScope(scope, place); + p::CPUDeviceContext ctx(place); + + // Create channel variable + scope.Var("Channel"); + + // Create Variables, x0 will be put into channel, + // result will be pulled from channel + CreateVariable(scope, place, "Status", false); + CreateVariable(scope, place, "x0", 99); + CreateVariable(scope, place, "result", 0); framework::Executor executor(place); ProgramDesc program; @@ -118,5 +215,78 @@ TEST(Concurrency, Go_Op) { auto *finalData = tensor.data(); EXPECT_EQ(finalData[0], 99); } + +/** + * This test implements the fibonacci function using go_op and select_op + */ +TEST(Concurrency, Select) { + Scope scope; + p::CPUPlace place; + + // Initialize scope variables + p::CPUDeviceContext ctx(place); + + CreateVariable(scope, place, "Status", false); + CreateVariable(scope, place, "result", 0); + CreateVariable(scope, place, "currentXFib", 0); + + framework::Executor executor(place); + ProgramDesc program; + BlockDesc *block = program.MutableBlock(0); + + // Create channel OP + std::string dataChanName = "Channel"; + scope.Var(dataChanName); + AddOp("channel_create", {}, {{"Out", {dataChanName}}}, + {{"capacity", 0}, {"data_type", f::proto::VarType::LOD_TENSOR}}, block); + + std::string quitChanName = "Quit"; + scope.Var(quitChanName); + AddOp("channel_create", {}, {{"Out", {quitChanName}}}, + {{"capacity", 0}, {"data_type", f::proto::VarType::LOD_TENSOR}}, block); + + // Create Go Op routine, which loops 10 times over fibonacci sequence + CreateVariable(scope, place, "xReceiveVar", 0); + + BlockDesc *goOpBlock = program.AppendBlock(program.Block(0)); + for (int i = 0; i < 10; ++i) { + AddOp("channel_recv", {{"Channel", {dataChanName}}}, + {{"Status", {"Status"}}, {"Out", {"currentXFib"}}}, {}, goOpBlock); + AddOp("print", {{"In", {"currentXFib"}}}, {{"Out", {"currentXFib"}}}, + {{"first_n", 100}, + {"summarize", -1}, + {"print_tensor_name", false}, + {"print_tensor_type", true}, + {"print_tensor_shape", false}, + {"print_tensor_lod", false}, + {"print_phase", std::string("FORWARD")}, + {"message", std::string("X: ")}}, + goOpBlock); + } + + CreateVariable(scope, place, "quitSignal", 0); + AddOp("channel_send", {{"Channel", {quitChanName}}, {"X", {"quitSignal"}}}, + {{"Status", {"Status"}}}, {}, goOpBlock); + + // Create Go Op + AddOp("go", {{"X", {dataChanName, quitChanName}}}, {}, + {{"sub_block", goOpBlock}}, block); + + AddFibonacciSelect(&scope, &place, &program, block, dataChanName, + quitChanName); + + // Create Channel Close Op + AddOp("channel_close", {{"Channel", {dataChanName}}}, {}, {}, block); + AddOp("channel_close", {{"Channel", {quitChanName}}}, {}, {}, block); + + executor.Run(program, &scope, 0, true, true); + + // After we call executor.run, "result" variable should be equal to 34 + // (which is 10 loops through fibonacci sequence) + const LoDTensor &tensor = (scope.FindVar("currentXFib"))->Get(); + auto *finalData = tensor.data(); + EXPECT_EQ(finalData[0], 34); +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/data_type_transform.cc b/paddle/fluid/framework/data_type_transform.cc index 554cd58916c5a1ba09a411b4dc0b3a834ccc486a..c0523f3c795b103c0c27081ec5dc717f6a0f11e0 100644 --- a/paddle/fluid/framework/data_type_transform.cc +++ b/paddle/fluid/framework/data_type_transform.cc @@ -53,6 +53,7 @@ struct CastDataType { auto* context = static_cast(ctx_); trans(*context, in_begin, in_end, out_begin, CastDataTypeFunctor()); + context->Wait(); #endif } else { PADDLE_THROW("Unsupported place!"); diff --git a/paddle/fluid/framework/data_type_transform_test.cc b/paddle/fluid/framework/data_type_transform_test.cc index c992cba9a3611d50839a8ec056ee6ab954cd88b6..6b9a8f5e28b372c45abfaa2c20575a55d9a9dd03 100644 --- a/paddle/fluid/framework/data_type_transform_test.cc +++ b/paddle/fluid/framework/data_type_transform_test.cc @@ -50,13 +50,13 @@ TEST(DataTypeTransform, CPUTransform) { TransDataType(kernel_fp32, kernel_fp64, in, &out); double* out_data_double = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_double[i], static_cast(i / 3)); + EXPECT_EQ(out_data_double[i], static_cast(i / 3)); } TransDataType(kernel_fp32, kernel_int32, in, &out); int* out_data_int = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_int[i], static_cast(i / 3)); + EXPECT_EQ(out_data_int[i], static_cast(i / 3)); } } @@ -76,31 +76,31 @@ TEST(DataTypeTransform, CPUTransform) { TransDataType(kernel_fp16, kernel_fp32, in, &out); float* out_data_float = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_float[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_float[i], static_cast(ptr[i])); } TransDataType(kernel_fp16, kernel_fp64, in, &out); double* out_data_double = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_double[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_double[i], static_cast(ptr[i])); } TransDataType(kernel_fp16, kernel_int32, in, &out); int* out_data_int = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_int[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_int[i], static_cast(ptr[i])); } TransDataType(kernel_fp16, kernel_int64, in, &out); int64_t* out_data_int64 = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_int64[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_int64[i], static_cast(ptr[i])); } TransDataType(kernel_fp16, kernel_bool, in, &out); bool* out_data_bool = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_bool[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_bool[i], static_cast(ptr[i])); } // transform float to float16 @@ -112,7 +112,7 @@ TEST(DataTypeTransform, CPUTransform) { TransDataType(kernel_fp32, kernel_fp16, in, &out); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_float[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_float[i]).x); } // transform double to float16 @@ -124,7 +124,7 @@ TEST(DataTypeTransform, CPUTransform) { TransDataType(kernel_fp64, kernel_fp16, in, &out); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_double[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_double[i]).x); } // transform int to float16 @@ -136,7 +136,7 @@ TEST(DataTypeTransform, CPUTransform) { TransDataType(kernel_int32, kernel_fp16, in, &out); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_int[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_int[i]).x); } // transform int64 to float16 @@ -148,7 +148,7 @@ TEST(DataTypeTransform, CPUTransform) { TransDataType(kernel_int64, kernel_fp16, in, &out); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_int64[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_int64[i]).x); } // transform bool to float16 @@ -160,7 +160,7 @@ TEST(DataTypeTransform, CPUTransform) { TransDataType(kernel_bool, kernel_fp16, in, &out); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_bool[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_bool[i]).x); } } } diff --git a/paddle/fluid/framework/data_type_transform_test.cu b/paddle/fluid/framework/data_type_transform_test.cu index 3939bc5e754cd3c2829cfcb3353f83969af055a9..de389ddabcb86de0155757406a406e44086c5474 100644 --- a/paddle/fluid/framework/data_type_transform_test.cu +++ b/paddle/fluid/framework/data_type_transform_test.cu @@ -49,15 +49,16 @@ TEST(DataTypeTransform, GPUTransform) { float arr[6] = {0, 1, 2, 3, 4, 5}; int data_number = sizeof(arr) / sizeof(arr[0]); memcpy(in_ptr, arr, sizeof(arr)); - TensorCopy(in, gpu_place, context, &in_gpu); + TensorCopy(in, gpu_place, context, &in_gpu); + context.Wait(); TransDataType(kernel_fp32, kernel_fp64, in_gpu, &out_gpu); TensorCopy(out_gpu, cpu_place, context, &out); context.Wait(); double* out_data_double = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_double[i], static_cast(arr[i])); + EXPECT_EQ(out_data_double[i], static_cast(arr[i])); } TransDataType(kernel_fp32, kernel_int32, in_gpu, &out_gpu); @@ -66,7 +67,7 @@ TEST(DataTypeTransform, GPUTransform) { int* out_data_int = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_int[i], static_cast(arr[i])); + EXPECT_EQ(out_data_int[i], static_cast(arr[i])); } } @@ -83,6 +84,7 @@ TEST(DataTypeTransform, GPUTransform) { int data_number = sizeof(arr) / sizeof(arr[0]); memcpy(ptr, arr, sizeof(arr)); TensorCopy(in, gpu_place, context, &in_gpu); + context.Wait(); // transform from float16 to other data types TransDataType(kernel_fp16, kernel_fp32, in_gpu, &out_gpu); @@ -91,7 +93,7 @@ TEST(DataTypeTransform, GPUTransform) { float* out_data_float = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_float[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_float[i], static_cast(ptr[i])); } TransDataType(kernel_fp16, kernel_fp64, in_gpu, &out_gpu); @@ -100,7 +102,7 @@ TEST(DataTypeTransform, GPUTransform) { double* out_data_double = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_double[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_double[i], static_cast(ptr[i])); } TransDataType(kernel_fp16, kernel_int32, in_gpu, &out_gpu); @@ -109,7 +111,7 @@ TEST(DataTypeTransform, GPUTransform) { int* out_data_int = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_int[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_int[i], static_cast(ptr[i])); } TransDataType(kernel_fp16, kernel_int64, in_gpu, &out_gpu); @@ -118,7 +120,7 @@ TEST(DataTypeTransform, GPUTransform) { int64_t* out_data_int64 = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_int64[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_int64[i], static_cast(ptr[i])); } TransDataType(kernel_fp16, kernel_bool, in_gpu, &out_gpu); @@ -127,7 +129,7 @@ TEST(DataTypeTransform, GPUTransform) { bool* out_data_bool = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(out_data_bool[i], static_cast(ptr[i])); + EXPECT_EQ(out_data_bool[i], static_cast(ptr[i])); } // transform float to float16 @@ -137,13 +139,14 @@ TEST(DataTypeTransform, GPUTransform) { } TensorCopy(in, gpu_place, context, &in_gpu); + context.Wait(); TransDataType(kernel_fp32, kernel_fp16, in_gpu, &out_gpu); TensorCopy(out_gpu, cpu_place, context, &out); context.Wait(); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_float[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_float[i]).x); } // transform double to float16 @@ -154,13 +157,14 @@ TEST(DataTypeTransform, GPUTransform) { } TensorCopy(in, gpu_place, context, &in_gpu); + context.Wait(); TransDataType(kernel_fp64, kernel_fp16, in_gpu, &out_gpu); TensorCopy(out_gpu, cpu_place, context, &out); context.Wait(); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_double[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_double[i]).x); } // transform int to float16 @@ -170,13 +174,14 @@ TEST(DataTypeTransform, GPUTransform) { } TensorCopy(in, gpu_place, context, &in_gpu); + context.Wait(); TransDataType(kernel_int32, kernel_fp16, in_gpu, &out_gpu); TensorCopy(out_gpu, cpu_place, context, &out); context.Wait(); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_int[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_int[i]).x); } // transform int64 to float16 @@ -187,13 +192,14 @@ TEST(DataTypeTransform, GPUTransform) { } TensorCopy(in, gpu_place, context, &in_gpu); + context.Wait(); TransDataType(kernel_int64, kernel_fp16, in_gpu, &out_gpu); TensorCopy(out_gpu, cpu_place, context, &out); context.Wait(); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_int64[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_int64[i]).x); } // transform bool to float16 @@ -203,13 +209,14 @@ TEST(DataTypeTransform, GPUTransform) { } TensorCopy(in, gpu_place, context, &in_gpu); + context.Wait(); TransDataType(kernel_bool, kernel_fp16, in_gpu, &out_gpu); TensorCopy(out_gpu, cpu_place, context, &out); context.Wait(); ptr = out.data(); for (int i = 0; i < data_number; ++i) { - ASSERT_EQ(ptr[i].x, static_cast(in_data_bool[i]).x); + EXPECT_EQ(ptr[i].x, static_cast(in_data_bool[i]).x); } } } diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 961e3e22f278d6e0346defd90190c53fd31ede08..7155d5ef2febc20aaa684c04a7a59f781857c9e5 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -25,6 +25,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" #include "paddle/fluid/platform/place.h" +#include "paddle/fluid/platform/profiler.h" DECLARE_bool(benchmark); DEFINE_bool(check_nan_inf, false, @@ -33,6 +34,20 @@ DEFINE_bool(check_nan_inf, false, namespace paddle { namespace framework { +namespace { +// block id starts from 0. This id is used to represent the codeblock +// wrapping the first block 0. +int kProgramId = -1; +} // namespace + +struct ExecutorPrepareContext { + ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id) + : prog_(prog), block_id_(block_id) {} + + const framework::ProgramDesc& prog_; + size_t block_id_; + std::vector> ops_; +}; Executor::Executor(const platform::Place& place) : place_(place) {} @@ -85,72 +100,10 @@ static void CheckTensorNANOrInf(const std::string& name, void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id, bool create_local_scope, bool create_vars) { - // TODO(tonyyang-svail): - // - only runs on the first device (i.e. no interdevice communication) - // - will change to use multiple blocks for RNN op and Cond Op - PADDLE_ENFORCE_LT(static_cast(block_id), pdesc.Size()); - auto& block = pdesc.Block(block_id); - - Scope* local_scope = scope; - if (create_vars) { - if (create_local_scope) { - local_scope = &scope->NewScope(); - for (auto& var : block.AllVars()) { - if (var->Name() == framework::kEmptyVarName) { - continue; - } - - if (var->Persistable()) { - auto* ptr = scope->Var(var->Name()); - CreateTensor(ptr, var->GetType()); - VLOG(3) << "Create Variable " << var->Name() - << " global, which pointer is " << ptr; - } else { - auto* ptr = local_scope->Var(var->Name()); - CreateTensor(ptr, var->GetType()); - VLOG(3) << "Create Variable " << var->Name() - << " locally, which pointer is " << ptr; - } - } - } else { - for (auto& var : block.AllVars()) { - auto* ptr = local_scope->Var(var->Name()); - CreateTensor(ptr, var->GetType()); - VLOG(3) << "Create variable " << var->Name() << ", which pointer is " - << ptr; - } - } // if (create_local_scope) - } // if (create_vars) - - for (auto& op_desc : block.AllOps()) { - auto op = paddle::framework::OpRegistry::CreateOp(*op_desc); - - VLOG(3) << place_ << " " << op->DebugStringEx(local_scope); - op->Run(*local_scope, place_); - - if (FLAGS_benchmark) { - VLOG(2) << "Memory used after operator " + op->Type() + " running: " - << memory::memory_usage(place_); - } - if (FLAGS_check_nan_inf) { - for (auto& vname : op->OutputVars(true)) { - auto* var = local_scope->FindVar(vname); - if (var == nullptr) continue; - if (var->IsType()) { - CheckTensorNANOrInf(vname, var->Get()); - } - } - } - } - if (create_vars && create_local_scope) { - scope->DeleteScope(local_scope); - } - if (FLAGS_benchmark) { - VLOG(2) << "-------------------------------------------------------"; - VLOG(2) << "Memory used after deleting local scope: " - << memory::memory_usage(place_); - VLOG(2) << "-------------------------------------------------------"; - } + platform::RecordBlock b(block_id); + auto* ctx = Prepare(pdesc, block_id); + RunPreparedContext(ctx, scope, create_local_scope, create_vars); + delete ctx; } // Check whether the block already has feed operators and feed_holder. @@ -160,10 +113,11 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id, // and feed_holder_name. Raise exception when any mismatch is found. // Return true if the block has feed operators and holder of matching info. static bool has_feed_operators( - BlockDesc* block, std::map& feed_targets, + const BlockDesc& block, + std::map& feed_targets, const std::string& feed_holder_name) { size_t feed_count = 0; - for (auto* op : block->AllOps()) { + for (auto* op : block.AllOps()) { if (op->Type() == kFeedOpType) { feed_count++; PADDLE_ENFORCE_EQ(op->Input("X")[0], feed_holder_name, @@ -182,7 +136,7 @@ static bool has_feed_operators( "The number of feed operators should match 'feed_targets'"); // When feed operator are present, so should be feed_holder - auto var = block->FindVar(feed_holder_name); + auto var = block.FindVar(feed_holder_name); PADDLE_ENFORCE_NOT_NULL(var, "Block should already have a '%s' variable", feed_holder_name); PADDLE_ENFORCE_EQ(var->GetType(), proto::VarType::FEED_MINIBATCH, @@ -200,10 +154,10 @@ static bool has_feed_operators( // and fetch_holder_name. Raise exception when any mismatch is found. // Return true if the block has fetch operators and holder of matching info. static bool has_fetch_operators( - BlockDesc* block, std::map& fetch_targets, + const BlockDesc& block, std::map& fetch_targets, const std::string& fetch_holder_name) { size_t fetch_count = 0; - for (auto* op : block->AllOps()) { + for (auto* op : block.AllOps()) { if (op->Type() == kFetchOpType) { fetch_count++; PADDLE_ENFORCE_EQ(op->Output("Out")[0], fetch_holder_name, @@ -222,7 +176,7 @@ static bool has_fetch_operators( "The number of fetch operators should match 'fetch_targets'"); // When fetch operator are present, so should be fetch_holder - auto var = block->FindVar(fetch_holder_name); + auto var = block.FindVar(fetch_holder_name); PADDLE_ENFORCE_NOT_NULL(var, "Block should already have a '%s' variable", fetch_holder_name); PADDLE_ENFORCE_EQ(var->GetType(), proto::VarType::FETCH_LIST, @@ -238,10 +192,20 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, std::map& fetch_targets, const std::string& feed_holder_name, const std::string& fetch_holder_name) { - auto* copy_program = new ProgramDesc(program); + platform::RecordBlock b(kProgramId); + bool has_feed_ops = + has_feed_operators(program.Block(0), feed_targets, feed_holder_name); + bool has_fetch_ops = + has_fetch_operators(program.Block(0), fetch_targets, fetch_holder_name); + + ProgramDesc* copy_program = const_cast(&program); + if (!has_feed_ops || !has_fetch_ops) { + copy_program = std::unique_ptr(new ProgramDesc(program)).get(); + } + auto* global_block = copy_program->MutableBlock(0); - if (!has_feed_operators(global_block, feed_targets, feed_holder_name)) { + if (!has_feed_ops) { // create feed_holder variable auto* feed_holder = global_block->Var(feed_holder_name); feed_holder->SetType(proto::VarType::FEED_MINIBATCH); @@ -274,7 +238,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, } } - if (!has_fetch_operators(global_block, fetch_targets, fetch_holder_name)) { + if (!has_fetch_ops) { // create fetch_holder variable auto* fetch_holder = global_block->Var(fetch_holder_name); fetch_holder->SetType(proto::VarType::FETCH_LIST); @@ -308,8 +272,81 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, GetFetchVariable(*scope, fetch_holder_name, idx); } } +} + +ExecutorPrepareContext* Executor::Prepare(const ProgramDesc& program, + int block_id) { + auto* ctx = new ExecutorPrepareContext(program, block_id); + PADDLE_ENFORCE_LT(static_cast(block_id), program.Size()); + auto& block = program.Block(block_id); + for (auto& op_desc : block.AllOps()) { + ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc)); + } + return ctx; +} + +void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, + bool create_local_scope, bool create_vars) { + auto& block = ctx->prog_.Block(ctx->block_id_); + + Scope* local_scope = scope; + if (create_vars) { + if (create_local_scope) { + local_scope = &scope->NewScope(); + for (auto& var : block.AllVars()) { + if (var->Name() == framework::kEmptyVarName) { + continue; + } - delete copy_program; + if (var->Persistable()) { + auto* ptr = scope->Var(var->Name()); + CreateTensor(ptr, var->GetType()); + VLOG(3) << "Create Variable " << var->Name() + << " global, which pointer is " << ptr; + } else { + auto* ptr = local_scope->Var(var->Name()); + CreateTensor(ptr, var->GetType()); + VLOG(3) << "Create Variable " << var->Name() + << " locally, which pointer is " << ptr; + } + } + } else { + for (auto& var : block.AllVars()) { + auto* ptr = local_scope->Var(var->Name()); + CreateTensor(ptr, var->GetType()); + VLOG(3) << "Create variable " << var->Name() << ", which pointer is " + << ptr; + } + } // if (create_local_scope) + } // if (create_vars) + + for (auto& op : ctx->ops_) { + VLOG(3) << place_ << " " << op->DebugStringEx(local_scope); + op->Run(*local_scope, place_); + + if (FLAGS_benchmark) { + VLOG(2) << "Memory used after operator " + op->Type() + " running: " + << memory::memory_usage(place_); + } + if (FLAGS_check_nan_inf) { + for (auto& vname : op->OutputVars(true)) { + auto* var = local_scope->FindVar(vname); + if (var == nullptr) continue; + if (var->IsType()) { + CheckTensorNANOrInf(vname, var->Get()); + } + } + } + } + if (create_vars && create_local_scope) { + scope->DeleteScope(local_scope); + } + if (FLAGS_benchmark) { + VLOG(2) << "-------------------------------------------------------"; + VLOG(2) << "Memory used after deleting local scope: " + << memory::memory_usage(place_); + VLOG(2) << "-------------------------------------------------------"; + } } } // namespace framework diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index c1f4d4e02a951e8b127b66cae125309e4798cc76..28ce3315154cea45412984df4daf7385ce2cf572 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -22,7 +22,7 @@ limitations under the License. */ namespace paddle { namespace framework { - +struct ExecutorPrepareContext; class Executor { public: // TODO(dzhwinter) : Do not rely on this function, it will be removed @@ -38,8 +38,8 @@ class Executor { * ProgramDesc * Scope */ - void Run(const ProgramDesc&, Scope*, int, bool create_local_scope = true, - bool create_vars = true); + void Run(const ProgramDesc& prog, Scope* scope, int block_id, + bool create_local_scope = true, bool create_vars = true); void Run(const ProgramDesc& program, Scope* scope, std::map& feed_targets, @@ -47,6 +47,13 @@ class Executor { const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); + static ExecutorPrepareContext* Prepare(const ProgramDesc& program, + int block_id); + + void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, + bool create_local_scope = true, + bool create_vars = true); + private: const platform::Place place_; }; diff --git a/paddle/fluid/framework/init.cc b/paddle/fluid/framework/init.cc index 2e0a224ff5df749fd8c809dc88a85a1643542abf..3c0d93642ac41e8d90f9a248e81cea7a4fe12293 100644 --- a/paddle/fluid/framework/init.cc +++ b/paddle/fluid/framework/init.cc @@ -26,6 +26,7 @@ namespace paddle { namespace framework { std::once_flag gflags_init_flag; +std::once_flag p2p_init_flag; void InitGflags(std::vector &argv) { std::call_once(gflags_init_flag, [&]() { @@ -42,6 +43,27 @@ void InitGflags(std::vector &argv) { }); } +void InitP2P(int count) { +#ifdef PADDLE_WITH_CUDA + std::call_once(p2p_init_flag, [&]() { + for (int i = 0; i < count; ++i) { + for (int j = 0; j < count; ++j) { + if (i == j) continue; + int can_acess = -1; + PADDLE_ENFORCE(cudaDeviceCanAccessPeer(&can_acess, i, j), + "Failed to test P2P access."); + if (can_acess != 1) { + LOG(WARNING) << "Cannot enable P2P access from " << i << " to " << j; + } else { + cudaSetDevice(i); + cudaDeviceEnablePeerAccess(j, 0); + } + } + } + }); +#endif +} + void InitDevices() { /*Init all avaiable devices by default */ @@ -63,7 +85,7 @@ void InitDevices() { for (int i = 0; i < count; ++i) { places.emplace_back(platform::CUDAPlace(i)); } - + InitP2P(count); platform::DeviceContextPool::Init(places); } diff --git a/paddle/fluid/framework/lod_tensor.cc b/paddle/fluid/framework/lod_tensor.cc index e2f4e9cad1996578b7c51257785e1273d126f80f..8155cb55a468a09320b1196b49fc3e34cea261b1 100644 --- a/paddle/fluid/framework/lod_tensor.cc +++ b/paddle/fluid/framework/lod_tensor.cc @@ -19,6 +19,9 @@ limitations under the License. */ #include "paddle/fluid/memory/memcpy.h" #include "paddle/fluid/memory/memory.h" +#include "paddle/fluid/recordio/scanner.h" +#include "paddle/fluid/recordio/writer.h" + #include #include #include @@ -291,6 +294,31 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor, TensorFromStream(is, static_cast(tensor), dev_ctx); } +void WriteToRecordIO(recordio::Writer &writer, + const std::vector &tensor, + const platform::DeviceContext &dev_ctx) { + std::stringstream buffer; + size_t sz = tensor.size(); + buffer.write(reinterpret_cast(&sz), sizeof(uint32_t)); + for (auto &each : tensor) { + SerializeToStream(buffer, each, dev_ctx); + } + writer.Write(buffer.str()); +} + +std::vector ReadFromRecordIO( + recordio::Scanner &scanner, const platform::DeviceContext &dev_ctx) { + std::istringstream sin(scanner.Next()); + uint32_t sz; + sin.read(reinterpret_cast(&sz), sizeof(uint32_t)); + std::vector result; + result.resize(sz); + for (uint32_t i = 0; i < sz; ++i) { + DeserializeFromStream(sin, &result[i], dev_ctx); + } + return result; +} + std::vector LoDTensor::SplitLoDTensor( const std::vector places) const { check_memory_size(); diff --git a/paddle/fluid/framework/lod_tensor.h b/paddle/fluid/framework/lod_tensor.h index 94d5a6e9fd9b68d3d8230a8c258316efadda5a05..dee505fee0dccd8d60bb290a8bec4df243e504a2 100644 --- a/paddle/fluid/framework/lod_tensor.h +++ b/paddle/fluid/framework/lod_tensor.h @@ -29,6 +29,12 @@ limitations under the License. */ #include "paddle/fluid/platform/place.h" namespace paddle { + +namespace recordio { +class Writer; +class Scanner; +} + namespace framework { /* @@ -209,5 +215,12 @@ void SerializeToStream(std::ostream& os, const LoDTensor& tensor, void DeserializeFromStream(std::istream& is, LoDTensor* tensor, const platform::DeviceContext& dev_ctx); +extern void WriteToRecordIO(recordio::Writer& writer, + const std::vector& tensor, + const platform::DeviceContext& dev_ctx); + +extern std::vector ReadFromRecordIO( + recordio::Scanner& scanner, const platform::DeviceContext& dev_ctx); + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/lod_tensor_test.cc b/paddle/fluid/framework/lod_tensor_test.cc index 5e135192ce774ab5c351b89164be9d7600ae3640..e691e29383d4842b80769021e0e494967d38e9bb 100644 --- a/paddle/fluid/framework/lod_tensor_test.cc +++ b/paddle/fluid/framework/lod_tensor_test.cc @@ -14,6 +14,9 @@ #include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/recordio/scanner.h" +#include "paddle/fluid/recordio/writer.h" + #include #include #include @@ -224,5 +227,43 @@ TEST(LoD, CheckAbsLoD) { abs_lod0.push_back(std::vector({0})); ASSERT_FALSE(CheckAbsLoD(abs_lod0)); } + +TEST(LoDTensor, RecordIO) { + LoDTensor tensor; + int* tmp = tensor.mutable_data(make_ddim({4, 5}), platform::CPUPlace()); + for (int i = 0; i < 20; ++i) { + tmp[i] = i; + } + + std::stringstream* stream = new std::stringstream(); + auto& ctx = + *platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); + { + recordio::Writer writer(stream, recordio::Compressor::kSnappy); + WriteToRecordIO(writer, {tensor, tensor}, ctx); + WriteToRecordIO(writer, {tensor, tensor}, ctx); + writer.Flush(); + } + + auto assert_tensor_ok = [](const LoDTensor& tensor) { + for (int i = 0; i < 20; ++i) { + ASSERT_EQ(tensor.data()[i], i); + } + }; + + { + std::unique_ptr stream_ptr(stream); + recordio::Scanner scanner(std::move(stream_ptr)); + auto tensors = ReadFromRecordIO(scanner, ctx); + ASSERT_EQ(tensors.size(), 2); + assert_tensor_ok(tensors[0]); + assert_tensor_ok(tensors[1]); + tensors = ReadFromRecordIO(scanner, ctx); + ASSERT_EQ(tensors.size(), 2); + assert_tensor_ok(tensors[0]); + assert_tensor_ok(tensors[1]); + } +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index ac6289c5abe8f40ae9ee32aa3d58cdef3ff0e836..b39a1164dbd9877d9f45cc6415d74f930921a42f 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -442,15 +442,7 @@ class RuntimeInferShapeContext : public InferShapeContext { } std::vector GetRepeatedDims(const std::string& name) const override { - Variable* var = scope_.FindVar(name); - if (var->IsType()) { - return var->Get().shapes(); - } else { - PADDLE_THROW( - "Only ReaderHolder support 'GetRepeatedDims', but Variable %s's " - "type_id is %s.", - name, var->Type().name()); - } + PADDLE_THROW("Only compile time support this method"); } void SetDim(const std::string& name, const DDim& dim) override { @@ -467,15 +459,7 @@ class RuntimeInferShapeContext : public InferShapeContext { void SetRepeatedDims(const std::string& name, const std::vector& dims) override { - Variable* var = scope_.FindVar(name); - if (var->IsType()) { - var->GetMutable()->set_shapes(dims); - } else { - PADDLE_THROW( - "Only ReaderHolder support 'SetRepeatedDims', but Variable %s's " - "type_id is %s.", - name, var->Type().name()); - } + PADDLE_THROW("Only compile time support this method"); } proto::VarType::Type GetVarType(const std::string& name) const override { @@ -497,8 +481,10 @@ void OperatorWithKernel::RunImpl(const Scope& scope, RuntimeInferShapeContext infer_shape_ctx(*this, scope); this->InferShape(&infer_shape_ctx); platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto dev_ctx = pool.Get(place); - // profile + auto* dev_ctx = pool.Get(place); + + // For profiling, don't move out of this function because that will result + // in the failure of multi-GPU profiling. platform::RecordEvent record_event(Type(), dev_ctx); // check if op[type] has kernel registered. auto& all_op_kernels = AllOpKernels(); diff --git a/paddle/fluid/framework/reader.cc b/paddle/fluid/framework/reader.cc index 91879d6d45868bb37ca44baafb8b0e8677cd6d1a..fa00c08e0d5791ee1187aed38b4d140564b7c97d 100644 --- a/paddle/fluid/framework/reader.cc +++ b/paddle/fluid/framework/reader.cc @@ -16,14 +16,22 @@ namespace paddle { namespace framework { +ReaderBase::~ReaderBase() {} -DDim ReaderBase::shape(size_t idx) const { - PADDLE_ENFORCE_LT( - idx, shapes_.size(), - "Cannot get the %d'th shape, 'shapes_' only has %d elements.", idx, - shapes_.size()); - return shapes_[idx]; -} +FileReader::FileReader(const std::vector &dims) : dims_(dims) {} + +void FileReader::ReadNext(std::vector *out) { + ReadNextImpl(out); + PADDLE_ENFORCE_EQ(out->size(), dims_.size()); + for (size_t i = 0; i < dims_.size(); ++i) { + auto &actual = out->at(i).dims(); + auto &expect = dims_[i]; + PADDLE_ENFORCE_EQ(actual.size(), expect.size()); + for (int j = 0; j < actual.size(); ++j) { + PADDLE_ENFORCE(actual[i] == expect[i] || expect[i] == -1); + } + } +} } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/reader.h b/paddle/fluid/framework/reader.h index 27ab6e750c2e665fa5055a3ecfb2f315cb4000c0..3573b99becf6d657c680c5fec0bda4bdde5dd7a2 100644 --- a/paddle/fluid/framework/reader.h +++ b/paddle/fluid/framework/reader.h @@ -16,50 +16,53 @@ #include "paddle/fluid/framework/ddim.h" #include "paddle/fluid/framework/lod_tensor_array.h" +#include "paddle/fluid/platform/place.h" + +#include +#include +#include namespace paddle { namespace framework { class ReaderBase { public: - explicit ReaderBase(const std::vector& shapes) : shapes_(shapes) { - PADDLE_ENFORCE(!shapes_.empty()); - } virtual void ReadNext(std::vector* out) = 0; - virtual bool HasNext() const = 0; virtual void ReInit() = 0; - DDim shape(size_t idx) const; - std::vector shapes() const { return shapes_; } - void set_shapes(const std::vector& shapes) { shapes_ = shapes; } - - virtual ~ReaderBase() {} + virtual bool HasNext() const = 0; - protected: - std::vector shapes_; -}; - -class FileReader : public ReaderBase { - public: - explicit FileReader(const std::vector& shapes) : ReaderBase(shapes) {} + virtual ~ReaderBase(); }; class DecoratedReader : public ReaderBase { public: - explicit DecoratedReader(ReaderBase* reader) - : ReaderBase(reader->shapes()), reader_(reader) { + explicit DecoratedReader(ReaderBase* reader) : ReaderBase(), reader_(reader) { PADDLE_ENFORCE_NOT_NULL(reader_); } - bool HasNext() const override { return reader_->HasNext(); } - void ReInit() override { reader_->ReInit(); } + bool HasNext() const override { return reader_->HasNext(); } + protected: ReaderBase* reader_; }; +class FileReader : public ReaderBase { + public: + explicit FileReader(const std::vector& dims); + + void ReadNext(std::vector* out) override; + + protected: + virtual void ReadNextImpl(std::vector* out) = 0; + + private: + std::vector dims_; +}; + // The ReaderHolder is used as reader' unified wrapper, // making it easier to access different type reader in Variables. class ReaderHolder { @@ -68,15 +71,16 @@ class ReaderHolder { ReaderBase* Get() const { return reader_.get(); } - void ReadNext(std::vector* out) { reader_->ReadNext(out); } - bool HasNext() const { return reader_->HasNext(); } - void ReInit() { reader_->ReInit(); } - - DDim shape(size_t idx) const { return reader_->shape(idx); } - std::vector shapes() const { return reader_->shapes(); } - void set_shapes(const std::vector& shapes) { - reader_->set_shapes(shapes); + void ReadNext(std::vector* out) { + PADDLE_ENFORCE_NOT_NULL(reader_); + reader_->ReadNext(out); } + void ReInit() { + PADDLE_ENFORCE_NOT_NULL(reader_); + reader_->ReInit(); + } + + bool HasNext() const { return reader_->HasNext(); } private: std::unique_ptr reader_; diff --git a/paddle/fluid/framework/scope.cc b/paddle/fluid/framework/scope.cc index ea6c8cebd3ff16451c974bf3a0ded9d822a9caf8..17e38b1cf042657834b4d0d1c12cbbb92f19fa45 100644 --- a/paddle/fluid/framework/scope.cc +++ b/paddle/fluid/framework/scope.cc @@ -16,6 +16,7 @@ limitations under the License. */ #include // for unique_ptr #include // for call_once +#include #include "glog/logging.h" #include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/string/printf.h" @@ -102,6 +103,18 @@ void Scope::DeleteScope(Scope* scope) { } } +void Scope::EraseVars(std::vector& var_names) { + std::set var_set(var_names.begin(), var_names.end()); + for (auto it = vars_.begin(); it != vars_.end();) { + if (var_set.find(it->first) != var_set.end()) { + delete it->second; + it = vars_.erase(it); + } else { + ++it; + } + } +} + void Scope::Rename(const std::string& origin_name, const std::string& new_name) const { auto origin_it = vars_.find(origin_name); diff --git a/paddle/fluid/framework/scope.h b/paddle/fluid/framework/scope.h index d8fad162e59c8ea6465b2dc78c90709a758ace24..c1e1f49caaa5a60df0e97289aada465b45213971 100644 --- a/paddle/fluid/framework/scope.h +++ b/paddle/fluid/framework/scope.h @@ -51,6 +51,8 @@ class Scope { /// Create a variable with a scope-unique name. Variable* Var(std::string* name = nullptr); + void EraseVars(std::vector& var_names); + /// Find a variable in the scope or any of its ancestors. Returns /// nullptr if cannot find. Variable* FindVar(const std::string& name) const; diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc index 9b465b85b0a02ffe990ab669a22f78e923e24f99..8b7533ce712b0a01060842b6f71449ed6bd23e2c 100644 --- a/paddle/fluid/framework/tensor_util.cc +++ b/paddle/fluid/framework/tensor_util.cc @@ -187,7 +187,6 @@ bool TensorContainsInf(const framework::Tensor& tensor) { void TensorToStream(std::ostream& os, const Tensor& tensor, const platform::DeviceContext& dev_ctx) { - // TODO(typhoonzero): serialize to ostream { // the 1st field, uint32_t version constexpr uint32_t version = 0; os.write(reinterpret_cast(&version), sizeof(version)); diff --git a/paddle/fluid/framework/threadpool.h b/paddle/fluid/framework/threadpool.h index 3adc260caf5eaaa1408a26ed91457666e6c4adce..df51fb24a588c84788d7d0b671f932ff4c40f9c2 100644 --- a/paddle/fluid/framework/threadpool.h +++ b/paddle/fluid/framework/threadpool.h @@ -67,10 +67,10 @@ class ThreadPool { } catch (platform::EnforceNotMet ex) { return std::unique_ptr( new platform::EnforceNotMet(ex)); - } catch (...) { - LOG(FATAL) - << "Unexpected exception is catched in thread pool. All " - "throwable exception in Fluid should be an EnforceNotMet."; + } catch (const std::exception& e) { + LOG(FATAL) << "Unexpected exception is catched in thread pool. All " + "throwable exception in Fluid should be an EnforceNotMet." + << e.what(); } return nullptr; }); diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index bdb147955ca0700dc0854b54c38d961caf8845f3..aff427310f15be72f5c8d0fa1537ffa6bbe2881d 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -5,13 +5,19 @@ cc_library(paddle_fluid_api DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) # Create static library -cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) +get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES) +cc_library(paddle_fluid DEPS ${fluid_modules}) # Create shared library cc_library(paddle_fluid_shared SHARED SRCS io.cc DEPS ARCHIVE_START ${GLOB_OP_LIB} ${FLUID_CORE_MODULES} ARCHIVE_END) set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid) +if(NOT APPLE) + # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac. + set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map") + set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}") +endif() if(WITH_TESTING) add_subdirectory(tests/book) diff --git a/paddle/fluid/inference/io.cc b/paddle/fluid/inference/io.cc index 80eb9889670744ae527ea29609b33631a021bfa8..52e9c0baa64508f82d0a86a88c8c5f8d23f9f7f2 100644 --- a/paddle/fluid/inference/io.cc +++ b/paddle/fluid/inference/io.cc @@ -22,14 +22,14 @@ namespace paddle { namespace inference { void ReadBinaryFile(const std::string& filename, std::string& contents) { - VLOG(3) << "loading model from " << filename; - std::ifstream inputfs(filename, std::ios::in | std::ios::binary); - inputfs.seekg(0, std::ios::end); + std::ifstream fin(filename, std::ios::in | std::ios::binary); + PADDLE_ENFORCE(static_cast(fin), "Cannot open file %s", filename); + fin.seekg(0, std::ios::end); contents.clear(); - contents.resize(inputfs.tellg()); - inputfs.seekg(0, std::ios::beg); - inputfs.read(&contents[0], contents.size()); - inputfs.close(); + contents.resize(fin.tellg()); + fin.seekg(0, std::ios::beg); + fin.read(&contents[0], contents.size()); + fin.close(); } bool IsPersistable(const framework::VarDesc* var) { @@ -97,6 +97,7 @@ std::unique_ptr Load(framework::Executor& executor, const std::string& dirname) { std::string model_filename = dirname + "/__model__"; std::string program_desc_str; + VLOG(3) << "loading model from " << model_filename; ReadBinaryFile(model_filename, program_desc_str); std::unique_ptr main_program( diff --git a/paddle/fluid/inference/paddle_fluid.map b/paddle/fluid/inference/paddle_fluid.map new file mode 100644 index 0000000000000000000000000000000000000000..5203784dc1fcb672eb6a26d9dfd3ffbe02e08038 --- /dev/null +++ b/paddle/fluid/inference/paddle_fluid.map @@ -0,0 +1,6 @@ +{ + global: + *paddle*; + local: + *; +}; diff --git a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc index d6fc51301bc45d3ce01142c0f015e4bb69e519bf..e9a27171f1cd68e7b10c860fb4a1417b930ed565 100644 --- a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc +++ b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc @@ -17,10 +17,13 @@ limitations under the License. */ #include "paddle/fluid/inference/tests/test_helper.h" DEFINE_string(dirname, "", "Directory of the inference model."); +DEFINE_int32(batch_size, 1, "Batch size of input data"); +DEFINE_int32(repeat, 1, "Running the inference program repeat times"); TEST(inference, image_classification) { - if (FLAGS_dirname.empty()) { - LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model"; + if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) { + LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model " + "--batch_size=1 --repeat=1"; } LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl; @@ -29,13 +32,11 @@ TEST(inference, image_classification) { // 0. Call `paddle::framework::InitDevices()` initialize all the devices // In unittests, this is done in paddle/testing/paddle_gtest_main.cc - int64_t batch_size = 1; - paddle::framework::LoDTensor input; // Use normilized image pixels as input data, // which should be in the range [0.0, 1.0]. SetupTensor(input, - {batch_size, 3, 32, 32}, + {FLAGS_batch_size, 3, 32, 32}, static_cast(0), static_cast(1)); std::vector cpu_feeds; @@ -46,7 +47,9 @@ TEST(inference, image_classification) { cpu_fetchs1.push_back(&output1); // Run inference on CPU - TestInference(dirname, cpu_feeds, cpu_fetchs1); + LOG(INFO) << "--- CPU Runs: ---"; + TestInference( + dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat); LOG(INFO) << output1.dims(); #ifdef PADDLE_WITH_CUDA @@ -55,7 +58,9 @@ TEST(inference, image_classification) { cpu_fetchs2.push_back(&output2); // Run inference on CUDA GPU - TestInference(dirname, cpu_feeds, cpu_fetchs2); + LOG(INFO) << "--- GPU Runs: ---"; + TestInference( + dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat); LOG(INFO) << output2.dims(); CheckError(output1, output2); diff --git a/paddle/fluid/inference/tests/book/test_inference_recognize_digits.cc b/paddle/fluid/inference/tests/book/test_inference_recognize_digits.cc index 99bee94cb82633219df7a4a5c5bada15d2d3ce64..1fb0f9e77797cf6e61e918700763ee33a495cb96 100644 --- a/paddle/fluid/inference/tests/book/test_inference_recognize_digits.cc +++ b/paddle/fluid/inference/tests/book/test_inference_recognize_digits.cc @@ -17,10 +17,13 @@ limitations under the License. */ #include "paddle/fluid/inference/tests/test_helper.h" DEFINE_string(dirname, "", "Directory of the inference model."); +DEFINE_int32(batch_size, 1, "Batch size of input data"); +DEFINE_int32(repeat, 1, "Running the inference program repeat times"); TEST(inference, recognize_digits) { - if (FLAGS_dirname.empty()) { - LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model"; + if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) { + LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model " + "--batch_size=1 --repeat=1"; } LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl; @@ -29,77 +32,39 @@ TEST(inference, recognize_digits) { // 0. Call `paddle::framework::InitDevices()` initialize all the devices // In unittests, this is done in paddle/testing/paddle_gtest_main.cc - int64_t batch_size = 1; - paddle::framework::LoDTensor input; // Use normilized image pixels as input data, // which should be in the range [-1.0, 1.0]. SetupTensor(input, - {batch_size, 1, 28, 28}, + {FLAGS_batch_size, 1, 28, 28}, static_cast(-1), static_cast(1)); std::vector cpu_feeds; cpu_feeds.push_back(&input); - paddle::framework::LoDTensor output1; - std::vector cpu_fetchs1; - cpu_fetchs1.push_back(&output1); + for (auto is_combined : {false, true}) { + paddle::framework::LoDTensor output1; + std::vector cpu_fetchs1; + cpu_fetchs1.push_back(&output1); - // Run inference on CPU - TestInference(dirname, cpu_feeds, cpu_fetchs1); - LOG(INFO) << output1.dims(); + // Run inference on CPU + LOG(INFO) << "--- CPU Runs: is_combined=" << is_combined << " ---"; + TestInference( + dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined); + LOG(INFO) << output1.dims(); #ifdef PADDLE_WITH_CUDA - paddle::framework::LoDTensor output2; - std::vector cpu_fetchs2; - cpu_fetchs2.push_back(&output2); + paddle::framework::LoDTensor output2; + std::vector cpu_fetchs2; + cpu_fetchs2.push_back(&output2); - // Run inference on CUDA GPU - TestInference(dirname, cpu_feeds, cpu_fetchs2); - LOG(INFO) << output2.dims(); + // Run inference on CUDA GPU + LOG(INFO) << "--- GPU Runs: is_combined=" << is_combined << " ---"; + TestInference( + dirname, cpu_feeds, cpu_fetchs2, FLAGS_repeat, is_combined); + LOG(INFO) << output2.dims(); - CheckError(output1, output2); + CheckError(output1, output2); #endif -} - -TEST(inference, recognize_digits_combine) { - if (FLAGS_dirname.empty()) { - LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model"; } - - LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl; - std::string dirname = FLAGS_dirname; - - // 0. Call `paddle::framework::InitDevices()` initialize all the devices - // In unittests, this is done in paddle/testing/paddle_gtest_main.cc - - paddle::framework::LoDTensor input; - // Use normilized image pixels as input data, - // which should be in the range [-1.0, 1.0]. - SetupTensor( - input, {1, 1, 28, 28}, static_cast(-1), static_cast(1)); - std::vector cpu_feeds; - cpu_feeds.push_back(&input); - - paddle::framework::LoDTensor output1; - std::vector cpu_fetchs1; - cpu_fetchs1.push_back(&output1); - - // Run inference on CPU - TestInference( - dirname, cpu_feeds, cpu_fetchs1); - LOG(INFO) << output1.dims(); - -#ifdef PADDLE_WITH_CUDA - paddle::framework::LoDTensor output2; - std::vector cpu_fetchs2; - cpu_fetchs2.push_back(&output2); - - // Run inference on CUDA GPU - TestInference( - dirname, cpu_feeds, cpu_fetchs2); - LOG(INFO) << output2.dims(); - - CheckError(output1, output2); -#endif } diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 49518e50d8541477234f17ac5b8709aeb57662ff..dce541c0971a6ff9a3728e915fe8c3d009c23550 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -15,6 +15,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/inference/io.h" +#include "paddle/fluid/platform/profiler.h" template void SetupTensor(paddle::framework::LoDTensor& input, @@ -87,32 +88,65 @@ void CheckError(paddle::framework::LoDTensor& output1, EXPECT_EQ(count, 0U) << "There are " << count << " different elements."; } -template +template void TestInference(const std::string& dirname, const std::vector& cpu_feeds, - std::vector& cpu_fetchs) { + std::vector& cpu_fetchs, + const int repeat = 1, + const bool is_combined = false) { // 1. Define place, executor, scope auto place = Place(); auto executor = paddle::framework::Executor(place); auto* scope = new paddle::framework::Scope(); + // Profile the performance + paddle::platform::ProfilerState state; + if (paddle::platform::is_cpu_place(place)) { + state = paddle::platform::ProfilerState::kCPU; + } else { +#ifdef PADDLE_WITH_CUDA + state = paddle::platform::ProfilerState::kCUDA; + // The default device_id of paddle::platform::CUDAPlace is 0. + // Users can get the device_id using: + // int device_id = place.GetDeviceId(); + paddle::platform::SetDeviceId(0); +#else + PADDLE_THROW("'CUDAPlace' is not supported in CPU only device."); +#endif + } + // 2. Initialize the inference_program and load parameters std::unique_ptr inference_program; - if (IsCombined) { - // All parameters are saved in a single file. - // Hard-coding the file names of program and parameters in unittest. - // The file names should be consistent with that used in Python API - // `fluid.io.save_inference_model`. - std::string prog_filename = "__model_combined__"; - std::string param_filename = "__params_combined__"; - inference_program = paddle::inference::Load(executor, - *scope, - dirname + "/" + prog_filename, - dirname + "/" + param_filename); - } else { - // Parameters are saved in separate files sited in the specified `dirname`. - inference_program = paddle::inference::Load(executor, *scope, dirname); + + // Enable the profiler + paddle::platform::EnableProfiler(state); + { + paddle::platform::RecordEvent record_event( + "init_program", + paddle::platform::DeviceContextPool::Instance().Get(place)); + + if (is_combined) { + // All parameters are saved in a single file. + // Hard-coding the file names of program and parameters in unittest. + // The file names should be consistent with that used in Python API + // `fluid.io.save_inference_model`. + std::string prog_filename = "__model_combined__"; + std::string param_filename = "__params_combined__"; + inference_program = + paddle::inference::Load(executor, + *scope, + dirname + "/" + prog_filename, + dirname + "/" + param_filename); + } else { + // Parameters are saved in separate files sited in the specified + // `dirname`. + inference_program = paddle::inference::Load(executor, *scope, dirname); + } } + // Disable the profiler and print the timing information + paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault, + "load_program_profiler.txt"); + paddle::platform::ResetProfiler(); // 3. Get the feed_target_names and fetch_target_names const std::vector& feed_target_names = @@ -134,7 +168,28 @@ void TestInference(const std::string& dirname, } // 6. Run the inference program - executor.Run(*inference_program, scope, feed_targets, fetch_targets); + { + // Ignore the profiling results of the first run + executor.Run(*inference_program, scope, feed_targets, fetch_targets); + + // Enable the profiler + paddle::platform::EnableProfiler(state); + + // Run repeat times to profile the performance + for (int i = 0; i < repeat; ++i) { + paddle::platform::RecordEvent record_event( + "run_inference", + paddle::platform::DeviceContextPool::Instance().Get(place)); + + executor.Run(*inference_program, scope, feed_targets, fetch_targets); + } + + // Disable the profiler and print the timing information + paddle::platform::DisableProfiler( + paddle::platform::EventSortingKey::kDefault, + "run_inference_profiler.txt"); + paddle::platform::ResetProfiler(); + } delete scope; } diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index 5d436a7e0c3752c889c19820507589f34d3bee94..d30124d4a3b89b802a4abaae07a33b76526f163d 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -165,7 +165,6 @@ op_library(cond_op DEPS framework_proto tensor net_op) op_library(cross_entropy_op DEPS cross_entropy) op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax) op_library(softmax_op DEPS softmax) -op_library(detection_output_op DEPS softmax) op_library(sequence_softmax_op DEPS softmax) op_library(sum_op DEPS selected_rows_functor) op_library(sgd_op DEPS selected_rows_functor) @@ -203,6 +202,11 @@ op_library(save_combine_op DEPS lod_tensor) op_library(load_combine_op DEPS lod_tensor) op_library(concat_op DEPS concat) +# FIXME(thuan): Move CSP operators to paddle/fluid/framework/operators/concurrency +add_subdirectory(concurrency) +op_library(channel_send_op DEPS concurrency) +op_library(channel_recv_op DEPS concurrency) + list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS}) foreach(src ${GENERAL_OPS}) op_library(${src}) @@ -222,8 +226,6 @@ cc_test(scatter_test SRCS scatter_test.cc DEPS tensor) cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor) cc_test(beam_search_op_test SRCS beam_search_op_test.cc DEPS lod_tensor beam_search_op) cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory) -if(WITH_GPU) - cc_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) -endif() cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op) cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op) +nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) diff --git a/paddle/fluid/operators/assign_op.cc b/paddle/fluid/operators/assign_op.cc index 39ae3c0040d04a6d901f1d6c992d547a6778c28e..d372213e1b6008b0c4227103dd40730f86a84301 100644 --- a/paddle/fluid/operators/assign_op.cc +++ b/paddle/fluid/operators/assign_op.cc @@ -56,6 +56,7 @@ class AssignFunctor { private: void copy_tensor(const framework::LoDTensor &lod_tensor, framework::LoDTensor *out) const { + if (lod_tensor.numel() == 0) return; auto &out_tensor = *out; TensorCopy(lod_tensor, lod_tensor.place(), dev_ctx_, &out_tensor); out_tensor.set_lod(lod_tensor.lod()); diff --git a/paddle/fluid/operators/cast_op.cc b/paddle/fluid/operators/cast_op.cc index a5ec47d84fe423f77de494e342931a6d278049d2..dd0068d571f72c9c22334e523cd091fe4c8da5a6 100644 --- a/paddle/fluid/operators/cast_op.cc +++ b/paddle/fluid/operators/cast_op.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/cast_op.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { @@ -63,15 +64,30 @@ class CastOpGradMaker : public framework::SingleGradOpDescMaker { } }; +class CastOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + framework::OpKernelType kt = OperatorWithKernel::GetExpectedKernelType(ctx); + // CastOp kernel's device type is decided by input tensor place + kt.place_ = ctx.Input("X")->place(); + return kt; + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; using CPU = paddle::platform::CPUDeviceContext; -REGISTER_OP_WITH_KERNEL(cast, ops::CastOpGradMaker, ops::CastOpInferShape, - ops::CastOpProtoMaker); +REGISTER_OPERATOR(cast, ops::CastOp, ops::CastOpGradMaker, + ops::CastOpInferShape, ops::CastOpProtoMaker); REGISTER_OP_CPU_KERNEL(cast, ops::CastOpKernel, ops::CastOpKernel, ops::CastOpKernel, ops::CastOpKernel, - ops::CastOpKernel); + ops::CastOpKernel, + ops::CastOpKernel); diff --git a/paddle/fluid/operators/cast_op.cu b/paddle/fluid/operators/cast_op.cu index 507e9a531aae70e60bc6748bfab800310d6e0c21..c486c5850e25fcf4370f02cb145c244743a4cc4b 100644 --- a/paddle/fluid/operators/cast_op.cu +++ b/paddle/fluid/operators/cast_op.cu @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/cast_op.h" +#include "paddle/fluid/platform/float16.h" template using CastOpKernel = @@ -20,4 +21,5 @@ using CastOpKernel = REGISTER_OP_CUDA_KERNEL(cast, CastOpKernel, CastOpKernel, CastOpKernel, CastOpKernel, - CastOpKernel); + CastOpKernel, + CastOpKernel); diff --git a/paddle/fluid/operators/channel_recv_op.cc b/paddle/fluid/operators/channel_recv_op.cc index c12b88e7a91c4ea7044223464a2f902db494d1a8..844b3ae3b7bf87c9b253128165b3c938801d5d60 100644 --- a/paddle/fluid/operators/channel_recv_op.cc +++ b/paddle/fluid/operators/channel_recv_op.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/concurrency/channel_util.h" #include "paddle/fluid/operators/math/math_function.h" static constexpr char Channel[] = "Channel"; @@ -36,25 +37,6 @@ void SetReceiveStatus(const platform::Place &dev_place, status_tensor[0] = status; } -bool ChannelReceive(framework::ChannelHolder *ch, framework::Variable *var) { - // Get type of channel and use that to call mutable data for Variable - auto type = framework::ToVarType(ch->Type()); - if (type == framework::proto::VarType_Type_LOD_TENSOR) - return ch->Receive(var->GetMutable()); - else if (type == framework::proto::VarType_Type_LOD_RANK_TABLE) - return ch->Receive(var->GetMutable()); - else if (type == framework::proto::VarType_Type_LOD_TENSOR_ARRAY) - return ch->Receive(var->GetMutable()); - else if (type == framework::proto::VarType_Type_SELECTED_ROWS) - return ch->Receive(var->GetMutable()); - else if (type == framework::proto::VarType_Type_READER) - return ch->Receive(var->GetMutable()); - else if (type == framework::proto::VarType_Type_CHANNEL) - return ch->Receive(var->GetMutable()); - else - PADDLE_THROW("ChannelReceive:Unsupported type"); -} - class ChannelRecvOp : public framework::OperatorBase { public: ChannelRecvOp(const std::string &type, @@ -81,7 +63,7 @@ class ChannelRecvOp : public framework::OperatorBase { scope.FindVar(Input(Channel))->GetMutable(); auto output_var = scope.FindVar(Output(Out)); // Receive the data from the channel. - bool ok = ChannelReceive(ch, output_var); + bool ok = concurrency::ChannelReceive(ch, output_var); // Set the status output of the `ChannelReceive` call. SetReceiveStatus(dev_place, *scope.FindVar(Output(Status)), ok); diff --git a/paddle/fluid/operators/channel_send_op.cc b/paddle/fluid/operators/channel_send_op.cc index 6d7715ad229e821f02437246e3326063cb1ee757..47cf7d7efc9996e8a8db11b79c0310f77c2435a4 100644 --- a/paddle/fluid/operators/channel_send_op.cc +++ b/paddle/fluid/operators/channel_send_op.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/concurrency/channel_util.h" #include "paddle/fluid/operators/math/math_function.h" static constexpr char Channel[] = "Channel"; @@ -37,24 +38,6 @@ void SetSendStatus(const platform::Place &dev_place, status_tensor[0] = status; } -bool ChannelSend(framework::ChannelHolder *ch, framework::Variable *var) { - auto type = framework::ToVarType(var->Type()); - if (type == framework::proto::VarType_Type_LOD_TENSOR) - return ch->Send(var->GetMutable()); - else if (type == framework::proto::VarType_Type_LOD_RANK_TABLE) - return ch->Send(var->GetMutable()); - else if (type == framework::proto::VarType_Type_LOD_TENSOR_ARRAY) - return ch->Send(var->GetMutable()); - else if (type == framework::proto::VarType_Type_SELECTED_ROWS) - return ch->Send(var->GetMutable()); - else if (type == framework::proto::VarType_Type_READER) - return ch->Send(var->GetMutable()); - else if (type == framework::proto::VarType_Type_CHANNEL) - return ch->Send(var->GetMutable()); - else - PADDLE_THROW("ChannelSend:Unsupported type"); -} - class ChannelSendOp : public framework::OperatorBase { public: ChannelSendOp(const std::string &type, @@ -82,7 +65,7 @@ class ChannelSendOp : public framework::OperatorBase { auto input_var = scope.FindVar(Input(X)); // Send the input data through the channel. - bool ok = ChannelSend(ch, input_var); + bool ok = concurrency::ChannelSend(ch, input_var); // Set the status output of the `ChannelSend` call. SetSendStatus(dev_place, *scope.FindVar(Output(Status)), ok); diff --git a/paddle/fluid/operators/concurrency/CMakeLists.txt b/paddle/fluid/operators/concurrency/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..e4617440d152b4c15d09e81cd19c76739b95b979 --- /dev/null +++ b/paddle/fluid/operators/concurrency/CMakeLists.txt @@ -0,0 +1 @@ +cc_library(concurrency SRCS channel_util.cc DEPS device_context framework_proto boost eigen3) diff --git a/paddle/fluid/operators/concurrency/channel_util.cc b/paddle/fluid/operators/concurrency/channel_util.cc new file mode 100644 index 0000000000000000000000000000000000000000..a483af7affd824da7d18676d934dc959167ef71f --- /dev/null +++ b/paddle/fluid/operators/concurrency/channel_util.cc @@ -0,0 +1,111 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +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 "channel_util.h" +#include "paddle/fluid/framework/var_type.h" + +namespace poc = paddle::operators::concurrency; + +bool poc::ChannelSend(framework::ChannelHolder *ch, framework::Variable *var) { + auto type = framework::ToVarType(var->Type()); + if (type == framework::proto::VarType_Type_LOD_TENSOR) + return ch->Send(var->GetMutable()); + else if (type == framework::proto::VarType_Type_LOD_RANK_TABLE) + return ch->Send(var->GetMutable()); + else if (type == framework::proto::VarType_Type_LOD_TENSOR_ARRAY) + return ch->Send(var->GetMutable()); + else if (type == framework::proto::VarType_Type_SELECTED_ROWS) + return ch->Send(var->GetMutable()); + else if (type == framework::proto::VarType_Type_READER) + return ch->Send(var->GetMutable()); + else if (type == framework::proto::VarType_Type_CHANNEL) + return ch->Send(var->GetMutable()); + else + PADDLE_THROW("ChannelSend:Unsupported type"); +} + +bool poc::ChannelReceive(framework::ChannelHolder *ch, + framework::Variable *var) { + // Get type of channel and use that to call mutable data for Variable + auto type = framework::ToVarType(ch->Type()); + if (type == framework::proto::VarType_Type_LOD_TENSOR) + return ch->Receive(var->GetMutable()); + else if (type == framework::proto::VarType_Type_LOD_RANK_TABLE) + return ch->Receive(var->GetMutable()); + else if (type == framework::proto::VarType_Type_LOD_TENSOR_ARRAY) + return ch->Receive(var->GetMutable()); + else if (type == framework::proto::VarType_Type_SELECTED_ROWS) + return ch->Receive(var->GetMutable()); + else if (type == framework::proto::VarType_Type_READER) + return ch->Receive(var->GetMutable()); + else if (type == framework::proto::VarType_Type_CHANNEL) + return ch->Receive(var->GetMutable()); + else + PADDLE_THROW("ChannelReceive:Unsupported type"); +} + +void poc::ChannelAddToSendQ(framework::ChannelHolder *ch, const void *referrer, + framework::Variable *var, + std::shared_ptr cond, + std::function cb) { + auto type = framework::ToVarType(var->Type()); + if (type == framework::proto::VarType_Type_LOD_TENSOR) { + ch->AddToSendQ(referrer, var->GetMutable(), cond, cb); + } else if (type == framework::proto::VarType_Type_LOD_RANK_TABLE) { + ch->AddToSendQ(referrer, var->GetMutable(), cond, + cb); + } else if (type == framework::proto::VarType_Type_LOD_TENSOR_ARRAY) { + ch->AddToSendQ(referrer, var->GetMutable(), cond, + cb); + } else if (type == framework::proto::VarType_Type_SELECTED_ROWS) { + ch->AddToSendQ(referrer, var->GetMutable(), cond, + cb); + } else if (type == framework::proto::VarType_Type_READER) { + ch->AddToSendQ(referrer, var->GetMutable(), cond, + cb); + } else if (type == framework::proto::VarType_Type_CHANNEL) { + ch->AddToSendQ(referrer, var->GetMutable(), cond, + cb); + } else { + PADDLE_THROW("ChannelAddToSendQ:Unsupported type"); + } +} + +void poc::ChannelAddToReceiveQ( + framework::ChannelHolder *ch, const void *referrer, + framework::Variable *var, std::shared_ptr cond, + std::function cb) { + auto type = framework::ToVarType(var->Type()); + if (type == framework::proto::VarType_Type_LOD_TENSOR) { + ch->AddToReceiveQ(referrer, var->GetMutable(), cond, + cb); + } else if (type == framework::proto::VarType_Type_LOD_RANK_TABLE) { + ch->AddToReceiveQ(referrer, var->GetMutable(), + cond, cb); + } else if (type == framework::proto::VarType_Type_LOD_TENSOR_ARRAY) { + ch->AddToReceiveQ(referrer, var->GetMutable(), + cond, cb); + } else if (type == framework::proto::VarType_Type_SELECTED_ROWS) { + ch->AddToReceiveQ(referrer, var->GetMutable(), + cond, cb); + } else if (type == framework::proto::VarType_Type_READER) { + ch->AddToReceiveQ(referrer, var->GetMutable(), + cond, cb); + } else if (type == framework::proto::VarType_Type_CHANNEL) { + ch->AddToReceiveQ(referrer, var->GetMutable(), + cond, cb); + } else { + PADDLE_THROW("ChannelAddToReceiveQ:Unsupported type"); + } +} diff --git a/paddle/fluid/operators/concurrency/channel_util.h b/paddle/fluid/operators/concurrency/channel_util.h new file mode 100644 index 0000000000000000000000000000000000000000..c3674bd9815df451751707bfa84d18dbb5fa0f6b --- /dev/null +++ b/paddle/fluid/operators/concurrency/channel_util.h @@ -0,0 +1,38 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +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/fluid/framework/channel.h" +#include "paddle/fluid/framework/variable.h" + +namespace paddle { +namespace operators { +namespace concurrency { + +bool ChannelSend(framework::ChannelHolder *ch, framework::Variable *var); +bool ChannelReceive(framework::ChannelHolder *ch, framework::Variable *var); + +void ChannelAddToSendQ(framework::ChannelHolder *ch, const void *referrer, + framework::Variable *var, + std::shared_ptr cond, + std::function cb); +void ChannelAddToReceiveQ(framework::ChannelHolder *ch, const void *referrer, + framework::Variable *var, + std::shared_ptr cond, + std::function cb); + +} // namespace concurrency +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/conv_cudnn_op.cu.cc b/paddle/fluid/operators/conv_cudnn_op.cu.cc index ff0fbf21f86269885df5491afab7443df813f13f..a32aba4c1ff2f5e775aeb41f25b02322dbc6a64a 100644 --- a/paddle/fluid/operators/conv_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_cudnn_op.cu.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include "paddle/fluid/operators/conv_op.h" #include "paddle/fluid/platform/assert.h" #include "paddle/fluid/platform/cudnn_helper.h" +#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { @@ -27,6 +28,8 @@ using ScopedTensorDescriptor = platform::ScopedTensorDescriptor; using ScopedFilterDescriptor = platform::ScopedFilterDescriptor; using ScopedConvolutionDescriptor = platform::ScopedConvolutionDescriptor; using DataLayout = platform::DataLayout; +template +using ScalingParamType = typename platform::CudnnDataType::ScalingParamType; static constexpr size_t kCONV_CUDNN_WORKSPACE_LIMIT_BYTES = static_cast(1024) * 1024 * 1024; @@ -133,7 +136,7 @@ class CUDNNConvOpKernel : public framework::OpKernel { platform::CUDAPlace gpu = boost::get(ctx.GetPlace()); cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes); // ------------------- cudnn conv forward --------------------- - T alpha = 1.0f, beta = 0.0f; + ScalingParamType alpha = 1.0f, beta = 0.0f; for (int i = 0; i < groups; i++) { PADDLE_ENFORCE(platform::dynload::cudnnConvolutionForward( handle, &alpha, cudnn_input_desc, input_data + i * group_offset_in, @@ -280,7 +283,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { platform::CUDAPlace gpu = boost::get(ctx.GetPlace()); cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes); // ------------------- cudnn conv backward data --------------------- - T alpha = 1.0f, beta = 0.0f; + ScalingParamType alpha = 1.0f, beta = 0.0f; if (input_grad) { T* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset input_grad. @@ -315,16 +318,18 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { } // namespace operators } // namespace paddle -REGISTER_OP_KERNEL(conv2d, CUDNN, ::paddle::platform::CUDAPlace, +namespace plat = paddle::platform; +REGISTER_OP_KERNEL(conv2d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel, - paddle::operators::CUDNNConvOpKernel); -REGISTER_OP_KERNEL(conv2d_grad, CUDNN, ::paddle::platform::CUDAPlace, + paddle::operators::CUDNNConvOpKernel, + paddle::operators::CUDNNConvOpKernel); +REGISTER_OP_KERNEL(conv2d_grad, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvGradOpKernel, paddle::operators::CUDNNConvGradOpKernel); -REGISTER_OP_KERNEL(conv3d, CUDNN, ::paddle::platform::CUDAPlace, +REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel, paddle::operators::CUDNNConvOpKernel); -REGISTER_OP_KERNEL(conv3d_grad, CUDNN, ::paddle::platform::CUDAPlace, +REGISTER_OP_KERNEL(conv3d_grad, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvGradOpKernel, paddle::operators::CUDNNConvGradOpKernel); diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc index d59cc2c9d424f067ca638cb76e52c2e95ae75182..0a8a5d4c71c4510f04eea2f7ef12f836d1fd9c9b 100644 --- a/paddle/fluid/operators/conv_mkldnn_op.cc +++ b/paddle/fluid/operators/conv_mkldnn_op.cc @@ -12,58 +12,21 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "mkldnn.hpp" -#include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/operators/conv_op.h" #include "paddle/fluid/platform/mkldnn_helper.h" namespace paddle { namespace operators { -using paddle::framework::Tensor; -using paddle::platform::MKLDNNDeviceContext; -using paddle::platform::MKLDNNMemDesc; - -using mkldnn::memory; // Note: paddle has also "memory" namespace -using mkldnn::primitive; -using mkldnn::convolution_forward; -using mkldnn::convolution_backward_weights; -using mkldnn::convolution_backward_data; -using mkldnn::convolution_direct; -using mkldnn::prop_kind; -using mkldnn::padding_kind; -using mkldnn::stream; - -namespace { -std::unique_ptr -ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights, - const memory::desc& dst, const std::vector& strides, - const std::vector& paddings, - const mkldnn::engine& engine); - -convolution_backward_weights::primitive_desc ConvBwdWeightsPrimitiveDesc( - const memory::desc& src, const memory::desc& diff_weights, - const memory::desc& diff_dst, const std::vector& strides, - const std::vector& paddings, - const convolution_forward::primitive_desc& conv_pd, - const mkldnn::engine& engine); - -convolution_backward_data::primitive_desc ConvBwdDataPrimitiveDesc( - const memory::desc& diff_src, const memory::desc& weights, - const memory::desc& diff_dst, const std::vector& strides, - const std::vector& paddings, - const convolution_forward::primitive_desc& conv_pd, - const mkldnn::engine& engine); -} // anonymous namespace - template -class ConvOpMkldnnKernel : public paddle::framework::OpKernel { +class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { public: void Compute(const paddle::framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), "It must use CPUPlace."); - auto& dev_ctx = ctx.template device_context(); + auto& dev_ctx = + ctx.template device_context(); const auto& mkldnn_engine = dev_ctx.GetEngine(); auto* input = ctx.Input("Input"); @@ -88,7 +51,6 @@ class ConvOpMkldnnKernel : public paddle::framework::OpKernel { const T* input_data = input->data(); const T* filter_data = filter->data(); - // allocate memory for output T* output_data = output->mutable_data(ctx.GetPlace()); PADDLE_ENFORCE(input->dims().size() == 4, @@ -102,48 +64,69 @@ class ConvOpMkldnnKernel : public paddle::framework::OpKernel { std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); // TODO(pzelazko-intel): support more formats - // memory descriptors for convolution src/weight/dst - auto conv_src_md = - MKLDNNMemDesc(src_tz, memory::data_type::f32, memory::format::nchw); - auto conv_weights_md = - MKLDNNMemDesc(weights_tz, memory::data_type::f32, memory::format::oihw); - auto conv_dst_md = - MKLDNNMemDesc(dst_tz, memory::data_type::f32, memory::format::nchw); - - // create memory primitives - auto conv_src_memory = - memory({conv_src_md, mkldnn_engine}, (void*)input_data); - auto conv_weights_memory = - memory({conv_weights_md, mkldnn_engine}, (void*)filter_data); - auto conv_dst_memory = memory({conv_dst_md, mkldnn_engine}, output_data); - - std::unique_ptr conv_pd = - ConvFwdPrimitiveDesc(conv_src_md, conv_weights_md, conv_dst_md, strides, - paddings, mkldnn_engine); - - // save p_conv_pd into dev_ctx to be referred in backward path - auto p_conv_pd = conv_pd.get(); - std::shared_ptr conv_pd_value = std::move(conv_pd); - dev_ctx.SetBlob(key_conv_pd, conv_pd_value); + auto src_md = platform::MKLDNNMemDesc( + src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + auto weights_md = + platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, + mkldnn::memory::format::oihw); + auto dst_md = platform::MKLDNNMemDesc( + dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + + auto src_memory = + mkldnn::memory({src_md, mkldnn_engine}, (void*)input_data); + auto weights_memory = + mkldnn::memory({weights_md, mkldnn_engine}, (void*)filter_data); + auto dst_memory = mkldnn::memory({dst_md, mkldnn_engine}, output_data); + + std::shared_ptr conv_pd = + ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings, + mkldnn_engine); + + // save conv_pd into global device context to be referred in backward path + dev_ctx.SetBlob(key_conv_pd, conv_pd); // create convolution op primitive - auto conv_prim = convolution_forward(*p_conv_pd, conv_src_memory, - conv_weights_memory, conv_dst_memory); + auto conv_prim = mkldnn::convolution_forward(*conv_pd, src_memory, + weights_memory, dst_memory); + + // push primitive to stream and wait until it's executed + std::vector pipeline{conv_prim}; + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + } - // push op to stream and wait MKLDNN until it's executed - std::vector pipeline{conv_prim}; - stream(stream::kind::eager).submit(pipeline).wait(); + private: + std::unique_ptr + ConvFwdPrimitiveDesc(const mkldnn::memory::desc& src, + const mkldnn::memory::desc& weights, + const mkldnn::memory::desc& dst, + const std::vector& strides, + const std::vector& paddings, + const mkldnn::engine& engine) const { + mkldnn::memory::dims stride_dims = {strides[0], strides[1]}; + mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]}; + + auto conv_desc = mkldnn::convolution_forward::desc( + mkldnn::prop_kind::forward, mkldnn::convolution_direct, src, weights, + dst, stride_dims, padding_dims, padding_dims, + mkldnn::padding_kind::zero); + + auto p_conv_pd = + new mkldnn::convolution_forward::primitive_desc(conv_desc, engine); + + return std::unique_ptr( + p_conv_pd); } }; template -class ConvGradOpMkldnnKernel : public paddle::framework::OpKernel { +class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel { public: void Compute(const paddle::framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), "It must use CPUPlace."); - auto& dev_ctx = ctx.template device_context(); + auto& dev_ctx = + ctx.template device_context(); const auto& mkldnn_engine = dev_ctx.GetEngine(); const Tensor* input = ctx.Input("Input"); @@ -170,7 +153,6 @@ class ConvGradOpMkldnnKernel : public paddle::framework::OpKernel { T* input_grad_data = nullptr; T* filter_grad_data = nullptr; - // allocate memory for gradient of input/filter if (input_grad) { input_grad_data = input_grad->mutable_data(ctx.GetPlace()); } @@ -184,130 +166,111 @@ class ConvGradOpMkldnnKernel : public paddle::framework::OpKernel { std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); // TODO(pzelazko-intel): support more formats - auto conv_src_md = - MKLDNNMemDesc(src_tz, memory::data_type::f32, memory::format::nchw); - auto conv_diff_src_md = - MKLDNNMemDesc(src_tz, memory::data_type::f32, memory::format::nchw); - auto conv_weights_md = - MKLDNNMemDesc(weights_tz, memory::data_type::f32, memory::format::oihw); - auto conv_diff_weights_md = - MKLDNNMemDesc(weights_tz, memory::data_type::f32, memory::format::oihw); - auto conv_diff_dst_md = - MKLDNNMemDesc(dst_tz, memory::data_type::f32, memory::format::nchw); + auto src_md = platform::MKLDNNMemDesc( + src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + auto diff_src_md = platform::MKLDNNMemDesc( + src_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + auto weights_md = + platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, + mkldnn::memory::format::oihw); + auto diff_weights_md = + platform::MKLDNNMemDesc(weights_tz, mkldnn::memory::data_type::f32, + mkldnn::memory::format::oihw); + auto diff_dst_md = platform::MKLDNNMemDesc( + dst_tz, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); // create memory - auto conv_diff_dst_memory = - memory({conv_diff_weights_md, mkldnn_engine}, (void*)output_grad_data); + auto diff_dst_memory = mkldnn::memory({diff_weights_md, mkldnn_engine}, + (void*)output_grad_data); // Retrieve conv_pd from device context - std::shared_ptr conv_pd; - convolution_forward::primitive_desc* p_conv_pd; - - conv_pd = dev_ctx.GetBlob(key_conv_pd); + auto conv_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_conv_pd)); PADDLE_ENFORCE(conv_pd != nullptr, "Fail to find conv_pd in device context"); - p_conv_pd = - static_cast(conv_pd.get()); // create backward conv primitive for weights if (filter_grad) { // create primitive descriptor - convolution_backward_weights::primitive_desc conv_bwd_weights_pd = - ConvBwdWeightsPrimitiveDesc(conv_src_md, conv_diff_weights_md, - conv_diff_dst_md, strides, paddings, - *p_conv_pd, mkldnn_engine); + mkldnn::convolution_backward_weights::primitive_desc conv_bwd_weights_pd = + ConvBwdWeightsPrimitiveDesc(src_md, diff_weights_md, diff_dst_md, + strides, paddings, *conv_pd, + mkldnn_engine); // create memory - auto conv_diff_weights_memory = memory( - {conv_diff_weights_md, mkldnn_engine}, (void*)filter_grad_data); - auto conv_src_memory = - memory({conv_src_md, mkldnn_engine}, (void*)input_data); + auto diff_weights_memory = mkldnn::memory( + {diff_weights_md, mkldnn_engine}, (void*)filter_grad_data); + auto src_memory = + mkldnn::memory({src_md, mkldnn_engine}, (void*)input_data); // create backward conv primitive for weights - auto conv_bwd_weights_prim = convolution_backward_weights( - conv_bwd_weights_pd, conv_src_memory, conv_diff_dst_memory, - conv_diff_weights_memory); + auto conv_bwd_weights_prim = mkldnn::convolution_backward_weights( + conv_bwd_weights_pd, src_memory, diff_dst_memory, + diff_weights_memory); // push primitive and execute it - std::vector pipeline{conv_bwd_weights_prim}; - stream(stream::kind::eager).submit(pipeline).wait(); + std::vector pipeline{conv_bwd_weights_prim}; + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } if (input_grad) { // create primitive descriptor - convolution_backward_data::primitive_desc conv_bwd_data_pd = - ConvBwdDataPrimitiveDesc(conv_diff_src_md, conv_weights_md, - conv_diff_dst_md, strides, paddings, - *p_conv_pd, mkldnn_engine); + mkldnn::convolution_backward_data::primitive_desc conv_bwd_data_pd = + ConvBwdDataPrimitiveDesc(diff_src_md, weights_md, diff_dst_md, + strides, paddings, *conv_pd, mkldnn_engine); // create memory - auto conv_diff_src_memory = - memory({conv_diff_src_md, mkldnn_engine}, (void*)input_grad_data); - auto conv_weights_memory = - memory({conv_weights_md, mkldnn_engine}, (void*)filter_data); + auto diff_src_memory = + mkldnn::memory({diff_src_md, mkldnn_engine}, (void*)input_grad_data); + auto weights_memory = + mkldnn::memory({weights_md, mkldnn_engine}, (void*)filter_data); // create backward conv primitive for data - auto conv_bwd_data_prim = - convolution_backward_data(conv_bwd_data_pd, conv_diff_dst_memory, - conv_weights_memory, conv_diff_src_memory); + auto conv_bwd_data_prim = mkldnn::convolution_backward_data( + conv_bwd_data_pd, diff_dst_memory, weights_memory, diff_src_memory); - // push primitive and execute it - std::vector pipeline{conv_bwd_data_prim}; - stream(stream::kind::eager).submit(pipeline).wait(); + // push primitive to stream and wait until it's executed + std::vector pipeline{conv_bwd_data_prim}; + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); } } // Compute() + + private: + mkldnn::convolution_backward_weights::primitive_desc + ConvBwdWeightsPrimitiveDesc( + const mkldnn::memory::desc& src, const mkldnn::memory::desc& diff_weights, + const mkldnn::memory::desc& diff_dst, const std::vector& strides, + const std::vector& paddings, + const mkldnn::convolution_forward::primitive_desc& conv_pd, + const mkldnn::engine& engine) const { + auto conv_bwd_weights_desc = mkldnn::convolution_backward_weights::desc( + mkldnn::convolution_direct, src, diff_weights, diff_dst, strides, + paddings, paddings, mkldnn::padding_kind::zero); + return mkldnn::convolution_backward_weights::primitive_desc( + conv_bwd_weights_desc, engine, conv_pd); + } + + mkldnn::convolution_backward_data::primitive_desc ConvBwdDataPrimitiveDesc( + const mkldnn::memory::desc& diff_src, const mkldnn::memory::desc& weights, + const mkldnn::memory::desc& diff_dst, const std::vector& strides, + const std::vector& paddings, + const mkldnn::convolution_forward::primitive_desc& conv_pd, + const mkldnn::engine& engine) const { + auto conv_bwd_data_desc = mkldnn::convolution_backward_data::desc( + mkldnn::convolution_direct, diff_src, weights, diff_dst, strides, + paddings, paddings, mkldnn::padding_kind::zero); + return mkldnn::convolution_backward_data::primitive_desc(conv_bwd_data_desc, + engine, conv_pd); + } }; -namespace { -std::unique_ptr ConvFwdPrimitiveDesc( - const memory::desc& src, const memory::desc& weights, - const memory::desc& dst, const std::vector& strides, - const std::vector& paddings, const mkldnn::engine& engine) { - mkldnn::memory::dims stride_dims = {strides[0], strides[1]}; - mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]}; - - auto conv_desc = mkldnn::convolution_forward::desc( - mkldnn::prop_kind::forward, mkldnn::convolution_direct, src, weights, dst, - stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero); - - auto p_conv_pd = new convolution_forward::primitive_desc(conv_desc, engine); - - return std::unique_ptr( - p_conv_pd); -} - -convolution_backward_weights::primitive_desc ConvBwdWeightsPrimitiveDesc( - const memory::desc& src, const memory::desc& diff_weights, - const memory::desc& diff_dst, const std::vector& strides, - const std::vector& paddings, - const convolution_forward::primitive_desc& conv_pd, - const mkldnn::engine& engine) { - auto conv_bwd_weights_desc = convolution_backward_weights::desc( - convolution_direct, src, diff_weights, diff_dst, strides, paddings, - paddings, padding_kind::zero); - return convolution_backward_weights::primitive_desc(conv_bwd_weights_desc, - engine, conv_pd); -} - -convolution_backward_data::primitive_desc ConvBwdDataPrimitiveDesc( - const memory::desc& diff_src, const memory::desc& weights, - const memory::desc& diff_dst, const std::vector& strides, - const std::vector& paddings, - const convolution_forward::primitive_desc& conv_pd, - const mkldnn::engine& engine) { - auto conv_bwd_data_desc = convolution_backward_data::desc( - convolution_direct, diff_src, weights, diff_dst, strides, paddings, - paddings, padding_kind::zero); - return convolution_backward_data::primitive_desc(conv_bwd_data_desc, engine, - conv_pd); -} -} // anonymous namespace } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OP_KERNEL(conv2d, MKLDNN, ::paddle::platform::CPUPlace, - ops::ConvOpMkldnnKernel); + ops::ConvMKLDNNOpKernel); REGISTER_OP_KERNEL(conv2d_grad, MKLDNN, ::paddle::platform::CPUPlace, - ops::ConvGradOpMkldnnKernel); + ops::ConvMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index 4b02b80d7772fa15d2333692551da5e59d93765f..650bc92be22af9ea8afcacf590a11190109e8811 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -70,25 +70,36 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { framework::OpKernelType ConvOp::GetExpectedKernelType( const framework::ExecutionContext& ctx) const { - framework::LibraryType library_{framework::LibraryType::kPlain}; + framework::LibraryType library{framework::LibraryType::kPlain}; #ifdef PADDLE_WITH_CUDA if (platform::CanCUDNNBeUsed(ctx)) { - library_ = framework::LibraryType::kCUDNN; + library = framework::LibraryType::kCUDNN; } #endif #ifdef PADDLE_WITH_MKLDNN - if (library_ == framework::LibraryType::kPlain && + if (library == framework::LibraryType::kPlain && platform::CanMKLDNNBeUsed(ctx)) { - library_ = framework::LibraryType::kMKLDNN; + library = framework::LibraryType::kMKLDNN; } #endif + auto input_data_type = + framework::ToDataType(ctx.Input("Input")->type()); + auto filter_data_type = + framework::ToDataType(ctx.Input("Filter")->type()); + PADDLE_ENFORCE_EQ(input_data_type, filter_data_type, + "input and filter data type should be consistent"); + + if (input_data_type == framework::proto::VarType::FP16) { + PADDLE_ENFORCE_EQ(library, framework::LibraryType::kCUDNN, + "float16 can only be used when CUDNN is used"); + } + std::string data_format = ctx.Attr("data_format"); // TODO(pzelazko-intel): enable MKLDNN layout when it's ready - framework::DataLayout layout_ = framework::StringToDataLayout(data_format); - return framework::OpKernelType( - framework::ToDataType(ctx.Input("Input")->type()), ctx.GetPlace(), - layout_, library_); + framework::DataLayout layout = framework::StringToDataLayout(data_format); + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout, + library); } Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker) diff --git a/paddle/fluid/operators/delete_var_op.cc b/paddle/fluid/operators/delete_var_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..1fe9404c00335edbe3594486f8c403e69f2ab08f --- /dev/null +++ b/paddle/fluid/operators/delete_var_op.cc @@ -0,0 +1,52 @@ +/* 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/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" + +namespace paddle { +namespace operators { +class DeleteVarOp : public framework::OperatorBase { + public: + DeleteVarOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + // get device context from pool + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(place); + dev_ctx.Wait(); + + auto delete_var_names = Inputs("X"); + const_cast(scope).EraseVars(delete_var_names); + } +}; + +class DeleteVarOpInfoMaker : public framework::OpProtoAndCheckerMaker { + public: + DeleteVarOpInfoMaker(OpProto *proto, OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input of delete op").AsDuplicable(); + AddComment(R"DOC( +Delete Operator. +It should not be configured by users directly. +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OPERATOR(delete_var, paddle::operators::DeleteVarOp, + paddle::framework::EmptyGradOpMaker, + paddle::operators::DeleteVarOpInfoMaker); diff --git a/paddle/fluid/operators/detail/CMakeLists.txt b/paddle/fluid/operators/detail/CMakeLists.txt index 0581bd2ac55218a2955fcb260d8b61cac0d210b5..94395ccfbcbd74ee40552a5c70dc8b8063a5f851 100644 --- a/paddle/fluid/operators/detail/CMakeLists.txt +++ b/paddle/fluid/operators/detail/CMakeLists.txt @@ -1,3 +1,6 @@ if(WITH_DISTRIBUTE) - grpc_library(sendrecvop_grpc SRCS sendrecvop_utils.cc grpc_client.cc grpc_server.cc PROTO send_recv.proto DEPS lod_tensor selected_rows) + grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc grpc_server.cc PROTO send_recv.proto DEPS lod_tensor selected_rows) + set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") + set_source_files_properties(test_serde.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + cc_test(serde_test SRCS test_serde.cc DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc) endif() diff --git a/paddle/fluid/operators/detail/bytebuffer_stream.cc b/paddle/fluid/operators/detail/bytebuffer_stream.cc new file mode 100644 index 0000000000000000000000000000000000000000..741dd51de9e75feb608161579e56cb160b058ebb --- /dev/null +++ b/paddle/fluid/operators/detail/bytebuffer_stream.cc @@ -0,0 +1,88 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +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. */ + +// NOTE: This file was originally created by tensorflow +// (https://github.com/tensorflow/tensorflow/) we borrow this +// file and did some modifications so that we can send gRPC +// requests without too much copying of the tensor data. + +#include "bytebuffer_stream.h" + +namespace paddle { +namespace operators { +namespace detail { + +GrpcByteBufferSource::GrpcByteBufferSource() {} + +bool GrpcByteBufferSource::Init(const grpc::ByteBuffer& src) { + cur_ = -1; + left_ = 0; + ptr_ = nullptr; + byte_count_ = 0; + bool ok = src.Dump(&slices_).ok(); + if (!ok) { + slices_.clear(); + } + return ok; +} + +bool GrpcByteBufferSource::Next(const void** data, int* size) { + // Use loop instead of if in case buffer contained empty slices. + while (left_ == 0) { + // Advance to next slice. + cur_++; + if (cur_ >= slices_.size()) { + return false; + } + const ::grpc::Slice& s = slices_[cur_]; + left_ = s.size(); + ptr_ = reinterpret_cast(s.begin()); + } + + *data = ptr_; + *size = left_; + byte_count_ += left_; + ptr_ += left_; + left_ = 0; + return true; +} + +void GrpcByteBufferSource::BackUp(int count) { + ptr_ -= count; + left_ += count; + byte_count_ -= count; +} + +bool GrpcByteBufferSource::Skip(int count) { + const void* data; + int size; + while (Next(&data, &size)) { + if (size >= count) { + BackUp(size - count); + return true; + } + // size < count; + count -= size; + } + // error or we have too large count; + return false; +} + +google::protobuf::int64 GrpcByteBufferSource::ByteCount() const { + return byte_count_; +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/bytebuffer_stream.h b/paddle/fluid/operators/detail/bytebuffer_stream.h new file mode 100644 index 0000000000000000000000000000000000000000..099deb12d0e436427c147ab9b1eb553b712e14fb --- /dev/null +++ b/paddle/fluid/operators/detail/bytebuffer_stream.h @@ -0,0 +1,51 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +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. */ + +// NOTE: This file was originally created by tensorflow +// (https://github.com/tensorflow/tensorflow/) we borrow this +// file and did some modifications so that we can send gRPC +// requests without too much copying of the tensor data. + +#pragma once + +#include +#include "google/protobuf/io/coded_stream.h" +#include "google/protobuf/io/zero_copy_stream.h" + +namespace paddle { +namespace operators { +namespace detail { + +// A ZeroCopyInputStream that reads from a grpc::ByteBuffer. +class GrpcByteBufferSource + : public ::google::protobuf::io::ZeroCopyInputStream { + public: + GrpcByteBufferSource(); + bool Init(const ::grpc::ByteBuffer& src); // Can be called multiple times. + bool Next(const void** data, int* size) override; + void BackUp(int count) override; + bool Skip(int count) override; + ::google::protobuf::int64 ByteCount() const override; + + private: + std::vector<::grpc::Slice> slices_; + size_t cur_; // Current slice index. + int left_; // Number of bytes in slices_[cur_] left to yield. + const char* ptr_; // Address of next byte in slices_[cur_] to yield. + ::google::protobuf::int64 byte_count_; +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/detail/grpc_client.cc index 7266f3276477891d3c7b6827316a428ef7a31c6e..ddeeebec58e02f1686fd2e3d3e5ac1a4c4fd3c59 100644 --- a/paddle/fluid/operators/detail/grpc_client.cc +++ b/paddle/fluid/operators/detail/grpc_client.cc @@ -97,7 +97,7 @@ bool RPCClient::AsyncGetVariable(const std::string& ep, return true; } -bool RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) { +void RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) { const auto ch = GetChannel(ep); BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); @@ -108,8 +108,18 @@ bool RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) { auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, (void*)s); req_count_++; +} - return true; +void RPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) { + const auto ch = GetChannel(ep); + FetchBarrierProcessor* s = new FetchBarrierProcessor(ch); + s->Prepare(time_out); + + sendrecv::VariableMessage req; + req.set_varname(FETCH_BARRIER_MESSAGE); + auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); + rpc->Finish(&s->reply_, &s->status_, (void*)s); + req_count_++; } bool RPCClient::Wait() { @@ -154,7 +164,7 @@ bool RPCClient::Proceed() { PADDLE_ENFORCE(tag); // TODO(gongwb): add more retries. - ClientBase* c = static_cast(tag); + BaseProcessor* c = static_cast(tag); if (!c->status_.ok()) { LOG(ERROR) << "proc param error:" << c->var_h_.String() << " grpc error:" << c->status_.error_message(); @@ -174,6 +184,8 @@ std::shared_ptr RPCClient::GetChannel(const std::string& ep) { } grpc::ChannelArguments args; + args.SetInt("grpc.testing.fixed_reconnect_backoff_ms", 5000); + args.SetCompressionAlgorithm(GRPC_COMPRESS_NONE); args.SetMaxSendMessageSize(std::numeric_limits::max()); args.SetMaxReceiveMessageSize(std::numeric_limits::max()); diff --git a/paddle/fluid/operators/detail/grpc_client.h b/paddle/fluid/operators/detail/grpc_client.h index 669838810de240358857300c1014aa622c17f808..f520367dd981288416631fdad15241fb5d811d07 100644 --- a/paddle/fluid/operators/detail/grpc_client.h +++ b/paddle/fluid/operators/detail/grpc_client.h @@ -52,14 +52,14 @@ struct VarHandle { void ProcGetResponse(const VarHandle& var_h, const sendrecv::VariableMessage& msg); -class ClientBase { +class BaseProcessor { public: - explicit ClientBase(std::shared_ptr ch) { + explicit BaseProcessor(std::shared_ptr ch) { stub_ = sendrecv::SendRecvService::NewStub(ch); context_ = NULL; } - virtual ~ClientBase() {} + virtual ~BaseProcessor() {} virtual void Prepare(const VarHandle& var_info, int64_t time_out) { context_.reset(new grpc::ClientContext()); @@ -91,9 +91,10 @@ class ClientBase { typedef std::function RequestSendCallBack; -class SendProcessor : public ClientBase { +class SendProcessor : public BaseProcessor { public: - explicit SendProcessor(std::shared_ptr ch) : ClientBase(ch) {} + explicit SendProcessor(std::shared_ptr ch) + : BaseProcessor(ch) {} virtual ~SendProcessor() {} @@ -110,9 +111,10 @@ class SendProcessor : public ClientBase { typedef std::function RequestGetCallBack; -class GetProcessor : public ClientBase { +class GetProcessor : public BaseProcessor { public: - explicit GetProcessor(std::shared_ptr ch) : ClientBase(ch) {} + explicit GetProcessor(std::shared_ptr ch) + : BaseProcessor(ch) {} virtual ~GetProcessor() {} @@ -126,10 +128,10 @@ class GetProcessor : public ClientBase { RequestGetCallBack response_call_back_ = ProcGetResponse; }; -class BatchBarrierProcessor : public ClientBase { +class BatchBarrierProcessor : public BaseProcessor { public: explicit BatchBarrierProcessor(std::shared_ptr ch) - : ClientBase(ch) {} + : BaseProcessor(ch) {} virtual ~BatchBarrierProcessor() {} @@ -137,6 +139,17 @@ class BatchBarrierProcessor : public ClientBase { sendrecv::VoidMessage reply_; }; +class FetchBarrierProcessor : public BaseProcessor { + public: + explicit FetchBarrierProcessor(std::shared_ptr ch) + : BaseProcessor(ch) {} + + virtual ~FetchBarrierProcessor() {} + + virtual void Process() {} + sendrecv::VariableMessage reply_; +}; + class RPCClient { public: bool AsyncSendVariable(const std::string& ep, @@ -151,7 +164,10 @@ class RPCClient { const std::string& var_name, int64_t time_out = 600 * 1000); - bool AsyncSendBatchBarrier(const std::string& ep, + void AsyncSendBatchBarrier(const std::string& ep, + int64_t time_out = 600 * 1000); + + void AsyncSendFetchBarrier(const std::string& ep, int64_t time_out = 600 * 1000); bool Wait(); diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc index 2a567516614aac39da9d069824f598ed26c5fb0c..8fff430cc4890925e4edba2fadb8eb7fc647d181 100644 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ b/paddle/fluid/operators/detail/grpc_server.cc @@ -84,7 +84,7 @@ class RequestGet final : public RequestBase { explicit RequestGet(sendrecv::SendRecvService::AsyncService* service, grpc::ServerCompletionQueue* cq, framework::Scope* scope, const platform::DeviceContext* dev_ctx, - SimpleBlockQueue* queue) + SimpleBlockQueue* queue) : RequestBase(service, cq), responder_(&ctx_), scope_(scope), @@ -101,11 +101,16 @@ class RequestGet final : public RequestBase { // proc request. std::string var_name = request_.varname(); auto* var = scope_->FindVar(var_name); - SerializeToMessage(var_name, var, *dev_ctx_, &reply_); + if (var_name != FETCH_BARRIER_MESSAGE) { + SerializeToMessage(var_name, var, *dev_ctx_, &reply_); + } // TODO(gongwb): check var's info. responder_.Finish(reply_, grpc::Status::OK, this); status_ = FINISH; - queue_->Push('c'); + MessageWithName msg_with_name = + // request name reply + std::make_pair(var_name, std::move(reply_)); + queue_->Push(msg_with_name); } protected: @@ -114,12 +119,16 @@ class RequestGet final : public RequestBase { ServerAsyncResponseWriter responder_; framework::Scope* scope_; const platform::DeviceContext* dev_ctx_; - SimpleBlockQueue* queue_; + SimpleBlockQueue* queue_; }; void AsyncGRPCServer::WaitClientGet(int count) { - for (int i = 0; i < count; ++i) { - var_get_queue_.Pop(); + int fetch_barriers = 0; + while (fetch_barriers < count) { + auto msg = var_get_queue_.Pop(); + if (msg.first == FETCH_BARRIER_MESSAGE) { + fetch_barriers++; + } } } diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h index e9402ff6aafe73684b57442c03ee4c573a902a46..b6666bcf96e484b0b17b935c0efb2930f19b19f2 100644 --- a/paddle/fluid/operators/detail/grpc_server.h +++ b/paddle/fluid/operators/detail/grpc_server.h @@ -77,7 +77,7 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { const platform::DeviceContext *dev_ctx_; // received variable from RPC, operators fetch variable from this queue. SimpleBlockQueue var_recv_queue_; - SimpleBlockQueue var_get_queue_; + SimpleBlockQueue var_get_queue_; // condition of the sub program std::mutex barrier_mutex_; diff --git a/paddle/fluid/operators/detail/proto_encoder_helper.h b/paddle/fluid/operators/detail/proto_encoder_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..4a7bfb8bd586fe84c9243bc64117d146c4386674 --- /dev/null +++ b/paddle/fluid/operators/detail/proto_encoder_helper.h @@ -0,0 +1,147 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +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. */ + +// NOTE: This file was originally created by tensorflow +// (https://github.com/tensorflow/tensorflow/) we borrow this +// file and did some modifications so that we can send gRPC +// requests without too much copying of the tensor data. + +#pragma once + +#include +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace detail { + +char* EncodeVarint32(char* dst, uint32_t v) { + // Operate on characters as unsigneds + unsigned char* ptr = reinterpret_cast(dst); + static const int B = 128; + if (v < (1 << 7)) { + *(ptr++) = v; + } else if (v < (1 << 14)) { + *(ptr++) = v | B; + *(ptr++) = v >> 7; + } else if (v < (1 << 21)) { + *(ptr++) = v | B; + *(ptr++) = (v >> 7) | B; + *(ptr++) = v >> 14; + } else if (v < (1 << 28)) { + *(ptr++) = v | B; + *(ptr++) = (v >> 7) | B; + *(ptr++) = (v >> 14) | B; + *(ptr++) = v >> 21; + } else { + *(ptr++) = v | B; + *(ptr++) = (v >> 7) | B; + *(ptr++) = (v >> 14) | B; + *(ptr++) = (v >> 21) | B; + *(ptr++) = v >> 28; + } + return reinterpret_cast(ptr); +} + +char* EncodeVarint64(char* dst, uint64_t v) { + static const int B = 128; + unsigned char* ptr = reinterpret_cast(dst); + while (v >= B) { + *(ptr++) = (v & (B - 1)) | B; + v >>= 7; + } + *(ptr++) = static_cast(v); + return reinterpret_cast(ptr); +} + +int VarintLength(uint64_t v) { + int len = 1; + while (v >= 128) { + v >>= 7; + len++; + } + return len; +} + +class ProtoEncodeHelper { + public: + ProtoEncodeHelper(char* buf, int max_size) + : base_(buf), p_(buf), limit_(base_ + max_size) {} + + ~ProtoEncodeHelper() { + // Make sure callers didn't do operations that went over max_size promised + PADDLE_ENFORCE_LE(p_, limit_); + } + + const char* data() const { return base_; } + size_t size() const { return p_ - base_; } + + void WriteUint64(int tag, uint64_t v) { + Encode32(combine(tag, WIRETYPE_VARINT)); + Encode64(v); + } + void WriteBool(int tag, bool v) { + Encode32(combine(tag, WIRETYPE_VARINT)); + EncodeBool(v); + } + void WriteString(int tag, const std::string& v) { + Encode32(combine(tag, WIRETYPE_LENGTH_DELIMITED)); + Encode32(v.size()); + EncodeBytes(v.data(), v.size()); + } + void WriteVarlengthBeginning(int tag, uint32_t len) { + Encode32(combine(tag, WIRETYPE_LENGTH_DELIMITED)); + Encode32(len); + } + void WriteRawBytes(const std::string& v) { EncodeBytes(v.data(), v.size()); } + + private: + // Note: this module's behavior must match the protocol buffer wire encoding + // format. + enum { + WIRETYPE_VARINT = 0, + WIRETYPE_LENGTH_DELIMITED = 2, + }; + static uint32_t combine(uint32_t tag, uint32_t type) { + return ((tag << 3) | type); + } + inline void Encode32(uint32_t v) { + if (v < 128) { + // Fast path for single-byte values. Many of the calls will use a + // constant value for v, so the comparison will get optimized away + // when Encode32 is inlined into the caller. + *p_ = v; + p_++; + } else { + p_ = EncodeVarint32(p_, v); + } + } + void Encode64(uint64_t v) { p_ = EncodeVarint64(p_, v); } + void EncodeBool(bool v) { + *p_ = (v ? 1 : 0); // Equal to varint32 encoding of 0 or 1 + p_++; + } + void EncodeBytes(const char* bytes, int N) { + memcpy(p_, bytes, N); + p_ += N; + } + + char* base_; + char* p_; + char* limit_; // Just for CHECKs +}; + +} // detail +} // operators +} // paddle diff --git a/paddle/fluid/operators/detail/safe_ref.h b/paddle/fluid/operators/detail/safe_ref.h index 9cb5851deba6b16261d4499afcfb867d9d706498..48bdce740878ea486eda6821dc29885a3e480114 100644 --- a/paddle/fluid/operators/detail/safe_ref.h +++ b/paddle/fluid/operators/detail/safe_ref.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once +#include "paddle/fluid/platform/enforce.h" + namespace paddle { namespace operators { namespace detail { diff --git a/paddle/fluid/operators/detail/send_recv.proto b/paddle/fluid/operators/detail/send_recv.proto index 8f962b4c69cc83dc2ab98b7dc27e18bc4b42bf18..b0215d4a80c9440f09c35434903fd6166b03e8b0 100644 --- a/paddle/fluid/operators/detail/send_recv.proto +++ b/paddle/fluid/operators/detail/send_recv.proto @@ -33,10 +33,34 @@ enum VarType { } message VariableMessage { + enum Type { + // Pod Types + BOOL = 0; + INT16 = 1; + INT32 = 2; + INT64 = 3; + FP16 = 4; + FP32 = 5; + FP64 = 6; + } + + message LodData { repeated int64 lod_data = 1; } + string varname = 1; // TODO(Yancey1989): reference framework::proto::VarDesc::VarType VarType type = 2; - bytes serialized = 3; + // bool persistable is not needed for sending. + // tensor info: + Type data_type = 3; + repeated int64 dims = 4; + + // lod details: + int64 lod_level = 5; + repeated LodData lod = 6; + // tensor data + bytes serialized = 7; + // selected_rows data + bytes rows = 8; } message VoidMessage {} diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/detail/sendrecvop_utils.cc index 169fd40fd950a74e61a4ed06a370f25b533957db..39117eeeb611b025c426938c60ddf82c6af232ca 100644 --- a/paddle/fluid/operators/detail/sendrecvop_utils.cc +++ b/paddle/fluid/operators/detail/sendrecvop_utils.cc @@ -13,6 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "google/protobuf/io/coded_stream.h" +#include "google/protobuf/io/zero_copy_stream.h" +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/operators/detail/bytebuffer_stream.h" +#include "paddle/fluid/operators/detail/proto_encoder_helper.h" namespace paddle { namespace operators { @@ -63,6 +68,233 @@ void DeserializeFromMessage(const sendrecv::VariableMessage& msg, } } +void SerializeToByteBuffer(const std::string& name, framework::Variable* var, + const platform::DeviceContext& ctx, + ::grpc::ByteBuffer* msg) { + using VarMsg = sendrecv::VariableMessage; + sendrecv::VariableMessage request; + std::string header; + request.AppendToString(&header); + // When using GPU, need to free the copied CPU buffer + // when the ByteBuffer destroies + // TODO(typhoonzero): add unref here, if we have dependent + // parallelism execution, need to know when to free the tensor. + DestroyCallback destroy_callback = [](void* backing) {}; + + void* buf = malloc(1024); + void* payload = nullptr; + size_t payload_size; + ProtoEncodeHelper e((char*)buf, 1024); + e.WriteString(VarMsg::kVarnameFieldNumber, name); + if (var->IsType()) { + e.WriteUint64(VarMsg::kTypeFieldNumber, 0); + } else if (var->IsType()) { + e.WriteUint64(VarMsg::kTypeFieldNumber, 1); + } + + switch (framework::ToVarType(var->Type())) { + case framework::proto::VarType_Type_LOD_TENSOR: { + auto tensor = var->Get(); + e.WriteUint64(VarMsg::kDataTypeFieldNumber, + framework::ToDataType(tensor.type())); + for (auto& dim : framework::vectorize(tensor.dims())) { + e.WriteUint64(VarMsg::kDimsFieldNumber, dim); + } + auto lod = tensor.lod(); // std::vector> + if (lod.size() > 0) { + e.WriteUint64(VarMsg::kLodLevelFieldNumber, lod.size()); + + for (auto& each : lod) { + e.WriteVarlengthBeginning(VarMsg::kLodFieldNumber, + 2 + // tag + varintlength of submessage + 1 + // kLodDataFieldNumber + each.size()); + // auto copied from GPU + for (auto& d : each) { + e.WriteUint64(VarMsg::LodData::kLodDataFieldNumber, d); + } + } + } + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + PADDLE_ENFORCE(platform::is_gpu_place(tensor.place())); + platform::CPUPlace cpu; + auto& gpu_dev_ctx = + static_cast(ctx); + auto copy_size = tensor.memory_size(); + payload = memory::Alloc(cpu, copy_size); + memory::Copy(cpu, payload, + boost::get(tensor.place()), + reinterpret_cast(tensor.data()), + copy_size, gpu_dev_ctx.stream()); + ctx.Wait(); + destroy_callback = [](void* backing) { + platform::CPUPlace cpu; + memory::Free(cpu, backing); + }; +#endif + } else { + payload = tensor.data(); + } + payload_size = tensor.memory_size(); + e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); + } break; + case framework::proto::VarType_Type_SELECTED_ROWS: { + // TODO(typhoonzero): selectedrows implement should not use unique_ptr + auto* slr = var->GetMutable(); + e.WriteUint64(VarMsg::kDataTypeFieldNumber, + framework::ToDataType(slr->value().type())); + for (auto& dim : framework::vectorize(slr->value().dims())) { + e.WriteUint64(VarMsg::kDimsFieldNumber, dim); + } + e.WriteUint64(VarMsg::kLodLevelFieldNumber, 0); + auto* tensor = slr->mutable_value(); + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + platform::CPUPlace cpu; + auto& gpu_dev_ctx = + static_cast(ctx); + auto copy_size = tensor->memory_size(); + payload = memory::Alloc(cpu, copy_size); + memory::Copy(cpu, payload, + boost::get(tensor->place()), + reinterpret_cast(tensor->data()), + copy_size, gpu_dev_ctx.stream()); + ctx.Wait(); + destroy_callback = [](void* backing) { + platform::CPUPlace cpu; + memory::Free(cpu, backing); + }; +#endif + } else { + payload = slr->mutable_value()->data(); + } + payload_size = tensor->memory_size(); + e.WriteVarlengthBeginning(VarMsg::kSerializedFieldNumber, payload_size); + } break; + default: + PADDLE_THROW("Serialize does not support type: %s", + typeid(var->Type()).name()); + break; + } + // steal reference of tensor data + ::grpc::Slice slices[4]; // metadata, tensor, rows meta, rows + int num_slices = 2; // only SelectedRows have rows buffer + slices[0] = ::grpc::Slice(e.size()); + memcpy(const_cast(slices[0].begin()), e.data(), e.size()); + slices[1] = ::grpc::Slice( + grpc_slice_new_with_user_data(payload, payload_size, destroy_callback, + static_cast(payload)), + ::grpc::Slice::STEAL_REF); + + if (framework::ToVarType(var->Type()) == + framework::proto::VarType_Type_SELECTED_ROWS) { + auto* slr = var->GetMutable(); + + ProtoEncodeHelper e2((char*)buf, 128); + // NOTE: rows is of type int64_t + size_t rows_memory_size = + slr->rows().capacity() * framework::SizeOfType(typeid(int64_t)); + e2.WriteVarlengthBeginning(VarMsg::kRowsFieldNumber, rows_memory_size); + slices[2] = ::grpc::Slice(e2.size()); + memcpy(const_cast(slices[2].begin()), e2.data(), e2.size()); + + slices[3] = ::grpc::Slice( + grpc_slice_new_with_user_data( + const_cast( + reinterpret_cast(slr->rows().data())), + rows_memory_size, + [](void* backing) { + // TODO(typhoonzero): add unref here, same as above. + }, + const_cast( + reinterpret_cast(slr->rows().data()))), + ::grpc::Slice::STEAL_REF); + num_slices = 4; + } + + ::grpc::ByteBuffer tmp(&slices[0], num_slices); + msg->Swap(&tmp); +} + +void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, + const platform::DeviceContext& ctx, + framework::Variable* var) { + sendrecv::VariableMessage meta; + GrpcByteBufferSource source; + source.Init(msg); + ::google::protobuf::io::CodedInputStream input(&source); + // do zerocopy parsing + PADDLE_ENFORCE(meta.ParseFromCodedStream(&input)); + PADDLE_ENFORCE(input.ConsumedEntireMessage()); + // dims is needed by both tensor and selectedrows + std::vector vecdims; + for (auto& d : meta.dims()) { + vecdims.push_back(d); + } + framework::DDim dims = framework::make_ddim(vecdims); + + if (meta.type() == sendrecv::LOD_TENSOR) { + auto* tensor = var->GetMutable(); + tensor->Resize(dims); + void* tensor_data = tensor->mutable_data( + ctx.GetPlace(), + paddle::operators::detail::ToTypeIndex(meta.data_type())); + framework::LoD lod; + for (int i = 0; i < meta.lod_level(); ++i) { + framework::Vector v; + for (int j = 0; j < meta.lod(i).lod_data_size(); ++j) { + v.push_back(meta.lod(i).lod_data(j)); + } + lod.push_back(v); + } + tensor->set_lod(lod); + // How to avoid copying and use the message buffer directly? + // Maybe need to find a way to release all memory except tensor content. + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + platform::CPUPlace cpu; + auto& gpu_dev_ctx = static_cast(ctx); + memory::Copy(boost::get(tensor->place()), + tensor_data, cpu, + reinterpret_cast(meta.serialized().data()), + meta.serialized().size(), gpu_dev_ctx.stream()); + ctx.Wait(); +#endif + } else { + memcpy(tensor_data, + reinterpret_cast(meta.serialized().data()), + meta.serialized().size()); + } + } else if (meta.type() == sendrecv::SELECTED_ROWS) { + auto* slr = var->GetMutable(); + auto* tensor = slr->mutable_value(); + int64_t* rows_data = slr->mutable_rows()->data(); + tensor->Resize(dims); + void* tensor_data = tensor->mutable_data( + ctx.GetPlace(), + paddle::operators::detail::ToTypeIndex(meta.data_type())); + if (platform::is_gpu_place(ctx.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + platform::CPUPlace cpu; + auto& gpu_dev_ctx = static_cast(ctx); + memory::Copy(boost::get(tensor->place()), + tensor_data, cpu, + reinterpret_cast(meta.serialized().data()), + meta.serialized().size(), gpu_dev_ctx.stream()); + ctx.Wait(); +#endif + } else { + memcpy(tensor_data, + reinterpret_cast(meta.serialized().data()), + meta.serialized().size()); + } + // copy rows CPU data, GPU data will be copied lazly + memcpy(rows_data, reinterpret_cast(meta.rows().data()), + meta.rows().size()); + } +} + } // namespace detail } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.h b/paddle/fluid/operators/detail/sendrecvop_utils.h index 670d0e162473750d0a5f8e9025ef1cf9a9ef407c..4fa6aefd3e0b1bd45ac52b1eff3b29126d79f03a 100644 --- a/paddle/fluid/operators/detail/sendrecvop_utils.h +++ b/paddle/fluid/operators/detail/sendrecvop_utils.h @@ -32,6 +32,9 @@ namespace detail { #define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV" #define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV" +#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV" + +typedef void (*DestroyCallback)(void*); void SerializeToMessage(const std::string& name, const framework::Variable* var, const platform::DeviceContext& ctx, @@ -40,6 +43,32 @@ void SerializeToMessage(const std::string& name, const framework::Variable* var, void DeserializeFromMessage(const sendrecv::VariableMessage& msg, const platform::DeviceContext& ctx, framework::Variable* var); + +void SerializeToByteBuffer(const std::string& name, framework::Variable* var, + const platform::DeviceContext& ctx, + ::grpc::ByteBuffer* msg); + +void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, + const platform::DeviceContext& ctx, + framework::Variable* var); + +inline std::type_index ToTypeIndex(sendrecv::VariableMessage::Type type) { + switch (type) { + case sendrecv::VariableMessage::FP32: + return typeid(float); // NOLINT + case sendrecv::VariableMessage::FP64: + return typeid(double); // NOLINT + case sendrecv::VariableMessage::INT32: + return typeid(int); // NOLINT + case sendrecv::VariableMessage::INT64: + return typeid(int64_t); // NOLINT + case sendrecv::VariableMessage::BOOL: + return typeid(bool); // NOLINT + default: + PADDLE_THROW("Not support type %d", type); + } +} + } // namespace detail } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/test_serde.cc b/paddle/fluid/operators/detail/test_serde.cc new file mode 100644 index 0000000000000000000000000000000000000000..2f06e5a686b996858d21930a1afa2861efca4a9b --- /dev/null +++ b/paddle/fluid/operators/detail/test_serde.cc @@ -0,0 +1,186 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +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 +#include + +#include "gtest/gtest.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/framework/variable.h" +#include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/place.h" +#include "paddle/fluid/string/printf.h" + +namespace framework = paddle::framework; +namespace platform = paddle::platform; +namespace operators = paddle::operators; +namespace math = paddle::operators::math; +namespace memory = paddle::memory; + +void RunSerdeTestTensor(platform::Place place) { + // serialize var to ByteBuffer + framework::Variable var; + auto* tensor = var.GetMutable(); + tensor->Resize(framework::make_ddim({4, 8, 4, 2})); + framework::LoD lod; + lod.push_back(framework::Vector({1, 3, 8})); + tensor->set_lod(lod); + int tensor_numel = 4 * 8 * 4 * 2; + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + tensor->mutable_data(place); + math::set_constant(ctx, tensor, 31.9); + + ::grpc::ByteBuffer msg; + operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg); + EXPECT_GT(msg.Length(), 0); + + // deserialize + std::vector<::grpc::Slice> slices; + (void)msg.Dump(&slices); + std::string tmp; + for (const auto& s : slices) { + tmp.append(reinterpret_cast(s.begin()), s.size()); + } + sendrecv::VariableMessage varmsg; + EXPECT_TRUE(varmsg.ParseFromString(tmp)); + EXPECT_EQ(varmsg.varname(), "myvar"); + EXPECT_EQ(varmsg.type(), 0); + EXPECT_EQ(varmsg.dims()[0], 4); + EXPECT_EQ(varmsg.dims()[1], 8); + EXPECT_EQ(varmsg.dims()[2], 4); + EXPECT_EQ(varmsg.dims()[3], 2); + EXPECT_EQ(varmsg.lod_level(), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); + EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); + + const float* tensor_data = + reinterpret_cast(varmsg.serialized().data()); + for (int i = 0; i < tensor_numel; ++i) { + EXPECT_FLOAT_EQ(tensor_data[i], 31.9); + } + + // deserialize zero-copy + framework::Variable var2; + operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2); + auto tensor2 = var2.Get(); + float* tensor_data2 = nullptr; + framework::Tensor tmp_tensor; + + if (platform::is_gpu_place(ctx.GetPlace())) { + platform::CPUPlace cpu; + framework::TensorCopy(tensor2, cpu, &tmp_tensor); + tensor_data2 = tmp_tensor.data(); + } else { + tensor_data2 = const_cast(tensor2.data()); + } + + EXPECT_EQ(varmsg.lod_level(), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); + EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); + for (int i = 0; i < tensor_numel; ++i) EXPECT_FLOAT_EQ(tensor_data2[i], 31.9); +} + +void RunSerdeTestSelectedRows(platform::Place place) { + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + + // serialize var to ByteBuffer + framework::Variable var; + auto* slr = var.GetMutable(); + auto* tensor = slr->mutable_value(); + auto* rows = slr->mutable_rows(); + tensor->Resize(framework::make_ddim({2, 10})); + tensor->mutable_data(place); + int tensor_numel = 2 * 10; + math::set_constant(ctx, tensor, 32.7); + rows->push_back(3); + rows->push_back(10); + + ::grpc::ByteBuffer msg; + operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg); + EXPECT_GT(msg.Length(), 0); + + // deserialize + std::vector<::grpc::Slice> slices; + (void)msg.Dump(&slices); + std::string tmp; + for (const auto& s : slices) { + tmp.append(reinterpret_cast(s.begin()), s.size()); + } + sendrecv::VariableMessage varmsg; + EXPECT_TRUE(varmsg.ParseFromString(tmp)); + + EXPECT_EQ(varmsg.varname(), "myvar"); + EXPECT_EQ(varmsg.type(), 1); + + const float* tensor_data = + reinterpret_cast(varmsg.serialized().data()); + const int64_t* rows_data = + reinterpret_cast(varmsg.rows().data()); + for (int i = 0; i < tensor_numel; ++i) { + EXPECT_FLOAT_EQ(tensor_data[i], 32.7); + } + EXPECT_EQ(rows_data[0], 3); + EXPECT_EQ(rows_data[1], 10); + // deserialize zero-copy + framework::Variable var2; + operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2); + + auto* slr2 = var2.GetMutable(); + auto* tensor2 = slr2->mutable_value(); + auto* rows2 = slr2->mutable_rows(); + float* tensor_data2 = nullptr; + framework::Tensor tmp_tensor; + + if (platform::is_gpu_place(ctx.GetPlace())) { + platform::CPUPlace cpu; + framework::TensorCopy(*tensor2, cpu, &tmp_tensor); + tensor_data2 = tmp_tensor.data(); + } else { + tensor_data2 = const_cast(tensor2->data()); + } + const int64_t* rows_data2 = rows2->data(); + + for (int i = 0; i < tensor_numel; ++i) { + EXPECT_FLOAT_EQ(tensor_data2[i], 32.7); + } + EXPECT_EQ(rows_data2[0], 3); + EXPECT_EQ(rows_data2[1], 10); +} + +TEST(SelectedRows, CPU) { + platform::CPUPlace place; + RunSerdeTestSelectedRows(place); +} + +TEST(SelectedRows, GPU) { + platform::CUDAPlace place; + RunSerdeTestSelectedRows(place); +} + +TEST(Tensor, CPU) { + platform::CPUPlace place; + RunSerdeTestTensor(place); +} + +TEST(Tensor, GPU) { + platform::CUDAPlace place; + RunSerdeTestTensor(place); +} \ No newline at end of file diff --git a/paddle/fluid/operators/detection_map_op.cc b/paddle/fluid/operators/detection_map_op.cc index 9b8ca925373eb99df3fcb14ec7962768dd2eab04..73c84c2fe0155d21d7059938330e44fa3668c6df 100644 --- a/paddle/fluid/operators/detection_map_op.cc +++ b/paddle/fluid/operators/detection_map_op.cc @@ -71,7 +71,7 @@ class DetectionMAPOp : public framework::OperatorWithKernel { return framework::OpKernelType( framework::ToDataType( ctx.Input("DetectRes")->type()), - ctx.device_context()); + platform::CPUPlace()); } }; diff --git a/paddle/fluid/operators/detection_map_op.h b/paddle/fluid/operators/detection_map_op.h index 637f8368f888933a2ad8321ad809412cb66d4482..8c15bfa36bfe72586cfcbdbd8efc4542253adaca 100644 --- a/paddle/fluid/operators/detection_map_op.h +++ b/paddle/fluid/operators/detection_map_op.h @@ -144,6 +144,15 @@ class DetectionMAPOpKernel : public framework::OpKernel { } } + inline void ClipBBox(const Box& bbox, Box* clipped_bbox) const { + T one = static_cast(1.0); + T zero = static_cast(0.0); + clipped_bbox->xmin = std::max(std::min(bbox.xmin, one), zero); + clipped_bbox->ymin = std::max(std::min(bbox.ymin, one), zero); + clipped_bbox->xmax = std::max(std::min(bbox.xmax, one), zero); + clipped_bbox->ymax = std::max(std::min(bbox.ymax, one), zero); + } + void GetBoxes(const framework::LoDTensor& input_label, const framework::LoDTensor& input_detect, std::vector>>& gt_boxes, @@ -264,7 +273,6 @@ class DetectionMAPOpKernel : public framework::OpKernel { std::map>>& true_pos, std::map>>& false_pos, const int class_num) const { - constexpr T kEPS = static_cast(1e-6); const int* pos_count_data = input_pos_count.data(); for (int i = 0; i < class_num; ++i) { label_pos_count[i] = pos_count_data[i]; @@ -273,12 +281,11 @@ class DetectionMAPOpKernel : public framework::OpKernel { auto SetData = [](const framework::LoDTensor& pos_tensor, std::map>>& pos) { const T* pos_data = pos_tensor.data(); - auto pos_data_lod = pos_tensor.lod(); - for (size_t i = 0; i < pos_data_lod.size(); ++i) { - for (size_t j = pos_data_lod[0][i]; j < pos_data_lod[0][i + 1]; ++j) { + auto pos_data_lod = pos_tensor.lod()[0]; + for (size_t i = 0; i < pos_data_lod.size() - 1; ++i) { + for (size_t j = pos_data_lod[i]; j < pos_data_lod[i + 1]; ++j) { T score = pos_data[j * 2]; - int flag = 1; - if (pos_data[j * 2 + 1] < kEPS) flag = 0; + int flag = pos_data[j * 2 + 1]; pos[i].push_back(std::make_pair(score, flag)); } } @@ -360,7 +367,9 @@ class DetectionMAPOpKernel : public framework::OpKernel { size_t max_idx = 0; auto score = pred_boxes[i].first; for (size_t j = 0; j < matched_bboxes.size(); ++j) { - T overlap = JaccardOverlap(pred_boxes[i].second, matched_bboxes[j]); + Box& pred_box = pred_boxes[i].second; + ClipBBox(pred_box, &pred_box); + T overlap = JaccardOverlap(pred_box, matched_bboxes[j]); if (overlap > max_overlap) { max_overlap = overlap; max_idx = j; diff --git a/paddle/fluid/operators/detection_output_op.cc b/paddle/fluid/operators/detection_output_op.cc deleted file mode 100644 index f7520475917ff23535f11ccfde0ee915112bba30..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detection_output_op.cc +++ /dev/null @@ -1,89 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -Indicesou 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/fluid/operators/detection_output_op.h" -namespace paddle { -namespace operators { - -class DetectionOutputOpMaker : public framework::OpProtoAndCheckerMaker { - public: - DetectionOutputOpMaker(OpProto* proto, OpAttrChecker* op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("Loc", - "(Tensor) The input tensor of detection_output operator." - "The input predict locations" - "The format of input tensor is kNCHW. Where K is priorbox point " - "numbers," - "N is How many boxes are there on each point, " - "C is 4, H and W both are 1."); - AddInput("Conf", - "(Tensor) The input tensor of detection_output operator." - "The input priorbox confidence." - "The format of input tensor is kNCHW. Where K is priorbox point " - "numbers," - "N is How many boxes are there on each point, " - "C is the number of classes, H and W both are 1."); - AddInput("PriorBox", - "(Tensor) The input tensor of detection_output operator." - "The format of input tensor is the position and variance " - "of the boxes"); - AddOutput("Out", - "(Tensor) The output tensor of detection_output operator."); - AddAttr("background_label_id", "(int), The background class index."); - AddAttr("num_classes", "(int), The number of the classification."); - AddAttr("nms_threshold", - "(float), The Non-maximum suppression threshold."); - AddAttr("confidence_threshold", - "(float), The classification confidence threshold."); - AddAttr("top_k", "(int), The bbox number kept of the layer’s output."); - AddAttr("nms_top_k", - "(int), The bbox number kept of the NMS’s output."); - AddComment(R"DOC( - detection output for SSD(single shot multibox detector) - Apply the NMS to the output of network and compute the predict - bounding box location. The output’s shape of this layer could - be zero if there is no valid bounding box. - )DOC"); - } -}; - -class DetectionOutputOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("Loc"), - "Input(X) of DetectionOutputOp" - "should not be null."); - PADDLE_ENFORCE(ctx->HasInput("Conf"), - "Input(X) of DetectionOutputOp" - "should not be null."); - PADDLE_ENFORCE(ctx->HasInput("PriorBox"), - "Input(X) of DetectionOutputOp" - "should not be null."); - PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Output(Out) of DetectionOutputOp should not be null."); - std::vector output_shape({1, 7}); - ctx->SetOutputDim("Out", framework::make_ddim(output_shape)); - } -}; -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(detection_output, ops::DetectionOutputOp, - ops::DetectionOutputOpMaker); -REGISTER_OP_CPU_KERNEL( - detection_output, - ops::DetectionOutputKernel, - ops::DetectionOutputKernel); diff --git a/paddle/fluid/operators/detection_output_op.cu.cc b/paddle/fluid/operators/detection_output_op.cu.cc deleted file mode 100644 index 0f48765c9c67c1d3fa32b19d5e87b2acaa3c486a..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detection_output_op.cu.cc +++ /dev/null @@ -1,21 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -Indicesou 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/fluid/operators/detection_output_op.h" - -namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL( - detection_output, - ops::DetectionOutputKernel, - ops::DetectionOutputKernel); diff --git a/paddle/fluid/operators/detection_output_op.h b/paddle/fluid/operators/detection_output_op.h deleted file mode 100644 index af9081c93436776b6ca6ee7139e340054111e440..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/detection_output_op.h +++ /dev/null @@ -1,167 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - Indicesou 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/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/tensor.h" -#include "paddle/fluid/operators/math/detection_util.h" -#include "paddle/fluid/operators/math/math_function.h" -#include "paddle/fluid/operators/math/softmax.h" -#include "paddle/fluid/operators/strided_memcpy.h" -namespace paddle { -namespace operators { -template -inline void transpose_fun(const framework::ExecutionContext& context, - const framework::Tensor& src, - framework::Tensor* dst) { - int input_nums = src.dims()[0]; - int offset = 0; - for (int j = 0; j < input_nums; ++j) { - framework::Tensor in_p_tensor = src.Slice(j, j + 1); - std::vector shape_vec( - {in_p_tensor.dims()[0], in_p_tensor.dims()[1], in_p_tensor.dims()[3], - in_p_tensor.dims()[4], in_p_tensor.dims()[2]}); - framework::DDim shape(framework::make_ddim(shape_vec)); - framework::Tensor in_p_tensor_transpose; - in_p_tensor_transpose.mutable_data(shape, context.GetPlace()); - std::vector shape_axis({0, 1, 3, 4, 2}); - math::Transpose trans5; - trans5(context.template device_context(), in_p_tensor, - &in_p_tensor_transpose, shape_axis); - auto dst_stride = framework::stride(dst->dims()); - auto src_stride = framework::stride(in_p_tensor_transpose.dims()); - StridedMemcpy(context.device_context(), in_p_tensor_transpose.data(), - src_stride, in_p_tensor_transpose.dims(), dst_stride, - dst->data() + offset); - offset += in_p_tensor_transpose.dims()[4] * src_stride[4]; - } -} -template -class DetectionOutputKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - const framework::Tensor* in_loc = context.Input("Loc"); - const framework::Tensor* in_conf = context.Input("Conf"); - const framework::Tensor* in_priorbox = - context.Input("PriorBox"); - auto* out = context.Output("Out"); - int num_classes = context.template Attr("num_classes"); - int top_k = context.template Attr("top_k"); - int nms_top_k = context.template Attr("nms_top_k"); - int background_label_id = context.template Attr("background_label_id"); - float nms_threshold = context.template Attr("nms_threshold"); - float confidence_threshold = - context.template Attr("confidence_threshold"); - size_t batch_size = in_conf->dims()[1]; - int conf_sum_size = in_conf->numel(); - // for softmax - std::vector conf_shape_softmax_vec( - {conf_sum_size / num_classes, num_classes}); - framework::DDim conf_shape_softmax( - framework::make_ddim(conf_shape_softmax_vec)); - // for knchw => nhwc - std::vector loc_shape_vec({1, in_loc->dims()[1], in_loc->dims()[3], - in_loc->dims()[4], - in_loc->dims()[2] * in_loc->dims()[0]}); - std::vector conf_shape_vec( - {1, in_conf->dims()[1], in_conf->dims()[3], in_conf->dims()[4], - in_conf->dims()[2] * in_conf->dims()[0]}); - framework::DDim loc_shape(framework::make_ddim(loc_shape_vec)); - framework::DDim conf_shape(framework::make_ddim(conf_shape_vec)); - framework::Tensor loc_tensor; - framework::Tensor conf_tensor; - loc_tensor.mutable_data(loc_shape, context.GetPlace()); - conf_tensor.mutable_data(conf_shape, context.GetPlace()); - // for cpu - framework::Tensor loc_cpu; - framework::Tensor conf_cpu; - framework::Tensor priorbox_cpu; - const T* priorbox_data = in_priorbox->data(); - transpose_fun(context, *in_loc, &loc_tensor); - transpose_fun(context, *in_conf, &conf_tensor); - conf_tensor.Resize(conf_shape_softmax); - math::SoftmaxFunctor()( - context.template device_context(), &conf_tensor, - &conf_tensor); - T* loc_data = loc_tensor.data(); - T* conf_data = conf_tensor.data(); - if (platform::is_gpu_place(context.GetPlace())) { - loc_cpu.mutable_data(loc_tensor.dims(), platform::CPUPlace()); - framework::TensorCopy(loc_tensor, platform::CPUPlace(), - context.device_context(), &loc_cpu); - loc_data = loc_cpu.data(); - conf_cpu.mutable_data(conf_tensor.dims(), platform::CPUPlace()); - framework::TensorCopy(conf_tensor, platform::CPUPlace(), - context.device_context(), &conf_cpu); - conf_data = conf_cpu.data(); - priorbox_cpu.mutable_data(in_priorbox->dims(), platform::CPUPlace()); - framework::TensorCopy(*in_priorbox, platform::CPUPlace(), - context.device_context(), &priorbox_cpu); - priorbox_data = priorbox_cpu.data(); - } - // get decode bboxes - size_t num_priors = in_priorbox->numel() / 8; - std::vector>> all_decoded_bboxes; - for (size_t n = 0; n < batch_size; ++n) { - std::vector> decoded_bboxes; - for (size_t i = 0; i < num_priors; ++i) { - size_t prior_offset = i * 8; - size_t loc_pred_offset = n * num_priors * 4 + i * 4; - std::vector> prior_bbox_vec; - math::GetBBoxFromPriorData(priorbox_data + prior_offset, 1, - prior_bbox_vec); - std::vector> prior_bbox_var; - math::GetBBoxVarFromPriorData(priorbox_data + prior_offset, 1, - prior_bbox_var); - std::vector loc_pred_data; - for (size_t j = 0; j < 4; ++j) - loc_pred_data.push_back(*(loc_data + loc_pred_offset + j)); - math::BBox bbox = math::DecodeBBoxWithVar( - prior_bbox_vec[0], prior_bbox_var[0], loc_pred_data); - decoded_bboxes.push_back(bbox); - } - all_decoded_bboxes.push_back(decoded_bboxes); - } - std::vector>> all_indices; - int num_kept = math::GetDetectionIndices( - conf_data, num_priors, num_classes, background_label_id, batch_size, - confidence_threshold, nms_top_k, nms_threshold, top_k, - all_decoded_bboxes, &all_indices); - - if (num_kept <= 0) { - std::vector out_shape_vec({0, 0}); - framework::DDim out_shape(framework::make_ddim(out_shape_vec)); - out->Resize(out_shape); - return; - } - std::vector out_shape_vec({num_kept, 7}); - framework::DDim out_shape(framework::make_ddim(out_shape_vec)); - out->mutable_data(out_shape, context.GetPlace()); - framework::Tensor out_cpu; - T* out_data = out->data(); - if (platform::is_gpu_place(context.GetPlace())) { - out_cpu.mutable_data(out->dims(), platform::CPUPlace()); - out_data = out_cpu.data(); - } - math::GetDetectionOutput(conf_data, num_kept, num_priors, num_classes, - batch_size, all_indices, all_decoded_bboxes, - out_data); - if (platform::is_gpu_place(context.GetPlace())) { - framework::TensorCopy(out_cpu, platform::CUDAPlace(), - context.device_context(), out); - } - } -}; -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/elementwise_add_op.cc b/paddle/fluid/operators/elementwise_add_op.cc index e9068fcd50ba9309a37939788ca8f67f1f6e25cd..4aab54f60236ecc5fa7f70e22f1553c3bfe68198 100644 --- a/paddle/fluid/operators/elementwise_add_op.cc +++ b/paddle/fluid/operators/elementwise_add_op.cc @@ -29,8 +29,11 @@ class ElementwiseAddOpMaker : public ElementwiseOpMaker { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(elementwise_add, ops::ElementwiseOp, ops::ElementwiseAddOpMaker, - elementwise_add_grad, ops::ElementwiseOpGrad); +REGISTER_OPERATOR(elementwise_add, ops::ElementwiseOp, + ops::ElementwiseAddOpMaker, ops::ElementwiseOpInferVarType, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(elementwise_add_grad, ops::ElementwiseOpGrad); + REGISTER_OP_CPU_KERNEL( elementwise_add, ops::ElementwiseAddKernel, diff --git a/paddle/fluid/operators/elementwise_div_op.h b/paddle/fluid/operators/elementwise_div_op.h index 6bcc577456b13f7930e07c24564953e93c5339ed..95649ac46e6bd41b9e1a865794cdec3ae1e6e247 100644 --- a/paddle/fluid/operators/elementwise_div_op.h +++ b/paddle/fluid/operators/elementwise_div_op.h @@ -41,77 +41,14 @@ class ElementwiseDivKernel : public framework::OpKernel { }; template -struct ElementwiseDivGradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { - auto y_e = framework::EigenVector::Flatten(*y); - auto z_e = framework::EigenVector::Flatten(*z); - auto dz_e = framework::EigenVector::Flatten(*dz); - - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = dz_e / y_e; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = -1.0 * dz_e * z_e / y_e; - } - } -}; - -template -struct ElementwiseDivBroadCastGradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { - auto x_e = framework::EigenVector::Flatten(*x); - auto y_e = framework::EigenVector::Flatten(*y); - auto dz_e = framework::EigenVector::Flatten(*dz); - - auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n)) - .broadcast(Eigen::DSizes(pre, 1)) - .reshape(Eigen::DSizes(x_e.size())); - - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = dz_e / y_e_bcast; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast)) - .reshape(Eigen::DSizes(pre, n)) - .sum(Eigen::array{{0}}); - } - } +struct DivGradDX { + HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout / y; } }; template -struct ElementwiseDivBroadCast2GradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, - Post post) { - auto x_e = framework::EigenVector::Flatten(*x); - auto y_e = framework::EigenVector::Flatten(*y); - auto dz_e = framework::EigenVector::Flatten(*dz); - - auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) - .broadcast(Eigen::DSizes(pre, 1, post)) - .reshape(Eigen::DSizes(x_e.size())); - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = dz_e / y_e_bcast; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast)) - .reshape(Eigen::DSizes(pre, n, post)) - .sum(Eigen::array{{0, 2}}); - } +struct DivGradDY { + HOSTDEVICE T operator()(T x, T y, T out, T dout) const { + return -dout * x / (y * y); } }; @@ -128,10 +65,8 @@ class ElementwiseDivGradKernel : public framework::OpKernel { auto* dx = ctx.Output(framework::GradVarName("X")); auto* dy = ctx.Output(framework::GradVarName("Y")); int axis = ctx.Attr("axis"); - ElementwiseGradCompute, - ElementwiseDivBroadCastGradFunctor, - ElementwiseDivBroadCast2GradFunctor>( - ctx, x, y, out, dout, axis, dx, dy); + ElemwiseGradCompute, DivGradDY>( + ctx, *x, *y, *out, *dout, axis, dx, dy, DivGradDX(), DivGradDY()); } }; diff --git a/paddle/fluid/operators/elementwise_max_op.h b/paddle/fluid/operators/elementwise_max_op.h index ab3a3d58275e8a48a92aa0c8a50af072e83ac968..527a18ee3ba88a158a13266a7fbcdafe59ec69d9 100644 --- a/paddle/fluid/operators/elementwise_max_op.h +++ b/paddle/fluid/operators/elementwise_max_op.h @@ -41,76 +41,16 @@ class ElementwiseMaxKernel : public framework::OpKernel { }; template -struct ElementwiseMaxGradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { - auto x_e = framework::EigenVector::Flatten(*x); - auto y_e = framework::EigenVector::Flatten(*y); - auto dz_e = framework::EigenVector::Flatten(*dz); - - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = (x_e > y_e).template cast() * dz_e; - } - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = (x_e <= y_e).template cast() * dz_e; - } +struct MaxGradDx { + HOSTDEVICE T operator()(T x, T y, T out, T dout) const { + return dout * (x > y); } }; template -struct ElementwiseMaxBroadCastGradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { - auto x_e = framework::EigenVector::Flatten(*x); - auto y_e = framework::EigenVector::Flatten(*y); - auto dz_e = framework::EigenVector::Flatten(*dz); - - auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n)) - .broadcast(Eigen::DSizes(pre, 1)) - .reshape(Eigen::DSizes(x_e.size())); - - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = (x_e > y_e_bcast).template cast() * dz_e; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = ((x_e <= y_e_bcast).template cast() * dz_e) - .reshape(Eigen::DSizes(pre, n)) - .sum(Eigen::array{{0}}); - } - } -}; - -template -struct ElementwiseMaxBroadCast2GradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, - Post post) { - auto x_e = framework::EigenVector::Flatten(*x); - auto y_e = framework::EigenVector::Flatten(*y); - auto dz_e = framework::EigenVector::Flatten(*dz); - - auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) - .broadcast(Eigen::DSizes(pre, 1, post)) - .reshape(Eigen::DSizes(x_e.size())); - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = (x_e > y_e_bcast).template cast() * dz_e; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = ((x_e <= y_e_bcast).template cast() * dz_e) - .reshape(Eigen::DSizes(pre, n, post)) - .sum(Eigen::array{{0, 2}}); - } +struct MaxGradDy { + HOSTDEVICE T operator()(T x, T y, T out, T dout) const { + return dout * (x <= y); } }; @@ -127,12 +67,9 @@ class ElementwiseMaxGradKernel : public framework::OpKernel { auto* dx = ctx.Output(framework::GradVarName("X")); auto* dy = ctx.Output(framework::GradVarName("Y")); int axis = ctx.Attr("axis"); - ElementwiseGradCompute, - ElementwiseMaxBroadCastGradFunctor, - ElementwiseMaxBroadCast2GradFunctor>( - ctx, x, y, out, dout, axis, dx, dy); + ElemwiseGradCompute, MaxGradDy>( + ctx, *x, *y, *out, *dout, axis, dx, dy, MaxGradDx(), MaxGradDy()); } }; - } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/elementwise_min_op.h b/paddle/fluid/operators/elementwise_min_op.h index f0eec9d2468b519473f813d71127d670adb08250..d4e5831463f3e54c72789b6876ea696cf1b4ef4b 100644 --- a/paddle/fluid/operators/elementwise_min_op.h +++ b/paddle/fluid/operators/elementwise_min_op.h @@ -41,76 +41,16 @@ class ElementwiseMinKernel : public framework::OpKernel { }; template -struct ElementwiseMinGradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { - auto x_e = framework::EigenVector::Flatten(*x); - auto y_e = framework::EigenVector::Flatten(*y); - auto dz_e = framework::EigenVector::Flatten(*dz); - - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = (x_e < y_e).template cast() * dz_e; - } - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = (x_e >= y_e).template cast() * dz_e; - } +struct MinGradDx { + HOSTDEVICE T operator()(T x, T y, T out, T dout) const { + return dout * (x < y); } }; template -struct ElementwiseMinBroadCastGradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { - auto x_e = framework::EigenVector::Flatten(*x); - auto y_e = framework::EigenVector::Flatten(*y); - auto dz_e = framework::EigenVector::Flatten(*dz); - - auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n)) - .broadcast(Eigen::DSizes(pre, 1)) - .reshape(Eigen::DSizes(x_e.size())); - - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = (x_e < y_e_bcast).template cast() * dz_e; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = ((x_e >= y_e_bcast).template cast() * dz_e) - .reshape(Eigen::DSizes(pre, n)) - .sum(Eigen::array{{0}}); - } - } -}; - -template -struct ElementwiseMinBroadCast2GradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, - Post post) { - auto x_e = framework::EigenVector::Flatten(*x); - auto y_e = framework::EigenVector::Flatten(*y); - auto dz_e = framework::EigenVector::Flatten(*dz); - - auto y_e_bcast = y_e.reshape(Eigen::DSizes(1, n, 1)) - .broadcast(Eigen::DSizes(pre, 1, post)) - .reshape(Eigen::DSizes(x_e.size())); - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = (x_e < y_e_bcast).template cast() * dz_e; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = ((x_e >= y_e_bcast).template cast() * dz_e) - .reshape(Eigen::DSizes(pre, n, post)) - .sum(Eigen::array{{0, 2}}); - } +struct MinGradDy { + HOSTDEVICE T operator()(T x, T y, T out, T dout) const { + return dout * (x >= y); } }; @@ -127,12 +67,9 @@ class ElementwiseMinGradKernel : public framework::OpKernel { auto* dx = ctx.Output(framework::GradVarName("X")); auto* dy = ctx.Output(framework::GradVarName("Y")); int axis = ctx.Attr("axis"); - ElementwiseGradCompute, - ElementwiseMinBroadCastGradFunctor, - ElementwiseMinBroadCast2GradFunctor>( - ctx, x, y, out, dout, axis, dx, dy); + ElemwiseGradCompute, MinGradDy>( + ctx, *x, *y, *out, *dout, axis, dx, dy, MinGradDx(), MinGradDy()); } }; - } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/elementwise_mul_op.h b/paddle/fluid/operators/elementwise_mul_op.h index e2b59b31120964d49d2de29ed327ffe0e6d44ddc..dc73cb6f23614504640283af01981d3f69e89126 100644 --- a/paddle/fluid/operators/elementwise_mul_op.h +++ b/paddle/fluid/operators/elementwise_mul_op.h @@ -40,14 +40,15 @@ class ElementwiseMulKernel : public framework::OpKernel { }; template -struct IdentityGrad_DX { +struct MulGradDX { HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout * y; } }; template -struct IdentityGrad_DY { +struct MulGradDY { HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout * x; } }; + template class ElementwiseMulGradKernel : public framework::OpKernel { public: @@ -61,10 +62,8 @@ class ElementwiseMulGradKernel : public framework::OpKernel { auto* dx = ctx.Output(framework::GradVarName("X")); auto* dy = ctx.Output(framework::GradVarName("Y")); int axis = ctx.Attr("axis"); - ElemwiseGradCompute, - IdentityGrad_DY>(ctx, *x, *y, *out, *dout, axis, dx, - dy, IdentityGrad_DX(), - IdentityGrad_DY()); + ElemwiseGradCompute, MulGradDY>( + ctx, *x, *y, *out, *dout, axis, dx, dy, MulGradDX(), MulGradDY()); } }; } // namespace operators diff --git a/paddle/fluid/operators/elementwise_op.h b/paddle/fluid/operators/elementwise_op.h index fe31bbaed44fced68b7b51dd2c2031950ec4247d..f04d8d8fd82ed2336dff9c5b88808dc32de6630a 100644 --- a/paddle/fluid/operators/elementwise_op.h +++ b/paddle/fluid/operators/elementwise_op.h @@ -41,6 +41,16 @@ class ElementwiseOp : public framework::OperatorWithKernel { } }; +class ElementwiseOpInferVarType : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + auto x_var = op_desc.Input("X")[0]; + auto out_var = op_desc.Output("Out")[0]; + block->Var(out_var)->SetType(block->Var(x_var)->GetType()); + } +}; + class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { public: ElementwiseOpMaker(OpProto* proto, OpAttrChecker* op_checker) diff --git a/paddle/fluid/operators/elementwise_sub_op.h b/paddle/fluid/operators/elementwise_sub_op.h index a8fc242ed79ecee77dc5bb56882a1a15f872fccf..fe088b8203722a43b9aba7be3878b8f4ca68ba12 100644 --- a/paddle/fluid/operators/elementwise_sub_op.h +++ b/paddle/fluid/operators/elementwise_sub_op.h @@ -40,61 +40,13 @@ class ElementwiseSubKernel : public framework::OpKernel { }; template -struct ElementwiseSubGradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) { - auto dz_e = framework::EigenVector::Flatten(*dz); - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = dz_e; - } - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = (-1.0) * dz_e; - } - } +struct SubGradDX { + HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout; } }; template -struct ElementwiseSubBroadCastGradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) { - auto dz_e = framework::EigenVector::Flatten(*dz); - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = dz_e; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = (-1.0) * - dz_e.reshape(Eigen::DSizes(pre, n)) - .sum(Eigen::array{{0}}); - } - } -}; - -template -struct ElementwiseSubBroadCast2GradFunctor { - template - void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n, - Post post) { - auto dz_e = framework::EigenVector::Flatten(*dz); - if (dx) { - auto dx_e = framework::EigenVector::Flatten(*dx); - dx_e.device(d) = dz_e; - } - - if (dy) { - auto dy_e = framework::EigenVector::Flatten(*dy); - dy_e.device(d) = (-1.0) * - dz_e.reshape(Eigen::DSizes(pre, n, post)) - .sum(Eigen::array{{0, 2}}); - } - } +struct SubGradDY { + HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return -dout; } }; template @@ -110,12 +62,9 @@ class ElementwiseSubGradKernel : public framework::OpKernel { auto* dx = ctx.Output(framework::GradVarName("X")); auto* dy = ctx.Output(framework::GradVarName("Y")); int axis = ctx.Attr("axis"); - ElementwiseGradCompute, - ElementwiseSubBroadCastGradFunctor, - ElementwiseSubBroadCast2GradFunctor>( - ctx, x, y, out, dout, axis, dx, dy); + ElemwiseGradCompute, SubGradDY>( + ctx, *x, *y, *out, *dout, axis, dx, dy, SubGradDX(), SubGradDY()); } }; - } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/feed_op.cc b/paddle/fluid/operators/feed_op.cc index 90c31877f6a87d1e237283d489353b4aba26c97b..debacf07c360b9aa69000a0d891f04239ed08807 100644 --- a/paddle/fluid/operators/feed_op.cc +++ b/paddle/fluid/operators/feed_op.cc @@ -15,6 +15,7 @@ limitations under the License. */ #include "paddle/fluid/framework/feed_fetch_type.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -28,6 +29,10 @@ class FeedOp : public framework::OperatorBase { private: void RunImpl(const framework::Scope &scope, const platform::Place &place) const override { + // get device context from pool + auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place); + platform::RecordEvent record_event(Type(), dev_ctx); + auto feed_var_name = Input("X"); auto *feed_var = scope.FindVar(feed_var_name); @@ -50,14 +55,10 @@ class FeedOp : public framework::OperatorBase { auto &feed_item = feed_list.at(static_cast(col)); auto *out_item = out_var->GetMutable(); - // get device context from pool - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); - auto &dev_ctx = *pool.Get(place); - if (platform::is_same_place(feed_item.place(), place)) { out_item->ShareDataWith(feed_item); } else { - framework::TensorCopy(feed_item, place, dev_ctx, out_item); + framework::TensorCopy(feed_item, place, *dev_ctx, out_item); } out_item->set_lod(feed_item.lod()); } diff --git a/paddle/fluid/operators/fetch_op.cc b/paddle/fluid/operators/fetch_op.cc index d66f01d1b7ce8528a7c0177b2889aff7e0c5a12b..7c7f3e9059fbb1e3f2cca4f04edfff55c9452761 100644 --- a/paddle/fluid/operators/fetch_op.cc +++ b/paddle/fluid/operators/fetch_op.cc @@ -15,6 +15,7 @@ limitations under the License. */ #include "paddle/fluid/framework/feed_fetch_type.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -29,6 +30,9 @@ class FetchOp : public framework::OperatorBase { private: void RunImpl(const framework::Scope &scope, const platform::Place &place) const override { + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + platform::RecordEvent record_event(Type(), pool.Get(place)); + auto fetch_var_name = Input("X"); auto *fetch_var = scope.FindVar(fetch_var_name); PADDLE_ENFORCE(fetch_var != nullptr, @@ -53,7 +57,6 @@ class FetchOp : public framework::OperatorBase { // FIXME(yuyang18): Should we assume the fetch operator always generate // CPU outputs? - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); auto &dev_ctx = *pool.Get(src_item.place()); TensorCopy(src_item, platform::CPUPlace(), dev_ctx, &dst_item); diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index 8e9923c87ce22ed229f78ef15430e50cab16c947..4253300788462a3704076fc79241a864f2f130a0 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -128,8 +128,8 @@ class ListenAndServOp : public framework::OperatorBase { } } if (exit_flag) { - rpc_service_->ShutDown(); rpc_service_->SetCond(1); + rpc_service_->ShutDown(); break; } try { @@ -148,7 +148,7 @@ class ListenAndServOp : public framework::OperatorBase { } rpc_service_->SetCond(1); // FIXME(typhoonzero): use another condition to sync wait clients get. - rpc_service_->WaitClientGet(ins.size()); + rpc_service_->WaitClientGet(fan_in); sparse_vars.clear(); } // while(true) } diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc index 05f809ac5628420251957116bb2390b4502f11b8..6ffe0bec5e38432676ecadfa1abbbe70a1425bb1 100644 --- a/paddle/fluid/operators/load_op.cc +++ b/paddle/fluid/operators/load_op.cc @@ -15,6 +15,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -29,6 +30,9 @@ class LoadOp : public framework::OperatorBase { private: void RunImpl(const framework::Scope &scope, const platform::Place &place) const override { + auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place); + platform::RecordEvent record_event(Type(), dev_ctx); + auto filename = Attr("file_path"); std::ifstream fin(filename); PADDLE_ENFORCE(static_cast(fin), "Cannot open file %s for load op", @@ -41,9 +45,7 @@ class LoadOp : public framework::OperatorBase { auto *tensor = out_var->GetMutable(); - platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); - auto &dev_ctx = *pool.Get(place); - DeserializeFromStream(fin, tensor, dev_ctx); + DeserializeFromStream(fin, tensor, *dev_ctx); if (platform::is_gpu_place(place)) { // copy CPU to GPU @@ -55,7 +57,7 @@ class LoadOp : public framework::OperatorBase { out_var->Clear(); tensor = out_var->GetMutable(); tensor->set_lod(cpu_tensor.lod()); - TensorCopy(cpu_tensor, place, dev_ctx, tensor); + TensorCopy(cpu_tensor, place, *dev_ctx, tensor); } } }; diff --git a/paddle/fluid/operators/lod_reset_op.cc b/paddle/fluid/operators/lod_reset_op.cc index 6a66297cb843ead1a507a6867c1c562224861cbf..7d5687f2d0666d393d7bb1c1a2fdde6c95e6d615 100644 --- a/paddle/fluid/operators/lod_reset_op.cc +++ b/paddle/fluid/operators/lod_reset_op.cc @@ -22,17 +22,16 @@ class LoDResetOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - // input check PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of LoDResetOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of LoDResetOp should not be null."); - // If target LoD is not set form Input(), then it must be set from Attr(). - if (!ctx->HasInput("TargetLoD")) { + + if (!ctx->HasInput("Y")) { auto level0 = ctx->Attrs().Get>("target_lod"); - PADDLE_ENFORCE(level0.size() > 1, - "Target LoD is not found, should be set to be a valid one " - "through Input() or Attr()."); + PADDLE_ENFORCE_GT(level0.size(), 1, + "If Input(Y) not provided, the target lod should be " + "specified by attribute `target_lod`."); } ctx->SetOutputDim("Out", ctx->GetInputDim("X")); } @@ -50,36 +49,77 @@ class LoDResetOpMaker : public framework::OpProtoAndCheckerMaker { public: LoDResetOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "(LoDTensor) The input tensor of lod_reset operator."); - AddInput("TargetLoD", - "(Tensor, optional) The target level 0 LoD from Input().") + AddInput("X", + "(Tensor, LoDTensor) Input variable of LoDResetOp which " + "could be a Tensor or LoDTensor, where the data of output " + "variable inherits from."); + AddInput("Y", + "(Tensor, LoDTensor, optional) If provided and Y is LoDTensor, " + "lod of Input(Y) would be considered as the target lod first, " + "otherwise data of Input(Y) would be considered as the " + "target lod.") .AsDispensable(); - AddOutput("Out", "(LoDTensor) The output tensor of lod_reset operator."); + AddOutput("Out", + "(LoDTensor) Output variable of LoDResetOp which should be a " + "LoDTensor."); AddAttr>("target_lod", "The target level 0 LoD from Attr().") .SetDefault(std::vector{}); AddComment(R"DOC(LoDReset operator -Reset LoD of Input(X) into a new one specified by Input(TargetLoD) or -Attr(target_lod), or set LoD for Input(X) if it doesn't have one. -Currently the lod_reset operator only supports the reset of level 0 LoD. -At least one of Input(TargetLoD) and Attr(target_lod) must be set, -and if both of them are set, Input(TargetLoD) will be chosen as the -target LoD. +Set LoD of `X` to a new one specified by `Y` or attribute `target_lod`. When `Y` +provided and `Y` is a LoDTensor, `Y.lod` would be considered as target LoD +first, otherwise `Y.data` would be considered as target LoD. If `Y` is not +provided, target LoD should be specified by attribute `target_lod`. +If target LoD is specified by `Y.data` or `target_lod`, only one level LoD +is supported. + +Example 1: + +Given a 1-level LoDTensor input(X): + X.lod = [[ 0, 2, 5 6 ]] + X.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + X.dims = [6, 1] + +attr(target_lod): [0, 4, 6] + +then we get a 1-level LoDTensor: + Out.lod = [[ 0, 4, 6 ]] + Out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + Out.dims = [6, 1] + +Example 2: -An example: -Given a float LoDTensor X with shape (6, 1), its transpose form represents +Given a 1-level LoDTensor input(X): + X.lod = [[ 0, 2, 5 6 ]] + X.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + X.dims = [6, 1] - [1.0, 2.0, 3.0, 4.0, 5.0, 6.0], +input(Y) is a Tensor: + Y.data = [[0, 2, 6]] + Y.dims = [1, 3] -with LoD = [[0, 2, 5, 6]] and the three (transposed) sequences look like +then we get a 1-level LoDTensor: + Out.lod = [[ 0, 2, 6 ]] + Out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + Out.dims = [6, 1] - [1.0, 2.0], [3.0, 4.0, 5.0], [6.0]. +Example 3: -If target LoD = [0, 4, 6], the lod_reset operator will reset the LoD and -the sequences that the LoDTensor Output(Out) contains becomes: +Given a 1-level LoDTensor input(X): + X.lod = [[ 0, 2, 5 6 ]] + X.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + X.dims = [6, 1] - [1.0, 2.0, 3.0, 4.0], [5.0, 6.0]. +input(Y) is a 2-level LoDTensor: + Y.lod = [[0, 2, 4], [0, 2, 5, 6]] + Y.data = [[1.1], [2.1], [3.1], [4.1], [5.1], [6.1]] + Y.dims = [6, 1] + +then we get a 2-level LoDTensor: + Out.lod = [[0, 2, 4], [0, 2, 5, 6]] + Out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + Out.dims = [6, 1] )DOC"); } @@ -90,10 +130,16 @@ class LoDResetGradOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of LoDResetGradOp should not be null."); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), - "Input(Out@GRAD) shouldn't be null."); - ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + "Input(Out@Grad) of LoDResetGradOp should not be null."); + + auto x_grad_name = framework::GradVarName("X"); + if (ctx->HasOutput(x_grad_name)) { + ctx->SetOutputDim(x_grad_name, ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ x_grad_name); + } } protected: @@ -111,9 +157,13 @@ class LoDResetGradOp : public framework::OperatorWithKernel { namespace ops = paddle::operators; REGISTER_OP(lod_reset, ops::LoDResetOp, ops::LoDResetOpMaker, lod_reset_grad, ops::LoDResetGradOp); -REGISTER_OP_CPU_KERNEL(lod_reset, - ops::LoDResetKernel, - ops::LoDResetKernel); +REGISTER_OP_CPU_KERNEL( + lod_reset, ops::LoDResetKernel, + ops::LoDResetKernel, + ops::LoDResetKernel, + ops::LoDResetKernel); REGISTER_OP_CPU_KERNEL( lod_reset_grad, ops::LoDResetGradKernel, - ops::LoDResetGradKernel); + ops::LoDResetGradKernel, + ops::LoDResetGradKernel, + ops::LoDResetGradKernel); diff --git a/paddle/fluid/operators/lod_reset_op.cu b/paddle/fluid/operators/lod_reset_op.cu index b0e87a851a77a1cc98d419a63d4d9e5e1b9dd163..888d4c12eb4e3f4fd94d8dd4178c59acd0abb23b 100644 --- a/paddle/fluid/operators/lod_reset_op.cu +++ b/paddle/fluid/operators/lod_reset_op.cu @@ -18,8 +18,12 @@ namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( lod_reset, ops::LoDResetKernel, - ops::LoDResetKernel); + ops::LoDResetKernel, + ops::LoDResetKernel, + ops::LoDResetKernel); REGISTER_OP_CUDA_KERNEL( lod_reset_grad, ops::LoDResetGradKernel, - ops::LoDResetGradKernel); + ops::LoDResetGradKernel, + ops::LoDResetGradKernel, + ops::LoDResetGradKernel); diff --git a/paddle/fluid/operators/lod_reset_op.h b/paddle/fluid/operators/lod_reset_op.h index 8186d4f8262101edc723af390eee1aec4fa6f3a5..99f01c2a255ade81421c2bba95ff3d38ced6f87c 100644 --- a/paddle/fluid/operators/lod_reset_op.h +++ b/paddle/fluid/operators/lod_reset_op.h @@ -26,35 +26,46 @@ class LoDResetKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const { auto* out = ctx.Output("Out"); auto* in = ctx.Input("X"); - auto* lod_t = ctx.Input("TargetLoD"); + auto* lod_t = ctx.Input("Y"); + + out->ShareDataWith(*in); std::vector level0; if (lod_t) { - auto* lod = lod_t->data(); - if (platform::is_gpu_place(ctx.GetPlace())) { - framework::Tensor lod_cpu; - framework::TensorCopy(*lod_t, platform::CPUPlace(), - ctx.device_context(), &lod_cpu); - lod = lod_cpu.data(); + if (lod_t->lod().size() > 0) { + auto y_lod = lod_t->lod(); + auto last_level = y_lod[y_lod.size() - 1]; + PADDLE_ENFORCE_EQ(last_level.back(), in->dims()[0], + "Last value of `Y`'s last level LoD should be equal " + "to the first dimension of `X`"); + out->set_lod(y_lod); + return; // early return, since lod already set + } else { + auto* lod = lod_t->data(); + if (platform::is_gpu_place(ctx.GetPlace())) { + framework::Tensor lod_cpu; + framework::TensorCopy(*lod_t, platform::CPUPlace(), + ctx.device_context(), &lod_cpu); + lod = lod_cpu.data(); + } + level0 = std::vector(lod, lod + lod_t->numel()); } - level0 = std::vector(lod, lod + lod_t->numel()); } else { level0 = ctx.Attr>("target_lod"); } - PADDLE_ENFORCE(level0.size() > 1UL, - "The size of target LoD should be greater than 1."); - PADDLE_ENFORCE(level0[0] == 0, - "Target LoD should be a vector starting from 0."); - PADDLE_ENFORCE(level0.back() == in->dims()[0], - "Target LoD should be a vector end with the " - "first dimension of Input(X)."); + PADDLE_ENFORCE_GT(level0.size(), 1UL, + "Size of target LoD should be greater than 1."); + PADDLE_ENFORCE_EQ(level0[0], 0, + "Target LoD should be a vector starting from 0."); + PADDLE_ENFORCE_EQ(level0.back(), in->dims()[0], + "Target LoD should be a vector end with the " + "first dimension of Input(X)."); for (size_t i = 0; i < level0.size() - 1; ++i) { PADDLE_ENFORCE(level0[i + 1] > level0[i], "Target LoD should be an ascending vector."); } - out->ShareDataWith(*in); // cast level0 to size_t std::vector ulevel0(level0.size(), 0); std::transform(level0.begin(), level0.end(), ulevel0.begin(), diff --git a/paddle/fluid/operators/lookup_table_op.cc b/paddle/fluid/operators/lookup_table_op.cc index 3acdca17afc2fea05fb81871e6e03d72691fe91e..50eeadab72e71f39325c5eda69e9a3c3e6517d7d 100644 --- a/paddle/fluid/operators/lookup_table_op.cc +++ b/paddle/fluid/operators/lookup_table_op.cc @@ -33,8 +33,16 @@ class LookupTableOp : public framework::OperatorWithKernel { auto table_dims = ctx->GetInputDim("W"); auto ids_dims = ctx->GetInputDim("Ids"); - PADDLE_ENFORCE_EQ(ids_dims.size(), 2); - PADDLE_ENFORCE_EQ(ids_dims[1], 1); + auto ids_var_type = ctx->GetInputsVarType("Ids").front(); + // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type + // is LoDTensor, this tensor contains the ids to be looked up in W + // and it must be a column vector with rank = 2 while the 2nd dimension + // size must be 1, when Ids's type is SelectedRows, the rows of Ids + // contains the ids to be looked up in W; + if (ids_var_type == framework::proto::VarType::LOD_TENSOR) { + PADDLE_ENFORCE_EQ(ids_dims.size(), 2); + PADDLE_ENFORCE_EQ(ids_dims[1], 1); + } ctx->SetOutputDim("Out", {ids_dims[0], table_dims[1]}); ctx->ShareLoD("Ids", /*->*/ "Out"); @@ -54,17 +62,22 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { LookupTableOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { AddInput("W", - "An input represents embedding tensors, " + "(Tensor) The input represents embedding tensors, " "which is a learnable parameter."); - AddInput("Ids", - "An input with type int32 or int64 " - "contains the ids to be looked up in W. " - "Ids must be a column vector with rank = 2. " - "The 2nd dimension size must be 1."); - AddOutput("Out", "The lookup results, which have the same type as W."); + AddInput( + "Ids", + "(Tensor or SelectedRows) Ids's type can be Tensor or " + "SelectedRows, when Ids's type is Tensor, this tensor contains " + "the ids to be looked up in W and it must be a column vector with " + "rank = 2 while the 2nd dimension size must be 1; when Ids's type is " + "SelectedRows, the rows of Ids contains the ids to be looked up " + "in W."); + AddOutput("Out", + "(Tensor or SelectedRows) The lookup results, which have the " + "same type as W."); AddAttr("is_sparse", "(boolean, default false) " - "Sparse update") + "Sparse update.") .SetDefault(false); AddAttr("padding_idx", "(int64, default -1) " @@ -76,10 +89,15 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { Lookup Table Operator. This operator is used to perform lookups on the parameter W, -then concatenated into a dense tensor. +then concatenated into a dense or sparse tensor. + +The type of Ids(Input) is SelectedRows, Tensor or LoDTensor, when Ids's +type is SelectedRows, the rows of Ids contains the ids to be looked up in W; +when Ids's type is Tensor, this tensor contains the ids to be looked up in W +and it must be a column vector with rank = 2 while the 2nd dimension size must be 1, +at this time, Ids can carry the LoD (Level of Details) information, or not, and +the output only shares the LoD information with input Ids. -The input Ids can carry the LoD (Level of Details) information, -or not. And the output only shares the LoD information with input Ids. )DOC"); } diff --git a/paddle/fluid/operators/lookup_table_op.cu b/paddle/fluid/operators/lookup_table_op.cu index 923340f46102d888f549c79684ae0ae2f78ed038..6d81fccd2059c511f71d403229e04587e553e93d 100644 --- a/paddle/fluid/operators/lookup_table_op.cu +++ b/paddle/fluid/operators/lookup_table_op.cu @@ -74,14 +74,32 @@ class LookupTableCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto* table_t = context.Input("W"); - auto* ids_t = context.Input("Ids"); - auto* output_t = context.Output("Out"); int64_t padding_idx = context.Attr("padding_idx"); + auto* ids_var = context.InputVar("Ids"); + Tensor* output_t = context.Output("Out"); + + int64_t* ids; + int64_t K; + + // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type + // is LoDTensor, this tensor contains the ids to be looked up in W; + // when Ids's type is SelectedRows, the rows of Ids contains the + // ids to be looked up in W. + if (ids_var->IsType()) { + auto* ids_t = context.Input("Ids"); + ids = const_cast(ids_t->data()); + K = ids_t->numel(); + } else if (ids_var->IsType()) { + auto* ids_t = context.Input("Ids"); + ids = const_cast(ids_t->rows().CUDAData(context.GetPlace())); + K = ids_t->rows().size(); + output_t->Resize({K, table_t->dims()[1]}); + } else { + PADDLE_THROW("Unsupported Variable Type of Ids"); + } size_t N = table_t->dims()[0]; size_t D = table_t->dims()[1]; - size_t K = ids_t->numel(); - auto* ids = ids_t->data(); auto* table = table_t->data(); auto* output = output_t->mutable_data(context.GetPlace()); diff --git a/paddle/fluid/operators/lookup_table_op.h b/paddle/fluid/operators/lookup_table_op.h index d88b034e919f1127ac3c424e87e4a5f81a598dc8..c92ce78eeffb8f1517e61c6d6624d406e04d974d 100644 --- a/paddle/fluid/operators/lookup_table_op.h +++ b/paddle/fluid/operators/lookup_table_op.h @@ -22,6 +22,7 @@ limitations under the License. */ namespace paddle { namespace operators { +using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; using SelectedRows = framework::SelectedRows; @@ -29,25 +30,45 @@ template class LookupTableKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* table_t = context.Input("W"); // float tensor - auto* ids_t = context.Input("Ids"); // int tensor - auto* output_t = context.Output("Out"); // float tensor + auto* table_t = context.Input("W"); + auto* ids_var = context.InputVar("Ids"); + Tensor* output_t = context.Output("Out"); + + int64_t* ids; + int64_t ids_numel; + + // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type + // is LoDTensor, this tensor contains the ids to be looked up in W; + // when Ids's type is SelectedRows, the rows of Ids contains the + // ids to be looked up in W. + if (ids_var->IsType()) { + auto* ids_t = context.Input("Ids"); + ids = const_cast(ids_t->data()); + ids_numel = ids_t->numel(); + } else if (ids_var->IsType()) { + auto* ids_t = context.Input("Ids"); + ids = const_cast(ids_t->rows().data()); + ids_numel = ids_t->rows().size(); + output_t->Resize({ids_numel, table_t->dims()[1]}); + } else { + PADDLE_THROW("Unsupported Variable Type of Ids"); + } + int64_t padding_idx = context.Attr("padding_idx"); 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()); if (padding_idx == -1) { - for (int64_t i = 0; i < ids_t->numel(); ++i) { + for (int64_t i = 0; i < ids_numel; ++i) { PADDLE_ENFORCE_LT(ids[i], N); PADDLE_ENFORCE_GE(ids[i], 0); memcpy(output + i * D, table + ids[i] * D, D * sizeof(T)); } } else { - for (int64_t i = 0; i < ids_t->numel(); ++i) { + for (int64_t i = 0; i < ids_numel; ++i) { if (ids[i] == padding_idx) { memset(output + i * D, 0, D * sizeof(T)); } else { diff --git a/paddle/fluid/operators/lrn_op.cc b/paddle/fluid/operators/lrn_op.cc index b0c213d637c244e4cbacbe75218537973efed047..692e85dcffa583abcb22a1629953badc67489efa 100644 --- a/paddle/fluid/operators/lrn_op.cc +++ b/paddle/fluid/operators/lrn_op.cc @@ -36,7 +36,7 @@ struct LRNFunctor { auto e_x = framework::EigenTensor::From(input); for (int m = 0; m < N; m++) { for (int i = 0; i < C; i++) { - for (int c = start; c <= end; c++) { + for (int c = start; c < end; c++) { int ch = i + c; if (ch >= 0 && ch < C) { auto s = e_mid.slice(Eigen::array({{m, i, 0, 0}}), @@ -92,7 +92,7 @@ struct LRNGradFunctor { Eigen::array({{1, 1, H, W}})); i_x_g = i_mid.pow(-beta) * i_out_g; - for (int c = start; c <= end; c++) { + for (int c = start; c < end; c++) { int ch = i + c; if (ch < 0 || ch >= C) { continue; diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt index a181d802262d15b188060dae4330cec0e24714ab..fba1612d10f0494f4ab06fabdd0e799a74dafd53 100644 --- a/paddle/fluid/operators/math/CMakeLists.txt +++ b/paddle/fluid/operators/math/CMakeLists.txt @@ -38,7 +38,7 @@ math_library(lstm_compute DEPS activation_functions) math_library(math_function DEPS cblas) math_library(maxouting) math_library(pooling) -math_library(selected_rows_functor DEPS selected_rows) +math_library(selected_rows_functor DEPS selected_rows math_function) math_library(sequence2batch) math_library(sequence_padding) math_library(sequence_pooling DEPS math_function) diff --git a/paddle/fluid/operators/math/detection_util.h b/paddle/fluid/operators/math/detection_util.h deleted file mode 100644 index c31764cfaf5bbdfea2f3ed06f31f97965a8858ed..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/math/detection_util.h +++ /dev/null @@ -1,300 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -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 "paddle/fluid/framework/selected_rows.h" -#include "paddle/fluid/platform/device_context.h" - -namespace paddle { -namespace operators { -namespace math { -template -struct BBox { - BBox(T x_min, T y_min, T x_max, T y_max) - : x_min(x_min), - y_min(y_min), - x_max(x_max), - y_max(y_max), - is_difficult(false) {} - - BBox() {} - - T get_width() const { return x_max - x_min; } - - T get_height() const { return y_max - y_min; } - - T get_center_x() const { return (x_min + x_max) / 2; } - - T get_center_y() const { return (y_min + y_max) / 2; } - - T get_area() const { return get_width() * get_height(); } - - // coordinate of bounding box - T x_min; - T y_min; - T x_max; - T y_max; - // whether difficult object (e.g. object with heavy occlusion is difficult) - bool is_difficult; -}; -// KNCHW ==> NHWC -// template -template -void GetBBoxFromPriorData(const T* prior_data, const size_t num_bboxes, - std::vector>& bbox_vec); -template -void GetBBoxVarFromPriorData(const T* prior_data, const size_t num, - std::vector>& var_vec); -template -BBox DecodeBBoxWithVar(BBox& prior_bbox, - const std::vector& prior_bbox_var, - const std::vector& loc_pred_data); -template -bool SortScorePairDescend(const std::pair& pair1, - const std::pair& pair2); -template -bool SortScorePairDescend(const std::pair>& pair1, - const std::pair>& pair2); -template -T jaccard_overlap(const BBox& bbox1, const BBox& bbox2); - -template -void ApplyNmsFast(const std::vector>& bboxes, const T* conf_score_data, - size_t class_idx, size_t top_k, T conf_threshold, - T nms_threshold, size_t num_priors, size_t num_classes, - std::vector* indices); -template -int GetDetectionIndices( - const T* conf_data, const size_t num_priors, const size_t num_classes, - const size_t background_label_id, const size_t batch_size, - const T conf_threshold, const size_t nms_top_k, const T nms_threshold, - const size_t top_k, - const std::vector>>& all_decoded_bboxes, - std::vector>>* all_detection_indices); -template -BBox ClipBBox(const BBox& bbox); -template -void GetDetectionOutput( - const T* conf_data, const size_t num_kept, const size_t num_priors, - const size_t num_classes, const size_t batch_size, - const std::vector>>& all_indices, - const std::vector>>& all_decoded_bboxes, T* out_data); -template -void GetBBoxFromPriorData(const T* prior_data, const size_t num_bboxes, - std::vector>& bbox_vec) { - size_t out_offset = bbox_vec.size(); - bbox_vec.resize(bbox_vec.size() + num_bboxes); - for (size_t i = 0; i < num_bboxes; ++i) { - BBox bbox; - bbox.x_min = *(prior_data + i * 8); - bbox.y_min = *(prior_data + i * 8 + 1); - bbox.x_max = *(prior_data + i * 8 + 2); - bbox.y_max = *(prior_data + i * 8 + 3); - bbox_vec[out_offset + i] = bbox; - } -} -template -void GetBBoxVarFromPriorData(const T* prior_data, const size_t num, - std::vector>& var_vec) { - size_t out_offset = var_vec.size(); - var_vec.resize(var_vec.size() + num); - for (size_t i = 0; i < num; ++i) { - std::vector var; - var.push_back(*(prior_data + i * 8 + 4)); - var.push_back(*(prior_data + i * 8 + 5)); - var.push_back(*(prior_data + i * 8 + 6)); - var.push_back(*(prior_data + i * 8 + 7)); - var_vec[out_offset + i] = var; - } -} -template -BBox DecodeBBoxWithVar(BBox& prior_bbox, - const std::vector& prior_bbox_var, - const std::vector& loc_pred_data) { - T prior_bbox_width = prior_bbox.get_width(); - T prior_bbox_height = prior_bbox.get_height(); - T prior_bbox_center_x = prior_bbox.get_center_x(); - T prior_bbox_center_y = prior_bbox.get_center_y(); - - T decoded_bbox_center_x = - prior_bbox_var[0] * loc_pred_data[0] * prior_bbox_width + - prior_bbox_center_x; - T decoded_bbox_center_y = - prior_bbox_var[1] * loc_pred_data[1] * prior_bbox_height + - prior_bbox_center_y; - T decoded_bbox_width = - std::exp(prior_bbox_var[2] * loc_pred_data[2]) * prior_bbox_width; - T decoded_bbox_height = - std::exp(prior_bbox_var[3] * loc_pred_data[3]) * prior_bbox_height; - - BBox decoded_bbox; - decoded_bbox.x_min = decoded_bbox_center_x - decoded_bbox_width / 2; - decoded_bbox.y_min = decoded_bbox_center_y - decoded_bbox_height / 2; - decoded_bbox.x_max = decoded_bbox_center_x + decoded_bbox_width / 2; - decoded_bbox.y_max = decoded_bbox_center_y + decoded_bbox_height / 2; - - return decoded_bbox; -} -template -bool SortScorePairDescend(const std::pair& pair1, - const std::pair& pair2) { - return pair1.first > pair2.first; -} -template -T jaccard_overlap(const BBox& bbox1, const BBox& bbox2) { - if (bbox2.x_min > bbox1.x_max || bbox2.x_max < bbox1.x_min || - bbox2.y_min > bbox1.y_max || bbox2.y_max < bbox1.y_min) { - return 0.0; - } else { - T inter_x_min = std::max(bbox1.x_min, bbox2.x_min); - T inter_y_min = std::max(bbox1.y_min, bbox2.y_min); - T interX_max = std::min(bbox1.x_max, bbox2.x_max); - T interY_max = std::min(bbox1.y_max, bbox2.y_max); - - T inter_width = interX_max - inter_x_min; - T inter_height = interY_max - inter_y_min; - T inter_area = inter_width * inter_height; - - T bbox_area1 = bbox1.get_area(); - T bbox_area2 = bbox2.get_area(); - - return inter_area / (bbox_area1 + bbox_area2 - inter_area); - } -} - -template -void ApplyNmsFast(const std::vector>& bboxes, const T* conf_score_data, - size_t class_idx, size_t top_k, T conf_threshold, - T nms_threshold, size_t num_priors, size_t num_classes, - std::vector* indices) { - std::vector> scores; - for (size_t i = 0; i < num_priors; ++i) { - size_t conf_offset = i * num_classes + class_idx; - if (conf_score_data[conf_offset] > conf_threshold) - scores.push_back(std::make_pair(conf_score_data[conf_offset], i)); - } - std::stable_sort(scores.begin(), scores.end(), - SortScorePairDescend); - if (top_k > 0 && top_k < scores.size()) scores.resize(top_k); - while (scores.size() > 0) { - const size_t idx = scores.front().second; - bool keep = true; - for (size_t i = 0; i < indices->size(); ++i) { - if (keep) { - const size_t saved_idx = (*indices)[i]; - T overlap = jaccard_overlap(bboxes[idx], bboxes[saved_idx]); - keep = overlap <= nms_threshold; - } else { - break; - } - } - if (keep) indices->push_back(idx); - scores.erase(scores.begin()); - } -} -template -int GetDetectionIndices( - const T* conf_data, const size_t num_priors, const size_t num_classes, - const size_t background_label_id, const size_t batch_size, - const T conf_threshold, const size_t nms_top_k, const T nms_threshold, - const size_t top_k, - const std::vector>>& all_decoded_bboxes, - std::vector>>* all_detection_indices) { - int total_keep_num = 0; - for (size_t n = 0; n < batch_size; ++n) { - const std::vector>& decoded_bboxes = all_decoded_bboxes[n]; - size_t num_detected = 0; - std::map> indices; - size_t conf_offset = n * num_priors * num_classes; - for (size_t c = 0; c < num_classes; ++c) { - if (c == background_label_id) continue; - ApplyNmsFast(decoded_bboxes, conf_data + conf_offset, c, nms_top_k, - conf_threshold, nms_threshold, num_priors, num_classes, - &(indices[c])); - num_detected += indices[c].size(); - } - if (top_k > 0 && num_detected > top_k) { - // std::vector> score_index_pairs; - std::vector>> score_index_pairs; - for (size_t c = 0; c < num_classes; ++c) { - const std::vector& label_indices = indices[c]; - for (size_t i = 0; i < label_indices.size(); ++i) { - size_t idx = label_indices[i]; - score_index_pairs.push_back( - std::make_pair((conf_data + conf_offset)[idx * num_classes + c], - std::make_pair(c, idx))); - } - } - std::sort(score_index_pairs.begin(), score_index_pairs.end(), - SortScorePairDescend>); - score_index_pairs.resize(top_k); - std::map> new_indices; - for (size_t i = 0; i < score_index_pairs.size(); ++i) { - size_t label = score_index_pairs[i].second.first; - size_t idx = score_index_pairs[i].second.second; - new_indices[label].push_back(idx); - } - all_detection_indices->push_back(new_indices); - total_keep_num += top_k; - } else { - all_detection_indices->push_back(indices); - total_keep_num += num_detected; - } - } - return total_keep_num; -} -template -BBox ClipBBox(const BBox& bbox) { - T one = static_cast(1.0); - T zero = static_cast(0.0); - BBox clipped_bbox; - clipped_bbox.x_min = std::max(std::min(bbox.x_min, one), zero); - clipped_bbox.y_min = std::max(std::min(bbox.y_min, one), zero); - clipped_bbox.x_max = std::max(std::min(bbox.x_max, one), zero); - clipped_bbox.y_max = std::max(std::min(bbox.y_max, one), zero); - return clipped_bbox; -} -template -void GetDetectionOutput( - const T* conf_data, const size_t num_kept, const size_t num_priors, - const size_t num_classes, const size_t batch_size, - const std::vector>>& all_indices, - const std::vector>>& all_decoded_bboxes, T* out_data) { - size_t count = 0; - for (size_t n = 0; n < batch_size; ++n) { - for (std::map>::const_iterator it = - all_indices[n].begin(); - it != all_indices[n].end(); ++it) { - size_t label = it->first; - const std::vector& indices = it->second; - const std::vector>& decoded_bboxes = all_decoded_bboxes[n]; - for (size_t i = 0; i < indices.size(); ++i) { - size_t idx = indices[i]; - size_t conf_offset = n * num_priors * num_classes + idx * num_classes; - out_data[count * 7] = n; - out_data[count * 7 + 1] = label; - out_data[count * 7 + 2] = (conf_data + conf_offset)[label]; - BBox clipped_bbox = ClipBBox(decoded_bboxes[idx]); - out_data[count * 7 + 3] = clipped_bbox.x_min; - out_data[count * 7 + 4] = clipped_bbox.y_min; - out_data[count * 7 + 5] = clipped_bbox.x_max; - out_data[count * 7 + 6] = clipped_bbox.y_max; - ++count; - } - } - } -} -} // namespace math -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc index f7f33917d7ef5bbcc7fb5d6e3d0a7f3ae63cde34..35d251f71a0cb631d5900498ea3188b5ddeae334 100644 --- a/paddle/fluid/operators/math/math_function.cc +++ b/paddle/fluid/operators/math/math_function.cc @@ -15,11 +15,23 @@ limitations under the License. */ #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/operators/math/math_function_impl.h" +#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { namespace math { +using float16 = paddle::platform::float16; + +template <> +void gemm( + const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, + const float16 alpha, const float16* A, const float16* B, const float16 beta, + float16* C) { + PADDLE_THROW("float16 GEMM not supported on CPU"); +} + template <> void gemm( const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, @@ -46,6 +58,15 @@ void gemm( beta, C, ldc); } +template <> +void gemm( + const platform::CPUDeviceContext& context, const bool transA, + const bool transB, const int M, const int N, const int K, + const float16 alpha, const float16* A, const int lda, const float16* B, + const int ldb, const float16 beta, float16* C, const int ldc) { + PADDLE_THROW("float16 GEMM not supported on CPU"); +} + template <> void gemm( const platform::CPUDeviceContext& context, const bool transA, @@ -68,6 +89,15 @@ void gemm( lda, B, ldb, beta, C, ldc); } +template <> +void matmul( + const platform::CPUDeviceContext& context, + const framework::Tensor& matrix_a, bool trans_a, + const framework::Tensor& matrix_b, bool trans_b, float16 alpha, + framework::Tensor* matrix_out, float16 beta) { + PADDLE_THROW("float16 matmul not supported on CPU"); +} + template <> void matmul( const platform::CPUDeviceContext& context, @@ -126,6 +156,15 @@ void matmul( matrix_b.data(), beta, matrix_out->data()); } +template <> +void batched_gemm( + const platform::CPUDeviceContext& context, const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, + const float16 alpha, const float16* A, const float16* B, const float16 beta, + float16* C, const int batchCount, const int strideA, const int strideB) { + PADDLE_THROW("float16 batched_gemm not supported on CPU"); +} + #ifdef PADDLE_WITH_MKLML // Use cblas_{s,d}gemm_batched if available: Run with 1 group of size batchSize. template <> diff --git a/paddle/fluid/operators/math/math_function.cu b/paddle/fluid/operators/math/math_function.cu index f8d0349ac5cd96791bcb508230f5ff6d594c0e05..3abbcdb71d03eaf6f8eba3d97150d27ac5a5405e 100644 --- a/paddle/fluid/operators/math/math_function.cu +++ b/paddle/fluid/operators/math/math_function.cu @@ -16,11 +16,43 @@ limitations under the License. */ #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/math_function_impl.h" +#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { namespace math { +using float16 = paddle::platform::float16; + +template <> +void gemm( + const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, + const float16 alpha, const float16* A, const float16* B, const float16 beta, + float16* C) { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + cublasOperation_t cuTransA = + (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = + (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + + const half h_alpha = static_cast(alpha); + const half h_beta = static_cast(beta); + const half* h_A = reinterpret_cast(A); + const half* h_B = reinterpret_cast(B); + half* h_C = reinterpret_cast(C); + + // TODO(kexinzhao): add processing code for compute capability < 53 case + PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53, + "cublas Hgemm requires GPU compute capability >= 53"); + PADDLE_ENFORCE(platform::dynload::cublasHgemm( + context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb, + h_A, lda, &h_beta, h_C, N)); +} + template <> void gemm( const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, @@ -60,6 +92,31 @@ void gemm( lda, &beta, C, N)); } +template <> +void gemm( + const platform::CUDADeviceContext& context, const bool transA, + const bool transB, const int M, const int N, const int K, + const float16 alpha, const float16* A, const int lda, const float16* B, + const int ldb, const float16 beta, float16* C, const int ldc) { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + cublasOperation_t cuTransA = transA == false ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = transB == false ? CUBLAS_OP_N : CUBLAS_OP_T; + + const half h_alpha = static_cast(alpha); + const half h_beta = static_cast(beta); + const half* h_A = reinterpret_cast(A); + const half* h_B = reinterpret_cast(B); + half* h_C = reinterpret_cast(C); + + // TODO(kexinzhao): add processing code for compute capability < 53 case + PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53, + "cublas Hgemm requires GPU compute capability >= 53"); + PADDLE_ENFORCE(platform::dynload::cublasHgemm( + context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb, + h_A, lda, &h_beta, h_C, ldc)); +} + template <> void gemm( const platform::CUDADeviceContext& context, const bool transA, @@ -90,6 +147,35 @@ void gemm( lda, &beta, C, ldc)); } +template <> +void matmul( + const platform::CUDADeviceContext& context, + const framework::Tensor& matrix_a, bool trans_a, + const framework::Tensor& matrix_b, bool trans_b, float16 alpha, + framework::Tensor* matrix_out, float16 beta) { + auto dim_a = matrix_a.dims(); + auto dim_b = matrix_b.dims(); + auto dim_out = matrix_out->dims(); + PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2, + "The input and output of matmul be matrix"); + + PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) && + platform::is_gpu_place(matrix_b.place()) && + platform::is_gpu_place(matrix_out->place()), + "Matrix must all be in CUDAPlace"); + + int M = dim_out[0]; + int N = dim_out[1]; + int K = (trans_a == false) ? dim_a[1] : dim_a[0]; + + CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans; + CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans; + + gemm( + context, transA, transB, M, N, K, alpha, matrix_a.data(), + matrix_b.data(), beta, matrix_out->data()); +} + template <> void matmul( const platform::CUDADeviceContext& context, @@ -148,6 +234,37 @@ void matmul( matrix_b.data(), beta, matrix_out->data()); } +template <> +void batched_gemm( + const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, + const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, + const float16 alpha, const float16* A, const float16* B, const float16 beta, + float16* C, const int batchCount, const int strideA, const int strideB) { + // Note that cublas follows fortran order, so the order is different from + // the cblas convention. + int lda = (transA == CblasNoTrans) ? K : M; + int ldb = (transB == CblasNoTrans) ? N : K; + int ldc = N; + cublasOperation_t cuTransA = + (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + cublasOperation_t cuTransB = + (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T; + const int strideC = M * N; + + const half h_alpha = static_cast(alpha); + const half h_beta = static_cast(beta); + const half* h_A = reinterpret_cast(A); + const half* h_B = reinterpret_cast(B); + half* h_C = reinterpret_cast(C); + + // TODO(kexinzhao): add processing code for compute capability < 53 case + PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53, + "cublas Hgemm requires GPU compute capability >= 53"); + PADDLE_ENFORCE(platform::dynload::cublasHgemmStridedBatched( + context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb, + strideB, h_A, lda, strideA, &h_beta, h_C, ldc, strideC, batchCount)); +} + template <> void batched_gemm( const platform::CUDADeviceContext& context, const CBLAS_TRANSPOSE transA, diff --git a/paddle/fluid/operators/math/math_function_test.cu b/paddle/fluid/operators/math/math_function_test.cu index 207d6a87bce178e446253ae2d880b6dd743cfc83..8982d9d066165a9da0461288685baa0c60e5f114 100644 --- a/paddle/fluid/operators/math/math_function_test.cu +++ b/paddle/fluid/operators/math/math_function_test.cu @@ -14,30 +14,41 @@ #include "gtest/gtest.h" #include "paddle/fluid/operators/math/math_function.h" -TEST(math_function, notrans_mul_trans) { - paddle::framework::Tensor input1; - paddle::framework::Tensor input1_gpu; - paddle::framework::Tensor input2_gpu; - paddle::framework::Tensor out_gpu; - paddle::framework::Tensor out; - - auto* cpu_place = new paddle::platform::CPUPlace(); - float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); +void fill_fp16_data(paddle::platform::float16* in_ptr, size_t size, + const std::vector& data) { + PADDLE_ENFORCE_EQ(size, data.size()); + for (size_t i = 0; i < data.size(); ++i) { + in_ptr[i] = paddle::platform::float16(data[i]); + } +} + +TEST(math_function, notrans_mul_trans_fp32) { + using namespace paddle::framework; + using namespace paddle::platform; + + Tensor input1; + Tensor input1_gpu; + Tensor input2_gpu; + Tensor out_gpu; + Tensor out; + + CPUPlace cpu_place; + CUDAPlace gpu_place(0); + CUDADeviceContext context(gpu_place); + + float* input1_ptr = input1.mutable_data({2, 3}, cpu_place); float arr[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr, 6 * sizeof(float)); - auto* gpu_place = new paddle::platform::CUDAPlace(0); - paddle::platform::CUDADeviceContext context(*gpu_place); - - paddle::framework::TensorCopy(input1, *gpu_place, context, &input1_gpu); - paddle::framework::TensorCopy(input1, *gpu_place, context, &input2_gpu); + TensorCopy(input1, gpu_place, context, &input1_gpu); + TensorCopy(input1, gpu_place, context, &input2_gpu); - out_gpu.mutable_data({2, 2}, *gpu_place); + out_gpu.mutable_data({2, 2}, gpu_place); - paddle::operators::math::matmul( + paddle::operators::math::matmul( context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0); - paddle::framework::TensorCopy(out_gpu, *cpu_place, context, &out); + TensorCopy(out_gpu, cpu_place, context, &out); float* out_ptr = out.data(); context.Wait(); @@ -45,33 +56,76 @@ TEST(math_function, notrans_mul_trans) { EXPECT_EQ(out_ptr[1], 14); EXPECT_EQ(out_ptr[2], 14); EXPECT_EQ(out_ptr[3], 50); - delete gpu_place; } -TEST(math_function, trans_mul_notrans) { - paddle::framework::Tensor input1; - paddle::framework::Tensor input1_gpu; - paddle::framework::Tensor input2_gpu; - paddle::framework::Tensor out_gpu; - paddle::framework::Tensor out; +TEST(math_function, notrans_mul_trans_fp16) { + using namespace paddle::framework; + using namespace paddle::platform; + + Tensor input1; + Tensor input1_gpu; + Tensor input2_gpu; + Tensor out_gpu; + Tensor out; + + CPUPlace cpu_place; + CUDAPlace gpu_place(0); + CUDADeviceContext context(gpu_place); + + // fp16 GEMM in cublas requires GPU compute capability >= 53 + if (context.GetComputeCapability() < 53) { + return; + } + + float16* input1_ptr = input1.mutable_data({2, 3}, cpu_place); + fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5}); + + TensorCopy(input1, gpu_place, context, &input1_gpu); + TensorCopy(input1, gpu_place, context, &input2_gpu); + + out_gpu.mutable_data({2, 2}, gpu_place); + + paddle::operators::math::matmul( + context, input1_gpu, false, input2_gpu, true, float16(1), &out_gpu, + float16(0)); + + TensorCopy(out_gpu, cpu_place, context, &out); + + float16* out_ptr = out.data(); + context.Wait(); + EXPECT_EQ(static_cast(out_ptr[0]), 5); + EXPECT_EQ(static_cast(out_ptr[1]), 14); + EXPECT_EQ(static_cast(out_ptr[2]), 14); + EXPECT_EQ(static_cast(out_ptr[3]), 50); +} + +TEST(math_function, trans_mul_notrans_fp32) { + using namespace paddle::framework; + using namespace paddle::platform; + + Tensor input1; + Tensor input1_gpu; + Tensor input2_gpu; + Tensor out_gpu; + Tensor out; + + CPUPlace cpu_place; + CUDAPlace gpu_place(0); + CUDADeviceContext context(gpu_place); - auto* cpu_place = new paddle::platform::CPUPlace(); - float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); + float* input1_ptr = input1.mutable_data({2, 3}, cpu_place); float arr[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr, 6 * sizeof(float)); - auto* gpu_place = new paddle::platform::CUDAPlace(0); - paddle::platform::CUDADeviceContext context(*gpu_place); + TensorCopy(input1, gpu_place, context, &input1_gpu); + TensorCopy(input1, gpu_place, context, &input2_gpu); - paddle::framework::TensorCopy(input1, *gpu_place, context, &input1_gpu); - paddle::framework::TensorCopy(input1, *gpu_place, context, &input2_gpu); - - out_gpu.mutable_data({3, 3}, *gpu_place); + out_gpu.mutable_data({3, 3}, gpu_place); paddle::operators::math::matmul( context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0); - paddle::framework::TensorCopy(out_gpu, *cpu_place, context, &out); + TensorCopy(out_gpu, cpu_place, context, &out); float* out_ptr = out.data(); context.Wait(); @@ -84,45 +138,93 @@ TEST(math_function, trans_mul_notrans) { EXPECT_EQ(out_ptr[6], 15); EXPECT_EQ(out_ptr[7], 22); EXPECT_EQ(out_ptr[8], 29); - delete gpu_place; } -TEST(math_function, gemm_notrans_cublas) { - paddle::framework::Tensor input1; - paddle::framework::Tensor input2; - paddle::framework::Tensor input3; - paddle::framework::Tensor input1_gpu; - paddle::framework::Tensor input2_gpu; - paddle::framework::Tensor input3_gpu; +TEST(math_function, trans_mul_notrans_fp16) { + using namespace paddle::framework; + using namespace paddle::platform; + + Tensor input1; + Tensor input1_gpu; + Tensor input2_gpu; + Tensor out_gpu; + Tensor out; + + CPUPlace cpu_place; + CUDAPlace gpu_place(0); + CUDADeviceContext context(gpu_place); + + // fp16 GEMM in cublas requires GPU compute capability >= 53 + if (context.GetComputeCapability() < 53) { + return; + } + + float16* input1_ptr = input1.mutable_data({2, 3}, cpu_place); + fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5}); + + TensorCopy(input1, gpu_place, context, &input1_gpu); + TensorCopy(input1, gpu_place, context, &input2_gpu); + + out_gpu.mutable_data({3, 3}, gpu_place); + + paddle::operators::math::matmul( + context, input1_gpu, true, input2_gpu, false, float16(1), &out_gpu, + float16(0)); + + TensorCopy(out_gpu, cpu_place, context, &out); + + float16* out_ptr = out.data(); + context.Wait(); + EXPECT_EQ(static_cast(out_ptr[0]), 9); + EXPECT_EQ(static_cast(out_ptr[1]), 12); + EXPECT_EQ(static_cast(out_ptr[2]), 15); + EXPECT_EQ(static_cast(out_ptr[3]), 12); + EXPECT_EQ(static_cast(out_ptr[4]), 17); + EXPECT_EQ(static_cast(out_ptr[5]), 22); + EXPECT_EQ(static_cast(out_ptr[6]), 15); + EXPECT_EQ(static_cast(out_ptr[7]), 22); + EXPECT_EQ(static_cast(out_ptr[8]), 29); +} + +TEST(math_function, gemm_notrans_cublas_fp32) { + using namespace paddle::framework; + using namespace paddle::platform; + + Tensor input1; + Tensor input2; + Tensor input3; + Tensor input1_gpu; + Tensor input2_gpu; + Tensor input3_gpu; + + CPUPlace cpu_place; + CUDAPlace gpu_place(0); + CUDADeviceContext context(gpu_place); int m = 2; int n = 3; int k = 3; - auto* cpu_place = new paddle::platform::CPUPlace(); - float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); + float* input1_ptr = input1.mutable_data({2, 3}, cpu_place); float arr1[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr1, 6 * sizeof(float)); - float* input2_ptr = input2.mutable_data({3, 4}, *cpu_place); + float* input2_ptr = input2.mutable_data({3, 4}, cpu_place); float arr2[12] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}; memcpy(input2_ptr, arr2, 12 * sizeof(float)); - float* input3_ptr = input3.mutable_data({2, 4}, *cpu_place); + float* input3_ptr = input3.mutable_data({2, 4}, cpu_place); float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; memcpy(input3_ptr, arr3, 8 * sizeof(float)); - auto* gpu_place = new paddle::platform::CUDAPlace(0); - paddle::platform::CUDADeviceContext context(*gpu_place); - - paddle::framework::TensorCopy(input1, *gpu_place, context, &input1_gpu); - paddle::framework::TensorCopy(input2, *gpu_place, context, &input2_gpu); - paddle::framework::TensorCopy(input3, *gpu_place, context, &input3_gpu); + TensorCopy(input1, gpu_place, context, &input1_gpu); + TensorCopy(input2, gpu_place, context, &input2_gpu); + TensorCopy(input3, gpu_place, context, &input3_gpu); float* a = input1_gpu.data(); float* b = input2_gpu.data(); - float* c = input3_gpu.mutable_data(*gpu_place); + float* c = input3_gpu.mutable_data(gpu_place); paddle::operators::math::gemm( context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4); - paddle::framework::TensorCopy(input3_gpu, *cpu_place, context, &input3); + TensorCopy(input3_gpu, cpu_place, context, &input3); // numpy code: // a = np.arange(6).reshape(2, 3) @@ -139,47 +241,110 @@ TEST(math_function, gemm_notrans_cublas) { EXPECT_EQ(input3_ptr[5], 73); EXPECT_EQ(input3_ptr[6], 86); EXPECT_EQ(input3_ptr[7], 99); - delete gpu_place; } -TEST(math_function, gemm_trans_cublas) { - paddle::framework::Tensor input1; - paddle::framework::Tensor input2; - paddle::framework::Tensor input3; - paddle::framework::Tensor input1_gpu; - paddle::framework::Tensor input2_gpu; - paddle::framework::Tensor input3_gpu; +TEST(math_function, gemm_notrans_cublas_fp16) { + using namespace paddle::framework; + using namespace paddle::platform; + + Tensor input1; + Tensor input2; + Tensor input3; + Tensor input1_gpu; + Tensor input2_gpu; + Tensor input3_gpu; + + CPUPlace cpu_place; + CUDAPlace gpu_place(0); + CUDADeviceContext context(gpu_place); + + // fp16 GEMM in cublas requires GPU compute capability >= 53 + if (context.GetComputeCapability() < 53) { + return; + } + + int m = 2; + int n = 3; + int k = 3; + float16* input1_ptr = input1.mutable_data({2, 3}, cpu_place); + fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5}); + float16* input2_ptr = input2.mutable_data({3, 4}, cpu_place); + fill_fp16_data(input2_ptr, input2.numel(), + {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11}); + float16* input3_ptr = input3.mutable_data({2, 4}, cpu_place); + fill_fp16_data(input3_ptr, input3.numel(), {0, 1, 2, 3, 4, 5, 6, 7}); + + TensorCopy(input1, gpu_place, context, &input1_gpu); + TensorCopy(input2, gpu_place, context, &input2_gpu); + TensorCopy(input3, gpu_place, context, &input3_gpu); + float16* a = input1_gpu.data(); + float16* b = input2_gpu.data(); + float16* c = input3_gpu.mutable_data(gpu_place); + + paddle::operators::math::gemm( + context, false, false, m, n, k, float16(1), a, 3, b + 1, 4, float16(1), + c + 1, 4); + + TensorCopy(input3_gpu, cpu_place, context, &input3); + + // numpy code: + // a = np.arange(6).reshape(2, 3) + // b = np.arange(12).reshape(3, 4)[:, 1:] + // c = np.arange(8).reshape(2, 4)[:, 1:] + // out = np.arange(8).reshape(2, 4) + // out[:, 1:] = np.dot(a, b) + c + context.Wait(); + EXPECT_EQ(static_cast(input3_ptr[0]), 0); + EXPECT_EQ(static_cast(input3_ptr[1]), 24); + EXPECT_EQ(static_cast(input3_ptr[2]), 28); + EXPECT_EQ(static_cast(input3_ptr[3]), 32); + EXPECT_EQ(static_cast(input3_ptr[4]), 4); + EXPECT_EQ(static_cast(input3_ptr[5]), 73); + EXPECT_EQ(static_cast(input3_ptr[6]), 86); + EXPECT_EQ(static_cast(input3_ptr[7]), 99); +} + +TEST(math_function, gemm_trans_cublas_fp32) { + using namespace paddle::framework; + using namespace paddle::platform; + + Tensor input1; + Tensor input2; + Tensor input3; + Tensor input1_gpu; + Tensor input2_gpu; + Tensor input3_gpu; + + CPUPlace cpu_place; + CUDAPlace gpu_place(0); + CUDADeviceContext context(gpu_place); int m = 2; int n = 3; int k = 3; - auto* cpu_place = new paddle::platform::CPUPlace(); - float* input1_ptr = input1.mutable_data({2, 3}, *cpu_place); + float* input1_ptr = input1.mutable_data({2, 3}, cpu_place); float arr1[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr1, 6 * sizeof(float)); - float* input2_ptr = input2.mutable_data({4, 3}, *cpu_place); + float* input2_ptr = input2.mutable_data({4, 3}, cpu_place); float arr2[12] = {0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11}; memcpy(input2_ptr, arr2, 12 * sizeof(float)); - float* input3_ptr = input3.mutable_data({2, 4}, *cpu_place); + float* input3_ptr = input3.mutable_data({2, 4}, cpu_place); float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; memcpy(input3_ptr, arr3, 8 * sizeof(float)); - auto* gpu_place = new paddle::platform::CUDAPlace(0); - paddle::platform::CUDADeviceContext context(*gpu_place); - - paddle::framework::TensorCopy(input1, *gpu_place, context, &input1_gpu); - paddle::framework::TensorCopy(input2, *gpu_place, context, &input2_gpu); - paddle::framework::TensorCopy(input3, *gpu_place, context, &input3_gpu); + TensorCopy(input1, gpu_place, context, &input1_gpu); + TensorCopy(input2, gpu_place, context, &input2_gpu); + TensorCopy(input3, gpu_place, context, &input3_gpu); float* a = input1_gpu.data(); float* b = input2_gpu.data(); - float* c = input3_gpu.mutable_data(*gpu_place); + float* c = input3_gpu.mutable_data(gpu_place); paddle::operators::math::gemm( context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4); - paddle::framework::TensorCopy(input3_gpu, *cpu_place, context, &input3); - context.Wait(); + TensorCopy(input3_gpu, cpu_place, context, &input3); + context.Wait(); EXPECT_EQ(input3_ptr[0], 0); EXPECT_EQ(input3_ptr[1], 24); EXPECT_EQ(input3_ptr[2], 28); @@ -188,27 +353,86 @@ TEST(math_function, gemm_trans_cublas) { EXPECT_EQ(input3_ptr[5], 73); EXPECT_EQ(input3_ptr[6], 86); EXPECT_EQ(input3_ptr[7], 99); - delete gpu_place; +} + +TEST(math_function, gemm_trans_cublas_fp16) { + using namespace paddle::framework; + using namespace paddle::platform; + + Tensor input1; + Tensor input2; + Tensor input3; + Tensor input1_gpu; + Tensor input2_gpu; + Tensor input3_gpu; + + CPUPlace cpu_place; + CUDAPlace gpu_place(0); + CUDADeviceContext context(gpu_place); + + // fp16 GEMM in cublas requires GPU compute capability >= 53 + if (context.GetComputeCapability() < 53) { + return; + } + + int m = 2; + int n = 3; + int k = 3; + float16* input1_ptr = input1.mutable_data({2, 3}, cpu_place); + fill_fp16_data(input1_ptr, input1.numel(), {0, 1, 2, 3, 4, 5}); + float16* input2_ptr = input2.mutable_data({4, 3}, cpu_place); + fill_fp16_data(input2_ptr, input2.numel(), + {0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11}); + float16* input3_ptr = input3.mutable_data({2, 4}, cpu_place); + fill_fp16_data(input3_ptr, input3.numel(), {0, 1, 2, 3, 4, 5, 6, 7}); + + TensorCopy(input1, gpu_place, context, &input1_gpu); + TensorCopy(input2, gpu_place, context, &input2_gpu); + TensorCopy(input3, gpu_place, context, &input3_gpu); + float16* a = input1_gpu.data(); + float16* b = input2_gpu.data(); + float16* c = input3_gpu.mutable_data(gpu_place); + + paddle::operators::math::gemm( + context, false, true, m, n, k, float16(1), a, 3, b + 3, 3, float16(1), + c + 1, 4); + + TensorCopy(input3_gpu, cpu_place, context, &input3); + + context.Wait(); + EXPECT_EQ(static_cast(input3_ptr[0]), 0); + EXPECT_EQ(static_cast(input3_ptr[1]), 24); + EXPECT_EQ(static_cast(input3_ptr[2]), 28); + EXPECT_EQ(static_cast(input3_ptr[3]), 32); + EXPECT_EQ(static_cast(input3_ptr[4]), 4); + EXPECT_EQ(static_cast(input3_ptr[5]), 73); + EXPECT_EQ(static_cast(input3_ptr[6]), 86); + EXPECT_EQ(static_cast(input3_ptr[7]), 99); } template void GemvTest(int m, int n, bool trans) { - paddle::framework::Tensor mat_a; - paddle::framework::Tensor vec_b; - paddle::framework::Tensor vec_c; - auto* cpu_place = new paddle::platform::CPUPlace(); - - T* data_a = mat_a.mutable_data({m, n}, *cpu_place); - T* data_b = vec_b.mutable_data({trans ? m : n}, *cpu_place); - T* data_c = vec_c.mutable_data({trans ? n : m}, *cpu_place); - - auto* gpu_place = new paddle::platform::CUDAPlace(0); - paddle::framework::Tensor g_mat_a; - paddle::framework::Tensor g_vec_b; - paddle::framework::Tensor g_vec_c; - T* g_data_a = g_mat_a.mutable_data(mat_a.dims(), *gpu_place); - T* g_data_b = g_vec_b.mutable_data(vec_b.dims(), *gpu_place); - T* g_data_c = g_vec_c.mutable_data(vec_c.dims(), *gpu_place); + using namespace paddle::framework; + using namespace paddle::platform; + + Tensor mat_a; + Tensor vec_b; + Tensor vec_c; + + CPUPlace cpu_place; + CUDAPlace gpu_place(0); + CUDADeviceContext context(gpu_place); + + T* data_a = mat_a.mutable_data({m, n}, cpu_place); + T* data_b = vec_b.mutable_data({trans ? m : n}, cpu_place); + T* data_c = vec_c.mutable_data({trans ? n : m}, cpu_place); + + Tensor g_mat_a; + Tensor g_vec_b; + Tensor g_vec_c; + T* g_data_a = g_mat_a.mutable_data(mat_a.dims(), gpu_place); + T* g_data_b = g_vec_b.mutable_data(vec_b.dims(), gpu_place); + T* g_data_c = g_vec_c.mutable_data(vec_c.dims(), gpu_place); for (int i = 0; i < mat_a.numel(); ++i) { data_a[i] = static_cast(i); @@ -217,16 +441,14 @@ void GemvTest(int m, int n, bool trans) { data_b[i] = static_cast(i); } - paddle::platform::CUDADeviceContext context(*gpu_place); - paddle::framework::TensorCopy(mat_a, *gpu_place, context, &g_mat_a); - paddle::framework::TensorCopy(vec_b, *gpu_place, context, &g_vec_b); + TensorCopy(mat_a, gpu_place, context, &g_mat_a); + TensorCopy(vec_b, gpu_place, context, &g_vec_b); - paddle::operators::math::gemv( + paddle::operators::math::gemv( context, trans, static_cast(m), static_cast(n), 1., g_data_a, g_data_b, 0., g_data_c); - paddle::framework::TensorCopy(g_vec_c, paddle::platform::CPUPlace(), context, - &vec_c); + TensorCopy(g_vec_c, cpu_place, context, &vec_c); if (!trans) { for (int i = 0; i < m; ++i) { diff --git a/paddle/fluid/operators/math/softmax.cu b/paddle/fluid/operators/math/softmax.cu index 38e93fdf15d99eb447948378a599891074c10fc5..34ea6a91ce7743462d378cf471a5ec3a12ca51d1 100644 --- a/paddle/fluid/operators/math/softmax.cu +++ b/paddle/fluid/operators/math/softmax.cu @@ -14,13 +14,86 @@ limitations under the License. */ #define EIGEN_USE_GPU +#include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/softmax.h" #include "paddle/fluid/operators/math/softmax_impl.h" +#include "paddle/fluid/platform/cudnn_helper.h" namespace paddle { namespace operators { namespace math { +using Tensor = framework::Tensor; +using ScopedTensorDescriptor = platform::ScopedTensorDescriptor; +using DataLayout = platform::DataLayout; +template +using CudnnDataType = platform::CudnnDataType; + +template +void SoftmaxCUDNNFunctor::operator()( + const platform::CUDADeviceContext& context, const framework::Tensor* X, + framework::Tensor* Y) { + // ------------------- cudnn descriptors --------------------- + ScopedTensorDescriptor xDesc; + ScopedTensorDescriptor yDesc; + std::vector cudnn_tensor_dims = framework::vectorize2int(X->dims()); + DataLayout layout = DataLayout::kNCHW; + if (cudnn_tensor_dims.size() == 5) { + layout = DataLayout::kNCDHW; + } + // NOTE(*) : cudnn softmax only support >= 4D Tensor, + // fill 1 at unused dims + if (cudnn_tensor_dims.size() <= 2) { + cudnn_tensor_dims.resize(4, 1); + } + cudnnTensorDescriptor_t cudnn_x_desc = + xDesc.descriptor(layout, cudnn_tensor_dims); + cudnnTensorDescriptor_t cudnn_y_desc = + xDesc.descriptor(layout, cudnn_tensor_dims); + PADDLE_ENFORCE(platform::dynload::cudnnSoftmaxForward( + context.cudnn_handle(), CUDNN_SOFTMAX_ACCURATE, + CUDNN_SOFTMAX_MODE_INSTANCE, CudnnDataType::kOne(), cudnn_x_desc, + X->data(), CudnnDataType::kZero(), cudnn_y_desc, + Y->mutable_data(context.GetPlace()))); +} + +template +void SoftmaxGradCUDNNFunctor::operator()( + const platform::CUDADeviceContext& context, const framework::Tensor* Y, + const framework::Tensor* YGrad, framework::Tensor* XGrad) { + // ------------------- cudnn descriptors --------------------- + ScopedTensorDescriptor yDesc; + ScopedTensorDescriptor dyDesc; + ScopedTensorDescriptor dxDesc; + std::vector cudnn_tensor_dims = framework::vectorize2int(Y->dims()); + DataLayout layout = DataLayout::kNCHW; + if (cudnn_tensor_dims.size() == 5) { + layout = DataLayout::kNCDHW; + } + // NOTE(*) : cudnn softmax only support >= 4D Tensor, + // fill 1 at unused dims + if (cudnn_tensor_dims.size() <= 2) { + cudnn_tensor_dims.resize(4, 1); + } + cudnnTensorDescriptor_t cudnn_y_desc = + yDesc.descriptor(layout, cudnn_tensor_dims); + cudnnTensorDescriptor_t cudnn_xgrad_desc = + dxDesc.descriptor(layout, cudnn_tensor_dims); + cudnnTensorDescriptor_t cudnn_ygrad_desc = + dyDesc.descriptor(layout, cudnn_tensor_dims); + PADDLE_ENFORCE(platform::dynload::cudnnSoftmaxBackward( + context.cudnn_handle(), CUDNN_SOFTMAX_ACCURATE, + CUDNN_SOFTMAX_MODE_INSTANCE, CudnnDataType::kOne(), cudnn_y_desc, + Y->data(), cudnn_ygrad_desc, YGrad->data(), + CudnnDataType::kZero(), cudnn_xgrad_desc, + XGrad->mutable_data(context.GetPlace()))); +} + +template class SoftmaxCUDNNFunctor; +template class SoftmaxCUDNNFunctor; +template class SoftmaxGradCUDNNFunctor; +template class SoftmaxGradCUDNNFunctor; + template class SoftmaxFunctor; template class SoftmaxFunctor; template class SoftmaxGradFunctor; diff --git a/paddle/fluid/operators/math/softmax.h b/paddle/fluid/operators/math/softmax.h index 14b2690c2a4e764058270953214a07aee8053444..da1f0b672d3a5fb5da8f4d72892be21964bdbc0d 100644 --- a/paddle/fluid/operators/math/softmax.h +++ b/paddle/fluid/operators/math/softmax.h @@ -33,6 +33,23 @@ class SoftmaxGradFunctor { const framework::Tensor* y_grad, framework::Tensor* x_grad); }; +#ifdef PADDLE_WITH_CUDA +template +class SoftmaxCUDNNFunctor { + public: + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor* X, framework::Tensor* Y); +}; + +template +class SoftmaxGradCUDNNFunctor { + public: + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor* Y, const framework::Tensor* y_grad, + framework::Tensor* x_grad); +}; +#endif + } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/mul_op.cc b/paddle/fluid/operators/mul_op.cc index e7bed2c39735b66c19e738c91f4977e46571143b..90af1e2d602ac039b4d98a69a889ff8b1b85ffc6 100644 --- a/paddle/fluid/operators/mul_op.cc +++ b/paddle/fluid/operators/mul_op.cc @@ -17,11 +17,14 @@ limitations under the License. */ namespace paddle { namespace operators { +using framework::OpKernelType; using framework::Tensor; -class MulOpShapeInference : public framework::InferShapeBase { +class MulOp : public framework::OperatorWithKernel { public: - void operator()(framework::InferShapeContext* ctx) const override { + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of MulOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) of MulOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), @@ -122,7 +125,7 @@ or not. But the output only shares the LoD information with input $X$. } }; -class MulOpGrad : public framework::OperatorWithKernel { +class MulGradOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -156,10 +159,7 @@ class MulOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(mul, paddle::framework::OperatorWithKernel, ops::MulOpMaker, - ops::MulOpShapeInference, - paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(mul_grad, ops::MulOpGrad); +REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulGradOp); REGISTER_OP_CPU_KERNEL( mul, ops::MulKernel); REGISTER_OP_CPU_KERNEL( diff --git a/paddle/fluid/operators/mul_op.cu.cc b/paddle/fluid/operators/mul_op.cu.cc index 0667530e943856576ae8c9fe4856cb6aa1448e4e..757f9c3ee2665c7ac654659416fe8dd727dca16d 100644 --- a/paddle/fluid/operators/mul_op.cu.cc +++ b/paddle/fluid/operators/mul_op.cu.cc @@ -13,9 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/mul_op.h" +#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL( - mul, ops::MulKernel); -REGISTER_OP_CUDA_KERNEL( - mul_grad, ops::MulGradKernel); +namespace plat = paddle::platform; +REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel, + ops::MulKernel); +REGISTER_OP_CUDA_KERNEL(mul_grad, + ops::MulGradKernel); diff --git a/paddle/fluid/operators/mul_op.h b/paddle/fluid/operators/mul_op.h index 38311cf87265ad0f1f815734cbf69bd682d62e62..b1260d36ebe11f65529ac274c959479dcb38ee5f 100644 --- a/paddle/fluid/operators/mul_op.h +++ b/paddle/fluid/operators/mul_op.h @@ -48,7 +48,7 @@ class MulKernel : public framework::OpKernel { } math::matmul( context.template device_context(), x_matrix, false, - y_matrix, false, 1, z, 0); + y_matrix, false, static_cast(1), z, static_cast(0)); if (z_dim.size() != 2) { z->Resize(z_dim); } diff --git a/paddle/fluid/operators/nccl_op.cc b/paddle/fluid/operators/nccl_op.cc index 9185666c56c4621d42429c9cfdb079001c6336f1..5e4ed886b10bd48bf991ce84a9099611cf5d1d26 100644 --- a/paddle/fluid/operators/nccl_op.cc +++ b/paddle/fluid/operators/nccl_op.cc @@ -14,7 +14,6 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/nccl/nccl_gpu_common.h" -#include "paddle/fluid/operators/nccl/nccl_gpu_common.h" namespace paddle { namespace operators { @@ -105,19 +104,38 @@ class NCCLAllReduceOp : public framework::OperatorWithKernel { " Input(Communicator) of AllReduce op input should not be NULL"); PADDLE_ENFORCE(ctx->HasOutput("Out"), " Output(Out) of AllReduce op output should not be NULL"); - - auto x_dims = ctx->GetInputsDim("X"); - std::string reduction = ctx->Attrs().Get("reduction"); PADDLE_ENFORCE((reduction == "ncclSum" || reduction == "ncclProd" || reduction == "ncclMin" || reduction == "ncclMax"), "invalid reduction."); + auto x_dims = ctx->GetInputsDim("X"); ctx->SetOutputsDim("Out", x_dims); ctx->ShareLoD("X", /*->*/ "Out"); } }; +// AllReduceOp +class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker { + public: + NCCLAllReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "The input of AllReduce op"); + AddInput("Communicator", "Communicator for communicating between gpus"); + AddOutput("Out", "The output of AllReduce op"); + AddAttr("reduction", + "(string, default 'ncclSum') " + "{'ncclMin', 'ncclMax', 'ncclProd', 'ncclSum'}.") + .SetDefault("ncclSum"); + AddComment(R"DOC( +NCCLAllReduce Operator. + +AllReduce the input tensors. + +)DOC"); + } +}; + // ReduceOp class NCCLReduceOp : public framework::OperatorWithKernel { public: @@ -144,50 +162,6 @@ class NCCLReduceOp : public framework::OperatorWithKernel { } }; -// BcastOp -class NCCLBcastOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), - " Input(X) of Bcast op input should not be NULL"); - PADDLE_ENFORCE(ctx->HasInput("Communicator"), - " Input(Communicator) of Bcast op input should not be NULL"); - PADDLE_ENFORCE(ctx->HasOutput("Out"), - " Output(Out) of Bcast op output should not be NULL"); - - int root = ctx->Attrs().Get("root"); - PADDLE_ENFORCE(root != platform::kInvalidGPUId, "Bcast root must be set."); - - auto x_dims = ctx->GetInputsDim("X"); - ctx->SetOutputsDim("Out", x_dims); - ctx->ShareLoD("X", /*->*/ "Out"); - } -}; - -// AllreduceOp -class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker { - public: - NCCLAllReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker) - : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The input of AllReduce op"); - AddInput("Communicator", "Communicator for communicating between gpus"); - AddOutput("Out", "The output of AllReduce op"); - AddAttr("reduction", - "(string, default 'ncclSum') " - "{'ncclMin', 'ncclMax', 'ncclProd', 'ncclSum'}.") - .SetDefault("ncclSum"); - AddComment(R"DOC( -NCCLAllReduce Operator. - -AllReduce the input tensors. - -)DOC"); - } -}; - // ReduceOp class NCCLReduceOpMaker : public framework::OpProtoAndCheckerMaker { public: @@ -214,6 +188,29 @@ Reduce the tensors. } }; +// BcastOp +class NCCLBcastOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + " Input(X) of Bcast op input should not be NULL"); + PADDLE_ENFORCE(ctx->HasInput("Communicator"), + " Input(Communicator) of Bcast op input should not be NULL"); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + " Output(Out) of Bcast op output should not be NULL"); + + int root = ctx->Attrs().Get("root"); + PADDLE_ENFORCE(root != platform::kInvalidGPUId, "Bcast root must be set."); + + auto x_dims = ctx->GetInputsDim("X"); + ctx->SetOutputsDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + // BcastOp class NCCLBcastOpMaker : public framework::OpProtoAndCheckerMaker { public: diff --git a/paddle/fluid/operators/nccl_op.cu.cc b/paddle/fluid/operators/nccl_op.cu.cc index 683a520e99fe72875d52393a18463324a8b6c3f2..ad623e1fe0f8941615b671a0c20bd3637ae6d407 100644 --- a/paddle/fluid/operators/nccl_op.cu.cc +++ b/paddle/fluid/operators/nccl_op.cu.cc @@ -43,13 +43,12 @@ class NCCLAllReduceKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), "This kernel only runs on GPU device."); - - auto ins = ctx.MultiInput("X"); - auto outs = ctx.MultiOutput("Out"); - + auto* x = ctx.Input("X"); + auto* out = ctx.Output("Out"); + auto* comm = ctx.Input("Communicator"); std::string reduction = ctx.Attr("reduction"); - ncclRedOp_t reduction_op_ = ncclSum; + ncclRedOp_t reduction_op_ = ncclSum; if (reduction == "ncclMin") { reduction_op_ = ncclMin; } else if (reduction == "ncclMax") { @@ -61,30 +60,19 @@ class NCCLAllReduceKernel : public framework::OpKernel { } else { PADDLE_THROW("Invalid reduction. default ncclSum."); } - - auto* comm = ctx.Input("Communicator"); - - auto stream = ctx.cuda_device_context().stream(); - // device id int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); int idx = comm->GetCommId(gpu_id); - - for (size_t i = 0; i < ins.size(); ++i) { - VLOG(1) << "gpu : " - << " invoke allreduce. send " << ins[i]->numel() << " recv " - << outs[i]->numel(); - - PADDLE_ENFORCE(platform::dynload::ncclAllReduce( - ins[i]->data(), outs[i]->mutable_data(ctx.GetPlace()), - outs[i]->numel(), NCCLTypeWrapper::type, reduction_op_, - comm->comms().at(idx), stream)); - PADDLE_ENFORCE(cudaStreamSynchronize(stream)); - - VLOG(1) << "gpu : " - << " finished allreduce. send " << ins[i]->numel() << " recv " - << outs[i]->numel(); - } + VLOG(3) << "gpu : " + << " invoke allreduce. send " << x->numel() << " recv " + << out->numel(); + PADDLE_ENFORCE(platform::dynload::ncclAllReduce( + x->data(), out->mutable_data(ctx.GetPlace()), out->numel(), + NCCLTypeWrapper::type, reduction_op_, comm->comms().at(idx), + ctx.cuda_device_context().stream())); + VLOG(3) << "gpu : " + << " finished allreduce. send " << x->numel() << " recv " + << out->numel(); } }; @@ -94,13 +82,13 @@ class NCCLReduceKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), "This kernel only runs on GPU device."); - - auto ins = ctx.MultiInput("X"); // x0, x1, x2 - auto outs = ctx.MultiOutput("Out"); - + auto x = ctx.Input("X"); // x0, x1, x2 + auto out = ctx.Output("Out"); + auto* comm = ctx.Input("Communicator"); + int root = ctx.Attr("root"); std::string reduction = ctx.Attr("reduction"); - ncclRedOp_t reduction_op_ = ncclSum; + ncclRedOp_t reduction_op_ = ncclSum; if (reduction == "ncclMin") { reduction_op_ = ncclMin; } else if (reduction == "ncclMax") { @@ -112,40 +100,23 @@ class NCCLReduceKernel : public framework::OpKernel { } else { PADDLE_THROW("Invalid reduction. default ncclSum."); } - - int root = ctx.Attr("root"); - auto* comm = ctx.Input("Communicator"); - - auto stream = reinterpret_cast( - ctx.device_context()) - .stream(); // device id int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); int idx = comm->GetCommId(gpu_id); - - auto ins_names = ctx.Inputs("X"); - std::hash hasher; - for (size_t i = 0; i < ins.size(); ++i) { - if (root == platform::kInvalidGPUId) { - root = hasher(ins_names[i]) % comm->comms().size(); - } - T* recvbuffer = nullptr; - if (root == gpu_id) { - recvbuffer = outs[i]->mutable_data(ctx.GetPlace()); - } - - VLOG(1) << "gpu : " << gpu_id << " invoke reduce. send " - << ins[i]->numel() << " recv " << outs[i]->numel(); - - PADDLE_ENFORCE(platform::dynload::ncclReduce( - ins[i]->data(), recvbuffer, ins[i]->numel(), - NCCLTypeWrapper::type, reduction_op_, root, comm->comms().at(idx), - stream)); - PADDLE_ENFORCE(cudaStreamSynchronize(stream)); - - VLOG(1) << "gpu : " << gpu_id << " finished reduce. send " - << ins[i]->numel() << " recv " << outs[i]->numel(); + T* recvbuffer = nullptr; + if (root == gpu_id) { + recvbuffer = out->mutable_data(ctx.GetPlace()); + } else { + out->Resize(framework::make_ddim({0})); } + VLOG(3) << "gpu : " << gpu_id << " invoke reduce. send " << x->numel() + << " recv " << out->numel(); + PADDLE_ENFORCE(platform::dynload::ncclReduce( + x->data(), recvbuffer, x->numel(), NCCLTypeWrapper::type, + reduction_op_, root, comm->comms().at(idx), + ctx.cuda_device_context().stream())); + VLOG(3) << "gpu : " << gpu_id << " finished reduce. send " << x->numel() + << " recv " << out->numel(); } }; @@ -155,47 +126,27 @@ class NCCLBcastKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), "This kernel only runs on GPU device."); - int root = ctx.Attr("root"); - auto* comm = ctx.Input("Communicator"); - - auto stream = reinterpret_cast( - ctx.device_context()) - .stream(); // device id int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); int idx = comm->GetCommId(gpu_id); - if (idx == root) { - auto ins = ctx.MultiInput("X"); - for (size_t i = 0; i < ins.size(); ++i) { - VLOG(1) << "gpu : " << gpu_id << " invoke Bcast. send " - << ins[i]->numel(); - - VLOG(1) << " before ncclBcast"; - PADDLE_ENFORCE(platform::dynload::ncclBcast( - (void*)ins[i]->data(), ins[i]->numel(), NCCLTypeWrapper::type, - root, comm->comms().at(idx), stream)); - VLOG(1) << " after ncclBcast"; - PADDLE_ENFORCE(cudaStreamSynchronize(stream)); - - VLOG(1) << "gpu : " << gpu_id << " finished Bcast."; - } + auto* x = ctx.Input("X"); + VLOG(3) << "gpu : " << gpu_id << " invoke Bcast. send " << x->numel(); + PADDLE_ENFORCE(platform::dynload::ncclBcast( + (void*)x->data(), x->numel(), NCCLTypeWrapper::type, root, + comm->comms().at(idx), ctx.cuda_device_context().stream())); + VLOG(3) << "gpu : " << gpu_id << " finished Bcast."; } else { - auto outs = ctx.MultiOutput("Out"); - for (size_t i = 0; i < outs.size(); ++i) { - VLOG(1) << "gpu : " << gpu_id << " invoke Bcast. recv buffer " - << framework::product(outs[i]->dims()); - - PADDLE_ENFORCE(platform::dynload::ncclBcast( - outs[i]->mutable_data(ctx.GetPlace()), outs[i]->numel(), - NCCLTypeWrapper::type, root, comm->comms().at(idx), stream)); - PADDLE_ENFORCE(cudaStreamSynchronize(stream)); - - VLOG(1) << "gpu : " << gpu_id << " finished Bcast. recv " - << outs[i]->numel(); - } + auto* out = ctx.Output("Out"); + VLOG(3) << "gpu : " << gpu_id << " invoke Bcast. recv buffer " + << framework::product(out->dims()); + PADDLE_ENFORCE(platform::dynload::ncclBcast( + out->mutable_data(ctx.GetPlace()), out->numel(), + NCCLTypeWrapper::type, root, comm->comms().at(idx), + ctx.cuda_device_context().stream())); + VLOG(3) << "gpu : " << gpu_id << " finished Bcast. recv " << out->numel(); } } }; diff --git a/paddle/fluid/operators/nccl_op_test.cu.cc b/paddle/fluid/operators/nccl_op_test.cu.cc index b4021a5dacd80a042042eadf9f5d1a932f0f00a2..90f6f955cea51ded2dbb2bde459113458d7749a4 100644 --- a/paddle/fluid/operators/nccl_op_test.cu.cc +++ b/paddle/fluid/operators/nccl_op_test.cu.cc @@ -14,19 +14,15 @@ limitations under the License. */ #include #include -#include #include #include #include -#include #include -#include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/op_desc.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/framework/var_desc.h" #include "paddle/fluid/operators/nccl/nccl_gpu_common.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/enforce.h" @@ -41,26 +37,35 @@ USE_CUDA_ONLY_OP(ncclBcast); namespace f = paddle::framework; namespace p = paddle::platform; -static std::vector gpu_list; - // test data amount -const f::DDim kDims = {100, 100}; +const f::DDim kDims = {20, 20}; // nccl op common tester, init communicator. class NCCLTester : public ::testing::Test { public: virtual void SetUp() override { + int count = p::GetCUDADeviceCount(); + if (count <= 1) { + LOG(WARNING) + << "Cannot test multi-gpu nccl, because the CUDA device count is " + << count; + exit(0); + } + for (int i = 0; i < count; ++i) { + gpu_list_.emplace_back(i); + } + paddle::platform::CPUPlace cpu_place; - for (size_t i = 0; i < gpu_list.size(); ++i) { + for (size_t i = 0; i < gpu_list_.size(); ++i) { p::CUDAPlace place(i); - dev_ctxs.emplace_back(new p::CUDADeviceContext(place)); + dev_ctxs_.emplace_back(new p::CUDADeviceContext(place)); } NCCLInitOp(); } virtual void TearDown() override { - for (auto &device_context : dev_ctxs) { + for (auto &device_context : dev_ctxs_) { delete device_context; } } @@ -70,36 +75,40 @@ class NCCLTester : public ::testing::Test { std::unique_ptr op1(new f::OpDesc); op1->SetType("ncclInit"); + op1->SetInput("parallel_scopes", {"p_scopes"}); op1->SetOutput("Communicator", {"comm"}); - op1->SetAttr("gpus", {gpu_list}); - auto *var = g_scope.Var("comm"); + auto *var = g_scope_.Var("comm"); var->GetMutable(); + auto *scope_var = g_scope_.Var("p_scopes"); + auto *p_scopes = scope_var->GetMutable>(); + (*p_scopes).resize(gpu_list_.size()); + auto op = f::OpRegistry::CreateOp(*op1); VLOG(1) << "invoke NCCLInitOp."; - op->Run(g_scope, cpu_place); + op->Run(g_scope_, cpu_place); VLOG(1) << "NCCLInitOp finished."; } + int GetGPUData(int gpu_id) { return gpu_id + 42; } + template void PerThreadProgram(int gpu_id, const f::OpDesc &op_desc, f::Scope *scope) { - std::unique_lock lk(mu); + std::unique_lock lk(mu_); const f::OpDesc *op1 = &op_desc; p::CUDAPlace place(gpu_id); - auto &ctx = dev_ctxs.at(gpu_id); + auto &ctx = dev_ctxs_.at(gpu_id); auto *send_tensor = scope->Var("st")->GetMutable(); auto *recv_tensor = scope->Var("rt")->GetMutable(); if (!send_tensor->numel()) { - send_tensor->Resize(kDims); send_tensor->mutable_data(kDims, place); - std::vector send_vector(f::product(kDims), gpu_id); + std::vector send_vector(f::product(kDims), GetGPUData(gpu_id)); paddle::framework::TensorFromVector(send_vector, *ctx, send_tensor); - ctx->Wait(); VLOG(1) << "Send Tensor filled with elements " << send_tensor->numel(); } @@ -118,30 +127,14 @@ class NCCLTester : public ::testing::Test { } public: - std::vector dev_ctxs; - f::Scope g_scope; - std::mutex mu; + std::vector dev_ctxs_; + f::Scope g_scope_; + std::mutex mu_; + std::vector gpu_list_; }; // ncclInitOp with desc -TEST(NCCL, ncclInitOp) { - std::unique_ptr op_desc(new f::OpDesc); - - op_desc->SetType("ncclInit"); - op_desc->SetOutput("Communicator", {"x1"}); - op_desc->SetAttr("gpus", {gpu_list}); - - f::Scope g_scope; - paddle::platform::CPUPlace cpu_place; - - auto *var = g_scope.Var("x1"); - var->GetMutable(); - - auto op = f::OpRegistry::CreateOp(*op_desc); - VLOG(1) << "invoke NCCLInitOp."; - op->Run(g_scope, cpu_place); - VLOG(1) << "NCCLInitOp finished."; -} +TEST_F(NCCLTester, ncclInitOp) {} // ncclAllReduceOp with desc TEST_F(NCCLTester, ncclAllReduceOp) { @@ -155,23 +148,25 @@ TEST_F(NCCLTester, ncclAllReduceOp) { std::vector ths; - for (size_t i = 0; i < gpu_list.size(); ++i) { - dev_scopes.emplace_back(&g_scope.NewScope()); - std::thread th(&NCCLTester::PerThreadProgram, this, gpu_list[i], + for (size_t i = 0; i < gpu_list_.size(); ++i) { + dev_scopes.emplace_back(&g_scope_.NewScope()); + std::thread th(&NCCLTester::PerThreadProgram, this, gpu_list_[i], *op2.get(), dev_scopes[i]); ths.emplace_back(std::move(th)); } - for (size_t i = 0; i < gpu_list.size(); ++i) { + for (size_t i = 0; i < gpu_list_.size(); ++i) { ths[i].join(); } - // check results - float result = std::accumulate(gpu_list.begin(), gpu_list.end(), 0); + float expected_result = 0.0; + for (int gpu_id : gpu_list_) { + expected_result = expected_result + GetGPUData(gpu_id); + } for (size_t i = 0; i < dev_scopes.size(); ++i) { p::CPUPlace cpu_place; - p::CUDAPlace gpu_place(gpu_list[i]); + p::CUDAPlace gpu_place(gpu_list_[i]); auto &recv_tensor = dev_scopes[i]->FindVar("rt")->Get(); auto *rt = recv_tensor.data(); @@ -180,12 +175,12 @@ TEST_F(NCCLTester, ncclAllReduceOp) { auto *ct = result_tensor->mutable_data(cpu_place); paddle::memory::Copy( - cpu_place, ct, p::CUDAPlace(gpu_list[i]), rt, + cpu_place, ct, p::CUDAPlace(gpu_list_[i]), rt, recv_tensor.numel() * sizeof(float), - static_cast(dev_ctxs[i])->stream()); + static_cast(dev_ctxs_[i])->stream()); for (int64_t j = 0; j < f::product(kDims); ++j) { - ASSERT_NEAR(ct[j], result, 1e-5); + ASSERT_NEAR(ct[j], expected_result, 1e-5); } } } @@ -204,22 +199,24 @@ TEST_F(NCCLTester, ncclReduceOp) { std::vector ths; - for (size_t i = 0; i < gpu_list.size(); ++i) { - dev_scopes.emplace_back(&g_scope.NewScope()); - std::thread th(&NCCLTester::PerThreadProgram, this, gpu_list[i], + for (size_t i = 0; i < gpu_list_.size(); ++i) { + dev_scopes.emplace_back(&g_scope_.NewScope()); + std::thread th(&NCCLTester::PerThreadProgram, this, gpu_list_[i], *op2.get(), dev_scopes[i]); ths.emplace_back(std::move(th)); } - for (size_t i = 0; i < gpu_list.size(); ++i) { + for (size_t i = 0; i < gpu_list_.size(); ++i) { ths[i].join(); } - // check results on - float result = std::accumulate(gpu_list.begin(), gpu_list.end(), 0); + float expected_result = 0.0; + for (int gpu_id : gpu_list_) { + expected_result = expected_result + GetGPUData(gpu_id); + } p::CPUPlace cpu_place; - p::CUDAPlace gpu_place(gpu_list[kRoot]); + p::CUDAPlace gpu_place(gpu_list_[kRoot]); auto &recv_tensor = dev_scopes[kRoot]->FindVar("rt")->Get(); auto *rt = recv_tensor.data(); @@ -229,12 +226,12 @@ TEST_F(NCCLTester, ncclReduceOp) { auto *ct = result_tensor->mutable_data(cpu_place); paddle::memory::Copy( - cpu_place, ct, p::CUDAPlace(gpu_list[kRoot]), rt, + cpu_place, ct, p::CUDAPlace(gpu_list_[kRoot]), rt, recv_tensor.numel() * sizeof(float), - static_cast(dev_ctxs[kRoot])->stream()); + static_cast(dev_ctxs_[kRoot])->stream()); for (int64_t j = 0; j < f::product(kDims); ++j) { - ASSERT_NEAR(ct[j], result, 1e-5); + ASSERT_NEAR(ct[j], expected_result, 1e-5); } } @@ -252,23 +249,22 @@ TEST_F(NCCLTester, ncclBcastOp) { std::vector ths; - for (size_t i = 0; i < gpu_list.size(); ++i) { - dev_scopes.emplace_back(&g_scope.NewScope()); - std::thread th(&NCCLTester::PerThreadProgram, this, gpu_list[i], + for (size_t i = 0; i < gpu_list_.size(); ++i) { + dev_scopes.emplace_back(&g_scope_.NewScope()); + std::thread th(&NCCLTester::PerThreadProgram, this, gpu_list_[i], *op2.get(), dev_scopes[i]); ths.emplace_back(std::move(th)); } - for (size_t i = 0; i < gpu_list.size(); ++i) { + for (size_t i = 0; i < gpu_list_.size(); ++i) { ths[i].join(); } const int idx = 1; - // check results on - float result = kRoot; + float result = GetGPUData(kRoot); p::CPUPlace cpu_place; - p::CUDAPlace gpu_place(gpu_list[idx]); + p::CUDAPlace gpu_place(gpu_list_[idx]); auto &recv_tensor = dev_scopes[idx]->FindVar("rt")->Get(); auto *rt = recv_tensor.data(); @@ -277,42 +273,11 @@ TEST_F(NCCLTester, ncclBcastOp) { auto *ct = result_tensor->mutable_data(cpu_place); paddle::memory::Copy( - cpu_place, ct, p::CUDAPlace(gpu_list[idx]), rt, + cpu_place, ct, p::CUDAPlace(gpu_list_[idx]), rt, recv_tensor.numel() * sizeof(float), - static_cast(dev_ctxs[idx])->stream()); + static_cast(dev_ctxs_[idx])->stream()); for (int64_t j = 0; j < f::product(kDims); ++j) { ASSERT_NEAR(ct[j], result, 1e-5); } } - -int main(int argc, char **argv) { - // FIXME(tonyyang-svail): - // Due to the driver issue on our CI, disable for now - return 0; - const int dev_count = p::GetCUDADeviceCount(); - if (dev_count <= 1) { - LOG(WARNING) - << "Cannot test multi-gpu nccl, because the CUDA device count is " - << dev_count; - return 0; - } - - std::vector places; - - places.emplace_back(paddle::platform::CPUPlace()); - int count = paddle::platform::GetCUDADeviceCount(); - for (int i = 0; i < count; ++i) { - places.emplace_back(paddle::platform::CUDAPlace(i)); - gpu_list.emplace_back(i); - } - - VLOG(0) << " DeviceCount " << count; - paddle::platform::DeviceContextPool::Init(places); - - testing::InitGoogleTest(&argc, argv); - - // device context should be release before scope. - // otherwise driver will down. - return RUN_ALL_TESTS(); -} diff --git a/paddle/fluid/operators/parallel_do_op.cc b/paddle/fluid/operators/parallel_do_op.cc index bf4d0476df32d7454d4064cb6ee454e6ad5d6fc5..4001b9a130348b4e3ea99f3017eae6d85e41fc6e 100644 --- a/paddle/fluid/operators/parallel_do_op.cc +++ b/paddle/fluid/operators/parallel_do_op.cc @@ -18,6 +18,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { @@ -158,11 +159,14 @@ class ParallelDoOp : public framework::OperatorBase { auto &place = places[place_idx]; auto *cur_scope = sub_scopes[place_idx]; - workers.emplace_back(framework::Async([program, cur_scope, place, block] { - framework::Executor executor(place); - executor.Run(*program, cur_scope, block->ID(), - false /*create_local_scope*/); - })); + workers.emplace_back( + framework::Async([program, cur_scope, place, block, place_idx] { + // Give the thread an id to distinguish parallel block with same id. + platform::RecordThread rt(static_cast(place_idx) + 1); + framework::Executor executor(place); + executor.Run(*program, cur_scope, block->ID(), + false /*create_local_scope*/); + })); } for (auto &worker : workers) { worker.wait(); @@ -234,11 +238,14 @@ class ParallelDoGradOp : public framework::OperatorBase { auto *cur_scope = sub_scopes[i]; // execute - workers.emplace_back(framework::Async([program, cur_scope, place, block] { - framework::Executor executor(place); - executor.Run(*program, cur_scope, block->ID(), - false /*create_local_scope*/); - })); + workers.emplace_back( + framework::Async([program, cur_scope, place, block, i] { + // Give the thread an id to distinguish parallel block with same id. + platform::RecordThread rt(static_cast(i) + 1); + framework::Executor executor(place); + executor.Run(*program, cur_scope, block->ID(), + false /*create_local_scope*/); + })); } for (auto &worker : workers) { worker.wait(); diff --git a/paddle/fluid/operators/pool_cudnn_op.cu.cc b/paddle/fluid/operators/pool_cudnn_op.cu.cc index 781d96981e4c033d9287ab3de9860dfd9fcd2875..39c862b03ad497dca5c38ccecff20be510ab60e5 100644 --- a/paddle/fluid/operators/pool_cudnn_op.cu.cc +++ b/paddle/fluid/operators/pool_cudnn_op.cu.cc @@ -24,6 +24,8 @@ using ScopedTensorDescriptor = platform::ScopedTensorDescriptor; using ScopedPoolingDescriptor = platform::ScopedPoolingDescriptor; using DataLayout = platform::DataLayout; using PoolingMode = platform::PoolingMode; +template +using ScalingParamType = typename platform::CudnnDataType::ScalingParamType; template class PoolCUDNNOpKernel : public framework::OpKernel { @@ -78,8 +80,7 @@ class PoolCUDNNOpKernel : public framework::OpKernel { // ------------------- cudnn pool algorithm --------------------- auto handle = ctx.cuda_device_context().cudnn_handle(); - T alpha = 1.0f, beta = 0.0f; - + ScalingParamType alpha = 1.0f, beta = 0.0f; PADDLE_ENFORCE(platform::dynload::cudnnPoolingForward( handle, cudnn_pool_desc, &alpha, cudnn_input_desc, input_data, &beta, cudnn_output_desc, output_data)); @@ -144,8 +145,7 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel { // ------------------- cudnn pool algorithm --------------------- auto handle = ctx.cuda_device_context().cudnn_handle(); - T alpha = 1.0f, beta = 0.0f; - + ScalingParamType alpha = 1.0f, beta = 0.0f; if (input_grad) { T *input_grad_data = input_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset input_grad. @@ -162,17 +162,19 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; +namespace plat = paddle::platform; -REGISTER_OP_KERNEL(pool2d, CUDNN, ::paddle::platform::CUDAPlace, +REGISTER_OP_KERNEL(pool2d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel, - ops::PoolCUDNNOpKernel); -REGISTER_OP_KERNEL(pool2d_grad, CUDNN, ::paddle::platform::CUDAPlace, + ops::PoolCUDNNOpKernel, + ops::PoolCUDNNOpKernel); +REGISTER_OP_KERNEL(pool2d_grad, CUDNN, plat::CUDAPlace, ops::PoolCUDNNGradOpKernel, ops::PoolCUDNNGradOpKernel); -REGISTER_OP_KERNEL(pool3d, CUDNN, ::paddle::platform::CUDAPlace, +REGISTER_OP_KERNEL(pool3d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel, ops::PoolCUDNNOpKernel); -REGISTER_OP_KERNEL(pool3d_grad, CUDNN, ::paddle::platform::CUDAPlace, +REGISTER_OP_KERNEL(pool3d_grad, CUDNN, plat::CUDAPlace, ops::PoolCUDNNGradOpKernel, ops::PoolCUDNNGradOpKernel); diff --git a/paddle/fluid/operators/pool_mkldnn_op.cc b/paddle/fluid/operators/pool_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c88578570c1acdecaa97dd8b12a702778fef2b7e --- /dev/null +++ b/paddle/fluid/operators/pool_mkldnn_op.cc @@ -0,0 +1,217 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +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/fluid/operators/pool_op.h" +#include "paddle/fluid/platform/mkldnn_helper.h" + +namespace paddle { +namespace operators { + +template +class PoolMKLDNNOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + + auto& dev_ctx = + ctx.template device_context(); + const auto& mkldnn_engine = dev_ctx.GetEngine(); + + const Tensor* input = ctx.Input("X"); + Tensor* output = ctx.Output("Out"); + + // Get an unique name from "argument" name of "Out" variable + // This name will be used as key when saving info into device context + const std::string key = ctx.op().Output("Out"); + const std::string key_pool_pd = key + "@pool_pd"; + const std::string key_pool_workspace_memory = + key + "@pool_workspace_memory"; + + std::string pooling_type = ctx.Attr("pooling_type"); + std::vector ksize = ctx.Attr>("ksize"); + std::vector strides = ctx.Attr>("strides"); + std::vector paddings = ctx.Attr>("paddings"); + if (ctx.Attr("global_pooling")) { + for (size_t i = 0; i < ksize.size(); ++i) { + paddings[i] = 0; + ksize[i] = static_cast(input->dims()[i + 2]); + } + } + + // Only 2D pooling is supported now + PADDLE_ENFORCE(ksize.size() == 2, "ksize must be 2D, i.e. 2D pooling"); + PADDLE_ENFORCE(pooling_type == "max" || pooling_type == "avg", + "pooling_type must be 'max' or 'avg'"); + PADDLE_ENFORCE(input->dims().size() == 4, + "Input dim must be with 4, i.e. NCHW"); + + const T* input_data = input->data(); + T* output_data = output->mutable_data(ctx.GetPlace()); + + std::vector src_tz = paddle::framework::vectorize2int(input->dims()); + std::vector dst_tz = paddle::framework::vectorize2int(output->dims()); + + // TODO(pzelazko-intel): support more formats + auto src_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + auto dst_md = platform::MKLDNNMemDesc(dst_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + + std::shared_ptr pool_pd = + CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize, + pooling_type, mkldnn_engine); + + // save pool_pd into global device context to be referred in backward path + dev_ctx.SetBlob(key_pool_pd, pool_pd); + + std::shared_ptr workspace_memory = + CreateWorkspaceMemory(pool_pd, pooling_type, mkldnn_engine); + + // save pool_workspace_memory to be referred in backward path + dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory); + + auto src_memory = + mkldnn::memory({src_md, mkldnn_engine}, (void*)input_data); + auto dst_memory = + mkldnn::memory({dst_md, mkldnn_engine}, (void*)output_data); + + auto pool_prim = mkldnn::pooling_forward(*pool_pd, src_memory, dst_memory, + *workspace_memory); + + // push primitive to stream and wait until it's executed + std::vector pipeline{pool_prim}; + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + } + + private: + std::unique_ptr CreatePrimitiveDesc( + const mkldnn::memory::desc& src, const mkldnn::memory::desc& dst, + const std::vector& stride, const std::vector& padding, + const std::vector& kernel, const std::string& pooling_type, + const mkldnn::engine& engine) const { + auto pool_desc = mkldnn::pooling_forward::desc( + mkldnn::prop_kind::forward, + pooling_type == "max" ? mkldnn::algorithm::pooling_max + : mkldnn::algorithm::pooling_avg, + src, dst, stride, kernel, padding, padding, mkldnn::padding_kind::zero); + + auto p_pool_pd = + new mkldnn::pooling_forward::primitive_desc(pool_desc, engine); + return std::unique_ptr(p_pool_pd); + } + + std::unique_ptr CreateWorkspaceMemory( + std::shared_ptr pool_pd, + const std::string& pooling_type, const mkldnn::engine& engine) const { + mkldnn::memory::primitive_desc workspace_md = + pooling_type == "max" + ? pool_pd->workspace_primitive_desc() + : mkldnn::memory::primitive_desc( + {{}, mkldnn::memory::f32, mkldnn::memory::format::nchw}, + engine); + + auto p_workspace_memory = new mkldnn::memory(workspace_md); + return std::unique_ptr(p_workspace_memory); + } +}; + +template +class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + + const Tensor* in_x = ctx.Input("X"); + const Tensor* out_grad = ctx.Input(framework::GradVarName("Out")); + Tensor* in_x_grad = ctx.Output(framework::GradVarName("X")); + + // Get an unique name from "argument" name of "Out" variable + // This name will be used as key when referring info from device context + const std::string key = ctx.op().Input("Out"); + const std::string key_pool_pd = key + "@pool_pd"; + const std::string key_pool_workspace_memory = + key + "@pool_workspace_memory"; + + std::string pooling_type = ctx.Attr("pooling_type"); + std::vector ksize = ctx.Attr>("ksize"); + std::vector strides = ctx.Attr>("strides"); + std::vector paddings = ctx.Attr>("paddings"); + + if (ctx.Attr("global_pooling")) { + for (size_t i = 0; i < ksize.size(); ++i) { + paddings[i] = 0; + ksize[i] = static_cast(in_x->dims()[i + 2]); + } + } + + auto& dev_ctx = + ctx.template device_context(); + const mkldnn::engine& mkldnn_engine = dev_ctx.GetEngine(); + + const T* out_grad_data = out_grad->data(); + T* in_x_grad_data = in_x_grad->mutable_data(ctx.GetPlace()); + + std::vector diff_src_tz = + paddle::framework::vectorize2int(in_x_grad->dims()); + std::vector diff_dst_tz = + paddle::framework::vectorize2int(out_grad->dims()); + + auto diff_src_md = platform::MKLDNNMemDesc(diff_src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + auto diff_dst_md = platform::MKLDNNMemDesc(diff_dst_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + + // Retrieve pool_pd/pool_workspace_memory from device context + auto pool_pd = + std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_pd)); + PADDLE_ENFORCE(pool_pd != nullptr, + "Fail to find pool_pd in device context"); + + auto workspace_memory = std::static_pointer_cast( + dev_ctx.GetBlob(key_pool_workspace_memory)); + PADDLE_ENFORCE(workspace_memory != nullptr, + "Fail to find workspace_memory in device context"); + + auto pool_bwd_desc = mkldnn::pooling_backward::desc( + pooling_type == "max" ? mkldnn::algorithm::pooling_max + : mkldnn::algorithm::pooling_avg, + diff_src_md, diff_dst_md, strides, ksize, paddings, paddings, + mkldnn::padding_kind::zero); + auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc( + pool_bwd_desc, mkldnn_engine, *pool_pd); + + auto diff_src_memory = + mkldnn::memory({diff_src_md, mkldnn_engine}, (void*)in_x_grad_data); + auto diff_dst_memory = + mkldnn::memory({diff_dst_md, mkldnn_engine}, (void*)out_grad_data); + + auto bwd_prim = mkldnn::pooling_backward( + pool_bwd_pd, diff_dst_memory, *workspace_memory, diff_src_memory); + + // push primitive to stream and wait until it's executed + std::vector pipeline{bwd_prim}; + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + } // Compute() +}; + +} // namespace operators +} // namespace paddle + +REGISTER_OP_KERNEL(pool2d, MKLDNN, ::paddle::platform::CPUPlace, + paddle::operators::PoolMKLDNNOpKernel); +REGISTER_OP_KERNEL(pool2d_grad, MKLDNN, ::paddle::platform::CPUPlace, + paddle::operators::PoolMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/pool_op.cc b/paddle/fluid/operators/pool_op.cc index ac22acb25a7ab33a26de49804667703e84b78a8a..b144ec5f7d315cb340dcd94b4a519bfcfd2a0e66 100644 --- a/paddle/fluid/operators/pool_op.cc +++ b/paddle/fluid/operators/pool_op.cc @@ -13,6 +13,12 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/pool_op.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/cudnn_helper.h" +#endif +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif namespace paddle { namespace operators { @@ -76,20 +82,18 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const { framework::OpKernelType PoolOp::GetExpectedKernelType( const framework::ExecutionContext &ctx) const { - bool use_cudnn = ctx.Attr("use_cudnn"); - use_cudnn &= platform::is_gpu_place(ctx.GetPlace()); + framework::LibraryType library_{framework::LibraryType::kPlain}; #ifdef PADDLE_WITH_CUDA - if (platform::is_gpu_place(ctx.GetPlace())) { - auto &dev_ctx = ctx.template device_context(); - use_cudnn &= dev_ctx.cudnn_handle() != nullptr; + if (platform::CanCUDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kCUDNN; } #endif - framework::LibraryType library_; - if (use_cudnn) { - library_ = framework::LibraryType::kCUDNN; - } else { - library_ = framework::LibraryType::kPlain; +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; } +#endif std::string data_format = ctx.Attr("data_format"); framework::DataLayout layout_ = framework::StringToDataLayout(data_format); @@ -107,26 +111,28 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const { framework::OpKernelType PoolOpGrad::GetExpectedKernelType( const framework::ExecutionContext &ctx) const { - bool use_cudnn = ctx.Attr("use_cudnn"); - use_cudnn &= platform::is_gpu_place(ctx.GetPlace()); + framework::LibraryType library_{framework::LibraryType::kPlain}; #ifdef PADDLE_WITH_CUDA - if (platform::is_gpu_place(ctx.GetPlace())) { - auto &dev_ctx = ctx.template device_context(); - use_cudnn &= dev_ctx.cudnn_handle() != nullptr; + if (platform::CanCUDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kCUDNN; } #endif - framework::LibraryType library_; - if (use_cudnn) { - library_ = framework::LibraryType::kCUDNN; - } else { - library_ = framework::LibraryType::kPlain; +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; } +#endif + auto input_data_type = framework::ToDataType(ctx.Input("X")->type()); + if (input_data_type == framework::proto::VarType::FP16) { + PADDLE_ENFORCE_EQ(library_, framework::LibraryType::kCUDNN, + "float16 can only be used when CUDNN is used"); + } std::string data_format = ctx.Attr("data_format"); framework::DataLayout layout_ = framework::StringToDataLayout(data_format); - return framework::OpKernelType( - framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), - layout_, library_); + return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_, + library_); } Pool2dOpMaker::Pool2dOpMaker(OpProto *proto, OpAttrChecker *op_checker) @@ -181,6 +187,9 @@ Pool2dOpMaker::Pool2dOpMaker(OpProto *proto, OpAttrChecker *op_checker) "output height and width. False is the default. If it is set to False, " "the floor function will be used.") .SetDefault(false); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddAttr( "data_format", "(string, default NCHW) Only used in " @@ -276,6 +285,9 @@ Pool3dOpMaker::Pool3dOpMaker(OpProto *proto, OpAttrChecker *op_checker) "output height and width. False is the default. If it is set to False, " "the floor function will be used.") .SetDefault(false); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddAttr( "data_format", "(string, default NCHW) Only used in " diff --git a/paddle/fluid/operators/prior_box_op.cc b/paddle/fluid/operators/prior_box_op.cc index be7898c22190339e0717317807b91e038f4949f6..7ba55437cb20f802cc12ceea7777d7d78bba62a6 100644 --- a/paddle/fluid/operators/prior_box_op.cc +++ b/paddle/fluid/operators/prior_box_op.cc @@ -111,7 +111,8 @@ class PriorBoxOpMaker : public framework::OpProtoAndCheckerMaker { }); AddAttr>( "max_sizes", - "(vector) List of max sizes of generated prior boxes."); + "(vector) List of max sizes of generated prior boxes.") + .SetDefault(std::vector{}); AddAttr>( "aspect_ratios", "(vector) List of aspect ratios of generated prior boxes."); diff --git a/paddle/fluid/operators/prior_box_op.h b/paddle/fluid/operators/prior_box_op.h index 0113d2f09af90eec32632180741db3a1fc40c724..18bb2deb6b5acf626dfb2883a5771d9d195d45c0 100644 --- a/paddle/fluid/operators/prior_box_op.h +++ b/paddle/fluid/operators/prior_box_op.h @@ -97,9 +97,6 @@ class PriorBoxOpKernel : public framework::OpKernel { boxes->mutable_data(ctx.GetPlace()); vars->mutable_data(ctx.GetPlace()); - T inv_img_width = 1.0 / img_width; - T inv_img_height = 1.0 / img_height; - auto e_boxes = framework::EigenTensor::From(*boxes); for (int h = 0; h < feature_height; ++h) { for (int w = 0; w < feature_width; ++w) { @@ -110,36 +107,30 @@ class PriorBoxOpKernel : public framework::OpKernel { for (size_t s = 0; s < min_sizes.size(); ++s) { auto min_size = min_sizes[s]; // first prior: aspect_ratio = 1, size = min_size - box_width = box_height = min_size; + box_width = box_height = min_size / 2.; // xmin - e_boxes(h, w, idx, 0) = (center_x - box_width * 0.5) * inv_img_width; + e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; // ymin - e_boxes(h, w, idx, 1) = - (center_y - box_height * 0.5) * inv_img_height; + e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; // xmax - e_boxes(h, w, idx, 2) = (center_x + box_width * 0.5) * inv_img_width; + e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; // ymax - e_boxes(h, w, idx, 3) = - (center_y + box_height * 0.5) * inv_img_height; + e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; idx++; if (max_sizes.size() > 0) { auto max_size = max_sizes[s]; // second prior: aspect_ratio = 1, // size = sqrt(min_size * max_size) - box_width = box_height = sqrt(min_size * max_size); + box_width = box_height = sqrt(min_size * max_size) / 2.; // xmin - e_boxes(h, w, idx, 0) = - (center_x - box_width * 0.5) * inv_img_width; + e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; // ymin - e_boxes(h, w, idx, 1) = - (center_y - box_height * 0.5) * inv_img_height; + e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; // xmax - e_boxes(h, w, idx, 2) = - (center_x + box_width * 0.5) * inv_img_width; + e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; // ymax - e_boxes(h, w, idx, 3) = - (center_y + box_height * 0.5) * inv_img_height; + e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; idx++; } @@ -149,20 +140,16 @@ class PriorBoxOpKernel : public framework::OpKernel { if (fabs(ar - 1.) < 1e-6) { continue; } - box_width = min_size * sqrt(ar); - box_height = min_size / sqrt(ar); + box_width = min_size * sqrt(ar) / 2.; + box_height = min_size / sqrt(ar) / 2.; // xmin - e_boxes(h, w, idx, 0) = - (center_x - box_width * 0.5) * inv_img_width; + e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width; // ymin - e_boxes(h, w, idx, 1) = - (center_y - box_height * 0.5) * inv_img_height; + e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height; // xmax - e_boxes(h, w, idx, 2) = - (center_x + box_width * 0.5) * inv_img_width; + e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width; // ymax - e_boxes(h, w, idx, 3) = - (center_y + box_height * 0.5) * inv_img_height; + e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height; idx++; } } diff --git a/paddle/fluid/operators/read_op.cc b/paddle/fluid/operators/read_op.cc index 62beab82d4f2b0b795d5d32f50352172de6870cc..2a5605e0d378a184ae132e657b2872279784855d 100644 --- a/paddle/fluid/operators/read_op.cc +++ b/paddle/fluid/operators/read_op.cc @@ -60,15 +60,16 @@ class ReadOp : public framework::OperatorBase { const platform::Place& dev_place) const override { framework::ReaderHolder* reader = scope.FindVar(Input("Reader"))->GetMutable(); - if (!reader->HasNext()) { + std::vector out_arg_names = Outputs("Out"); + std::vector ins; + reader->ReadNext(&ins); + if (ins.empty()) { reader->ReInit(); + reader->ReadNext(&ins); PADDLE_ENFORCE( - reader->HasNext(), + !ins.empty(), "Reader can not read the next data even it has been re-initialized."); } - std::vector out_arg_names = Outputs("Out"); - std::vector ins; - reader->ReadNext(&ins); PADDLE_ENFORCE_EQ(ins.size(), out_arg_names.size()); for (size_t i = 0; i < ins.size(); ++i) { auto* out = diff --git a/paddle/fluid/operators/reader/CMakeLists.txt b/paddle/fluid/operators/reader/CMakeLists.txt index 06489f32d64d69030c084a038acb78ac2bac6200..744bd3b7ef71f83ad82979eb966369c2e9456a7d 100644 --- a/paddle/fluid/operators/reader/CMakeLists.txt +++ b/paddle/fluid/operators/reader/CMakeLists.txt @@ -1,5 +1,24 @@ cc_library(reader_op_registry SRCS reader_op_registry.cc DEPS operator op_registry reader) -op_library(create_random_data_generator_op SRCS create_random_data_generator_op.cc DEPS reader_op_registry) -op_library(create_shuffle_reader_op SRCS create_shuffle_reader_op.cc DEPS reader_op_registry) -op_library(create_batch_reader_op SRCS create_batch_reader_op.cc DEPS reader_op_registry) -set(READER_LIBRARY create_random_data_generator_op create_shuffle_reader_op create_batch_reader_op PARENT_SCOPE) +set(LOCAL_READER_LIBS) + +function(reader_library TARGET_NAME) + set(oneValueArgs "") + set(multiValueArgs SRCS DEPS) + set(options "") + set(common_deps reader_op_registry) + cmake_parse_arguments(reader_library "${options}" "${oneValueArgs}" + "${multiValueArgs}" ${ARGN}) + op_library(${TARGET_NAME} SRCS ${reader_library_SRCS} DEPS ${common_deps} ${reader_library_DEPS}) + set(LOCAL_READER_LIBS + ${TARGET_NAME} + ${LOCAL_READER_LIBS} + PARENT_SCOPE) +endfunction() + +reader_library(create_random_data_generator_op SRCS create_random_data_generator_op.cc) +reader_library(create_shuffle_reader_op SRCS create_shuffle_reader_op.cc) +reader_library(create_batch_reader_op SRCS create_batch_reader_op.cc) +reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_op.cc) +reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc) +# Export local libraries to parent +set(READER_LIBRARY ${LOCAL_READER_LIBS} PARENT_SCOPE) diff --git a/paddle/fluid/operators/reader/create_batch_reader_op.cc b/paddle/fluid/operators/reader/create_batch_reader_op.cc index bac043a5529d877dba79c03f07b9d43c9b71d7aa..277f2856c07b3fec2113486539aec1d9139fae92 100644 --- a/paddle/fluid/operators/reader/create_batch_reader_op.cc +++ b/paddle/fluid/operators/reader/create_batch_reader_op.cc @@ -68,10 +68,10 @@ void BatchReader::ReadNext(std::vector* out) { buffer_.clear(); buffer_.reserve(batch_size_); for (int i = 0; i < batch_size_; ++i) { - if (reader_->HasNext()) { - buffer_.push_back(std::vector()); - reader_->ReadNext(&buffer_.back()); - } else { + buffer_.push_back(std::vector()); + reader_->ReadNext(&buffer_.back()); + if (buffer_.back().empty()) { + buffer_.pop_back(); break; } } diff --git a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..d0de092947eb04a1b7d06dedea919f6b1094dd06 --- /dev/null +++ b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc @@ -0,0 +1,185 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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 "paddle/fluid/framework/channel.h" +#include "paddle/fluid/operators/reader/reader_op_registry.h" + +namespace paddle { +namespace operators { +namespace reader { + +static constexpr size_t kDoubleBufferSize = 2; + +class DoubleBufferReader : public framework::DecoratedReader { + public: + struct Item { + Item() : ctx_(nullptr) {} + + std::vector payloads_; + platform::DeviceContext* ctx_; + }; + + explicit DoubleBufferReader( + ReaderBase* reader, platform::Place target_place = platform::CPUPlace()) + : DecoratedReader(reader), place_(target_place) { + for (size_t i = 0; i < kDoubleBufferSize; ++i) { + if (platform::is_gpu_place(place_)) { +#ifdef PADDLE_WITH_CUDA + ctxs_.emplace_back(new platform::CUDADeviceContext( + boost::get(place_))); +#endif + } + } + + start_thread(); + } + + void start_thread() { + buffer_ = framework::MakeChannel(kDoubleBufferSize); + std::thread prefetch([this] { PrefetchThreadFunc(); }); + prefetch.detach(); + } + + void ReadNext(std::vector* out) override; + void ReInit() override; + + ~DoubleBufferReader() { buffer_->Close(); } + + bool HasNext() const override; + + private: + void PrefetchThreadFunc(); + + framework::Channel* buffer_; + platform::Place place_; + std::vector> ctxs_; + mutable Item local_buffer_; +}; + +class CreateDoubleBufferReaderOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope& scope, + const platform::Place& dev_place) const override { + const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader")) + ->Get(); + auto* out = scope.FindVar(Output("Out")) + ->template GetMutable(); + + auto place_str = Attr("place"); + platform::Place place; + if (place_str == "CPU") { + place = platform::CPUPlace(); + } else { + std::istringstream sin(place_str); + sin.seekg(std::string("CUDA:").size(), std::ios::beg); + size_t num; + sin >> num; + place = platform::CUDAPlace(static_cast(num)); + } + + out->Reset(new DoubleBufferReader(underlying_reader.Get(), place)); + } +}; + +class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase { + public: + CreateDoubleBufferReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) + : DecoratedReaderMakerBase(op_proto, op_checker) { + AddComment(R"DOC( + CreateDoubleBufferReader Operator + + A double buffer reader takes another reader as its 'underlying reader'. + It launches another thread to execute the 'underlying reader' asynchronously, + which prevents reading process from blocking subsequent training. + )DOC"); + std::unordered_set enum_range; + constexpr size_t kMaxCUDADevs = 128; + for (size_t i = 0; i < kMaxCUDADevs; ++i) { + enum_range.insert(string::Sprintf("CUDA:%d", i)); + } + enum_range.insert("CPU"); + AddAttr("place", "The double buffer place, default is CPU") + .SetDefault("CPU") + .InEnum({enum_range}); + } +}; + +void DoubleBufferReader::ReadNext(std::vector* out) { + if (local_buffer_.payloads_.empty()) { + buffer_->Receive(&local_buffer_); + } + + *out = local_buffer_.payloads_; + local_buffer_.payloads_.clear(); + if (local_buffer_.ctx_) { + local_buffer_.ctx_->Wait(); + } +} + +void DoubleBufferReader::ReInit() { + reader_->ReInit(); + buffer_->Close(); + start_thread(); +} + +void DoubleBufferReader::PrefetchThreadFunc() { + VLOG(5) << "A new prefetch thread starts."; + size_t gpu_ctx_offset = 0; + while (reader_->HasNext()) { + Item batch; + reader_->ReadNext(&batch.payloads_); + if (platform::is_gpu_place(place_)) { + std::vector gpu_batch; + auto& gpu_ctx = this->ctxs_[gpu_ctx_offset++]; + gpu_ctx_offset %= this->ctxs_.size(); + gpu_batch.resize(batch.payloads_.size()); + for (size_t i = 0; i < batch.payloads_.size(); ++i) { + framework::TensorCopy(batch.payloads_[i], place_, *gpu_ctx, + &gpu_batch[i]); + gpu_batch[i].set_lod(batch.payloads_[i].lod()); + } + batch.ctx_ = gpu_ctx.get(); + std::swap(gpu_batch, batch.payloads_); + } + + if (!buffer_->Send(&batch)) { + VLOG(5) << "WARNING: The double buffer channel has been closed. The " + "prefetch thread terminates."; + break; + } + } + buffer_->Close(); +} + +bool DoubleBufferReader::HasNext() const { + if (local_buffer_.payloads_.empty()) { + bool ok = buffer_->Receive(&local_buffer_); + return ok; + } else { + return true; + } +} + +} // namespace reader +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators::reader; +REGISTER_DECORATED_READER_OPERATOR(create_double_buffer_reader, + ops::CreateDoubleBufferReaderOp, + ops::CreateDoubleBufferReaderOpMaker); diff --git a/paddle/fluid/operators/reader/create_random_data_generator_op.cc b/paddle/fluid/operators/reader/create_random_data_generator_op.cc index f77ab8ab196dae4cf9351cee9bc5566ec2c04e4b..95d8674c08b63e872926ff8708d0c734da33684c 100644 --- a/paddle/fluid/operators/reader/create_random_data_generator_op.cc +++ b/paddle/fluid/operators/reader/create_random_data_generator_op.cc @@ -19,11 +19,11 @@ namespace operators { namespace reader { template -class RandomDataGenerator : public framework::FileReader { +class RandomDataGenerator : public framework::ReaderBase { public: RandomDataGenerator(const std::vector& shapes, float min, float max) - : FileReader(shapes), min_(min), max_(max) { + : framework::ReaderBase(), min_(min), max_(max), shapes_(shapes) { PADDLE_ENFORCE_LE( min, max, "'min' shouldn't be greater than 'max'.(%f vs %f)", min, max); unsigned int seed = std::random_device()(); @@ -50,15 +50,16 @@ class RandomDataGenerator : public framework::FileReader { } } - bool HasNext() const override { return true; } - void ReInit() override { return; } + bool HasNext() const override { return true; } + private: float min_; float max_; std::minstd_rand engine_; std::uniform_real_distribution dist_; + std::vector shapes_; }; template diff --git a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c4aa29c7206dbd3fe6a99b2a6c5ac6f083621944 --- /dev/null +++ b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc @@ -0,0 +1,90 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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/fluid/operators/reader/reader_op_registry.h" +#include "paddle/fluid/recordio/scanner.h" + +namespace paddle { +namespace operators { +namespace reader { +class RecordIOFileReader : public framework::FileReader { + public: + explicit RecordIOFileReader(const std::string& filename, + const std::vector& dims) + : FileReader(dims), + scanner_(filename), + dev_ctx_(*platform::DeviceContextPool::Instance().Get( + platform::CPUPlace())) {} + + bool HasNext() const override { return scanner_.HasNext(); } + + void ReInit() override { scanner_.Reset(); } + + protected: + void ReadNextImpl(std::vector* out) override { + *out = framework::ReadFromRecordIO(scanner_, dev_ctx_); + } + + private: + recordio::Scanner scanner_; + const platform::DeviceContext& dev_ctx_; +}; + +class CreateRecordIOReaderOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope& scope, + const platform::Place& dev_place) const override { + const auto& shape_concat = Attr>("shape_concat"); + const auto& ranks = Attr>("ranks"); + PADDLE_ENFORCE(!shape_concat.empty() && !ranks.empty()); + PADDLE_ENFORCE_EQ(std::accumulate(ranks.begin(), ranks.end(), 0), + int(shape_concat.size()), + "The accumulate of all ranks should be equal to the " + "shape concat's length."); + std::string filename = Attr("filename"); + + auto* out = scope.FindVar(Output("Out")) + ->template GetMutable(); + out->Reset( + new RecordIOFileReader(filename, RestoreShapes(shape_concat, ranks))); + } +}; + +class CreateRecordIOReaderOpMaker : public FileReaderMakerBase { + public: + CreateRecordIOReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker) + : FileReaderMakerBase(op_proto, op_checker) { + AddAttr("filename", "The filename of record io reader"); + AddComment(R"DOC( + CreateRecordIOReader Operator + + Create a reader from a record io file + )DOC"); + } +}; + +} // namespace reader +} // namespace operators +} // namespace paddle + +namespace reader = paddle::operators::reader; + +REGISTER_FILE_READER_OPERATOR(create_recordio_file_reader, + reader::CreateRecordIOReaderOp, + reader::CreateRecordIOReaderOpMaker); + +REGISTER_FILE_READER(recordio, reader::RecordIOFileReader); diff --git a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc index 3e8b463efc99e4a962e5ae14ab133cf634548756..70e2f587dc414a850ddc341b98f26ae54636755c 100644 --- a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc +++ b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc @@ -12,6 +12,9 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include +#include "glog/logging.h" +#include "paddle/fluid/operators/detail/safe_ref.h" #include "paddle/fluid/operators/reader/reader_op_registry.h" namespace paddle { @@ -20,43 +23,53 @@ namespace reader { class ShuffleReader : public framework::DecoratedReader { public: - ShuffleReader(ReaderBase* reader, int buffer_size) - : DecoratedReader(reader), buffer_size_(buffer_size), iteration_pos_(0) { - buffer_.reserve(buffer_size); + ShuffleReader(ReaderBase* reader, size_t buffer_size, size_t seed = 0) + : DecoratedReader(reader), buffer_size_(buffer_size), seed_(seed) { + VLOG(10) << "Create shuffle reader of " << reader_; + if (seed_ == 0) { + std::random_device device; + seed_ = device(); + } + ReadIntoBuffers(); } - void ReadNext(std::vector* out) override; + void ReadNext(std::vector* out) override { + if (iteration_pos_ >= buffer_.size()) { + VLOG(10) << "Resetting shuffle buffer"; + ReadIntoBuffers(); + } + *out = buffer_[iteration_pos_++]; + } - private: - int buffer_size_; - std::vector> buffer_; - size_t iteration_pos_; -}; + bool HasNext() const override { + return iteration_pos_ < buffer_.size() || reader_->HasNext(); + } -void ShuffleReader::ReadNext(std::vector* out) { - if (iteration_pos_ >= buffer_.size()) { - // Reload buffer with new data + private: + void ReadIntoBuffers() { buffer_.clear(); buffer_.reserve(buffer_size_); - for (int i = 0; i < buffer_size_; ++i) { - if (reader_->HasNext()) { - buffer_.push_back(std::vector()); - reader_->ReadNext(&buffer_.back()); - } else { + iteration_pos_ = 0; + PADDLE_ENFORCE(reader_->HasNext()); + for (size_t i = 0; i < buffer_size_; ++i) { + if (!reader_->HasNext()) { break; } + buffer_.emplace_back(); + reader_->ReadNext(&buffer_.back()); } - // TODO(fengjiayi): 'std::random_shuffle' can be very slow. It needs to be - // optimize. - std::random_shuffle(buffer_.begin(), buffer_.end()); - iteration_pos_ = 0; + std::mt19937 g(seed_); + std::shuffle(buffer_.begin(), buffer_.end(), g); + seed_ = g(); // update seed_; + VLOG(10) << "random buffer size = " << buffer_.size(); } - out->clear(); - if (!buffer_.empty()) { - std::swap(*out, buffer_[iteration_pos_++]); - } - // if buffer_ is empty, the 'out' will return as an empty vector. -} + + size_t buffer_size_; + std::vector> buffer_; + + size_t iteration_pos_; + size_t seed_; +}; class CreateShuffleReaderOp : public framework::OperatorBase { public: @@ -67,10 +80,10 @@ class CreateShuffleReaderOp : public framework::OperatorBase { const platform::Place& dev_place) const override { const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader")) ->Get(); - auto* out = scope.FindVar(Output("Out")) - ->template GetMutable(); - out->Reset( - new ShuffleReader(underlying_reader.Get(), Attr("buffer_size"))); + auto& var = detail::Ref(scope.FindVar(Output("Out"))); + var.GetMutable()->Reset( + new ShuffleReader(underlying_reader.Get(), + static_cast(Attr("buffer_size")))); } }; diff --git a/paddle/fluid/operators/reader/reader_op_registry.cc b/paddle/fluid/operators/reader/reader_op_registry.cc index 7ea4f4b8d9feecac5bc2d0338bbbe9ab7a532040..0ba4f3854431742eb354f8c90eb395f5d7b32b2e 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.cc +++ b/paddle/fluid/operators/reader/reader_op_registry.cc @@ -31,11 +31,16 @@ std::vector RestoreShapes(const std::vector& shape_concat, return res; } +std::unordered_map& FileReaderRegistry() { + static std::unordered_map regs; + return regs; +} + FileReaderMakerBase::FileReaderMakerBase( framework::OpProtoAndCheckerMaker::OpProto* op_proto, framework::OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(op_proto, op_checker) { - AddOutput("Out", "(ReaderHolder) The created random reader."); + AddOutput("Out", "(ReaderHolder) The created random reader.").AsDuplicable(); AddAttr>("shape_concat", "The concat of all data's shapes."); AddAttr>( "ranks", @@ -49,6 +54,10 @@ FileReaderMakerBase::FileReaderMakerBase( } void FileReaderInferShape::operator()(framework::InferShapeContext* ctx) const { + PADDLE_ENFORCE( + !ctx->IsRuntime(), + "'FileReaderInferShape' should only be invoked during compile time."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "The output file reader should not be null."); const auto shape_concat = ctx->Attrs().Get>("shape_concat"); @@ -56,16 +65,14 @@ void FileReaderInferShape::operator()(framework::InferShapeContext* ctx) const { std::vector shapes = RestoreShapes(shape_concat, ranks); ctx->SetReaderDims("Out", shapes); - if (ctx->IsRuntime()) { - const auto lod_levels = ctx->Attrs().Get>("lod_levels"); - PADDLE_ENFORCE_EQ(lod_levels.size(), shapes.size(), - "The number of 'lod_levels'(%d) doesn't match the number " - "of 'shapes'(%d).", - lod_levels.size(), shapes.size()); - framework::VarDesc* reader = - boost::get(ctx->GetOutputVarPtrs("Out")[0]); - reader->SetLoDLevels(lod_levels); - } + const auto lod_levels = ctx->Attrs().Get>("lod_levels"); + PADDLE_ENFORCE_EQ(lod_levels.size(), shapes.size(), + "The number of 'lod_levels'(%d) doesn't match the number " + "of 'shapes'(%d).", + lod_levels.size(), shapes.size()); + framework::VarDesc* reader = + boost::get(ctx->GetOutputVarPtrs("Out")[0]); + reader->SetLoDLevels(lod_levels); } void FileReaderInferVarType::operator()(const framework::OpDesc& op_desc, @@ -77,19 +84,21 @@ void FileReaderInferVarType::operator()(const framework::OpDesc& op_desc, void DecoratedReaderInferShape::operator()( framework::InferShapeContext* ctx) const { + PADDLE_ENFORCE(!ctx->IsRuntime(), + "'DecoratedReaderInferShape' should only be invoked during " + "compile time."); + PADDLE_ENFORCE(ctx->HasInput("UnderlyingReader"), "Input(UnderlyingReader) should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "The output decorated reader should not be null."); ctx->SetReaderDims("Out", ctx->GetReaderDims("UnderlyingReader")); - if (ctx->IsRuntime()) { - framework::VarDesc* in_reader = boost::get( - ctx->GetInputVarPtrs("UnderlyingReader")[0]); - framework::VarDesc* out_reader = - boost::get(ctx->GetOutputVarPtrs("Out")[0]); - out_reader->SetLoDLevels(in_reader->GetLoDLevels()); - } + framework::VarDesc* in_reader = boost::get( + ctx->GetInputVarPtrs("UnderlyingReader")[0]); + framework::VarDesc* out_reader = + boost::get(ctx->GetOutputVarPtrs("Out")[0]); + out_reader->SetLoDLevels(in_reader->GetLoDLevels()); } void DecoratedReaderInferVarType::operator()( const framework::OpDesc& op_desc, framework::BlockDesc* block) const { diff --git a/paddle/fluid/operators/reader/reader_op_registry.h b/paddle/fluid/operators/reader/reader_op_registry.h index d1f0498f4692247cda72fbcbdd5070ddfaa11553..58f9b4ba35546571fd3b1d0c3ce128f18e248f01 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.h +++ b/paddle/fluid/operators/reader/reader_op_registry.h @@ -21,6 +21,20 @@ namespace paddle { namespace operators { namespace reader { +using FileReaderCreator = std::function&)>; + +std::unordered_map& FileReaderRegistry(); + +template +int RegisterFileReader(const std::string& filetype) { + FileReaderRegistry()[filetype] = []( + const std::string& fn, const std::vector& dim) { + return new Reader(fn, dim); + }; + return 0; +} + extern std::vector RestoreShapes( const std::vector& shape_concat, const std::vector& ranks); @@ -73,3 +87,15 @@ class DecoratedReaderMakerBase : public framework::OpProtoAndCheckerMaker { paddle::operators::reader::DecoratedReaderInferShape, \ paddle::framework::EmptyGradOpMaker, \ paddle::operators::reader::DecoratedReaderInferVarType) + +#define REGISTER_FILE_READER(_filetype, _reader) \ + STATIC_ASSERT_GLOBAL_NAMESPACE( \ + _reg_file_reader_##_filetype, \ + "Must use REGISTER_FILE_READER in global namespace"); \ + int TouchFileReader##_filetype() { return 0; } \ + int _reg_file_reader_entry_##filetype = \ + paddle::operators::reader::RegisterFileReader<_reader>(#_filetype) + +#define USE_FILE_READER(filetype) \ + extern int TouchFileReader##filetype(); \ + static int _use_##filetype = TouchFileReader##filetype() diff --git a/paddle/fluid/operators/reduce_op.cc b/paddle/fluid/operators/reduce_op.cc index 69e8f8081e93cb74177eac1a57e0eaf284951e3f..7879367830216cdd875f9f95f95e2a88f282ac64 100644 --- a/paddle/fluid/operators/reduce_op.cc +++ b/paddle/fluid/operators/reduce_op.cc @@ -173,6 +173,15 @@ class ReduceMinOpMaker : public ReduceOpMaker { } }; +class ReduceProdOpMaker : public ReduceOpMaker { + public: + ReduceProdOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : ReduceOpMaker(proto, op_checker) { + SetComment("ReduceProd", "production"); + AddComment(comment_); + } +}; + } // namespace operators } // namespace paddle @@ -190,6 +199,9 @@ REGISTER_OP(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_max_grad, REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, reduce_min_grad, ops::ReduceGradOp); +REGISTER_OP(reduce_prod, ops::ReduceOp, ops::ReduceProdOpMaker, + reduce_prod_grad, ops::ReduceGradOp); + #define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor) \ REGISTER_OP_CPU_KERNEL(reduce_type, \ ops::ReduceKernel + void operator()(const DeviceContext& place, X& x, Y& y, const Dim& dim) { + y.device(place) = x.prod(dim); + } +}; + +struct ProdGradFunctor { + template + void operator()(const DeviceContext& place, X& x, Y& y, DX& dx, DY& dy, + const Dim& dim, int size) { + dx.device(place) = dy.broadcast(dim) * y.broadcast(dim) * x.inverse(); + } +}; + template class ReduceKernel : public framework::OpKernel { public: @@ -254,4 +270,5 @@ class ReduceGradKernel : public framework::OpKernel { __macro(reduce_sum, SumFunctor, SumGradFunctor); \ __macro(reduce_mean, MeanFunctor, MeanGradFunctor); \ __macro(reduce_max, MaxFunctor, MaxOrMinGradFunctor); \ - __macro(reduce_min, MinFunctor, MaxOrMinGradFunctor); + __macro(reduce_min, MinFunctor, MaxOrMinGradFunctor); \ + __macro(reduce_prod, ProdFunctor, ProdGradFunctor); diff --git a/paddle/fluid/operators/scatter_op.cc b/paddle/fluid/operators/scatter_op.cc index 3fb8b56d2676f90ff7e1cefa46c459ee37f63ca8..d6fd6214711f4ee66b1daffa4db2e84aa7201e79 100644 --- a/paddle/fluid/operators/scatter_op.cc +++ b/paddle/fluid/operators/scatter_op.cc @@ -23,24 +23,24 @@ class ScatterOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("Ref"), - "Input(Ref) of ScatterOp should not be null."); - PADDLE_ENFORCE(ctx->HasInput("Index"), - "Input(Index) of ScatterOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ScatterOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Ids"), + "Input(Ids) of ScatterOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Updates"), "Input(Updates) of ScatterOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of ScatterOp should not be null."); auto updates_dims = ctx->GetInputDim("Updates"); - auto ref_dims = ctx->GetInputDim("Ref"); - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Index").size(), 1, - "Update Index should be 1-D."); + auto ref_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Ids").size(), 1, + "Update Ids should be 1-D."); PADDLE_ENFORCE_EQ(ref_dims.size(), updates_dims.size(), - "Reference and Updates should have the same shape size"); + "Xerence and Updates should have the same shape size"); PADDLE_ENFORCE_EQ(ctx->GetInputDim("Updates")[0], - ctx->GetInputDim("Index")[0], - "Updates and Index should have same batch-size."); + ctx->GetInputDim("Ids")[0], + "Updates and Ids should have same batch-size."); framework::DDim data_dim(updates_dims); for (int i = 1; i < data_dim.size(); ++i) { PADDLE_ENFORCE_EQ(data_dim[i], updates_dims[i]); @@ -52,7 +52,7 @@ class ScatterOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType(ctx.Input("Ref")->type()), + framework::ToDataType(ctx.Input("X")->type()), ctx.device_context()); } }; @@ -64,14 +64,14 @@ class ScatterGradOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override { ctx->SetOutputDim(framework::GradVarName("Updates"), ctx->GetInputDim("Updates")); - ctx->SetOutputDim(framework::GradVarName("Ref"), ctx->GetInputDim("Ref")); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( - framework::ToDataType(ctx.Input("Ref")->type()), + framework::ToDataType(ctx.Input("X")->type()), ctx.device_context()); } }; @@ -80,9 +80,8 @@ class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { public: ScatterOpMaker(OpProto* proto, OpAttrChecker* op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("Ref", "The source input of scatter op"); - AddInput("Index", - "The index input of scatter op where Ref will be updated"); + AddInput("X", "The source input of scatter op"); + AddInput("Ids", "The index input of scatter op where X will be updated"); AddInput("Updates", "The updated value of updates op"); AddOutput("Out", "The output of add op"); AddComment(R"DOC( @@ -91,8 +90,8 @@ Scatter Operator. This operator obtains output by updating the input on selected indices on the first axis: $$ -Out = Ref \\ -Out[Index] = Ref[Index] + Updates +Out = X \\ +Out[Ids] = X[Ids] + Updates $$ )DOC"); diff --git a/paddle/fluid/operators/scatter_op.cu b/paddle/fluid/operators/scatter_op.cu index bdabb29fa680f8f87873b4381acf0dbd2b6195d0..ef7d700659d8d713715a10910baf739954ba0786 100644 --- a/paddle/fluid/operators/scatter_op.cu +++ b/paddle/fluid/operators/scatter_op.cu @@ -25,14 +25,14 @@ class ScatterOpCUDAKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext &ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), "This kernel only runs on GPU device."); - auto *Ref = ctx.Input("Ref"); - auto *Index = ctx.Input("Index"); + auto *X = ctx.Input("X"); + auto *Ids = ctx.Input("Ids"); auto *Updates = ctx.Input("Updates"); auto *Out = ctx.Output("Out"); - Out->ShareDataWith(*Ref); + Out->ShareDataWith(*X); - GPUScatterAssign(ctx.device_context(), *Updates, *Index, Out); + GPUScatterAssign(ctx.device_context(), *Updates, *Ids, Out); } }; @@ -42,16 +42,16 @@ class ScatterGradOpCUDAKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext &ctx) const override { PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), "This kernel only runs on GPU device."); - auto *dRef = ctx.Output(framework::GradVarName("Ref")); + auto *dX = ctx.Output(framework::GradVarName("X")); auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); - auto *Index = ctx.Input("Index"); + auto *Ids = ctx.Input("Ids"); auto *dOut = ctx.Input(framework::GradVarName("Out")); - // In place gradient: dRef = dO - dRef->ShareDataWith(*dOut); + // In place gradient: dX = dO + dX->ShareDataWith(*dOut); dUpdates->mutable_data(ctx.GetPlace()); - // Gradient by Gather: dUpdates = dO[Index] - GPUGather(ctx.device_context(), *dOut, *Index, dUpdates); + // Gradient by Gather: dUpdates = dO[Ids] + GPUGather(ctx.device_context(), *dOut, *Ids, dUpdates); } }; diff --git a/paddle/fluid/operators/scatter_op.h b/paddle/fluid/operators/scatter_op.h index 3c6e7ece320229e1a311ef6d7a27387d40be3c2a..2151d8a9240fc88966533f4a07d5cf56b6c1c3bc 100644 --- a/paddle/fluid/operators/scatter_op.h +++ b/paddle/fluid/operators/scatter_op.h @@ -29,15 +29,15 @@ class ScatterOpKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext &ctx) const override { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), "This kernel only runs on CPU."); - auto *Ref = ctx.Input("Ref"); - auto *Index = ctx.Input("Index"); + auto *X = ctx.Input("X"); + auto *Ids = ctx.Input("Ids"); auto *Updates = ctx.Input("Updates"); auto *Out = ctx.Output("Out"); - // In place output: Out = Ref, Out[Index] += Updates - Out->ShareDataWith(*Ref); + // In place output: Out = X, Out[Ids] += Updates + Out->ShareDataWith(*X); // Apply ScatterUpdate: Out[index] += Updates[:] - ScatterAssign(ctx.device_context(), *Updates, *Index, Out); + ScatterAssign(ctx.device_context(), *Updates, *Ids, Out); } }; @@ -47,16 +47,16 @@ class ScatterGradientOpKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext &ctx) const override { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()), "This kernel only runs on CPU."); - auto *dRef = ctx.Output(framework::GradVarName("Ref")); + auto *dX = ctx.Output(framework::GradVarName("X")); auto *dUpdates = ctx.Output(framework::GradVarName("Updates")); - auto *Index = ctx.Input("Index"); + auto *Ids = ctx.Input("Ids"); auto *dOut = ctx.Input(framework::GradVarName("Out")); - // In place gradient: dRef = dO - dRef->ShareDataWith(*dOut); + // In place gradient: dX = dO + dX->ShareDataWith(*dOut); dUpdates->mutable_data(ctx.GetPlace()); - // Gradient by Gather: dUpdates += dO[Index] - CPUGather(ctx.device_context(), *dOut, *Index, dUpdates); + // Gradient by Gather: dUpdates += dO[Ids] + CPUGather(ctx.device_context(), *dOut, *Ids, dUpdates); } }; diff --git a/paddle/fluid/operators/select_op.cc b/paddle/fluid/operators/select_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..8344a239df7b3fcbe91f91a17a3c5958013b55a6 --- /dev/null +++ b/paddle/fluid/operators/select_op.cc @@ -0,0 +1,414 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +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 +#include +#include +#include "paddle/fluid/framework/channel.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/concurrency/channel_util.h" + +namespace paddle { +namespace operators { + +static constexpr char kX[] = "X"; +static constexpr char kCaseToExecute[] = "case_to_execute"; + +static constexpr char kCases[] = "cases"; +static constexpr char kCasesBlock[] = "sub_block"; + +class SelectOp : public framework::OperatorBase { + public: + SelectOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : framework::OperatorBase(type, inputs, outputs, attrs) {} + + private: + enum class SelectOpCaseType { + DEFAULT = 0, + SEND = 1, + RECEIVE = 2, + }; + + struct SelectOpCase { + int caseIndex; + SelectOpCaseType caseType; + std::string channelName; + std::string varName; + + SelectOpCase() {} + + SelectOpCase(int caseIndex, SelectOpCaseType caseType, + std::string channelName, std::string varName) + : caseIndex(caseIndex), + caseType(caseType), + channelName(channelName), + varName(varName) {} + }; + + void RunImpl(const framework::Scope &scope, + const platform::Place &dev_place) const override { + std::vector casesConfigs = + Attr>(kCases); + + framework::BlockDesc *casesBlock = + Attr(kCasesBlock); + + framework::Scope &casesBlockScope = scope.NewScope(); + + std::string caseToExecuteVarName = Input(kCaseToExecute); + framework::Variable *caseToExecuteVar = + casesBlockScope.FindVar(caseToExecuteVarName); + + // Construct cases from "conditional_block_op"(s) in the casesBlock + std::vector> cases = + ParseAndShuffleCases(&casesConfigs); + + // Get all unique channels involved in select + std::set channelsSet; + for (auto c : cases) { + if (!c->channelName.empty()) { + auto channelVar = scope.FindVar(c->channelName); + framework::ChannelHolder *ch = + channelVar->GetMutable(); + + if (channelsSet.find(ch) == channelsSet.end()) { + channelsSet.insert(ch); + } + } + } + + // Order all channels by their pointer address + std::vector channels(channelsSet.begin(), + channelsSet.end()); + std::sort(channels.begin(), channels.end()); + + // Poll all cases + int32_t caseToExecute = pollCases(&scope, &cases, channels); + + // At this point, the case to execute has already been determined, + // so we can proceed with executing the cases block + framework::LoDTensor *caseToExecuteTensor = + caseToExecuteVar->GetMutable(); + caseToExecuteTensor->data()[0] = caseToExecute; + + // Execute the cases block, only one case will be executed since we set the + // case_to_execute value to the index of the case we want to execute + framework::Executor executor(dev_place); + framework::ProgramDesc *program = casesBlock->Program(); + executor.Run(*program, &casesBlockScope, casesBlock->ID(), + false /*create_local_scope*/); + } + + /** + * Goes through all operators in the casesConfigs and processes + * "conditional_block" operators. These operators are mapped to our + * SelectOpCase objects. We randomize the case orders, and set the + * default case (if any exists) as the last case) + * @param casesBlock + * @return + */ + std::vector> ParseAndShuffleCases( + std::vector *casesConfigs) const { + std::vector> cases; + std::shared_ptr defaultCase; + + if (casesConfigs != nullptr) { + boost::char_delimiters_separator sep(false, ",", ""); + for (std::vector::iterator itr = casesConfigs->begin(); + itr < casesConfigs->end(); ++itr) { + std::string caseConfig = *itr; + boost::tokenizer<> tokens(caseConfig, sep); + + boost::tokenizer<>::iterator tok_iter = tokens.begin(); + PADDLE_ENFORCE(tok_iter != tokens.end(), "Cannot get case index"); + std::string caseIndexString = *tok_iter; + int caseIndex = std::stoi(caseIndexString); + + ++tok_iter; + PADDLE_ENFORCE(tok_iter != tokens.end(), "Cannot get case type"); + std::string caseTypeString = *tok_iter; + SelectOpCaseType caseType = (SelectOpCaseType)std::stoi(caseTypeString); + + std::string caseChannel; + std::string caseChannelVar; + + ++tok_iter; + if (caseType != SelectOpCaseType::DEFAULT) { + PADDLE_ENFORCE(tok_iter != tokens.end(), "Cannot get case channel"); + caseChannel = *tok_iter; + + ++tok_iter; + PADDLE_ENFORCE(tok_iter != tokens.end(), + "Cannot get case channel variable"); + caseChannelVar = *tok_iter; + } + + auto c = std::make_shared(caseIndex, caseType, + caseChannel, caseChannelVar); + + if (caseType == SelectOpCaseType::DEFAULT) { + PADDLE_ENFORCE(defaultCase == nullptr, + "Select can only contain one default case."); + defaultCase = c; + } else { + cases.push_back(c); + } + } + } + + // Randomly sort cases, with default case being last + std::random_shuffle(cases.begin(), cases.end()); + if (defaultCase != nullptr) { + cases.push_back(defaultCase); + } + + return cases; + } + + /** + * This method will recursively poll the cases and determines if any case + * condition is true. + * If none of the cases conditions are true (and there is no default case), + * then block + * the thread. The thread may be woken up by a channel operation, at which + * point we + * execute the case. + * @param scope + * @param cases + * @param channels + * @return + */ + int32_t pollCases(const framework::Scope *scope, + std::vector> *cases, + std::vector channels) const { + // Lock all involved channels + lockChannels(channels); + + std::atomic caseToExecute(-1); + + std::vector>::iterator it = cases->begin(); + while (it != cases->end()) { + std::shared_ptr c = *it; + + auto chVar = scope->FindVar(c->channelName); + framework::ChannelHolder *ch = + chVar->GetMutable(); + + switch (c->caseType) { + case SelectOpCaseType::SEND: + PADDLE_ENFORCE(!ch->IsClosed(), "Cannot send to a closed channel"); + if (ch->CanSend()) { + // We can send to channel directly, send the data to channel + // and execute case + auto chVar = scope->FindVar(c->varName); + concurrency::ChannelSend(ch, chVar); + caseToExecute = c->caseIndex; + } + break; + case SelectOpCaseType::RECEIVE: + if (ch->CanReceive()) { + // We can receive from channel directly, send the data to channel + // and execute case + auto chVar = scope->FindVar(c->varName); + concurrency::ChannelReceive(ch, chVar); + caseToExecute = c->caseIndex; + } + break; + case SelectOpCaseType::DEFAULT: + caseToExecute = c->caseIndex; + break; + } + + if (caseToExecute != -1) { + // We found a case to execute, stop looking at other case statements + break; + } + + ++it; + } + + if (caseToExecute == -1) { + // None of the cases are eligible to execute, enqueue current thread + // into all the sending/receiving queue of each involved channel + std::atomic completed(false); + std::recursive_mutex mutex; + std::unique_lock lock{mutex}; + // std::condition_variable_any selectCond; + auto selectCond = std::make_shared(); + + std::recursive_mutex callbackMutex; + pushThreadOnChannelQueues(scope, cases, selectCond, caseToExecute, + completed, callbackMutex); + + // TODO(thuan): Atomically unlock all channels and sleep current thread + unlockChannels(channels); + selectCond->wait(lock, [&completed]() { return completed.load(); }); + + // Select has been woken up by case operation + lockChannels(channels); + removeThreadOnChannelQueues(scope, cases); + + if (caseToExecute == -1) { + // Recursively poll cases, since we were woken up by a channel close + // TODO(thuan): Need to test if this is a valid case + unlockChannels(channels); + return pollCases(scope, cases, channels); + } + } + + // At this point, caseToExecute != -1, and we can proceed with executing + // the case block + unlockChannels(channels); + + return caseToExecute; + } + + void lockChannels(std::vector chs) const { + std::vector::iterator it = chs.begin(); + while (it != chs.end()) { + framework::ChannelHolder *ch = *it; + ch->Lock(); + ++it; + } + } + + void unlockChannels(std::vector chs) const { + std::vector::reverse_iterator it = chs.rbegin(); + while (it != chs.rend()) { + framework::ChannelHolder *ch = *it; + ch->Unlock(); + ++it; + } + } + + void pushThreadOnChannelQueues( + const framework::Scope *scope, + std::vector> *cases, + std::shared_ptr rCond, + std::atomic &caseToExecute, std::atomic &completed, + std::recursive_mutex &callbackMutex) const { + std::vector>::iterator it = cases->begin(); + while (it != cases->end()) { + std::shared_ptr c = *it; + + auto chVar = scope->FindVar(c->channelName); + framework::ChannelHolder *ch = + chVar->GetMutable(); + + std::function cb = + [&caseToExecute, &completed, &callbackMutex, + c](framework::ChannelAction channelAction) { + std::lock_guard lock{callbackMutex}; + + bool canProcess = false; + if (!completed) { + // If the channel wasn't closed, we set the caseToExecute index + // as this current case + if (channelAction != framework::ChannelAction::CLOSE) { + caseToExecute = c->caseIndex; + } + // This will allow our conditional variable to break out of wait + completed = true; + canProcess = true; + } + + return canProcess; + }; + + switch (c->caseType) { + case SelectOpCaseType::SEND: { + auto chOutputVar = scope->FindVar(c->varName); + concurrency::ChannelAddToSendQ(ch, this, chOutputVar, rCond, cb); + break; + } + case SelectOpCaseType::RECEIVE: { + auto chOutputVar = scope->FindVar(c->varName); + concurrency::ChannelAddToReceiveQ(ch, this, chOutputVar, rCond, cb); + break; + } + default: + break; + } + ++it; + } + } + + void removeThreadOnChannelQueues( + const framework::Scope *scope, + std::vector> *cases) const { + std::vector>::iterator it = cases->begin(); + while (it != cases->end()) { + std::shared_ptr c = *it; + + auto chVar = scope->FindVar(c->channelName); + framework::ChannelHolder *ch = + chVar->GetMutable(); + switch (c->caseType) { + case SelectOpCaseType::SEND: { + ch->RemoveFromSendQ(this); + break; + } + case SelectOpCaseType::RECEIVE: { + ch->RemoveFromReceiveQ(this); + break; + } + default: + break; + } + ++it; + } + } +}; + +class SelectOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SelectOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput(kX, + "A set of variables, which are required by operators inside the " + "cases of Select Op") + .AsDuplicable(); + AddInput(kCaseToExecute, + "(Int) The variable the sets the index of the case to execute, " + "after evaluating the channels being sent to and received from") + .AsDuplicable(); + AddAttr>(kCases, + "(String vector) Serialized list of" + "all cases in the select op. Each" + "case is serialized as: " + "',,,'" + "where type is 0 for default, 1 for" + "send, and 2 for receive" + "No channel and values are needed for" + "default cases."); + AddAttr(kCasesBlock, + "The cases block inside select_op"); + AddComment(R"DOC( +)DOC"); + } +}; + +// TODO(thuan): Implement Gradient Operator for SELECT_OP + +} // namespace operators +} // namespace paddle + +REGISTER_OPERATOR(select, paddle::operators::SelectOp, + paddle::framework::EmptyGradOpMaker, + paddle::operators::SelectOpMaker); diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index 178976f96fdbd08cead7b7c518ea1fbaaa2a5db8..443f40e803ea31c3961ed77842bd0775e0f74f35 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -24,15 +24,15 @@ limitations under the License. */ namespace paddle { namespace operators { -static bool IsVariableInitialized(const framework::Scope& scope, - const std::string& varname) { +static bool NeedSend(const framework::Scope& scope, + const std::string& varname) { auto* var = scope.FindVar(varname); PADDLE_ENFORCE_NOT_NULL(var, "Can not find variable '%s' in the send side.", varname); if (var->IsType()) { return var->Get().IsInitialized(); } else if (var->IsType()) { - return var->Get().value().IsInitialized(); + return var->Get().rows().size() > 0UL; } else { PADDLE_THROW( "Variable type in send side should be in " @@ -67,7 +67,7 @@ class SendOp : public framework::OperatorBase { detail::RPCClient* rpc_client = client_var->GetMutable(); for (size_t i = 0; i < ins.size(); i++) { - if (IsVariableInitialized(scope, ins[i])) { + if (NeedSend(scope, ins[i])) { VLOG(3) << "sending " << ins[i] << " to " << epmap[i]; rpc_client->AsyncSendVariable(epmap[i], ctx, scope, ins[i]); } else { @@ -88,6 +88,12 @@ class SendOp : public framework::OperatorBase { rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]); } PADDLE_ENFORCE(rpc_client->Wait()); + // tell pservers that current trainer have called fetch + for (auto& ep : endpoints) { + VLOG(3) << "send fetch barrier, ep: " << ep; + rpc_client->AsyncSendFetchBarrier(ep); + } + PADDLE_ENFORCE(rpc_client->Wait()); } } }; diff --git a/paddle/fluid/operators/sequence_softmax_cudnn_op.cu.cc b/paddle/fluid/operators/sequence_softmax_cudnn_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..5661f4b42f37fed7f589c515e25fd66cfcede2c7 --- /dev/null +++ b/paddle/fluid/operators/sequence_softmax_cudnn_op.cu.cc @@ -0,0 +1,105 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +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/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/softmax.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +template +class SequenceSoftmaxCUDNNKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* out = ctx.Output("Out"); + + auto lod = x->lod(); + auto dims = x->dims(); + + const size_t level = lod.size() - 1; + PADDLE_ENFORCE_EQ(dims[0], static_cast(lod[level].back()), + "The first dimension of Input(X) should be equal to the " + "sum of all sequences' lengths."); + PADDLE_ENFORCE_EQ(dims[0], x->numel(), + "The width of each timestep in Input(X) of " + "SequenceSoftmaxOp should be 1."); + + out->mutable_data(ctx.GetPlace()); + for (int i = 0; i < static_cast(lod[level].size()) - 1; ++i) { + int start_pos = static_cast(lod[level][i]); + int end_pos = static_cast(lod[level][i + 1]); + Tensor x_i = x->Slice(start_pos, end_pos); + Tensor out_i = out->Slice(start_pos, end_pos); + + // Reshape from (end_pos - start_pos) x 1UL to 1UL x (end_pos - start_pos) + framework::DDim dims_i = + // framework::make_ddim({1UL, end_pos - start_pos, 1UL, 1UL}); + framework::make_ddim({1UL, end_pos - start_pos}); + x_i.Resize(dims_i); + out_i.Resize(dims_i); + math::SoftmaxCUDNNFunctor()( + ctx.template device_context(), &x_i, + &out_i); + } + } +}; + +template +class SequenceSoftmaxGradCUDNNKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* out = ctx.Input("Out"); + auto* out_grad = ctx.Input(framework::GradVarName("Out")); + auto* x = ctx.Input("X"); + auto* x_grad = ctx.Output(framework::GradVarName("X")); + + auto lod = x->lod(); + const size_t level = lod.size() - 1; + + x_grad->mutable_data(ctx.GetPlace()); + for (int i = 0; i < static_cast(lod[level].size()) - 1; ++i) { + int start_pos = static_cast(lod[level][i]); + int end_pos = static_cast(lod[level][i + 1]); + + Tensor out_i = out->Slice(start_pos, end_pos); + Tensor out_grad_i = out_grad->Slice(start_pos, end_pos); + Tensor x_grad_i = x_grad->Slice(start_pos, end_pos); + + // Reshape from (end_pos - start_pos) x 1UL to 1UL x (end_pos - start_pos) + framework::DDim dims_i = framework::make_ddim({1UL, end_pos - start_pos}); + out_i.Resize(dims_i); + out_grad_i.Resize(dims_i); + x_grad_i.Resize(dims_i); + math::SoftmaxGradCUDNNFunctor()( + ctx.template device_context(), &out_i, + &out_grad_i, &x_grad_i); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_KERNEL(sequence_softmax, CUDNN, ::paddle::platform::CUDAPlace, + ops::SequenceSoftmaxCUDNNKernel, + ops::SequenceSoftmaxCUDNNKernel) +REGISTER_OP_KERNEL(sequence_softmax_grad, CUDNN, ::paddle::platform::CUDAPlace, + ops::SequenceSoftmaxGradCUDNNKernel, + ops::SequenceSoftmaxGradCUDNNKernel) diff --git a/paddle/fluid/operators/sequence_softmax_op.cc b/paddle/fluid/operators/sequence_softmax_op.cc index 7e685eb3dc7b12ef38f06b37d99a1212cfbc992c..e8b4df04286d327f568f4c43886f9fcf89cc4a88 100644 --- a/paddle/fluid/operators/sequence_softmax_op.cc +++ b/paddle/fluid/operators/sequence_softmax_op.cc @@ -29,6 +29,29 @@ class SequenceSoftmaxOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Out", ctx->GetInputDim("X")); ctx->ShareLoD("X", /*->*/ "Out"); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + // choose cudnn kernel if the runtime supported. + bool use_cudnn = ctx.Attr("use_cudnn"); + bool runtime_cudnn_support = false; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + auto& dev_ctx = + ctx.template device_context(); + runtime_cudnn_support = dev_ctx.cudnn_handle() != nullptr ? true : false; + } +#endif + framework::LibraryType library_ = framework::LibraryType::kPlain; + if (use_cudnn && runtime_cudnn_support) { + library_ = framework::LibraryType::kCUDNN; + } + std::string data_format = ctx.Attr("data_format"); + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + framework::StringToDataLayout(data_format), library_); + } }; class SequenceSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { @@ -41,6 +64,17 @@ class SequenceSoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("Out", "(LoDTensor) 1-D or 2-D output LoDTensor with the 2-nd dimension " "of length 1."); + AddAttr( + "use_cudnn", + "(bool, default false) Only used in cudnn kernel, need install cudnn") + .SetDefault(false); + AddAttr( + "data_format", + "(string, default NCHW) Only used in " + "An optional string from: \"NHWC\", \"NCHW\". " + "Defaults to \"NHWC\". Specify the data format of the output data, " + "the input will be transformed automatically. ") + .SetDefault("AnyLayout"); AddComment(R"DOC( Sequence Softmax Operator. @@ -91,6 +125,29 @@ class SequenceSoftmaxGradOp : public framework::OperatorWithKernel { ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + // choose cudnn kernel if the runtime supported. + bool use_cudnn = ctx.Attr("use_cudnn"); + bool runtime_cudnn_support = false; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + auto& dev_ctx = + ctx.template device_context(); + runtime_cudnn_support = dev_ctx.cudnn_handle() != nullptr ? true : false; + } +#endif + framework::LibraryType library_ = framework::LibraryType::kPlain; + if (use_cudnn && runtime_cudnn_support) { + library_ = framework::LibraryType::kCUDNN; + } + std::string data_format = ctx.Attr("data_format"); + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + framework::StringToDataLayout(data_format), library_); + } }; } // namespace operators @@ -102,7 +159,9 @@ REGISTER_OP(sequence_softmax, ops::SequenceSoftmaxOp, ops::SequenceSoftmaxGradOp); REGISTER_OP_CPU_KERNEL( sequence_softmax, - ops::SequenceSoftmaxKernel); + ops::SequenceSoftmaxKernel, + ops::SequenceSoftmaxKernel); REGISTER_OP_CPU_KERNEL( sequence_softmax_grad, - ops::SequenceSoftmaxGradKernel); + ops::SequenceSoftmaxGradKernel, + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/fluid/operators/sequence_softmax_op.cu.cc b/paddle/fluid/operators/sequence_softmax_op.cu.cc index 295c68c5b936d6522666a4cc4e621db6f5f5f3ed..57adea3a1b9dbcbb5787d005e4d3ec595f61d4b2 100644 --- a/paddle/fluid/operators/sequence_softmax_op.cu.cc +++ b/paddle/fluid/operators/sequence_softmax_op.cu.cc @@ -17,7 +17,10 @@ limitations under the License. */ namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( sequence_softmax, - ops::SequenceSoftmaxKernel) + ops::SequenceSoftmaxKernel, + ops::SequenceSoftmaxKernel) REGISTER_OP_CUDA_KERNEL( sequence_softmax_grad, - ops::SequenceSoftmaxGradKernel); + ops::SequenceSoftmaxGradKernel, + ops::SequenceSoftmaxGradKernel); diff --git a/paddle/fluid/operators/sgd_op.cc b/paddle/fluid/operators/sgd_op.cc index 7cc73de8788e3dceb763b6f5a1519459d0fb05dd..d0aa2f9cbadaadf4e7e625628d9db5677d50d277 100644 --- a/paddle/fluid/operators/sgd_op.cc +++ b/paddle/fluid/operators/sgd_op.cc @@ -39,6 +39,14 @@ class SGDOp : public framework::OperatorWithKernel { // and run time. ctx->SetOutputDim("ParamOut", param_dim); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Param")->type()), + ctx.GetPlace()); + } }; class SGDOpMaker : public framework::OpProtoAndCheckerMaker { diff --git a/paddle/fluid/operators/sgd_op.h b/paddle/fluid/operators/sgd_op.h index 2fec84815a9ecc63675de88816b23cfaa75aca65..0ad801079400f1830d85a945e57a434a86adeb00 100644 --- a/paddle/fluid/operators/sgd_op.h +++ b/paddle/fluid/operators/sgd_op.h @@ -47,6 +47,12 @@ class SGDOpKernel : public framework::OpKernel { PADDLE_ENFORCE_EQ(param, param_out); auto* grad = ctx.Input("Grad"); + // for distributed training, a sparse var may be empty, + // just skip updating. + if (grad->rows().size() == 0) { + return; + } + auto in_height = grad->height(); auto out_dims = param_out->dims(); PADDLE_ENFORCE_EQ(in_height, out_dims[0]); @@ -60,13 +66,15 @@ class SGDOpKernel : public framework::OpKernel { auto* in_data = in_value.data(); auto* out_data = param_out->data(); auto* lr = learning_rate->data(); - for (size_t i = 0; i < in_rows.size(); i++) { + PADDLE_ENFORCE(in_rows[i] < in_height, + "Input rows index should less than height"); for (int64_t j = 0; j < in_row_numel; j++) { out_data[in_rows[i] * in_row_numel + j] -= lr[0] * in_data[i * in_row_numel + j]; } } + } else { PADDLE_THROW("Unsupported Variable Type of Grad"); } diff --git a/paddle/fluid/operators/softmax_cudnn_op.cu.cc b/paddle/fluid/operators/softmax_cudnn_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..47cb336d87f8627d86ac33d6ac32c04d5d93f753 --- /dev/null +++ b/paddle/fluid/operators/softmax_cudnn_op.cu.cc @@ -0,0 +1,62 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +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/fluid/operators/math/softmax.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class SoftmaxCUDNNKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* X = context.Input("X"); + auto* Out = context.Output("Out"); + + // allocate memory on device. + Out->mutable_data(context.GetPlace()); + + math::SoftmaxCUDNNFunctor()( + context.template device_context(), X, Out); + } +}; + +template +class SoftmaxGradCUDNNKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* Out = context.Input("Out"); + auto* dOut = context.Input(framework::GradVarName("Out")); + auto* dX = context.Output(framework::GradVarName("X")); + + // allocate memory on device. + dX->mutable_data(context.GetPlace()); + + math::SoftmaxGradCUDNNFunctor()( + context.template device_context(), Out, + dOut, dX); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_KERNEL(softmax, CUDNN, ::paddle::platform::CUDAPlace, + ops::SoftmaxCUDNNKernel); +REGISTER_OP_KERNEL(softmax_grad, CUDNN, ::paddle::platform::CUDAPlace, + ops::SoftmaxGradCUDNNKernel); diff --git a/paddle/fluid/operators/softmax_op.cc b/paddle/fluid/operators/softmax_op.cc index 09275ef290e8c78dc0902033e904cc4e7ccd7adb..1b63f8a499e5d20d2f10c3cd1024d1bcf78764d4 100644 --- a/paddle/fluid/operators/softmax_op.cc +++ b/paddle/fluid/operators/softmax_op.cc @@ -33,6 +33,29 @@ class SoftmaxOp : public framework::OperatorWithKernel { ctx->SetOutputDim("Out", x_dims); ctx->ShareLoD("X", /*->*/ "Out"); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + // choose cudnn kernel if the runtime supported. + bool use_cudnn = ctx.Attr("use_cudnn"); + bool runtime_cudnn_support = false; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + auto& dev_ctx = + ctx.template device_context(); + runtime_cudnn_support = dev_ctx.cudnn_handle() != nullptr ? true : false; + } +#endif + framework::LibraryType library_ = framework::LibraryType::kPlain; + if (use_cudnn && runtime_cudnn_support) { + library_ = framework::LibraryType::kCUDNN; + } + std::string data_format = ctx.Attr("data_format"); + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + framework::StringToDataLayout(data_format), library_); + } }; class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { @@ -43,6 +66,17 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { "The input tensor of softmax. " "2-D with shape [batch_size, input_feature_dimensions]."); AddOutput("Out", "The normalized values with the same shape as X."); + AddAttr( + "use_cudnn", + "(bool, default false) Only used in cudnn kernel, need install cudnn") + .SetDefault(false); + AddAttr( + "data_format", + "(string, default NCHW) Only used in " + "An optional string from: \"NHWC\", \"NCHW\". " + "Defaults to \"NHWC\". Specify the data format of the output data, " + "the input will be transformed automatically. ") + .SetDefault("AnyLayout"); AddComment(R"DOC( Softmax Operator. @@ -80,6 +114,29 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel { ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + // choose cudnn kernel if the runtime supported. + bool use_cudnn = ctx.Attr("use_cudnn"); + bool runtime_cudnn_support = false; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + auto& dev_ctx = + ctx.template device_context(); + runtime_cudnn_support = dev_ctx.cudnn_handle() != nullptr ? true : false; + } +#endif + framework::LibraryType library_ = framework::LibraryType::kPlain; + if (use_cudnn && runtime_cudnn_support) { + library_ = framework::LibraryType::kCUDNN; + } + std::string data_format = ctx.Attr("data_format"); + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + framework::StringToDataLayout(data_format), library_); + } }; } // namespace operators diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu index 39b246a5bedb2819fc9b7fc407cfe03e59af0b68..8f7840cee1dd95a828fd4ac8815e335a5db47e3d 100644 --- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu +++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu @@ -23,21 +23,21 @@ using Tensor = framework::Tensor; namespace { template -__global__ void CrossEntropyGrad(T* logit_grad, const T* loss_grad, - const int64_t* labels, const int batch_size, - const int class_num) { - int tid = blockIdx.x * blockDim.x + threadIdx.x; - int sample_idx = tid / class_num; - - if (tid < batch_size) { - PADDLE_ASSERT(labels[sample_idx] >= 0 && labels[sample_idx] < class_num); - logit_grad[tid * class_num + labels[tid]] -= static_cast(1.); +__global__ void CrossEntropyGrad(T* logit_grad, const int64_t* labels, + const int batch_size, const int class_num) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < batch_size; + i += blockDim.x * gridDim.x) { + int idx = i * class_num + labels[i]; + logit_grad[idx] -= static_cast(1.); } +} - __syncthreads(); - - if (tid < batch_size * class_num) { - logit_grad[tid] *= loss_grad[sample_idx]; +template +__global__ void Scale(T* logit_grad, const T* loss_grad, const int num, + const int class_num) { + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; + i += blockDim.x * gridDim.x) { + logit_grad[i] *= loss_grad[i / class_num]; } } @@ -94,22 +94,22 @@ class SoftmaxWithCrossEntropyGradCUDAKernel : public framework::OpKernel { const int batch_size = logit_grad->dims()[0]; const int class_num = logit_grad->dims()[1]; int block = 512; - int grid = (batch_size * class_num + block - 1) / block; + auto stream = context.cuda_device_context().stream(); if (context.Attr("soft_label")) { + int grid = (batch_size * class_num + block - 1) / block; const T* label_data = labels->data(); - SoftCrossEntropyGradientKernel< - T><<() - .stream()>>>(logit_grad_data, loss_grad_data, label_data, - batch_size, class_num); + SoftCrossEntropyGradientKernel<<>>( + logit_grad_data, loss_grad_data, label_data, batch_size, class_num); } else { + int grid = (batch_size + block - 1) / block; const int64_t* label_data = labels->data(); - CrossEntropyGrad< - T><<() - .stream()>>>(logit_grad_data, loss_grad_data, label_data, - batch_size, class_num); + CrossEntropyGrad<<>>( + logit_grad_data, label_data, batch_size, class_num); + int num = batch_size * class_num; + grid = (num + block - 1) / block; + Scale<<>>(logit_grad_data, loss_grad_data, num, + class_num); } } }; diff --git a/paddle/fluid/operators/split_selected_rows_op.h b/paddle/fluid/operators/split_selected_rows_op.h index 23baf8e72eca87f3865fc2b63ce2de96f799dce3..0e9ce165b98845f4745ee70b028513ea31cc6657 100644 --- a/paddle/fluid/operators/split_selected_rows_op.h +++ b/paddle/fluid/operators/split_selected_rows_op.h @@ -21,15 +21,24 @@ limitations under the License. */ namespace paddle { namespace operators { -static int FindOutIdx(int row, const std::vector& height_sections) { - int offset = 0; - for (size_t i = 0; i < height_sections.size(); ++i) { - if (row >= offset && row < (offset + height_sections[i])) { - return i; +static int FindOutIdx(int row, const std::vector& abs_sections) { + for (size_t i = 1; i < abs_sections.size(); ++i) { + if (row < abs_sections[i]) { + return i - 1; } - offset += height_sections[i]; } - return -1; + return abs_sections.size() - 1; +} + +static std::vector ToAbsoluteSection( + const std::vector& height_sections) { + std::vector abs_sections; + abs_sections.resize(height_sections.size()); + abs_sections[0] = 0; + for (size_t i = 1; i < height_sections.size(); ++i) { + abs_sections[i] = height_sections[i - 1] + abs_sections[i - 1]; + } + return abs_sections; } template @@ -40,16 +49,23 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel { auto outs = ctx.MultiOutput("Out"); auto height_sections = ctx.Attr>("height_sections"); + auto abs_sections = ToAbsoluteSection(height_sections); + auto x_rows = x->rows(); std::vector> outs_rows_idx; + std::vector> outs_dense_idx; + outs_rows_idx.resize(outs.size()); + outs_dense_idx.resize(outs.size()); auto row_numel = x->value().numel() / x->value().dims()[0]; auto src = x->value().data(); + // split rows index into output sparse vars for (size_t i = 0; i < x_rows.size(); ++i) { - int out_idx = FindOutIdx(x_rows[i], height_sections); - outs_rows_idx[out_idx].push_back(i); + int out_idx = FindOutIdx(x_rows[i], abs_sections); + outs_rows_idx[out_idx].push_back(x_rows[i]); + outs_dense_idx[out_idx].push_back(i); } auto place = ctx.GetPlace(); @@ -61,19 +77,20 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel { dims[0] = rows_idx.size(); outs[i]->mutable_value()->mutable_data(dims, x->place()); for (auto idx : rows_idx) { - outs[i]->mutable_rows()->push_back(x_rows[idx]); + outs[i]->mutable_rows()->push_back(idx - abs_sections[i]); } auto dst = outs[i]->mutable_value()->mutable_data(ctx.GetPlace()); for (size_t j = 0; j < rows_idx.size(); j++) { if (platform::is_cpu_place(place)) { - memory::Copy(platform::CPUPlace(), dst + j * row_numel, - platform::CPUPlace(), src + rows_idx[j] * row_numel, - sizeof(T) * row_numel); + memory::Copy( + platform::CPUPlace(), dst + j * row_numel, platform::CPUPlace(), + src + outs_dense_idx[i][j] * row_numel, sizeof(T) * row_numel); } else { #ifdef PADDLE_WITH_CUDA auto stream = ctx.cuda_device_context().stream(); memory::Copy(platform::CUDAPlace(), dst + j * row_numel, - platform::CUDAPlace(), src + rows_idx[j] * row_numel, + platform::CUDAPlace(), + src + outs_dense_idx[i][j] * row_numel, sizeof(T) * row_numel, stream); #else PADDLE_THROW("Paddle is not compiled with GPU"); diff --git a/paddle/fluid/operators/sum_op.cc b/paddle/fluid/operators/sum_op.cc index c3abb3ea4a53126c22c817069e8ad955b202f09d..d3d5c8a3429e2070c5472355b4440401eaa699cb 100644 --- a/paddle/fluid/operators/sum_op.cc +++ b/paddle/fluid/operators/sum_op.cc @@ -76,10 +76,16 @@ class SumOp : public framework::OperatorWithKernel { static_cast(dtype), ctx.device_context()); } else if (x_vars[0]->IsType()) { - return framework::OpKernelType( - framework::ToDataType( - x_vars[0]->Get().value().type()), - ctx.device_context()); + for (auto& var : x_vars) { + auto& value = var->Get().value(); + if (value.IsInitialized()) { + return framework::OpKernelType(framework::ToDataType(value.type()), + ctx.device_context()); + } + } + // if input sparse vars are not initialized, use an default kernel type. + return framework::OpKernelType(framework::proto::VarType::FP32, + ctx.device_context()); } else if (x_vars[0]->IsType()) { for (auto& x_var : x_vars) { auto& array = x_var->Get(); diff --git a/paddle/fluid/operators/sum_op.h b/paddle/fluid/operators/sum_op.h index 48b2d2779aeeb168cf87e61557e01d5cbde476b3..e7e5346cdca5efaf81c2b0fddedde7406e3b874d 100644 --- a/paddle/fluid/operators/sum_op.h +++ b/paddle/fluid/operators/sum_op.h @@ -109,6 +109,12 @@ class SumKernel : public framework::OpKernel { in_dim[0] = static_cast(first_dim); out_value->Resize(framework::make_ddim(in_dim)); + + // if all the input sparse vars are empty, no need to + // merge these vars. + if (first_dim == 0UL) { + return; + } out_value->mutable_data(context.GetPlace()); math::SelectedRowsAddTo functor; @@ -116,7 +122,7 @@ class SumKernel : public framework::OpKernel { int64_t offset = 0; for (int i = 0; i < N; i++) { auto &sel_row = get_selected_row(i); - if (!sel_row.value().IsInitialized() || sel_row.rows().size() == 0) { + if (sel_row.rows().size() == 0) { continue; } PADDLE_ENFORCE_EQ(out->height(), sel_row.height()); diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt index a1a743d94e204ca506c994f8fabb6bbf8c22cea5..7eec6ab657723c6390dfa14a78d6c49a76f2a279 100644 --- a/paddle/fluid/platform/CMakeLists.txt +++ b/paddle/fluid/platform/CMakeLists.txt @@ -48,7 +48,6 @@ nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_ nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda) nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context) -nv_test(nccl_test SRCS nccl_test.cu DEPS dynload_cuda gpu_info device_context) cc_library(device_tracer SRCS device_tracer.cc DEPS profiler_proto ${GPU_CTX_DEPS}) cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer) diff --git a/paddle/fluid/platform/cudnn_helper.h b/paddle/fluid/platform/cudnn_helper.h index 1842ecd745e3f5cb75600ce00d89018f81682632..7e001ecc56173db76e8c576e7efd66f41192f292 100644 --- a/paddle/fluid/platform/cudnn_helper.h +++ b/paddle/fluid/platform/cudnn_helper.h @@ -19,6 +19,7 @@ limitations under the License. */ #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/platform/dynload/cudnn.h" #include "paddle/fluid/platform/enforce.h" +#include "paddle/fluid/platform/float16.h" #include "paddle/fluid/platform/macros.h" namespace paddle { @@ -80,6 +81,22 @@ enum class PoolingMode { template class CudnnDataType; +template <> +class CudnnDataType { + public: + static const cudnnDataType_t type = CUDNN_DATA_HALF; + // The scaling param type is float for HALF and FLOAT tensors + typedef const float ScalingParamType; + static ScalingParamType* kOne() { + static ScalingParamType v = 1.0; + return &v; + } + static ScalingParamType* kZero() { + static ScalingParamType v = 0.0; + return &v; + } +}; + template <> class CudnnDataType { public: @@ -289,7 +306,7 @@ inline bool CanCUDNNBeUsed(const framework::ExecutionContext& ctx) { use_cudnn &= paddle::platform::is_gpu_place(ctx.GetPlace()); #ifdef PADDLE_WITH_CUDA if (use_cudnn) { - auto& dev_ctx = ctx.template device_context(); + auto& dev_ctx = ctx.device_context(); use_cudnn &= dev_ctx.cudnn_handle() != nullptr; } #endif diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc index bb9fbd468f38fffc94107e321e777fc0e772fbe6..98b4178177b0a8bafd6fe34a92be2a07a2fbc5a7 100644 --- a/paddle/fluid/platform/device_context.cc +++ b/paddle/fluid/platform/device_context.cc @@ -127,6 +127,7 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface { CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place) { SetDeviceId(place_.device); + compute_capability = GetCUDAComputeCapability(place_.device); multi_process = GetCUDAMultiProcessors(place_.device); max_threads_per_mp = GetCUDAMaxThreadsPerMultiProcessor(place_.device); PADDLE_ENFORCE(cudaStreamCreate(&stream_)); @@ -162,6 +163,10 @@ void CUDADeviceContext::Wait() const { PADDLE_ENFORCE(cudaGetLastError()); } +int CUDADeviceContext::GetComputeCapability() const { + return compute_capability; +} + int CUDADeviceContext::GetMaxPhysicalThreadCount() const { return multi_process * max_threads_per_mp; } diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h index e779644190de1246cd650fbf91eeaeb03494643f..603b890af13b529c490c29112a73a09cc815d07a 100644 --- a/paddle/fluid/platform/device_context.h +++ b/paddle/fluid/platform/device_context.h @@ -79,6 +79,9 @@ class CUDADeviceContext : public DeviceContext { /*! \brief Return place in the device context. */ Place GetPlace() const override; + /*! \brief Return compute capability in the device context. */ + int GetComputeCapability() const; + /*! \brief Return the max physical thread count in the device context */ int GetMaxPhysicalThreadCount() const; @@ -104,6 +107,7 @@ class CUDADeviceContext : public DeviceContext { cudnnHandle_t cudnn_handle_; cublasHandle_t cublas_handle_; + int compute_capability; int multi_process; int max_threads_per_mp; }; diff --git a/paddle/fluid/platform/device_tracer.cc b/paddle/fluid/platform/device_tracer.cc index e2fd8e90b682a24b06a237cb0e2ba52c6dd2f95c..3b4437f576e1c2e931a86ec6d5e823ec1f344c52 100644 --- a/paddle/fluid/platform/device_tracer.cc +++ b/paddle/fluid/platform/device_tracer.cc @@ -26,8 +26,14 @@ limitations under the License. */ namespace paddle { namespace platform { namespace { +// Current thread's id. Note, we don't distinguish nested threads +// for now. +thread_local int cur_thread_id = 0; +// Tracking the nested block stacks of each thread. +thread_local std::deque block_id_stack; +// Tracking the nested event stacks. +thread_local std::deque annotation_stack; -thread_local const char *cur_annotation = nullptr; std::once_flag tracer_once_flag; DeviceTracer *tracer = nullptr; } // namespace @@ -191,28 +197,37 @@ class DeviceTracerImpl : public DeviceTracer { correlations_[id] = anno; } - void AddCPURecords(const char *anno, uint64_t start_ns, uint64_t end_ns) { - if (!anno) { - // TODO(panyx0718): Currently, it doesn't support nested situation - // Up-level can be cleared by low-level and therefore get nullptr - // here. + void AddCPURecords(const std::string &anno, uint64_t start_ns, + uint64_t end_ns, int64_t device_id, int64_t thread_id) { + if (anno.empty()) { + VLOG(1) << "Empty timeline annotation."; return; } std::lock_guard l(trace_mu_); cpu_records_.push_back( - CPURecord{anno, start_ns, end_ns, - std::hash{}(std::this_thread::get_id())}); + CPURecord{anno, start_ns, end_ns, device_id, thread_id}); } void AddMemRecords(const std::string &name, uint64_t start_ns, - uint64_t end_ns, uint32_t device_id, uint32_t stream_id, + uint64_t end_ns, int64_t device_id, int64_t stream_id, uint32_t correlation_id, uint64_t bytes) { + // 0 means timestamp information could not be collected for the kernel. + if (start_ns == 0 || end_ns == 0) { + VLOG(3) << name << " cannot be traced"; + return; + } + std::lock_guard l(trace_mu_); mem_records_.push_back(MemRecord{name, start_ns, end_ns, device_id, stream_id, correlation_id, bytes}); } - void AddKernelRecords(uint64_t start, uint64_t end, uint32_t device_id, - uint32_t stream_id, uint32_t correlation_id) { + void AddKernelRecords(uint64_t start, uint64_t end, int64_t device_id, + int64_t stream_id, uint32_t correlation_id) { + // 0 means timestamp information could not be collected for the kernel. + if (start == 0 || end == 0) { + VLOG(3) << correlation_id << " cannot be traced"; + return; + } std::lock_guard l(trace_mu_); kernel_records_.push_back( KernelRecord{start, end, device_id, stream_id, correlation_id}); @@ -261,34 +276,37 @@ class DeviceTracerImpl : public DeviceTracer { continue; } auto *event = profile_pb.add_events(); + event->set_type(proto::Event::GPUKernel); event->set_name(correlations_.at(r.correlation_id)); event->set_start_ns(r.start_ns); event->set_end_ns(r.end_ns); - event->set_stream_id(r.stream_id); + event->set_sub_device_id(r.stream_id); event->set_device_id(r.device_id); } for (const CPURecord &r : cpu_records_) { auto *event = profile_pb.add_events(); + event->set_type(proto::Event::CPU); event->set_name(r.name); event->set_start_ns(r.start_ns); event->set_end_ns(r.end_ns); - event->set_stream_id(r.thread_id); - event->set_device_id(-1); + event->set_sub_device_id(r.thread_id); + event->set_device_id(r.device_id); } for (const MemRecord &r : mem_records_) { auto *event = profile_pb.add_events(); + event->set_type(proto::Event::GPUKernel); event->set_name(r.name); event->set_start_ns(r.start_ns); event->set_end_ns(r.end_ns); - event->set_stream_id(r.stream_id); + event->set_sub_device_id(r.stream_id); event->set_device_id(r.device_id); event->mutable_memcopy()->set_bytes(r.bytes); } - std::string profile_str; - google::protobuf::TextFormat::PrintToString(profile_pb, &profile_str); std::ofstream profile_f; profile_f.open(profile_path, std::ios::out | std::ios::trunc); + std::string profile_str; + profile_pb.SerializeToString(&profile_str); profile_f << profile_str; profile_f.close(); return profile_pb; @@ -314,8 +332,9 @@ class DeviceTracerImpl : public DeviceTracer { if ((domain == CUPTI_CB_DOMAIN_DRIVER_API) && (cbid == CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel)) { if (cbInfo->callbackSite == CUPTI_API_ENTER) { - const std::string anno = - cur_annotation ? cur_annotation : cbInfo->symbolName; + const std::string anno = !annotation_stack.empty() + ? annotation_stack.back() + : cbInfo->symbolName; tracer->AddAnnotation(cbInfo->correlationId, anno); } } else { @@ -342,14 +361,15 @@ class DeviceTracerDummy : public DeviceTracer { void AddAnnotation(uint64_t id, const std::string &anno) {} - void AddCPURecords(const char *anno, uint64_t start_ns, uint64_t end_ns) {} + void AddCPURecords(const std::string &anno, uint64_t start_ns, + uint64_t end_ns, int64_t device_id, int64_t thread_id) {} void AddMemRecords(const std::string &name, uint64_t start_ns, - uint64_t end_ns, uint32_t device_id, uint32_t stream_id, + uint64_t end_ns, int64_t device_id, int64_t stream_id, uint32_t correlation_id, uint64_t bytes) {} - void AddKernelRecords(uint64_t start, uint64_t end, uint32_t device_id, - uint32_t stream_id, uint32_t correlation_id) {} + void AddKernelRecords(uint64_t start, uint64_t end, int64_t device_id, + int64_t stream_id, uint32_t correlation_id) {} bool IsEnabled() { return false; } @@ -375,11 +395,28 @@ DeviceTracer *GetDeviceTracer() { return tracer; } -void SetCurAnnotation(const char *anno) { cur_annotation = anno; } +void SetCurAnnotation(const std::string &anno) { + annotation_stack.push_back(anno); +} + +void ClearCurAnnotation() { annotation_stack.pop_back(); } + +std::string CurAnnotation() { + if (annotation_stack.empty()) return ""; + return annotation_stack.back(); +} + +void SetCurBlock(int block_id) { block_id_stack.push_back(block_id); } + +void ClearCurBlock() { block_id_stack.pop_back(); } + +int BlockDepth() { return block_id_stack.size(); } + +void SetCurThread(int thread_id) { cur_thread_id = thread_id; } -void ClearCurAnnotation() { cur_annotation = nullptr; } +void ClearCurThread() { cur_thread_id = 0; } -const char *CurAnnotation() { return cur_annotation; } +int CurThread() { return cur_thread_id; } } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/device_tracer.h b/paddle/fluid/platform/device_tracer.h index 23f7cdbdffc9f48ac5555455bf745233c81dd0cb..deb3d23f786353b8e7a2f28d094e364158885a34 100644 --- a/paddle/fluid/platform/device_tracer.h +++ b/paddle/fluid/platform/device_tracer.h @@ -32,22 +32,23 @@ class DeviceTracer { struct KernelRecord { uint64_t start_ns; uint64_t end_ns; - uint32_t device_id; - uint32_t stream_id; + int64_t device_id; + int64_t stream_id; uint32_t correlation_id; }; struct CPURecord { std::string name; uint64_t start_ns; uint64_t end_ns; - uint64_t thread_id; + int64_t device_id; + int64_t thread_id; }; struct MemRecord { std::string name; uint64_t start_ns; uint64_t end_ns; - uint32_t device_id; - uint32_t stream_id; + int64_t device_id; + int64_t stream_id; uint32_t correlation_id; uint64_t bytes; }; @@ -64,18 +65,18 @@ class DeviceTracer { virtual void AddAnnotation(uint64_t id, const std::string& anno) = 0; virtual void AddMemRecords(const std::string& name, uint64_t start_ns, - uint64_t end_ns, uint32_t device_id, - uint32_t stream_id, uint32_t correlation_id, + uint64_t end_ns, int64_t device_id, + int64_t stream_id, uint32_t correlation_id, uint64_t bytes) = 0; - virtual void AddCPURecords(const char* anno, uint64_t start_ns, - uint64_t end_ns) = 0; + virtual void AddCPURecords(const std::string& anno, uint64_t start_ns, + uint64_t end_ns, int64_t device_id, + int64_t thread_id) = 0; // Add a cuda kernel stats. `correlation_id` will be mapped to annotation // added before for human readability. - virtual void AddKernelRecords(uint64_t start, uint64_t end, - uint32_t device_id, uint32_t stream_id, - uint32_t correlation_id) = 0; + virtual void AddKernelRecords(uint64_t start, uint64_t end, int64_t device_id, + int64_t stream_id, uint32_t correlation_id) = 0; // Generate a proto after done (Disabled). virtual proto::Profile GenProfile(const std::string& profile_path) = 0; @@ -87,10 +88,18 @@ class DeviceTracer { DeviceTracer* GetDeviceTracer(); // Set a name for the cuda kernel operation being launched by the thread. -void SetCurAnnotation(const char* anno); +void SetCurAnnotation(const std::string& anno); // Clear the name after the operation is done. void ClearCurAnnotation(); // Current name of the operation being run in the thread. -const char* CurAnnotation(); +std::string CurAnnotation(); + +void SetCurBlock(int block_id); +void ClearCurBlock(); +int BlockDepth(); + +void SetCurThread(int thread_id); +void ClearCurThread(); +int CurThread(); } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/dynload/cublas.h b/paddle/fluid/platform/dynload/cublas.h index 580ed9bb57fca942bb76b9f7bf76f48ea281a0de..fa9041134d863ebfd8d1e00379da3b92323ae6e3 100644 --- a/paddle/fluid/platform/dynload/cublas.h +++ b/paddle/fluid/platform/dynload/cublas.h @@ -68,6 +68,8 @@ extern void *cublas_dso_handle; __macro(cublasDgemv_v2); \ __macro(cublasSgemm_v2); \ __macro(cublasDgemm_v2); \ + __macro(cublasHgemm); \ + __macro(cublasSgemmEx); \ __macro(cublasSgeam_v2); \ __macro(cublasDgeam_v2); \ __macro(cublasCreate_v2); \ @@ -83,6 +85,7 @@ extern void *cublas_dso_handle; __macro(cublasDgemmStridedBatched); \ __macro(cublasCgemmStridedBatched); \ __macro(cublasZgemmStridedBatched); \ + __macro(cublasHgemmStridedBatched); \ __macro(cublasSgetrfBatched); \ __macro(cublasSgetriBatched); \ __macro(cublasDgetrfBatched); \ diff --git a/paddle/fluid/platform/gpu_info.cc b/paddle/fluid/platform/gpu_info.cc index da4041bad0d82fe1c8c7a12fd0c7177e6dbddef3..dd70ff9ff574b32bc96a9e8255b1bf77a5cc84e4 100644 --- a/paddle/fluid/platform/gpu_info.cc +++ b/paddle/fluid/platform/gpu_info.cc @@ -33,6 +33,15 @@ int GetCUDADeviceCount() { return count; } +int GetCUDAComputeCapability(int id) { + PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count"); + cudaDeviceProp device_prop; + PADDLE_ENFORCE(cudaGetDeviceProperties(&device_prop, id), + "cudaGetDeviceProperties failed in " + "paddle::platform::GetCUDAComputeCapability"); + return device_prop.major * 10 + device_prop.minor; +} + int GetCUDAMultiProcessors(int id) { PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count"); int count; diff --git a/paddle/fluid/platform/gpu_info.h b/paddle/fluid/platform/gpu_info.h index c38ccf0f2ade1d2405177b541b33fd84283726ff..fa469fa77f5ca780da153cc87da8d04f239711f3 100644 --- a/paddle/fluid/platform/gpu_info.h +++ b/paddle/fluid/platform/gpu_info.h @@ -30,6 +30,9 @@ const std::string kEnvFractionGpuMemoryToUse = //! Get the total number of GPU devices in system. int GetCUDADeviceCount(); +//! Get the compute capability of the ith GPU (format: major * 10 + minor) +int GetCUDAComputeCapability(int i); + //! Get the MultiProcessors of the ith GPU. int GetCUDAMultiProcessors(int i); diff --git a/paddle/fluid/platform/nccl_test.cu b/paddle/fluid/platform/nccl_test.cu deleted file mode 100644 index 32a293796c09e5254c5eb48d11fa74617b3465ac..0000000000000000000000000000000000000000 --- a/paddle/fluid/platform/nccl_test.cu +++ /dev/null @@ -1,153 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -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 -#include - -#include "glog/logging.h" -#include "gtest/gtest.h" - -#include "paddle/fluid/framework/init.h" -#include "paddle/fluid/platform/device_context.h" -#include "paddle/fluid/platform/dynload/nccl.h" -#include "paddle/fluid/platform/enforce.h" -#include "paddle/fluid/platform/gpu_info.h" - -static int dev_count = 0; - -namespace paddle { -namespace platform { - -TEST(NCCL, init) { - std::vector comms; - comms.resize(dev_count); - PADDLE_ENFORCE(dynload::ncclCommInitAll(comms.data(), dev_count, nullptr)); - - for (int i = 0; i < dev_count; ++i) { - dynload::ncclCommDestroy(comms[i]); - } -} - -template -struct PerThreadData { - thrust::device_vector send_buff; - thrust::device_vector recv_buff; - CUDADeviceContext dev_ctx; - - T* SendBuff() { return thrust::raw_pointer_cast(send_buff.data()); } - - T* RecvBuff() { return thrust::raw_pointer_cast(recv_buff.data()); } - - PerThreadData(int gpu_id, size_t size) : dev_ctx(CUDAPlace(gpu_id)) { - send_buff.resize(size); - for (size_t i = 0; i < size; ++i) { - send_buff[i] = static_cast(i); - } - recv_buff.resize(size); - } -}; - -static constexpr int ELEM_COUNT = 10000; - -TEST(NCCL, all_reduce) { - std::vector comms; - comms.resize(dev_count); - VLOG(1) << "Initializing ncclComm"; - dynload::ncclCommInitAll(comms.data(), dev_count, nullptr); - VLOG(1) << "ncclComm initialized"; - VLOG(1) << "Creating thread data"; - std::vector>> data; - data.reserve(dev_count); - for (int i = 0; i < dev_count; ++i) { - VLOG(1) << "Creating thread data for device " << i; - SetDeviceId(i); - data.emplace_back(new PerThreadData(i, ELEM_COUNT)); - } - VLOG(1) << "Thread data created"; - - VLOG(1) << "Check send_buf data"; - for (int i = 0; i < dev_count; ++i) { - VLOG(1) << "Check on device " << i; - SetDeviceId(i); - thrust::host_vector tmp = data[i]->send_buff; - for (size_t j = 0; j < tmp.size(); ++j) { - ASSERT_NEAR(static_cast(j), tmp[j], 1e-5); - } - } - - VLOG(1) << "Invoking ncclAllReduce"; - - dynload::ncclGroupStart(); - for (int i = 0; i < dev_count; ++i) { - VLOG(1) << "Invoking ncclAllReduce with device " << i; - SetDeviceId(i); - PADDLE_ENFORCE(dynload::ncclAllReduce( - data[i]->SendBuff(), data[i]->RecvBuff(), ELEM_COUNT, ncclDouble, - ncclSum, comms[i], data[i]->dev_ctx.stream())); - VLOG(1) << "Invoked ncclAllReduce for device " << i; - } - dynload::ncclGroupEnd(); - - VLOG(1) << "Invoked ncclAllReduce"; - - VLOG(1) << "Sync devices"; - for (int i = 0; i < dev_count; ++i) { - VLOG(1) << "Sync device " << i; - SetDeviceId(i); - data[i]->dev_ctx.Wait(); - } - VLOG(1) << "device synced"; - - for (int i = 0; i < dev_count; ++i) { - SetDeviceId(i); - VLOG(1) << "Checking vector on device " << i; - thrust::host_vector tmp = data[i]->recv_buff; - for (size_t j = 0; j < tmp.size(); ++j) { - auto elem = static_cast(j); - elem *= dev_count; - ASSERT_NEAR(tmp[j], elem, 1e-4); - } - } - - for (int i = 0; i < dev_count; ++i) { - dynload::ncclCommDestroy(comms[i]); - } -} -} // namespace platform -} // namespace paddle - -int main(int argc, char** argv) { - dev_count = paddle::platform::GetCUDADeviceCount(); - if (dev_count <= 1) { - LOG(WARNING) - << "Cannot test multi-gpu nccl, because the CUDA device count is " - << dev_count; - return 0; - } - - std::vector places; - - places.emplace_back(paddle::platform::CPUPlace()); - int count = paddle::platform::GetCUDADeviceCount(); - for (int i = 0; i < count; ++i) { - places.emplace_back(paddle::platform::CUDAPlace(i)); - } - - VLOG(0) << " DeviceCount " << count; - paddle::platform::DeviceContextPool::Init(places); - - testing::InitGoogleTest(&argc, argv); - return RUN_ALL_TESTS(); -} diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 094f9224f719f6656fd0c44aa4a620730432ccac..b25206ff35cc87dcdd363bc0de54530f629d73ed 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -147,19 +147,48 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx) name_ = name; PushEvent(name_, dev_ctx_); // Maybe need the same push/pop behavior. - SetCurAnnotation(name_.c_str()); + SetCurAnnotation(name_); } RecordEvent::~RecordEvent() { if (g_state == ProfilerState::kDisabled) return; DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { - tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec()); + tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec(), + BlockDepth(), CurThread()); } ClearCurAnnotation(); PopEvent(name_, dev_ctx_); } +RecordBlock::RecordBlock(int block_id) : start_ns_(PosixInNsec()) { + if (g_state == ProfilerState::kDisabled) return; + SetCurBlock(block_id); + name_ = string::Sprintf("block_%d", block_id); +} + +RecordBlock::~RecordBlock() { + if (g_state == ProfilerState::kDisabled) return; + DeviceTracer* tracer = GetDeviceTracer(); + if (tracer) { + // We try to put all blocks at the same nested depth in the + // same timeline lane. and distinguish the using thread_id. + tracer->AddCPURecords(name_, start_ns_, PosixInNsec(), BlockDepth(), + CurThread()); + } + ClearCurBlock(); +} + +RecordThread::RecordThread(int thread_id) { + if (g_state == ProfilerState::kDisabled) return; + SetCurThread(thread_id); +} + +RecordThread::~RecordThread() { + if (g_state == ProfilerState::kDisabled) return; + ClearCurThread(); +} + void EnableProfiler(ProfilerState state) { PADDLE_ENFORCE(state != ProfilerState::kDisabled, "Can't enbale profling, since the input state is ", @@ -178,7 +207,7 @@ void EnableProfiler(ProfilerState state) { } #ifdef PADDLE_WITH_CUDA if (g_state == ProfilerState::kCUDA) { - // Generate some dummy evenets first to reduce the startup overhead. + // Generate some dummy events first to reduce the startup overhead. for (int i = 0; i < 5; i++) { ForEachDevice([](int d) { DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(d)); diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index 3542ce6cda87e3b013d60393e4ba93da61921940..de9a5cc20d76bf84778e0933831f218abb66c465 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -118,6 +118,20 @@ struct RecordEvent { std::string full_name_; }; +struct RecordBlock { + explicit RecordBlock(int block_id); + ~RecordBlock(); + + private: + std::string name_; + uint64_t start_ns_; +}; + +struct RecordThread { + explicit RecordThread(int thread_id); + ~RecordThread(); +}; + // Return the event list of all threads. Assumed the returned value calls // event_lists, event_lists[i][j] represents the j-th Event of i-th thread. std::vector> GetAllEvents(); diff --git a/paddle/fluid/platform/profiler.proto b/paddle/fluid/platform/profiler.proto index 06db7ed63895d5385952169e0672bf9888d7abae..7b42aa785ec6ad5731e3adee1e9f189127a826a1 100644 --- a/paddle/fluid/platform/profiler.proto +++ b/paddle/fluid/platform/profiler.proto @@ -15,15 +15,20 @@ limitations under the License. */ syntax = "proto2"; package paddle.platform.proto; -message MemCopy { optional uint64 bytes = 3; } +message MemCopy { optional uint64 bytes = 1; } message Event { + enum EventType { + CPU = 0; + GPUKernel = 1; + } + optional EventType type = 8; optional string name = 1; optional uint64 start_ns = 2; optional uint64 end_ns = 3; // When positive, it represents gpu id. When -1, it represents CPU. optional int64 device_id = 5; - optional uint32 stream_id = 6; + optional int64 sub_device_id = 6; optional MemCopy memcopy = 7; } diff --git a/paddle/fluid/platform/profiler_test.cc b/paddle/fluid/platform/profiler_test.cc index 4a86d8ec622772d0b62d5f4755a7b6138529b1d1..fc77e0f3213da776e0b05ad5b5da9081665cdf6e 100644 --- a/paddle/fluid/platform/profiler_test.cc +++ b/paddle/fluid/platform/profiler_test.cc @@ -75,6 +75,7 @@ TEST(RecordEvent, RecordEvent) { * ... * PopEvent(evt_name, dev_ctx); */ + LOG(INFO) << "Usage 1: PushEvent & PopEvent"; for (int loop = 0; loop < 3; ++loop) { for (int i = 1; i < 5; ++i) { std::string name = "op_" + std::to_string(i); @@ -93,6 +94,7 @@ TEST(RecordEvent, RecordEvent) { * ... * } */ + LOG(INFO) << "Usage 2: RecordEvent"; for (int i = 1; i < 5; ++i) { std::string name = "evs_op_" + std::to_string(i); RecordEvent record_event(name, dev_ctx); @@ -100,6 +102,34 @@ TEST(RecordEvent, RecordEvent) { while (counter != i * 1000) counter++; } + /* Usage 3 + * { + * RecordEvent record_event(name1, dev_ctx); + * ... + * code to be analyzed + * ... + * { + * RecordEvent nested_record_event(name2, dev_ctx); + * ... + * code to be analyzed + * ... + * } + * } + */ + LOG(INFO) << "Usage 3: nested RecordEvent"; + for (int i = 1; i < 5; ++i) { + std::string name = "ano_evs_op_" + std::to_string(i); + RecordEvent record_event(name, dev_ctx); + int counter = 1; + while (counter != i * 100) counter++; + { + std::string nested_name = "nested_ano_evs_op_" + std::to_string(i); + RecordEvent nested_record_event(nested_name, dev_ctx); + int nested_counter = 1; + while (nested_counter != i * 100) nested_counter++; + } + } + // Bad Usage: PushEvent("event_without_pop", dev_ctx); PopEvent("event_without_push", dev_ctx); diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt index d62f34030894e2fa21925bbc44e24b4e7d738d15..8942b5c9430ffa4e499b0ad1d2b5acf6d18ec0ab 100644 --- a/paddle/fluid/pybind/CMakeLists.txt +++ b/paddle/fluid/pybind/CMakeLists.txt @@ -1,6 +1,6 @@ if(WITH_PYTHON) cc_library(paddle_pybind SHARED - SRCS pybind.cc exception.cc protobuf.cc const_value.cc + SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc DEPS pybind python backward proto_desc paddle_memory executor prune init profiler feed_fetch_method ${GLOB_OP_LIB}) if(NOT APPLE AND NOT ANDROID) diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc index b0a2497d919b65afbe5eeaf4fe47c19baa1aba1c..45a64f43846e79c27295e52c59dca6bdfaa120a3 100644 --- a/paddle/fluid/pybind/protobuf.cc +++ b/paddle/fluid/pybind/protobuf.cc @@ -161,6 +161,8 @@ void BindBlockDesc(py::module &m) { py::return_value_policy::reference) .def("prepend_op", &BlockDesc::PrependOp, py::return_value_policy::reference) + .def("insert_op", &BlockDesc::InsertOp, + py::return_value_policy::reference) .def("remove_op", &BlockDesc::RemoveOp) .def("var", [](BlockDesc &self, py::bytes byte_name) { diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index ac7d1efb577505b70e10a70cdcfd3ed9c5fe1f5c..6c05442466f5f3d8e04a8f0a2206443b1007a107 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -26,16 +26,20 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/lod_tensor_array.h" #include "paddle/fluid/framework/prune.h" +#include "paddle/fluid/framework/reader.h" #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/operators/cond_op.h" #include "paddle/fluid/operators/net_op.h" #include "paddle/fluid/platform/enforce.h" +#include "paddle/fluid/platform/gpu_info.h" #include "paddle/fluid/platform/place.h" #include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/pybind/const_value.h" #include "paddle/fluid/pybind/exception.h" #include "paddle/fluid/pybind/pybind.h" +#include "paddle/fluid/pybind/recordio.h" #include "paddle/fluid/pybind/tensor_py.h" + #include "paddle/fluid/string/to_string.h" #ifdef PADDLE_WITH_CUDA @@ -100,12 +104,14 @@ PYBIND11_PLUGIN(core) { .def("set", PyCPUTensorSetFromArray) .def("set", PyCPUTensorSetFromArray) .def("set", PyCPUTensorSetFromArray) + .def("set", PyCPUTensorSetFromArray) #ifdef PADDLE_WITH_CUDA .def("set", PyCUDATensorSetFromArray) .def("set", PyCUDATensorSetFromArray) .def("set", PyCUDATensorSetFromArray) .def("set", PyCUDATensorSetFromArray) .def("set", PyCUDATensorSetFromArray) + .def("set", PyCUDATensorSetFromArray) #endif .def("shape", [](Tensor &self) { return vectorize(self.dims()); }) .def("set_float_element", TensorSetElement) @@ -217,8 +223,18 @@ All parameter, weight, gradient are variables in Paddle. [](Variable &self) -> operators::NetOp * { return self.GetMutable(); }, + py::return_value_policy::reference) + .def("get_reader", + [](Variable &self) -> framework::ReaderHolder * { + PADDLE_ENFORCE(self.IsType()); + return self.GetMutable(); + }, py::return_value_policy::reference); + py::class_(m, "Reader", "") + .def("has_next", &framework::ReaderHolder::HasNext) + .def("reset", &framework::ReaderHolder::ReInit); + py::class_(m, "Scope", "") .def("var", [](Scope &self, const std::string &name) -> Variable * { @@ -302,7 +318,6 @@ All parameter, weight, gradient are variables in Paddle. #endif }); // clang-format on - #ifdef PADDLE_WITH_CUDA py::class_(m, "Communicator").def(py::init<>()); #endif @@ -410,6 +425,12 @@ All parameter, weight, gradient are variables in Paddle. m.def("init_devices", &framework::InitDevices); m.def("is_compiled_with_cuda", IsCompiledWithCUDA); +#ifdef PADDLE_WITH_CUDA + m.def("is_float16_supported", [](const platform::CUDAPlace &place) -> bool { + // Only GPUs with Compute Capability >= 53 support float16 + return platform::GetCUDAComputeCapability(place.device) >= 53; + }); +#endif m.def("set_feed_variable", framework::SetFeedVariable); m.def("get_fetch_variable", framework::GetFetchVariable); @@ -474,6 +495,8 @@ All parameter, weight, gradient are variables in Paddle. m.def("enable_profiler", platform::EnableProfiler); m.def("disable_profiler", platform::DisableProfiler); m.def("reset_profiler", platform::ResetProfiler); + + BindRecordIOWriter(m); return m.ptr(); } } // namespace pybind diff --git a/paddle/fluid/pybind/recordio.cc b/paddle/fluid/pybind/recordio.cc new file mode 100644 index 0000000000000000000000000000000000000000..16f8bfb1a2e3a840670594d3cc2970e690dce891 --- /dev/null +++ b/paddle/fluid/pybind/recordio.cc @@ -0,0 +1,70 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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/fluid/pybind/recordio.h" +#include +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/recordio/writer.h" + +namespace paddle { +namespace pybind { + +class RecordIOWriter { + public: + RecordIOWriter(const std::string& filename, recordio::Compressor compressor, + size_t max_num_record) + : stream_(filename), writer_(&stream_, compressor, max_num_record) {} + + void AppendTensor(const framework::LoDTensor& tensor) { + tensors_.push_back(tensor); + } + + void CompleteAppendTensor() { + auto& ctx = + *platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); + framework::WriteToRecordIO(writer_, tensors_, ctx); + tensors_.clear(); + } + + void Close() { + PADDLE_ENFORCE(tensors_.empty()); + writer_.Flush(); + stream_.close(); + } + + private: + std::vector tensors_; + std::ofstream stream_; + recordio::Writer writer_; +}; + +void BindRecordIOWriter(py::module& m) { + py::class_ writer(m, "RecordIOWriter", ""); + py::enum_(writer, "Compressor", "") + .value("Snappy", recordio::Compressor::kSnappy) + .value("NoCompress", recordio::Compressor::kNoCompress); + + writer + .def("__init__", + [](RecordIOWriter& self, const std::string& filename, + recordio::Compressor compressor, size_t max_num_record) { + new (&self) RecordIOWriter(filename, compressor, max_num_record); + }) + .def("append_tensor", &RecordIOWriter::AppendTensor) + .def("complete_append_tensor", &RecordIOWriter::CompleteAppendTensor) + .def("close", &RecordIOWriter::Close); +} + +} // namespace pybind +} // namespace paddle diff --git a/paddle/fluid/pybind/recordio.h b/paddle/fluid/pybind/recordio.h new file mode 100644 index 0000000000000000000000000000000000000000..60e6a9e8595614b38375fca8c13d520739af9aaf --- /dev/null +++ b/paddle/fluid/pybind/recordio.h @@ -0,0 +1,26 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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 "pybind11/pybind11.h" +#include "pybind11/stl.h" + +namespace py = pybind11; + +namespace paddle { +namespace pybind { + +extern void BindRecordIOWriter(py::module& m); +} // namespace pybind +} // namespace paddle diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h index 1b0916ea0370d95a0c7dd149ee3f7b294c5e2351..6f8c597f8e610594851c318c122563523e4e7ea6 100644 --- a/paddle/fluid/pybind/tensor_py.h +++ b/paddle/fluid/pybind/tensor_py.h @@ -17,6 +17,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/memory/memcpy.h" #include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/float16.h" #include "pybind11/numpy.h" #include "pybind11/pybind11.h" @@ -71,27 +72,39 @@ struct CastToPyBufferImpl { paddle::platform::GpuMemcpyAsync( dst_ptr, src_ptr, sizeof(CUR_TYPE) * tensor.numel(), cudaMemcpyDeviceToHost, dev_ctx->stream()); + dev_ctx->Wait(); #else PADDLE_THROW("'CUDAPlace' is not supported in CPU only device."); #endif } else if (paddle::platform::is_cpu_place(tensor.place())) { dst_tensor = tensor; } - return py::buffer_info(dst_tensor.data(), sizeof(CUR_TYPE), - py::format_descriptor::format(), - (size_t)framework::arity(dst_tensor.dims()), - dims_outside, strides); + + if (std::type_index(typeid(CUR_TYPE)) == + std::type_index(typeid(platform::float16))) { + return py::buffer_info(dst_tensor.data(), sizeof(CUR_TYPE), + "e", /* np.dtype('e') == np.float16 */ + (size_t)framework::arity(dst_tensor.dims()), + dims_outside, strides); + } else { + return py::buffer_info(dst_tensor.data(), sizeof(CUR_TYPE), + py::format_descriptor::format(), + (size_t)framework::arity(dst_tensor.dims()), + dims_outside, strides); + } } else { constexpr bool less = I + 1 < std::tuple_size>::value; return CastToPyBufferImpl()(tensor); } } }; + } // namespace details + inline py::buffer_info CastToPyBuffer(framework::Tensor &tensor) { auto buffer_info = - details::CastToPyBufferImpl()( - tensor); + details::CastToPyBufferImpl()(tensor); return buffer_info; } @@ -136,6 +149,22 @@ void PyCPUTensorSetFromArray( std::memcpy(dst, array.data(), sizeof(T) * array.size()); } +template <> +void PyCPUTensorSetFromArray( + framework::Tensor &self, + py::array_t array, + paddle::platform::CPUPlace &place) { + std::vector dims; + dims.reserve(array.ndim()); + for (size_t i = 0; i < array.ndim(); ++i) { + dims.push_back((int)array.shape()[i]); + } + + self.Resize(framework::make_ddim(dims)); + auto *dst = self.mutable_data(place); + std::memcpy(dst, array.data(), sizeof(uint16_t) * array.size()); +} + #ifdef PADDLE_WITH_CUDA template void PyCUDATensorSetFromArray( @@ -157,6 +186,28 @@ void PyCUDATensorSetFromArray( paddle::platform::GpuMemcpyAsync(dst, array.data(), sizeof(T) * array.size(), cudaMemcpyHostToDevice, dev_ctx->stream()); } + +template <> +void PyCUDATensorSetFromArray( + framework::Tensor &self, + py::array_t array, + paddle::platform::CUDAPlace &place) { + std::vector dims; + dims.reserve(array.ndim()); + for (size_t i = 0; i < array.ndim(); ++i) { + dims.push_back((int)array.shape()[i]); + } + + self.Resize(framework::make_ddim(dims)); + auto *dst = self.mutable_data(place); + + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto dev_ctx = + static_cast(pool.Get(place)); + paddle::platform::GpuMemcpyAsync(dst, array.data(), + sizeof(uint16_t) * array.size(), + cudaMemcpyHostToDevice, dev_ctx->stream()); +} #endif } // namespace pybind diff --git a/paddle/fluid/recordio/CMakeLists.txt b/paddle/fluid/recordio/CMakeLists.txt index e1e7c2cdb3d0c960d5cd408420b5aaead73e70d7..92e97a6c85d7c8f01c8473feb9772f2285d49673 100644 --- a/paddle/fluid/recordio/CMakeLists.txt +++ b/paddle/fluid/recordio/CMakeLists.txt @@ -3,4 +3,7 @@ cc_library(header SRCS header.cc) cc_test(header_test SRCS header_test.cc DEPS header) cc_library(chunk SRCS chunk.cc DEPS snappystream snappy header zlib) cc_test(chunk_test SRCS chunk_test.cc DEPS chunk) -cc_library(recordio DEPS chunk header) +cc_library(writer SRCS writer.cc DEPS chunk) +cc_library(scanner SRCS scanner.cc DEPS chunk) +cc_test(writer_scanner_test SRCS writer_scanner_test.cc DEPS writer scanner) +cc_library(recordio DEPS chunk header writer scanner) diff --git a/paddle/fluid/recordio/chunk.cc b/paddle/fluid/recordio/chunk.cc index 587fd375c38ca83e1c65cb3ccc20b3509b6348c7..187a6a4ea7bd9d3a8ae48fa262e18f71b0f7d20d 100644 --- a/paddle/fluid/recordio/chunk.cc +++ b/paddle/fluid/recordio/chunk.cc @@ -24,33 +24,52 @@ namespace paddle { namespace recordio { constexpr size_t kMaxBufSize = 1024; +/** + * Read Stream by a fixed sized buffer. + * @param in input stream + * @param limit read at most `limit` bytes from input stream. 0 means no limit + * @param callback A function object with (const char* buf, size_t size) -> void + * as its type. + */ template -static void ReadStreamByBuf(std::istream& in, int limit, Callback callback) { +static void ReadStreamByBuf(std::istream& in, size_t limit, Callback callback) { char buf[kMaxBufSize]; std::streamsize actual_size; size_t counter = 0; - do { - auto actual_max = - limit > 0 ? std::min(limit - counter, kMaxBufSize) : kMaxBufSize; - actual_size = in.readsome(buf, actual_max); + size_t actual_max; + while (!in.eof() || + (limit != 0 && counter >= limit)) { // End of file or reach limit + actual_max = + limit != 0 ? std::min(limit - counter, kMaxBufSize) : kMaxBufSize; + in.read(buf, actual_max); + actual_size = in.gcount(); if (actual_size == 0) { break; } callback(buf, actual_size); - if (limit > 0) { + if (limit != 0) { counter += actual_size; } - } while (actual_size == kMaxBufSize); + } + in.clear(); // unset eof state } +/** + * Copy stream in to another stream + */ static void PipeStream(std::istream& in, std::ostream& os) { ReadStreamByBuf( - in, -1, [&os](const char* buf, size_t len) { os.write(buf, len); }); + in, 0, [&os](const char* buf, size_t len) { os.write(buf, len); }); } -static uint32_t Crc32Stream(std::istream& in, int limit = -1) { - auto crc = crc32(0, nullptr, 0); + +/** + * Calculate CRC32 from an input stream. + */ +static uint32_t Crc32Stream(std::istream& in, size_t limit = 0) { + uint32_t crc = static_cast(crc32(0, nullptr, 0)); ReadStreamByBuf(in, limit, [&crc](const char* buf, size_t len) { - crc = crc32(crc, reinterpret_cast(buf), len); + crc = static_cast(crc32( + crc, reinterpret_cast(buf), static_cast(len))); }); return crc; } @@ -85,28 +104,29 @@ bool Chunk::Write(std::ostream& os, Compressor ct) const { compressed_stream.reset(); } - auto end_pos = sout.tellg(); + sout.seekg(0, std::ios::end); + uint32_t len = static_cast(sout.tellg()); sout.seekg(0, std::ios::beg); - uint32_t len = static_cast(end_pos - sout.tellg()); uint32_t crc = Crc32Stream(sout); - sout.seekg(0, std::ios::beg); - Header hdr(static_cast(records_.size()), crc, ct, len); hdr.Write(os); + sout.seekg(0, std::ios::beg); + sout.clear(); PipeStream(sout, os); return true; } -void Chunk::Parse(std::istream& sin) { +bool Chunk::Parse(std::istream& sin) { Header hdr; - hdr.Parse(sin); + bool ok = hdr.Parse(sin); + if (!ok) { + return ok; + } auto beg_pos = sin.tellg(); - auto crc = Crc32Stream(sin, hdr.CompressSize()); + uint32_t crc = Crc32Stream(sin, hdr.CompressSize()); PADDLE_ENFORCE_EQ(hdr.Checksum(), crc); - Clear(); - - sin.seekg(beg_pos, std::ios::beg); + sin.seekg(beg_pos, sin.beg); std::unique_ptr compressed_stream; switch (hdr.CompressType()) { case Compressor::kNoCompress: @@ -126,8 +146,10 @@ void Chunk::Parse(std::istream& sin) { std::string buf; buf.resize(rec_len); stream.read(&buf[0], rec_len); + PADDLE_ENFORCE_EQ(rec_len, stream.gcount()); Add(buf); } + return true; } } // namespace recordio diff --git a/paddle/fluid/recordio/chunk.h b/paddle/fluid/recordio/chunk.h index 0ba9c63abbe72e7a51ddb1af5f0d206aa9f6cc5b..bf20ebd455c26ddeebeeea8db04cf7103b0c085f 100644 --- a/paddle/fluid/recordio/chunk.h +++ b/paddle/fluid/recordio/chunk.h @@ -26,9 +26,9 @@ namespace recordio { class Chunk { public: Chunk() : num_bytes_(0) {} - void Add(std::string buf) { - records_.push_back(buf); + void Add(const std::string& buf) { num_bytes_ += buf.size(); + records_.emplace_back(buf); } // dump the chunk into w, and clears the chunk and makes it ready for // the next add invocation. @@ -37,10 +37,15 @@ public: records_.clear(); num_bytes_ = 0; } - void Parse(std::istream& sin); - size_t NumBytes() { return num_bytes_; } + + // returns true if ok, false if eof + bool Parse(std::istream& sin); + size_t NumBytes() const { return num_bytes_; } + size_t NumRecords() const { return records_.size(); } const std::string& Record(int i) const { return records_[i]; } + bool Empty() const { return records_.empty(); } + private: std::vector records_; // sum of record lengths in bytes. diff --git a/paddle/fluid/recordio/chunk_test.cc b/paddle/fluid/recordio/chunk_test.cc index a67ba32ed6ab8bda230d1414975c96a0be6d682b..1f0e36a14d373ca96167199d4582bc8f17290ae8 100644 --- a/paddle/fluid/recordio/chunk_test.cc +++ b/paddle/fluid/recordio/chunk_test.cc @@ -26,7 +26,7 @@ TEST(Chunk, SaveLoad) { ch.Add(std::string("123", 4)); std::stringstream ss; ch.Write(ss, Compressor::kNoCompress); - ch.Clear(); + ss.seekg(0); ch.Parse(ss); ASSERT_EQ(ch.NumBytes(), 10U); } diff --git a/paddle/fluid/recordio/header.cc b/paddle/fluid/recordio/header.cc index 3641caaa8981020519cbc31e5362348c02d3bbce..e50de15b7c2b480357f5f6c7daa2b4a676749679 100644 --- a/paddle/fluid/recordio/header.cc +++ b/paddle/fluid/recordio/header.cc @@ -13,6 +13,7 @@ // limitations under the License. #include "paddle/fluid/recordio/header.h" +#include "paddle/fluid/platform/enforce.h" namespace paddle { namespace recordio { @@ -26,23 +27,33 @@ Header::Header() Header::Header(uint32_t num, uint32_t sum, Compressor c, uint32_t cs) : num_records_(num), checksum_(sum), compressor_(c), compress_size_(cs) {} -void Header::Parse(std::istream& is) { +bool Header::Parse(std::istream& is) { + uint32_t magic; + size_t read_size = + is.readsome(reinterpret_cast(&magic), sizeof(uint32_t)); + if (read_size < sizeof(uint32_t)) { + return false; + } + PADDLE_ENFORCE_EQ(magic, kMagicNumber); + is.read(reinterpret_cast(&num_records_), sizeof(uint32_t)) .read(reinterpret_cast(&checksum_), sizeof(uint32_t)) .read(reinterpret_cast(&compressor_), sizeof(uint32_t)) .read(reinterpret_cast(&compress_size_), sizeof(uint32_t)); + return true; } void Header::Write(std::ostream& os) const { - os.write(reinterpret_cast(&num_records_), sizeof(uint32_t)) + os.write(reinterpret_cast(&kMagicNumber), sizeof(uint32_t)) + .write(reinterpret_cast(&num_records_), sizeof(uint32_t)) .write(reinterpret_cast(&checksum_), sizeof(uint32_t)) .write(reinterpret_cast(&compressor_), sizeof(uint32_t)) .write(reinterpret_cast(&compress_size_), sizeof(uint32_t)); } std::ostream& operator<<(std::ostream& os, Header h) { - os << h.NumRecords() << h.Checksum() - << static_cast(h.CompressType()) << h.CompressSize(); + os << "Header: " << h.NumRecords() << ", " << h.Checksum() << ", " + << static_cast(h.CompressType()) << ", " << h.CompressSize(); return os; } diff --git a/paddle/fluid/recordio/header.h b/paddle/fluid/recordio/header.h index cbd52642a668d1eaeeafb672e50af1a476975080..9200ac090de4514bef3704ac502039222eef2284 100644 --- a/paddle/fluid/recordio/header.h +++ b/paddle/fluid/recordio/header.h @@ -19,8 +19,6 @@ namespace paddle { namespace recordio { -// Default ChunkSize -constexpr size_t kDefaultMaxChunkSize = 32 * 1024 * 1024; // MagicNumber for memory checking constexpr uint32_t kMagicNumber = 0x01020304; @@ -44,7 +42,9 @@ public: Header(uint32_t num, uint32_t sum, Compressor ct, uint32_t cs); void Write(std::ostream& os) const; - void Parse(std::istream& is); + + // returns true if OK, false if eof + bool Parse(std::istream& is); uint32_t NumRecords() const { return num_records_; } uint32_t Checksum() const { return checksum_; } diff --git a/paddle/fluid/recordio/scanner.cc b/paddle/fluid/recordio/scanner.cc new file mode 100644 index 0000000000000000000000000000000000000000..d842f8fe5a4c9d1a2b564c738d97fffb02f3ccb5 --- /dev/null +++ b/paddle/fluid/recordio/scanner.cc @@ -0,0 +1,51 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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/fluid/recordio/scanner.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace recordio { +Scanner::Scanner(std::unique_ptr &&stream) + : stream_(std::move(stream)) { + Reset(); +} + +Scanner::Scanner(const std::string &filename) { + stream_.reset(new std::ifstream(filename)); + Reset(); +} + +void Scanner::Reset() { + stream_->seekg(0, std::ios::beg); + ParseNextChunk(); +} + +std::string Scanner::Next() { + PADDLE_ENFORCE(!eof_, "StopIteration"); + auto rec = cur_chunk_.Record(offset_++); + if (offset_ == cur_chunk_.NumRecords()) { + ParseNextChunk(); + } + return rec; +} + +void Scanner::ParseNextChunk() { + eof_ = !cur_chunk_.Parse(*stream_); + offset_ = 0; +} + +bool Scanner::HasNext() const { return !eof_; } +} // namespace recordio +} // namespace paddle diff --git a/paddle/fluid/recordio/scanner.h b/paddle/fluid/recordio/scanner.h new file mode 100644 index 0000000000000000000000000000000000000000..f3f17b69f195ddd92f5a39ead9755a7b8e2dd329 --- /dev/null +++ b/paddle/fluid/recordio/scanner.h @@ -0,0 +1,44 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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/fluid/recordio/chunk.h" +namespace paddle { +namespace recordio { + +class Scanner { +public: + explicit Scanner(std::unique_ptr&& stream); + + explicit Scanner(const std::string& filename); + + void Reset(); + + std::string Next(); + + bool HasNext() const; + +private: + std::unique_ptr stream_; + Chunk cur_chunk_; + size_t offset_; + bool eof_; + + void ParseNextChunk(); +}; +} // namespace recordio +} // namespace paddle diff --git a/paddle/fluid/recordio/writer.cc b/paddle/fluid/recordio/writer.cc new file mode 100644 index 0000000000000000000000000000000000000000..196d66edff8cc6000afcd74fb945c05dcab7106a --- /dev/null +++ b/paddle/fluid/recordio/writer.cc @@ -0,0 +1,35 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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/fluid/recordio/writer.h" +#include "paddle/fluid/platform/enforce.h" +namespace paddle { +namespace recordio { +void Writer::Write(const std::string& record) { + cur_chunk_.Add(record); + if (cur_chunk_.NumRecords() >= max_num_records_in_chunk_) { + Flush(); + } +} + +void Writer::Flush() { + cur_chunk_.Write(stream_, compressor_); + cur_chunk_.Clear(); +} + +Writer::~Writer() { + PADDLE_ENFORCE(cur_chunk_.Empty(), "Writer must be flushed when destroy."); +} + +} // namespace recordio +} // namespace paddle diff --git a/paddle/fluid/recordio/writer.h b/paddle/fluid/recordio/writer.h new file mode 100644 index 0000000000000000000000000000000000000000..0c478d507547b10b8ebaaf5e512557a5c8c13e65 --- /dev/null +++ b/paddle/fluid/recordio/writer.h @@ -0,0 +1,43 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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/fluid/recordio/chunk.h" +namespace paddle { +namespace recordio { + +class Writer { +public: + Writer(std::ostream* sout, + Compressor compressor, + size_t max_num_records_in_chunk = 1000) + : stream_(*sout), + max_num_records_in_chunk_(max_num_records_in_chunk), + compressor_(compressor) {} + + void Write(const std::string& record); + + void Flush(); + + ~Writer(); + +private: + std::ostream& stream_; + size_t max_num_records_in_chunk_; + Chunk cur_chunk_; + Compressor compressor_; +}; + +} // namespace recordio +} // namespace paddle diff --git a/paddle/fluid/recordio/writer_scanner_test.cc b/paddle/fluid/recordio/writer_scanner_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..7e764f0d9439709ad101af2b8864dc0158bd359b --- /dev/null +++ b/paddle/fluid/recordio/writer_scanner_test.cc @@ -0,0 +1,69 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// 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 "gtest/gtest.h" + +#include +#include "paddle/fluid/recordio/scanner.h" +#include "paddle/fluid/recordio/writer.h" + +TEST(WriterScanner, Normal) { + std::stringstream* stream = new std::stringstream(); + + { + paddle::recordio::Writer writer(stream, + paddle::recordio::Compressor::kSnappy); + writer.Write("ABC"); + writer.Write("BCD"); + writer.Write("CDE"); + writer.Flush(); + } + + { + stream->seekg(0, std::ios::beg); + std::unique_ptr stream_ptr(stream); + paddle::recordio::Scanner scanner(std::move(stream_ptr)); + ASSERT_TRUE(scanner.HasNext()); + ASSERT_EQ(scanner.Next(), "ABC"); + ASSERT_EQ("BCD", scanner.Next()); + ASSERT_TRUE(scanner.HasNext()); + ASSERT_EQ("CDE", scanner.Next()); + ASSERT_FALSE(scanner.HasNext()); + } +} + +TEST(WriterScanner, TinyChunk) { + std::stringstream* stream = new std::stringstream(); + { + paddle::recordio::Writer writer( + stream, paddle::recordio::Compressor::kNoCompress, 2 /*max chunk num*/); + writer.Write("ABC"); + writer.Write("BCD"); + writer.Write("CDE"); + writer.Write("DEFG"); + writer.Flush(); + } + + { + stream->seekg(0, std::ios::beg); + std::unique_ptr stream_ptr(stream); + paddle::recordio::Scanner scanner(std::move(stream_ptr)); + ASSERT_TRUE(scanner.HasNext()); + ASSERT_EQ(scanner.Next(), "ABC"); + ASSERT_EQ(scanner.Next(), "BCD"); + ASSERT_EQ(scanner.Next(), "CDE"); + ASSERT_EQ(scanner.Next(), "DEFG"); + ASSERT_FALSE(scanner.HasNext()); + } +} \ No newline at end of file diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh index 06319fc638984f8f8ed897c362f516e1534bc8db..6be2bd8fad9e33cf4e1dcafdd6b8f39111bdbe88 100644 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -213,7 +213,7 @@ function gen_fluid_inference_lib() { if [ ${WITH_C_API:-OFF} == "OFF" ] ; then cat < 0 else "Tensor", name=proto.name, @@ -62,20 +65,24 @@ def repr_lodtensor(proto): def repr_selected_rows(proto): - if not proto.selected_rows: return + if proto.type.type != framework_pb2.VarType.SELECTED_ROWS: + return + return reprtpl.format( ttype="SelectedRows", name=proto.name, - reprs=repr_tensor(proto.selected_rows)) + reprs=repr_tensor(proto.type.selected_rows)) def repr_tensor_array(proto): - if not proto.tensor_array: return + if proto.type.type != framework_pb2.VarType.LOD_TENSOR_ARRAY: + return + return reprtpl.format( ttype="TensorArray", name=proto.name, - reprs="level=%d, %s" % (proto.tensor_array.lod_level, - repr_tensor(proto.lod_tensor))) + reprs="level=%d, %s" % (proto.type.tensor_array.lod_level, + repr_tensor(proto.type.lod_tensor.tensor))) type_handlers = [ @@ -119,6 +126,7 @@ def pprint_block_codes(block_desc, show_backward=False): def is_var_backward(var_desc): return "@GRAD" in var_desc.name + #print(type(block_desc)) if type(block_desc) is not framework_pb2.BlockDesc: block_desc = framework_pb2.BlockDesc.FromString( block_desc.serialize_to_string()) diff --git a/python/paddle/fluid/distribute_transpiler.py b/python/paddle/fluid/distribute_transpiler.py index bb2ce4d45d5da6b2fbd097a94479f3696271e5ec..3d3a6c116eeb39fb7236d0e9707415cdd6b828bd 100644 --- a/python/paddle/fluid/distribute_transpiler.py +++ b/python/paddle/fluid/distribute_transpiler.py @@ -250,6 +250,8 @@ class DistributeTranspiler: def get_trainer_program(self): # remove optimize ops and add a send op to main_program self.program.global_block().delete_ops(self.optimize_ops) + # FIXME(typhoonzero): serialize once will fix error occurs when clone. + self.program.__str__() return self.program def get_pserver_program(self, endpoint): @@ -309,7 +311,8 @@ class DistributeTranspiler: for _, opt_op in enumerate(opt_op_on_pserver): if ufind.is_connected(op, opt_op): if self._is_opt_op(op): - self._append_pserver_ops(optimize_block, op, endpoint) + self._append_pserver_ops(optimize_block, op, endpoint, + default_main_program()) else: self._append_pserver_non_opt_ops(optimize_block, op) break @@ -520,7 +523,8 @@ class DistributeTranspiler: orig_var_name = varname[:suff_idx] return orig_var_name - def _append_pserver_ops(self, optimize_block, opt_op, endpoint): + def _append_pserver_ops(self, optimize_block, opt_op, endpoint, + origin_program): program = optimize_block.program pserver_block = program.global_block() new_inputs = dict() @@ -576,7 +580,17 @@ class DistributeTranspiler: elif key == "LearningRate": # leraning rate variable has already be created by non-optimize op, # don't create it once again. - new_inputs[key] = pserver_block.vars[opt_op.input(key)[0]] + lr_varname = opt_op.input(key)[0] + if pserver_block.vars.has_key(lr_varname): + new_inputs[key] = pserver_block.vars[opt_op.input(key)[0]] + else: + origin_var = origin_program.global_block().vars[lr_varname] + tmpvar = pserver_block.create_var( + name=origin_var.name, + persistable=origin_var.persistable, + dtype=origin_var.dtype, + shape=origin_var.shape) + new_inputs[key] = tmpvar for key in opt_op.input_names: new_shape = None diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index d14d6349b1bcf598e25bbeb9913d2d0da71a5054..70ecffd910a46570b5a8e576d88039fa5e22e726 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -487,7 +487,7 @@ class Operator(object): 'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv', 'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine', 'ncclInit', 'channel_create', 'channel_close', - 'channel_send', 'channel_recv' + 'channel_send', 'channel_recv', 'select' } if type not in no_kernel_op_set: self.desc.infer_var_type(self.block.desc) diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index 1817caa94275e4efa47ec1a5a0aa861255c75561..1c0f1f6eb415b1c05c1052c1f52743a19c49f017 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -47,7 +47,7 @@ def is_parameter(var): def is_persistable(var): if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \ - var.desc.type() == core.VarDesc.VarType.FETCH_LIST: + var.desc.type() == core.VarDesc.VarType.FETCH_LIST: return False return var.persistable @@ -102,6 +102,9 @@ def save_vars(executor, save_var_map = {} for each_var in vars: + # NOTE: don't save the variable which type is RAW + if each_var.type == core.VarDesc.VarType.RAW: + continue new_var = _clone_var_in_block_(save_block, each_var) if filename is None: save_block.append_op( diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 1bb1aa30ee1019c6f80eb64b6dc20459e7a3073b..02cd0a05a11d8d1d52d42c2b62799f1093d0abc2 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -16,7 +16,7 @@ import contextlib from layer_function_generator import autodoc from tensor import assign, fill_constant from .. import core -from ..framework import Program, Variable, Operator +from ..framework import Program, Variable, Operator, Block from ..layer_helper import LayerHelper, unique_name from ops import logical_and, logical_not, logical_or @@ -29,6 +29,7 @@ __all__ = [ 'WhileGuard', 'While', 'Switch', + 'Select', 'lod_rank_table', 'max_sequence_len', 'topk', @@ -1211,6 +1212,186 @@ class Switch(object): return True +class SelectCase(object): + DEFAULT = 0 + SEND = 1 + RECEIVE = 2 + + def __init__(self, + case_idx, + case_to_execute, + channel_action_fn=None, + channel=None, + value=None): + self.helper = LayerHelper('conditional_block') + self.main_program = self.helper.main_program + self.is_scalar_condition = True + + self.case_to_execute = case_to_execute + self.idx = case_idx + + # Since we aren't going to use the `channel_send` or `channel_recv` + # functions directly, we just need to capture the name. + self.action = (self.SEND + if channel_action_fn.__name__ == ('channel_send') else + self.RECEIVE) if channel_action_fn else (self.DEFAULT) + self.value = value + self.channel = channel + + def __enter__(self): + self.block = self.main_program.create_block() + + def construct_op(self): + main_program = self.helper.main_program + cases_block = main_program.current_block() + + inner_outputs = set() + input_set = set() + params = set() + + for op in self.block.ops: + # Iterate over all operators, get all the inputs + # and add as input to the SelectCase operator. + for iname in op.input_names: + for in_var_name in op.input(iname): + if in_var_name not in inner_outputs: + input_set.add(in_var_name) + + for oname in op.output_names: + for out_var_name in op.output(oname): + inner_outputs.add(out_var_name) + + param_list = [ + cases_block.var(each_name) for each_name in params + if each_name not in input_set + ] + + # Iterate over all operators, get all the outputs + # add to the output list of SelectCase operator only if + # they exist in the parent block. + out_vars = [] + for inner_out_name in inner_outputs: + if inner_out_name in cases_block.vars: + out_vars.append(cases_block.var(inner_out_name)) + + # First, create an op that will determine whether or not this is the + # conditional variable to execute. + should_execute_block = equal( + fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=self.idx), + self.case_to_execute) + + step_scope = cases_block.create_var( + type=core.VarDesc.VarType.STEP_SCOPES) + + cases_block.append_op( + type='conditional_block', + inputs={'X': [should_execute_block], + 'Params': param_list}, + outputs={'Out': out_vars, + 'Scope': [step_scope]}, + attrs={ + 'sub_block': self.block, + 'is_scalar_condition': self.is_scalar_condition + }) + + return '%s,%s,%s,%s' % (self.idx, self.action, self.channel.name + if self.channel else '', self.value.name + if self.value else '') + + def __exit__(self, exc_type, exc_val, exc_tb): + self.main_program.rollback() + if exc_type is not None: + return False # re-raise exception + return True + + +class Select(BlockGuard): + def __init__(self, name=None): + self.helper = LayerHelper('select', name=name) + self.cases = [] + + super(Select, self).__init__(self.helper.main_program) + self.case_to_execute = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=-1) + + def __enter__(self): + super(Select, self).__enter__() + return self + + def case(self, channel_action_fn, channel, value): + """Create a new block for this condition. + """ + select_case = SelectCase( + len(self.cases), self.case_to_execute, channel_action_fn, channel, + value) + + self.cases.append(select_case) + + return select_case + + def default(self): + """Create a default case block for this condition. + """ + default_case = SelectCase(len(self.cases), self.case_to_execute) + + self.cases.append(default_case) + + return default_case + + def __exit__(self, exc_type, exc_val, exc_tb): + if exc_type is not None: + return False + + # Create a select op and another block to wrap its + # case blocks. + select_block = self.helper.main_program.current_block() + parent_block = self.helper.main_program.block(select_block.parent_idx) + + # Construct each case op, inside the newly created select block. + serialized_cases = [] + for case in self.cases: + serialized_cases.append(case.construct_op()) + + intermediate = set() + params = set() + + for case_block in select_block.ops: + if case_block.attrs and 'sub_block' in case_block.attrs: + for each_op in case_block.attrs['sub_block'].ops: + assert isinstance(each_op, Operator) + for iname in each_op.input_names: + for in_var_name in each_op.input(iname): + if in_var_name not in intermediate: + params.add(in_var_name) + + for oname in each_op.output_names: + for out_var_name in each_op.output(oname): + intermediate.add(out_var_name) + + # TODO(varunarora): Figure out if defining output is needed. + out_list = [ + parent_block.var(var_name) for var_name in parent_block.vars + if var_name in intermediate + ] + + X = [select_block.var_recursive(x_name) for x_name in params] + + # Needs to be used by `equal` inside the cases block. + X.append(self.case_to_execute) + + # Construct the select op. + parent_block.append_op( + type='select', + inputs={'X': X, + 'case_to_execute': self.case_to_execute}, + attrs={'sub_block': select_block, + 'cases': serialized_cases}, + outputs={}) + + return super(Select, self).__exit__(exc_type, exc_val, exc_tb) + + class IfElseBlockGuard(object): def __init__(self, is_true, ifelse): if not isinstance(ifelse, IfElse): diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index 2bf7cf21ca94c742a15d980194b896d9ec8ad91b..a889ab6bdc6ac9494ef992a97292b7a2536c41c4 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -130,8 +130,13 @@ def detection_output(loc, target_box=loc, code_type='decode_center_size') - nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype) + old_shape = scores.shape + scores = ops.reshape(x=scores, shape=(-1, old_shape[-1])) + scores = nn.softmax(input=scores) + scores = ops.reshape(x=scores, shape=old_shape) scores = nn.transpose(scores, perm=[0, 2, 1]) + + nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype) helper.append_op( type="multiclass_nms", inputs={'Scores': scores, @@ -562,16 +567,16 @@ def multi_box_head(inputs, base_size, num_classes, aspect_ratios, - min_ratio, - max_ratio, + min_ratio=None, + max_ratio=None, min_sizes=None, max_sizes=None, steps=None, step_w=None, step_h=None, offset=0.5, - variance=[0.1, 0.1, 0.1, 0.1], - flip=False, + variance=[0.1, 0.1, 0.2, 0.2], + flip=True, clip=False, kernel_size=1, pad=0, @@ -614,7 +619,7 @@ def multi_box_head(inputs, the inputs[i] will be automatically calculated. Default: None. offset(float): Prior boxes center offset. Default: 0.5 variance(list|tuple): the variances to be encoded in prior boxes. - Default:[0.1, 0.1, 0.1, 0.1]. + Default:[0.1, 0.1, 0.2, 0.2]. flip(bool): Whether to flip aspect ratios. Default:False. clip(bool): Whether to clip out-of-boundary boxes. Default: False. kernel_size(int): The kernel size of conv2d. Default: 1. @@ -668,6 +673,19 @@ def multi_box_head(inputs, helper = LayerHelper("prior_box", **locals()) dtype = helper.input_dtype() + attrs = { + 'min_sizes': min_sizes, + 'aspect_ratios': aspect_ratios, + 'variances': variance, + 'flip': flip, + 'clip': clip, + 'step_w': step_w, + 'step_h': step_h, + 'offset': offset + } + if len(max_sizes) > 0 and max_sizes[0] > 0: + attrs['max_sizes'] = max_sizes + box = helper.create_tmp_variable(dtype) var = helper.create_tmp_variable(dtype) helper.append_op( @@ -676,17 +694,7 @@ def multi_box_head(inputs, "Image": image}, outputs={"Boxes": box, "Variances": var}, - attrs={ - 'min_sizes': min_sizes, - 'max_sizes': max_sizes, - 'aspect_ratios': aspect_ratios, - 'variances': variance, - 'flip': flip, - 'clip': clip, - 'step_w': step_w, - 'step_h': step_h, - 'offset': offset - }) + attrs=attrs, ) return box, var def _reshape_with_axis_(input, axis=1): @@ -714,7 +722,7 @@ def multi_box_head(inputs, if num_layer <= 2: assert min_sizes is not None and max_sizes is not None assert len(min_sizes) == num_layer and len(max_sizes) == num_layer - else: + elif min_sizes is None and max_sizes is None: min_sizes = [] max_sizes = [] step = int(math.floor(((max_ratio - min_ratio)) / (num_layer - 2))) @@ -759,9 +767,6 @@ def multi_box_head(inputs, min_size = [min_size] if not _is_list_or_tuple_(max_size): max_size = [max_size] - if not (len(max_size) == len(min_size)): - raise ValueError( - 'the length of max_size and min_size should be equal.') aspect_ratio = [] if aspect_ratios is not None: @@ -779,7 +784,7 @@ def multi_box_head(inputs, num_boxes = box.shape[2] - # get box_loc + # get loc num_loc_output = num_boxes * 4 mbox_loc = nn.conv2d( input=input, @@ -796,7 +801,7 @@ def multi_box_head(inputs, mbox_loc_flatten = ops.reshape(mbox_loc, shape=new_shape) mbox_locs.append(mbox_loc_flatten) - # get conf_loc + # get conf num_conf_output = num_boxes * num_classes conf_loc = nn.conv2d( input=input, diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index af3ae54248a744e7e2fed8190aeeb0eb481cb315..9c91f395e7c9d7ca76c1a5cc310bc3bbc06daec9 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -13,11 +13,16 @@ # limitations under the License. from .. import core -from ..layer_helper import LayerHelper +from ..framework import convert_np_dtype_to_dtype_, default_main_program, default_startup_program +from ..unique_name import generate as unique_name from control_flow import BlockGuard from ..layer_helper import LayerHelper +from ..executor import global_scope -__all__ = ['data', 'BlockGuardServ', 'ListenAndServ', 'Send'] +__all__ = [ + 'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'open_recordio_file', + 'read_file', 'create_shuffle_reader', 'create_double_buffer_reader' +] def data(name, @@ -224,3 +229,101 @@ def Recv(endpoints, get_vars): outputs={"Out": get_vars}, attrs={"endpoints": endpoints, "epmap": epmap}) + + +def monkey_patch_reader_methods(reader): + def __get_reader__(): + scope = global_scope() + var = scope.find_var(reader.name) + return var.get_reader() + + def eof(): + return not __get_reader__().has_next() + + def reset(): + return __get_reader__().reset() + + reader.eof = eof + reader.reset = reset + reader.stop_gradient = True + reader.persistable = True + return reader + + +def _copy_reader_var_(block, var): + new_var = block.create_var(name=var.name, type=core.VarDesc.VarType.READER) + new_var.desc.set_shapes(var.desc.shapes()) + new_var.desc.set_dtypes(var.desc.dtypes()) + new_var.persistable = True + return monkey_patch_reader_methods(new_var) + + +def open_recordio_file(filename, shapes, lod_levels, dtypes): + dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] + shape_concat = [] + ranks = [] + + for shape in shapes: + shape_concat.extend(shape) + ranks.append(len(shape)) + + var_name = unique_name('open_recordio_file') + + startup_blk = default_startup_program().current_block() + startup_var = startup_blk.create_var(name=var_name) + startup_blk.append_op( + type='create_recordio_file_reader', + outputs={'Out': [startup_var]}, + attrs={ + 'shape_concat': shape_concat, + 'lod_levels': lod_levels, + 'filename': filename, + 'ranks': ranks + }) + + startup_var.desc.set_dtypes(dtypes) + startup_var.persistable = True + return _copy_reader_var_(default_main_program().current_block(), + startup_var) + + +def __create_decorated_reader__(op_type, reader, attrs): + var_name = unique_name(op_type) + startup_blk = default_startup_program().current_block() + startup_var = startup_blk.create_var(name=var_name) + startup_blk.append_op( + type=op_type, + inputs={'UnderlyingReader': reader}, + outputs={'Out': [startup_var]}, + attrs=attrs) + startup_var.persistable = True + return _copy_reader_var_(default_main_program().current_block(), + startup_var) + + +def create_shuffle_reader(reader, buffer_size): + return __create_decorated_reader__('create_shuffle_reader', reader, + {'buffer_size': int(buffer_size)}) + + +def create_double_buffer_reader(reader, place=None): + attrs = dict() + if place is not None: + attrs['place'] = str(place).upper() + return __create_decorated_reader__('create_double_buffer_reader', reader, + attrs) + + +def read_file(file_obj): + helper = LayerHelper('read_file') + out = [ + helper.create_tmp_variable( + stop_gradient=True, dtype='float32') + for _ in range(len(file_obj.desc.shapes())) + ] + helper.append_op( + type='read', inputs={'Reader': [file_obj]}, outputs={'Out': out}) + if len(out) == 1: + return out[0] + else: + return out diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index bd79022a0c39cf18bd05d49ac62986d342a4ae06..35b01a79914b3427836d4abd51aa2e2eb471d517 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -16,10 +16,7 @@ import cStringIO import functools import warnings -from .. import proto - -framework_pb2 = proto.framework_pb2 - +from ..proto import framework_pb2 from ..framework import OpProtoHolder, Variable from ..layer_helper import LayerHelper diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index b4fa530aa66fa1b84673b504a922595b83f42268..9656dcf94f14ad9250bb7e79c1330c9bdd44d9d6 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -39,6 +39,8 @@ __all__ = [ 'sequence_conv', 'conv2d', 'sequence_pool', + 'sequence_softmax', + 'softmax', 'pool2d', 'batch_norm', 'beam_search_decode', @@ -49,6 +51,7 @@ __all__ = [ 'reduce_mean', 'reduce_max', 'reduce_min', + 'reduce_prod', 'sequence_first_step', 'sequence_last_step', 'dropout', @@ -70,6 +73,7 @@ __all__ = [ 'smooth_l1', 'one_hot', 'autoincreased_step_counter', + 'lod_reset', ] @@ -84,13 +88,12 @@ def fc(input, **Fully Connected Layer** The fully connected layer can take multiple tensors as its inputs. It - creates a variable (one for each input tensor) called weights for each - input tensor, which represents a fully connected weight matrix from - each input unit to each output unit. The fully connected layer - multiplies each input tensor with its coresponding weight to produce - an output Tensor. If multiple input tensors are given, the results of - multiple multiplications will be sumed up. If bias_attr is not None, - a biases variable will be created and added to the output. Finally, + creates a variable called weights for each input tensor, which represents + a fully connected weight matrix from each input unit to each output unit. + The fully connected layer multiplies each input tensor with its coresponding + weight to produce an output Tensor. If multiple input tensors are given, + the results of multiple multiplications will be sumed up. If bias_attr is + not None, a bias variable will be created and added to the output. Finally, if activation is not None, it will be applied to the output as well. This process can be formulated as follows: @@ -109,44 +112,27 @@ def fc(input, * :math:`Out`: The output tensor. Args: - input(Variable|list): The input tensor(s) to the fully connected layer. - size(int): The number of output units in the fully connected layer. - num_flatten_dims(int): The fc layer can accept an input tensor with more - than two dimensions. If this happens, the - multidimensional tensor will first be flattened - into a 2-dimensional matrix. The parameter - `num_flatten_dims` determines how the input tensor - is flattened: the first `num_flatten_dims` - (inclusive, index starts from 1) dimensions will - be flatten to form the first dimension of the - final matrix (height of the matrix), and the rest - `rank(X) - num_flatten_dims` dimensions are - flattened to form the second dimension of the - final matrix (width of the matrix). For example, - suppose `X` is a 6-dimensional tensor with a shape - [2, 3, 4, 5, 6], and `num_flatten_dims` = 3. Then, - the flattened matrix will have a shape - [2 x 3 x 4, 5 x 6] = [24, 30]. By default, - `num_flatten_dims` is set to 1. - param_attr(ParamAttr|list): The parameter attribute for learnable - parameters/weights of the fully connected - layer. - param_initializer(ParamAttr|list): The initializer used for the - weight/parameter. If set None, - XavierInitializer() will be used. - bias_attr(ParamAttr|list): The parameter attribute for the bias parameter - for this layer. If set None, no bias will be - added to the output units. - bias_initializer(ParamAttr|list): The initializer used for the bias. - If set None, then ConstantInitializer() - will be used. - act(str): Activation to be applied to the output of the fully connected - layer. - name(str): Name/alias of the fully connected layer. - + input (Variable|list of Variable): The input tensor(s) of this layer, and the dimension of + the input tensor(s) is at least 2. + size(int): The number of output units in this layer. + num_flatten_dims (int, default 1): The fc layer can accept an input tensor with more than + two dimensions. If this happens, the multidimensional tensor will first be flattened + into a 2-dimensional matrix. The parameter `num_flatten_dims` determines how the input + tensor is flattened: the first `num_flatten_dims` (inclusive, index starts from 1) + dimensions will be flatten to form the first dimension of the final matrix (height of + the matrix), and the rest `rank(X) - num_flatten_dims` dimensions are flattened to + form the second dimension of the final matrix (width of the matrix). For example, suppose + `X` is a 6-dimensional tensor with a shape [2, 3, 4, 5, 6], and `num_flatten_dims` = 3. + Then, the flattened matrix will have a shape [2 x 3 x 4, 5 x 6] = [24, 30]. + param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable + parameters/weights of this layer. + bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias + of this layer. If it is set to None, no bias will be added to the output units. + act (str, default None): Activation to be applied to the output of this layer. + name (str, default None): The name of this layer. Returns: - Variable: The output tensor variable. + A tensor variable storing the transformation result. Raises: ValueError: If rank of the input tensor is less than 2. @@ -1102,6 +1088,30 @@ def sequence_conv(input, return helper.append_activation(pre_act) +def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): + helper = LayerHelper('sequence_softmax', **locals()) + dtype = helper.input_dtype() + softmax_out = helper.create_tmp_variable(dtype) + helper.append_op( + type="sequence_softmax", + inputs={"X": input}, + outputs={"Out": softmax_out}, + attrs={"use_cudnn": use_cudnn}) + return softmax_out + + +def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): + helper = LayerHelper('softmax', **locals()) + dtype = helper.input_dtype() + softmax_out = helper.create_tmp_variable(dtype) + helper.append_op( + type="softmax", + inputs={"X": input}, + outputs={"Out": softmax_out}, + attrs={"use_cudnn": use_cudnn}) + return softmax_out + + def conv2d(input, num_filters, filter_size, @@ -1406,6 +1416,7 @@ def pool2d(input, global_pooling=False, use_cudnn=True, ceil_mode=False, + use_mkldnn=False, name=None): """ This function adds the operator for pooling in 2 dimensions, using the @@ -1443,7 +1454,8 @@ def pool2d(input, "strides": pool_stride, "paddings": pool_padding, "use_cudnn": use_cudnn, - "ceil_mode": ceil_mode + "ceil_mode": ceil_mode, + "use_mkldnn": use_mkldnn }) return pool_out @@ -2200,6 +2212,53 @@ def reduce_min(input, dim=None, keep_dim=False, name=None): return out +def reduce_prod(input, dim=None, keep_dim=False, name=None): + """ + Computes the product of tensor elements over the given dimension. + + Args: + input (Variable): The input variable which is a Tensor or LoDTensor. + dim (int|None): The dimension along which the product is performed. If + :attr:`None`, multipy all elements of :attr:`input` and return a + Tensor variable with a single element, otherwise must be in the + range :math:`[-rank(input), rank(input))`. If :math:`dim < 0`, + the dimension to reduce is :math:`rank + dim`. + keep_dim (bool|False): Whether to reserve the reduced dimension in the + output Tensor. The result tensor will have one fewer dimension + than the :attr:`input` unless :attr:`keep_dim` is true. + name(str|None): A name for this layer(optional). If set None, the + layer will be named automatically. + + Returns: + Variable: The reduced Tensor variable. + + Examples: + .. code-block:: python + + # x is a Tensor variable with following elements: + # [[0.2, 0.3, 0.5, 0.9] + # [0.1, 0.2, 0.6, 0.7]] + # Each example is followed by the correspending output tensor. + fluid.layers.reduce_prod(x) # [0.0002268] + fluid.layers.reduce_prod(x, dim=0) # [0.02, 0.06, 0.3, 0.63] + fluid.layers.reduce_prod(x, dim=-1) # [0.027, 0.0084] + fluid.layers.reduce_prod(x, dim=1, + keep_dim=True) # [[0.027], [0.0084]] + """ + helper = LayerHelper('reduce_prod', **locals()) + out = helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.append_op( + type='reduce_prod', + inputs={'X': input}, + outputs={'Out': out}, + attrs={ + 'dim': dim if dim != None else 0, + 'keep_dim': keep_dim, + 'reduce_all': True if dim == None else False + }) + return out + + def split(input, num_or_sections, dim=-1, name=None): """ Split the input tensor into multiple sub-tensors. @@ -3234,3 +3293,98 @@ def autoincreased_step_counter(counter_name=None, begin=1, step=1): counter.stop_gradient = True return counter + + +def lod_reset(x, y=None, target_lod=None): + """ + LoD Reset Operator. Set LoD of **x** to a new one specified by **y** or + **target_lod**. When **y** provided, **y.lod** would be considered as target + LoD first, otherwise **y.data** would be considered as target LoD. If **y** + is not provided, target LoD should be specified by **target_lod**. + If target LoD is specified by **Y.data** or **target_lod**, only one level + LoD is supported. + + .. code-block:: text + + * Example 1: + + Given a 1-level LoDTensor x: + x.lod = [[ 0, 2, 5 6 ]] + x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + x.dims = [6, 1] + + target_lod: [0, 4, 6] + + then we get a 1-level LoDTensor: + out.lod = [[ 0, 4, 6 ]] + out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + out.dims = [6, 1] + + * Example 2: + + Given a 1-level LoDTensor x: + x.lod = [[ 0, 2, 5 6 ]] + x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + x.dims = [6, 1] + + y is a Tensor: + y.data = [[0, 2, 6]] + y.dims = [1, 3] + + then we get a 1-level LoDTensor: + out.lod = [[ 0, 2, 6 ]] + out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + out.dims = [6, 1] + + * Example 3: + + Given a 1-level LoDTensor x: + x.lod = [[ 0, 2, 5 6 ]] + x.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + x.dims = [6, 1] + + y is a 2-level LoDTensor: + y.lod = [[0, 2, 4], [0, 2, 5, 6]] + y.data = [[1.1], [2.1], [3.1], [4.1], [5.1], [6.1]] + y.dims = [6, 1] + + then we get a 2-level LoDTensor: + out.lod = [[0, 2, 4], [0, 2, 5, 6]] + out.data = [[1.0], [2.0], [3.0], [4.0], [5.0], [6.0]] + out.dims = [6, 1] + + Args: + x (Variable): Input variable which could be a Tensor or LodTensor. + y (Variable|None): If provided, output's LoD would be derived from y. + target_lod (list|tuple|None): One level LoD which should be considered + as target LoD when y not provided. + + Returns: + Variable: Output variable with LoD specified by this operator. + + Raises: + ValueError: If y and target_lod are both None. + + Examples: + .. code-block:: python + + x = layers.data(name='x', shape=[10]) + y = layers.data(name='y', shape=[10, 20], lod_level=2) + out = layers.lod_reset(x=x, y=y) + """ + helper = LayerHelper("lod_reset", **locals()) + out = helper.create_tmp_variable(dtype=x.dtype) + if y is not None: + helper.append_op( + type="lod_reset", inputs={'X': x, + 'Y': y}, outputs={'Out': out}) + elif target_lod is not None: + helper.append_op( + type="lod_reset", + inputs={'X': x}, + attrs={'target_lod': target_lod}, + outputs={'Out': out}) + else: + raise ValueError("y and target_lod should not be both None.") + + return out diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index 0b88b639629ac73b16ec36aa5930c3d6a9665943..d7bad221c5fa7b18137bf317125195267437a644 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -59,8 +59,6 @@ __all__ = [ 'elementwise_pow', 'clip', 'clip_by_norm', - 'softmax', - 'sequence_softmax', 'logical_and', 'logical_or', 'logical_xor', @@ -70,6 +68,7 @@ __all__ = [ 'gaussian_random', 'gaussian_random_batch_size_like', 'cumsum', + 'scatter', ] + __activations__ for _OP in set(__all__): diff --git a/python/paddle/fluid/memory_optimization_transpiler.py b/python/paddle/fluid/memory_optimization_transpiler.py index 4fa2d03ef563b98b2eec576bf87d4b2e54ca0a36..41d1eca82e8b680977f44f1756c25c37340668a4 100644 --- a/python/paddle/fluid/memory_optimization_transpiler.py +++ b/python/paddle/fluid/memory_optimization_transpiler.py @@ -29,7 +29,10 @@ dtype_to_size = { core.VarDesc.VarType.BOOL: 1 } -sub_block_ops = ["while", "while_grad", "parallel_do", "parallel_do_grad"] +sub_block_ops = [ + "while", "while_grad", "parallel_do", "parallel_do_grad", + "conditional_block", "conditional_block_grad" +] PRINT_LOG = False @@ -122,36 +125,80 @@ class ControlFlowGraph(object): else: return block_desc.find_var_recursive(str(var_name)) - def memory_optimize(self): - def check_var_validity(block_desc, x, is_forward): - if str(x) == "@EMPTY@": - return False - if not self._has_var(block_desc, x, is_forward): - return False - if self._find_var(block_desc, x, is_forward).persistable(): - return False - if self._find_var( - block_desc, x, - is_forward).type() != core.VarDesc.VarType.LOD_TENSOR: - return False - if x in self._skip_opt: - return False - if not self._find_var(block_desc, x, is_forward).shape(): - return False - return True + def _check_var_validity(self, block_desc, x, is_forward): + if str(x) == "@EMPTY@": + return False + if not self._has_var(block_desc, x, is_forward): + return False + if self._find_var(block_desc, x, is_forward).persistable(): + return False + if self._find_var(block_desc, x, + is_forward).type() != core.VarDesc.VarType.LOD_TENSOR: + return False + if x in self._skip_opt: + return False + if not self._find_var(block_desc, x, is_forward).shape(): + return False + return True - self._build_graph() + def _update_skip_opt_set(self): + for i in range(self.op_size): + op = self._ops[i] + if op.type() == "fill_constant" and op.attr("force_cpu") == True: + self._skip_opt.update(op.output_arg_names()) + + def release_memory(self): self._dataflow_analyze() + self._update_skip_opt_set() + fwd_id = 0 + bwd_id = 0 + for i in range(self.op_size): + op = self._ops[i] + if op.type() in sub_block_ops: + continue + block_desc = op.block() + is_forward = i < self._forward_num + in_diff, out_diff = self._get_diff(self._live_in[i], + self._live_out[i]) + can_optimize = filter( + lambda x: self._check_var_validity(block_desc, x, is_forward), + in_diff) + if can_optimize: + index = i + fwd_id + 1 if is_forward else i - self._forward_num + bwd_id + 1 + delete_op = block_desc.insert_op(index) + delete_op.set_type("delete_var") + delete_op.set_input("X", can_optimize) + if is_forward: + fwd_id += 1 + else: + bwd_id += 1 + + def memory_optimize(self, level=0): + def compare_shape(x_shape, cache_shape, opt_level): + if opt_level == 0: + return x_shape == cache_shape + if opt_level == 1: + if (x_shape[0] == -1) ^ (cache_shape[0] == -1): + return False + x_size = abs(reduce(lambda x, y: x * y, x_shape)) + cache_size = abs(reduce(lambda x, y: x * y, cache_shape)) + if x_size <= cache_size: + return True + return False + + self._dataflow_analyze() + self._update_skip_opt_set() self.pool = [] for i in range(self.op_size): op = self._ops[i] if op.type() in sub_block_ops: continue block_desc = op.block() + self.current_block_desc = block_desc is_forward = i < self._forward_num if self.pool: defs_can_optimize = filter( - lambda x: check_var_validity(block_desc, x, is_forward), + lambda x: self._check_var_validity(block_desc, x, is_forward), self._defs[i]) out_pair = [ (x, self._find_var(block_desc, x, is_forward).shape()) @@ -164,7 +211,7 @@ class ControlFlowGraph(object): for index, cache_pair in enumerate(self.pool): cache_var = cache_pair[0] cache_shape = cache_pair[1] - if x_shape == cache_shape: + if compare_shape(x_shape, cache_shape, level): if self._has_var(block_desc, cache_var, is_forward): x_dtype = self._find_var(block_desc, x, is_forward).dtype() @@ -196,7 +243,7 @@ class ControlFlowGraph(object): in_diff, out_diff = self._get_diff(self._live_in[i], self._live_out[i]) can_optimize = filter( - lambda x: check_var_validity(block_desc, x, is_forward), + lambda x: self._check_var_validity(block_desc, x, is_forward), in_diff) if can_optimize: for var_name in can_optimize: @@ -270,7 +317,8 @@ def _get_cfgs(input_program): ([block_desc.op(i) for i in range(op_size)], op_size, set())) sub_block_pair = [("while", "while_grad"), ("parallel_do", - "parallel_do_grad")] + "parallel_do_grad"), + ("conditional_block", "conditional_block_grad")] ops_list.extend(_process_sub_block_pair(pdesc, sub_block_pair)) @@ -281,9 +329,15 @@ def _get_cfgs(input_program): return cfgs -def memory_optimize(input_program, print_log=False): +def memory_optimize(input_program, print_log=False, level=0): global PRINT_LOG PRINT_LOG = print_log cfgs = _get_cfgs(input_program) for cfg in cfgs: - cfg.memory_optimize() + cfg.memory_optimize(level) + + +def release_memory(input_program): + cfgs = _get_cfgs(input_program) + for cfg in cfgs: + cfg.release_memory() diff --git a/python/paddle/fluid/nets.py b/python/paddle/fluid/nets.py index 8c627ad55bcbaf0cb0e8dd74a3e67ed40a6245db..3b2e1a3073251a6d6460450dc957e1b5c7a873c5 100644 --- a/python/paddle/fluid/nets.py +++ b/python/paddle/fluid/nets.py @@ -45,7 +45,8 @@ def simple_img_conv_pool(input, pool_size=pool_size, pool_type=pool_type, pool_stride=pool_stride, - use_cudnn=use_cudnn) + use_cudnn=use_cudnn, + use_mkldnn=use_mkldnn) return pool_out @@ -107,7 +108,8 @@ def img_conv_group(input, pool_size=pool_size, pool_type=pool_type, pool_stride=pool_stride, - use_cudnn=use_cudnn) + use_cudnn=use_cudnn, + use_mkldnn=use_mkldnn) return pool_out diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index d104cc5cbdcb6ab92092ea5b3ab05abf7e903b1f..75e1de00cf38a8e35fc0cb1f5a36459b7983bd63 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -94,7 +94,10 @@ class Optimizer(object): # create learning rate variable for every parameter param = param_and_grad[0] param_lr = param.optimize_attr['learning_rate'] - return self.global_learning_rate() * param_lr + if param_lr == 1.0: + return self.global_learning_rate() + else: + return self.global_learning_rate() * param_lr def _create_accumulators(self, block, parameters): """Create all accumulators needed by the parameters @@ -222,6 +225,8 @@ class Optimizer(object): params_grads = append_backward(loss, parameter_list, no_grad_set, [error_clip_callback]) + params_grads = sorted(params_grads, key=lambda x: x[0].name) + params_grads = append_gradient_clip_ops(params_grads) # Add regularization if any diff --git a/python/paddle/fluid/recordio_writer.py b/python/paddle/fluid/recordio_writer.py new file mode 100644 index 0000000000000000000000000000000000000000..5accaacd5361165d30b92c71ae4fd62e23e44e07 --- /dev/null +++ b/python/paddle/fluid/recordio_writer.py @@ -0,0 +1,48 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# 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. + +import core +import contextlib + +__all__ = ['convert_reader_to_recordio_file'] + + +@contextlib.contextmanager +def create_recordio_writer(filename, + compressor=core.RecordIOWriter.Compressor.Snappy, + max_num_records=1000): + writer = core.RecordIOWriter(filename, compressor, max_num_records) + yield writer + writer.close() + + +def convert_reader_to_recordio_file( + filename, + reader_creator, + feeder, + compressor=core.RecordIOWriter.Compressor.Snappy, + max_num_records=1000, + feed_order=None): + if feed_order is None: + feed_order = feeder.feed_names + counter = 0 + with create_recordio_writer(filename, compressor, + max_num_records) as writer: + for batch in reader_creator(): + res = feeder.feed(batch) + for each in feed_order: + writer.append_tensor(res[each]) + writer.complete_append_tensor() + counter += 1 + return counter diff --git a/python/paddle/fluid/regularizer.py b/python/paddle/fluid/regularizer.py index a29f9a208ebefc75b531030c9f0de9487f2b136c..604c6f9ab36c2332223d1ba943d67113922615b3 100644 --- a/python/paddle/fluid/regularizer.py +++ b/python/paddle/fluid/regularizer.py @@ -13,6 +13,7 @@ # limitations under the License. import framework +from . import core __all__ = [ 'append_regularization_ops', @@ -43,16 +44,20 @@ def append_regularization_ops(parameters_and_grads, regularization=None): """ params_and_grads = [] for param, grad in parameters_and_grads: + # If no gradient then we don't need to do anything + if grad is None: + params_and_grads.append((param, grad)) + continue + regularization_term = None if param.regularizer is not None: # Add variable for regularization term in grad block - regularization_term = param.regularizer(param, grad.block) + regularization_term = param.regularizer(param, grad, grad.block) elif regularization is not None: - regularization_term = regularization(param, grad.block) + regularization_term = regularization(param, grad, grad.block) - # If no gradient or no regularization specified, - # then we don't need to do anything - if grad is None or regularization_term is None: + # If no regularization specified, then we don't need to do anything + if regularization_term is None: params_and_grads.append((param, grad)) continue @@ -82,7 +87,7 @@ class WeightDecayRegularizer(object): def __init__(self): pass - def __call__(self, param, block): + def __call__(self, param, grad, block): """Add corresponding weight decay operations to the network """ raise NotImplementedError() @@ -102,7 +107,7 @@ class L2DecayRegularizer(WeightDecayRegularizer): super(L2DecayRegularizer, self).__init__() self._regularization_coeff = regularization_coeff - def __call__(self, param, block): + def __call__(self, param, grad, block): """Add L2 weight decay ops to network Adds L2 weight decay ops. @@ -117,8 +122,23 @@ class L2DecayRegularizer(WeightDecayRegularizer): """ assert isinstance(param, framework.Parameter) assert isinstance(block, framework.Block) + decay = block.create_var( dtype="float32", shape=param.shape, lod_level=param.lod_level) + + if grad.type == core.VarDesc.VarType.SELECTED_ROWS: + decay = block.create_var( + dtype="float32", + shape=param.shape, + type=core.VarDesc.VarType.SELECTED_ROWS) + block.append_op( + type='lookup_table', + inputs={'W': param, + 'Ids': grad}, + outputs={'Out': decay}, + attrs={'is_sparse': True}) + param = decay + # Append Op to calculate decay block.append_op( type='scale', @@ -141,7 +161,7 @@ class L1DecayRegularizer(WeightDecayRegularizer): super(L1DecayRegularizer, self).__init__() self._regularization_coeff = regularization_coeff - def __call__(self, param, block): + def __call__(self, param, grad, block): """Add L1 weight decay ops to network Adds L1 weight decay ops. @@ -158,6 +178,19 @@ class L1DecayRegularizer(WeightDecayRegularizer): assert isinstance(block, framework.Block) decay = block.create_var( dtype="float32", shape=param.shape, lod_level=param.lod_level) + + if grad.type == core.VarDesc.VarType.SELECTED_ROWS: + decay = block.create_var( + dtype="float32", + shape=param.shape, + type=core.VarDesc.VarType.SELECTED_ROWS) + block.append_op( + type='lookup_table', + inputs={'W': param, + 'Ids': grad}, + outputs={'Out': decay}, + attrs={'is_sparse': True}) + # Append sign op block.append_op( type='sign', inputs={"X": param}, outputs={"Out": decay}) diff --git a/python/paddle/fluid/tests/book/test_machine_translation.py b/python/paddle/fluid/tests/book/test_machine_translation.py index caa9596a100de4f9364467690db1e80ee227c3c1..fa38bd3762423497b82c3b421b3a1db4cd87525b 100644 --- a/python/paddle/fluid/tests/book/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/test_machine_translation.py @@ -181,7 +181,10 @@ def train_main(use_cuda, is_sparse, is_local=True): cost = pd.cross_entropy(input=rnn_out, label=label) avg_cost = pd.mean(cost) - optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4) + optimizer = fluid.optimizer.Adagrad( + learning_rate=1e-4, + regularization=fluid.regularizer.L2DecayRegularizer( + regularization_coeff=0.1)) optimize_ops, params_grads = optimizer.minimize(avg_cost) train_data = paddle.batch( diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py index c9d2a5ecaab0669f308b5b9c5cf74d0212fa462a..ad79e96b958b36a06c8a3cc990dbe3608e32c9ac 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py @@ -50,6 +50,7 @@ sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.01) sgd_optimizer.minimize(avg_cost) fluid.memory_optimize(fluid.default_main_program(), print_log=True) +# fluid.release_memory(fluid.default_main_program()) BATCH_SIZE = 200 @@ -69,8 +70,6 @@ exe.run(fluid.default_startup_program()) PASS_NUM = 100 for pass_id in range(PASS_NUM): - fluid.io.save_persistables(exe, "./fit_a_line.model/") - fluid.io.load_persistables(exe, "./fit_a_line.model/") for data in train_reader(): avg_loss_value, = exe.run(fluid.default_main_program(), feed=feeder.feed(data), diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py index 80ff11f8d78b0a22fc6aefd722c9e6a2c23fbd5c..204669d7e6176e9e8250e8aebc2d10441fa24b67 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py @@ -125,9 +125,10 @@ opts = optimizer.minimize(avg_cost) batch_size = fluid.layers.create_tensor(dtype='int64') batch_acc = fluid.layers.accuracy(input=predict, label=label, total=batch_size) -fluid.memory_optimize(fluid.default_main_program()) +# fluid.memory_optimize(fluid.default_main_program(), level=0) +fluid.release_memory(fluid.default_main_program()) -BATCH_SIZE = 128 +BATCH_SIZE = 16 PASS_NUM = 1 # fix the order of training data @@ -159,8 +160,7 @@ for pass_id in range(PASS_NUM): print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str( pass_acc)) # this model is slow, so if we can train two mini batch, we think it works properly. - - if i > 2: + if i > 0: exit(0) if math.isnan(float(loss)): sys.exit("got NaN loss, training failed.") diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py index 689a75afc7ccdf84142f5531a438e1f9af7af4ca..a24834a6f0b19d1265f6c8d7089d31583af82d1f 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py @@ -105,7 +105,8 @@ def main(): optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4) optimizer.minimize(avg_cost) - fluid.memory_optimize(fluid.default_main_program()) + # fluid.memory_optimize(fluid.default_main_program()) + fluid.release_memory(fluid.default_main_program()) # fix the order of training data train_data = paddle.batch( diff --git a/python/paddle/fluid/tests/test_concurrency.py b/python/paddle/fluid/tests/test_concurrency.py index 9f7bf63c5e017251e87af94690ff32c47c538c6b..924895a9afac610059bac5f617c49712441339cc 100644 --- a/python/paddle/fluid/tests/test_concurrency.py +++ b/python/paddle/fluid/tests/test_concurrency.py @@ -15,9 +15,9 @@ import unittest import paddle.fluid as fluid import paddle.fluid.core as core -from paddle.fluid import framework, unique_name +from paddle.fluid import framework, unique_name, layer_helper from paddle.fluid.executor import Executor -from paddle.fluid.layers import fill_constant +from paddle.fluid.layers import fill_constant, assign, While, elementwise_add, Print class TestRoutineOp(unittest.TestCase): @@ -86,8 +86,7 @@ class TestRoutineOp(unittest.TestCase): self.assertEqual(leftmost_data[0][0], n + 1) def _create_one_dim_tensor(self, value): - one_dim_tensor = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.INT64, value=value) + one_dim_tensor = fill_constant(shape=[1], dtype='int', value=value) one_dim_tensor.stop_gradient = True return one_dim_tensor @@ -95,6 +94,180 @@ class TestRoutineOp(unittest.TestCase): return framework.default_main_program().current_block().create_var( name=unique_name.generate(name), type=type, dtype=dtype) + def _create_persistable_tensor(self, name, type, dtype): + return framework.default_main_program().current_block().create_var( + name=unique_name.generate(name), + type=type, + dtype=dtype, + persistable=True) + + def test_select(self): + with framework.program_guard(framework.Program()): + ch1 = fluid.make_channel( + dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) + + result1 = self._create_tensor('return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.FP64) + + input_value = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.FP64, value=10) + + with fluid.Select() as select: + with select.case(fluid.channel_send, ch1, input_value): + # Execute something. + pass + + with select.default(): + pass + + # This should not block because we are using a buffered channel. + result1, status = fluid.channel_recv(ch1, result1) + fluid.channel_close(ch1) + + cpu = core.CPUPlace() + exe = Executor(cpu) + + result = exe.run(fetch_list=[result1]) + self.assertEqual(result[0][0], 10) + + def test_fibonacci(self): + """ + Mimics Fibonacci Go example: https://tour.golang.org/concurrency/5 + """ + with framework.program_guard(framework.Program()): + quit_ch_input_var = self._create_persistable_tensor( + 'quit_ch_input', core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.INT32) + quit_ch_input = fill_constant( + shape=[1], + dtype=core.VarDesc.VarType.INT32, + value=0, + out=quit_ch_input_var) + + result = self._create_persistable_tensor( + 'result', core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.INT32) + fill_constant( + shape=[1], + dtype=core.VarDesc.VarType.INT32, + value=0, + out=result) + + x = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) + y = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=1) + + while_cond = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.BOOL, value=True) + + while_false = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.BOOL, value=False) + + x_tmp = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) + + def fibonacci(channel, quit_channel): + while_op = While(cond=while_cond) + with while_op.block(): + result2 = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) + x_to_send_tmp = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) + + # TODO(abhinav): Need to perform copy when doing a channel send. + # Once this is complete, we can remove these lines + assign(input=x, output=x_to_send_tmp) + + with fluid.Select() as select: + with select.case(fluid.channel_send, channel, + x_to_send_tmp): + assign(input=x, output=x_tmp) + assign(input=y, output=x) + assign(elementwise_add(x=x_tmp, y=y), output=y) + + with select.case(fluid.channel_recv, quit_channel, + result2): + # Quit + helper = layer_helper.LayerHelper('assign') + helper.append_op( + type='assign', + inputs={'X': [while_false]}, + outputs={'Out': [while_cond]}) + + ch1 = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) + quit_ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) + + with fluid.Go(): + for i in xrange(10): + fluid.channel_recv(ch1, result) + Print(result) + + fluid.channel_send(quit_ch, quit_ch_input) + + fibonacci(ch1, quit_ch) + + fluid.channel_close(ch1) + fluid.channel_close(quit_ch) + + cpu = core.CPUPlace() + exe = Executor(cpu) + + exe_result = exe.run(fetch_list=[result]) + self.assertEqual(exe_result[0][0], 34) + + def test_ping_pong(self): + """ + Mimics Ping Pong example: https://gobyexample.com/channel-directions + """ + with framework.program_guard(framework.Program()): + result = self._create_tensor('return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.FP64) + + ping_result = self._create_tensor('ping_return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.FP64) + + pong_result = self._create_tensor('pong_return_value', + core.VarDesc.VarType.LOD_TENSOR, + core.VarDesc.VarType.FP64) + + def ping(ch, message): + message_to_send_tmp = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.FP64, value=0) + + assign(input=message, output=message_to_send_tmp) + fluid.channel_send(ch, message_to_send_tmp) + + def pong(ch1, ch2): + fluid.channel_recv(ch1, ping_result) + assign(input=ping_result, output=pong_result) + fluid.channel_send(ch2, pong_result) + + pings = fluid.make_channel( + dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) + pongs = fluid.make_channel( + dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) + + msg = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.FP64, value=9) + + ping(pings, msg) + pong(pings, pongs) + + fluid.channel_recv(pongs, result) + + fluid.channel_close(pings) + fluid.channel_close(pongs) + + cpu = core.CPUPlace() + exe = Executor(cpu) + + exe_result = exe.run(fetch_list=[result]) + self.assertEqual(exe_result[0][0], 9) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/test_cpp_reader.py b/python/paddle/fluid/tests/test_cpp_reader.py index b65592057817cef83bf2157c55bacea5bbe34ea1..4b0d039b7e05a55980946a8949e32802e9e57c20 100644 --- a/python/paddle/fluid/tests/test_cpp_reader.py +++ b/python/paddle/fluid/tests/test_cpp_reader.py @@ -15,16 +15,30 @@ import paddle.v2 as paddle import paddle.fluid as fluid import numpy as np +import sys -prog = fluid.framework.Program() -block = prog.current_block() +startup_prog = fluid.framework.Program() +startup_block = startup_prog.current_block() -random_reader = block.create_var( +random_reader = startup_block.create_var( type=fluid.core.VarDesc.VarType.READER, name="RandomDataGenerator") random_reader.desc.set_dtypes( [fluid.core.VarDesc.VarType.FP32, fluid.core.VarDesc.VarType.FP32]) +random_reader.persistable = True +shuffle_reader = startup_block.create_var( + type=fluid.core.VarDesc.VarType.READER, name="ShuffleReader") +shuffle_reader.persistable = True +batch_reader = startup_block.create_var( + type=fluid.core.VarDesc.VarType.READER, name="BatchReader") +batch_reader.persistable = True +double_buffer = startup_block.create_var( + type=fluid.core.VarDesc.VarType.READER, name="DoubleBuffer") +double_buffer.persistable = True + +main_prog = startup_prog.clone() +main_block = main_prog.current_block() -create_random_data_generator_op = block.append_op( +create_random_data_generator_op = startup_block.append_op( type="create_random_data_generator", outputs={"Out": random_reader}, attrs={ @@ -34,37 +48,45 @@ create_random_data_generator_op = block.append_op( "max": 1.0, 'lod_levels': [0, 0] }) -shuffle_reader = block.create_var( - type=fluid.core.VarDesc.VarType.READER, name="ShuffleReader") -create_shuffle_reader_op = block.append_op( +create_shuffle_reader_op = startup_block.append_op( type="create_shuffle_reader", inputs={"UnderlyingReader": random_reader}, outputs={"Out": shuffle_reader}, attrs={"buffer_size": 7}) -batch_reader = block.create_var( - type=fluid.core.VarDesc.VarType.READER, name="BatchReader") - -create_batch_reader_op = block.append_op( +create_batch_reader_op = startup_block.append_op( type="create_batch_reader", inputs={"UnderlyingReader": shuffle_reader}, outputs={"Out": batch_reader}, attrs={"batch_size": 10}) -out1 = block.create_var(type=fluid.core.VarDesc.VarType.LOD_TENSOR, name="Out1") -out2 = block.create_var(type=fluid.core.VarDesc.VarType.LOD_TENSOR, name="Out2") +create_double_buffer_reader_op = startup_block.append_op( + type="create_double_buffer_reader", + inputs={"UnderlyingReader": batch_reader}, + outputs={"Out": double_buffer}) + +out1 = main_block.create_var( + type=fluid.core.VarDesc.VarType.LOD_TENSOR, name="Out1") +out2 = main_block.create_var( + type=fluid.core.VarDesc.VarType.LOD_TENSOR, name="Out2") -read_op = block.append_op( - type="read", inputs={"Reader": batch_reader}, +main_block.var("DoubleBuffer").desc.set_shapes(double_buffer.desc.shapes()) +main_block.var("DoubleBuffer").desc.set_dtypes(double_buffer.desc.dtypes()) +main_block.var("DoubleBuffer").desc.set_lod_levels( + double_buffer.desc.lod_levels()) + +read_op = main_block.append_op( + type="read", + inputs={"Reader": double_buffer}, outputs={"Out": [out1, out2]}) place = fluid.CPUPlace() exe = fluid.Executor(place) -[res1, res2] = exe.run(prog, fetch_list=[out1, out2]) - -if not (res1.shape == (10, 2) and res2.shape == (10, 1)): - exit(1) +exe.run(startup_prog) -exit(0) +for i in range(1, 100): + [res1, res2] = exe.run(main_prog, fetch_list=[out1, out2]) + if not (res1.shape == (10, 2) and res2.shape == (10, 1)): + exit(1) diff --git a/python/paddle/fluid/tests/unittests/.gitignore b/python/paddle/fluid/tests/unittests/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..6b3fc2a83c649c28d21c9a8a0b35c2f2fa04f269 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/.gitignore @@ -0,0 +1 @@ +mnist.recordio diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index f96c2ca4f0593b6c2624d449304f23425c69ab93..0ad273c7161977e18f91f952fd3a9dc144bf73f0 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -11,7 +11,6 @@ list(REMOVE_ITEM TEST_OPS test_lstm_unit_op) # # FIXME(qijun) https://github.com list(REMOVE_ITEM TEST_OPS test_nce) # IXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/7778 list(REMOVE_ITEM TEST_OPS test_recurrent_op) # FIXME(qijun) https://github.com/PaddlePaddle/Paddle/issues/6152 list(REMOVE_ITEM TEST_OPS test_cond_op) # FIXME(qijun): https://github.com/PaddlePaddle/Paddle/issues/5101#issuecomment-339814957 -list(REMOVE_ITEM TEST_OPS test_detection_output_op) # FIXME: detection_output_op will be rewritten. This unittest should be list(REMOVE_ITEM TEST_OPS op_test) # op_test is a helper python file, not a test list(REMOVE_ITEM TEST_OPS decorators) # decorators is a helper python file, not a test diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index f7e02595ec3b41ae7bb32353c258736968ca78d4..8393f7827b1c7d361ebea72f2cfc6033268772f0 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -469,6 +469,28 @@ class OpTest(unittest.TestCase): tensor.set_lod(lod) return tensor + @staticmethod + def np_dtype_to_fluid_dtype(input): + """Change the dtype of float16 numpy array + + numpy float16 is binded to paddle::platform::float16 + in tensor_py.h via the help of uint16 data type since + the internal memory representation of float16 is + uint16_t in paddle and np.uint16 in numpy, which are + themselves binded together by pybind. + + Args: + input: input numpy array + + Returns: + input: The dtype of input will be changed to np.uint16 if + it is originally np.float16, such that the internal memory + of input will be reinterpreted as of dtype np.uint16. + """ + if input.dtype == np.float16: + input.dtype = np.uint16 + return input + def _get_gradient(self, input_to_check, place, output_names, no_grad_set): prog = Program() block = prog.global_block() diff --git a/python/paddle/fluid/tests/unittests/test_cast_op.py b/python/paddle/fluid/tests/unittests/test_cast_op.py index 8fb8d03828393ccfe57c0848d79b960c641ad39a..b8d3ed3aa3eb0e47e79f46cdf681a3b9cca46036 100644 --- a/python/paddle/fluid/tests/unittests/test_cast_op.py +++ b/python/paddle/fluid/tests/unittests/test_cast_op.py @@ -18,7 +18,7 @@ import numpy as np import paddle.fluid.core as core -class TestCastOp(op_test.OpTest): +class TestCastOp1(op_test.OpTest): def setUp(self): ipt = np.random.random(size=[10, 10]) self.inputs = {'X': ipt.astype('float32')} @@ -36,5 +36,36 @@ class TestCastOp(op_test.OpTest): self.check_grad(['X'], ['Out']) +class TestCastOp2(op_test.OpTest): + def setUp(self): + ipt = np.random.random(size=[10, 10]) + # numpy float16 is binded to fluid float16 via uint16 + self.inputs = {'X': ipt.astype('float16').view(np.uint16)} + self.outputs = {'Out': ipt.astype('float32')} + self.attrs = { + 'in_dtype': int(core.VarDesc.VarType.FP16), + 'out_dtype': int(core.VarDesc.VarType.FP32) + } + self.op_type = 'cast' + + def test_check_output(self): + self.check_output(atol=1e-3) + + +class TestCastOp3(op_test.OpTest): + def setUp(self): + ipt = np.random.random(size=[10, 10]) + self.inputs = {'X': ipt.astype('float32')} + self.outputs = {'Out': ipt.astype('float16')} + self.attrs = { + 'in_dtype': int(core.VarDesc.VarType.FP32), + 'out_dtype': int(core.VarDesc.VarType.FP16) + } + self.op_type = 'cast' + + def test_check_output(self): + self.check_output(atol=1e-3) + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_op.py b/python/paddle/fluid/tests/unittests/test_conv2d_op.py index a49fecf09509f7b1d9f758eebcf90bf9fbf7669f..4b6e3fb69a12095c77f343515fe3b6d1f3fccb14 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv2d_op.py @@ -63,9 +63,11 @@ def conv2d_forward_naive(input, filter, group, conv_param): class TestConv2dOp(OpTest): def setUp(self): + self.op_type = "conv2d" self.use_cudnn = False self.use_mkldnn = False - self.init_op_type() + self.dtype = np.float32 + self.init_kernel_type() self.init_group() self.init_dilation() self.init_test_case() @@ -75,12 +77,16 @@ class TestConv2dOp(OpTest): 'pad': self.pad, 'dilation': self.dilations } - input = np.random.random(self.input_size).astype("float32") - filter = np.random.random(self.filter_size).astype("float32") + + input = np.random.random(self.input_size).astype(self.dtype) + filter = np.random.random(self.filter_size).astype(self.dtype) output = conv2d_forward_naive(input, filter, self.groups, - conv2d_param).astype('float32') + conv2d_param).astype(self.dtype) - self.inputs = {'Input': input, 'Filter': filter} + self.inputs = { + 'Input': OpTest.np_dtype_to_fluid_dtype(input), + 'Filter': OpTest.np_dtype_to_fluid_dtype(filter) + } self.attrs = { 'strides': self.stride, 'paddings': self.pad, @@ -99,6 +105,8 @@ class TestConv2dOp(OpTest): self.check_output() def test_check_grad(self): + if self.dtype == np.float16: + return if self.use_cudnn: place = core.CUDAPlace(0) self.check_grad_with_place( @@ -111,6 +119,8 @@ class TestConv2dOp(OpTest): set(['Input', 'Filter']), 'Output', max_relative_error=0.02) def test_check_grad_no_filter(self): + if self.dtype == np.float16: + return if self.use_cudnn: place = core.CUDAPlace(0) self.check_grad_with_place( @@ -126,6 +136,8 @@ class TestConv2dOp(OpTest): no_grad_set=set(['Filter'])) def test_check_grad_no_input(self): + if self.dtype == np.float16: + return if self.use_cudnn: place = core.CUDAPlace(0) self.check_grad_with_place( @@ -154,8 +166,8 @@ class TestConv2dOp(OpTest): def init_group(self): self.groups = 1 - def init_op_type(self): - self.op_type = "conv2d" + def init_kernel_type(self): + pass class TestWithPad(TestConv2dOp): @@ -227,39 +239,105 @@ class TestWithInput1x1Filter1x1(TestConv2dOp): #----------------Conv2dCUDNN---------------- class TestCUDNN(TestConv2dOp): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv2d" + + +class TestFP16CUDNN(TestConv2dOp): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestCUDNNWithPad(TestWithPad): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv2d" + + +class TestFP16CUDNNWithPad(TestWithPad): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestCUDNNWithStride(TestWithStride): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv2d" + + +class TestFP16CUDNNWithStride(TestWithStride): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestCUDNNWithGroup(TestWithGroup): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv2d" + + +class TestFP16CUDNNWithGroup(TestWithGroup): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestCUDNNWith1x1(TestWith1x1): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv2d" + + +class TestFP16CUDNNWith1x1(TestWith1x1): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestCUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "conv2d" + + +class TestFP16CUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) class TestDepthwiseConv(TestConv2dOp): @@ -295,21 +373,18 @@ class TestDepthwiseConv2(TestConv2dOp): #----------------Conv2dMKLDNN---------------- class TestMKLDNN(TestConv2dOp): - def init_op_type(self): + def init_kernel_type(self): self.use_mkldnn = True - self.op_type = "conv2d" class TestMKLDNNWithPad(TestWithPad): - def init_op_type(self): + def init_kernel_type(self): self.use_mkldnn = True - self.op_type = "conv2d" class TestMKLDNNWithStride(TestWithStride): - def init_op_type(self): + def init_kernel_type(self): self.use_mkldnn = True - self.op_type = "conv2d" if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_debugger.py b/python/paddle/fluid/tests/unittests/test_debugger.py new file mode 100644 index 0000000000000000000000000000000000000000..2b7bbf9218f9b8fd8f5b29ac3cbc2f9680f471eb --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_debugger.py @@ -0,0 +1,58 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# 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. + +import unittest +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.fluid import debuger +from paddle.fluid.framework import Program + + +class TestDebugger(unittest.TestCase): + def test_debug_str(self): + p = Program() + b = p.current_block() + + #selected_rows + b.create_var( + name='selected_rows', + dtype="float32", + shape=[5, 10], + type=core.VarDesc.VarType.SELECTED_ROWS) + + #tensor array + b.create_var( + name='tensor_array', + shape=[5, 10], + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY) + + #operator + mul_x = b.create_parameter( + dtype="float32", shape=[5, 10], lod_level=0, name="mul.x") + mul_y = b.create_var( + dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") + mul_out = b.create_var( + dtype="float32", shape=[5, 8], lod_level=0, name="mul.out") + b.append_op( + type="mul", + inputs={"X": mul_x, + "Y": mul_y}, + outputs={"Out": mul_out}, + attrs={"x_num_col_dims": 1}) + + print(debuger.pprint_program_codes(p.desc)) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_detection_map_op.py b/python/paddle/fluid/tests/unittests/test_detection_map_op.py index f3197a623efb1730d27467c5650bc90f2762e7b2..a905a854ad157ffa3d7816dfbd445f3e344a1249 100644 --- a/python/paddle/fluid/tests/unittests/test_detection_map_op.py +++ b/python/paddle/fluid/tests/unittests/test_detection_map_op.py @@ -166,8 +166,6 @@ class TestDetectionMAPOp(OpTest): elif not difficult: label_count[label] += 1 - true_pos = collections.defaultdict(list) - false_pos = collections.defaultdict(list) for (label, score, tp, fp) in tf_pos: true_pos[label].append([score, tp]) false_pos[label].append([score, fp]) diff --git a/python/paddle/fluid/tests/unittests/test_detection_output_op.py b/python/paddle/fluid/tests/unittests/test_detection_output_op.py deleted file mode 100644 index 94681319144ee3e0d51b57944f5692183578c01b..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/unittests/test_detection_output_op.py +++ /dev/null @@ -1,71 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# 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. - -import unittest -import numpy as np -from op_test import OpTest - - -class TestUnpoolOp(OpTest): - def setUp(self): - self.op_type = "detection_output" - self.init_test_case() - - #loc.shape ((1, 4, 4, 1, 1)) - #conf.shape ((1, 4, 2, 1, 1)) - - loc = np.array([[[[[0.1]], [[0.1]], [[0.1]], [[0.1]]], - [[[0.1]], [[0.1]], [[0.1]], [[0.1]]], - [[[0.1]], [[0.1]], [[0.1]], [[0.1]]], - [[[0.1]], [[0.1]], [[0.1]], [[0.1]]]]]) - conf = np.array([[[[[0.1]], [[0.9]]], [[[0.2]], [[0.8]]], - [[[0.3]], [[0.7]]], [[[0.4]], [[0.6]]]]]) - priorbox = np.array([ - 0.1, 0.1, 0.5, 0.5, 0.1, 0.1, 0.2, 0.2, 0.2, 0.2, 0.6, 0.6, 0.1, - 0.1, 0.2, 0.2, 0.3, 0.3, 0.7, 0.7, 0.1, 0.1, 0.2, 0.2, 0.4, 0.4, - 0.8, 0.8, 0.1, 0.1, 0.2, 0.2 - ]) - - output = np.array([ - 0, 1, 0.68997443, 0.099959746, 0.099959746, 0.50804031, 0.50804031 - ]) - self.inputs = { - 'Loc': loc.astype('float32'), - 'Conf': conf.astype('float32'), - 'PriorBox': priorbox.astype('float32') - } - self.attrs = { - 'num_classes': self.num_classes, - 'top_k': self.top_k, - 'nms_top_k': self.nms_top_k, - 'background_label_id': self.background_label_id, - 'nms_threshold': self.nms_threshold, - 'confidence_threshold': self.confidence_threshold, - } - self.outputs = {'Out': output.astype('float32')} - - def test_check_output(self): - self.check_output() - - def init_test_case(self): - self.num_classes = 2 - self.top_k = 10 - self.nms_top_k = 20 - self.background_label_id = 0 - self.nms_threshold = 0.01 - self.confidence_threshold = 0.01 - - -if __name__ == '__main__': - unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 6944cca394fbc1ddde09dfeb0bc82e357a3cd225..744a762ae7664f1f28713c505f9112ba712fd41d 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -220,7 +220,7 @@ class TestBook(unittest.TestCase): seq_data = layers.data( name='seq_data', shape=[10, 10], dtype='float32', lod_level=1) seq = layers.fc(input=seq_data, size=20) - self.assertIsNotNone(layers.sequence_softmax(x=seq)) + self.assertIsNotNone(layers.sequence_softmax(seq)) print(str(program)) def test_softmax(self): @@ -228,7 +228,7 @@ class TestBook(unittest.TestCase): with program_guard(program): data = layers.data(name='data', shape=[10], dtype='float32') hid = layers.fc(input=data, size=20) - self.assertIsNotNone(layers.softmax(x=hid)) + self.assertIsNotNone(layers.softmax(hid)) print(str(program)) def test_get_places(self): @@ -327,6 +327,15 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(loss) print(str(program)) + def test_lod_reset(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[10], dtype='float32') + y = layers.data( + name='y', shape=[10, 20], dtype='float32', lod_level=2) + print(layers.lod_reset(x=x, y=y)) + print(str(program)) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py b/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py index e75a6529e9fa265121ba187f3ed6bc0273c058d7..6382e290eb30c621da64d5c600be6d8a7c6254f1 100644 --- a/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py +++ b/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py @@ -19,6 +19,7 @@ import unittest import paddle.fluid as fluid import paddle.fluid.layers as layers import paddle.fluid.framework as framework +import paddle.fluid.core as core def exponential_decay(learning_rate, @@ -81,12 +82,25 @@ def piecewise_decay(global_step, boundaries, values): class TestLearningRateDecay(unittest.TestCase): def check_decay(self, python_decay_fn, fluid_decay_fn, kwargs): + places = [fluid.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(fluid.CUDAPlace(0)) + for place in places: + self.check_decay_with_place(place, python_decay_fn, fluid_decay_fn, + kwargs) + + def check_decay_with_place(self, place, python_decay_fn, fluid_decay_fn, + kwargs): + decayed_lr = fluid_decay_fn(**kwargs) place = fluid.CPUPlace() exe = fluid.Executor(place) exe.run(fluid.default_startup_program()) + + fluid.memory_optimize(fluid.default_main_program()) + for step in range(10): lr_val, = exe.run(fluid.default_main_program(), feed={}, diff --git a/python/paddle/fluid/tests/unittests/test_lod_reset_op.py b/python/paddle/fluid/tests/unittests/test_lod_reset_op.py index 3bf8230f8748dd87ec3c85b0cbd78df2e695a96b..6b6d4c824aeae319dacf224408ce96a0d9c5bb35 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_reset_op.py +++ b/python/paddle/fluid/tests/unittests/test_lod_reset_op.py @@ -42,7 +42,7 @@ class TestLodResetOpByInput(OpTest): target_lod_0 = [0, 4, 7, 10] self.inputs = { 'X': (x, lod), - 'TargetLoD': np.array([target_lod_0]).astype('int32') + 'Y': np.array([target_lod_0]).astype('int32') } self.outputs = {'Out': (x, [target_lod_0])} @@ -50,7 +50,7 @@ class TestLodResetOpByInput(OpTest): self.check_output() def test_check_grad(self): - self.check_grad(["X"], "Out", no_grad_set=set("TargetLoD")) + self.check_grad(["X"], "Out", no_grad_set=set("Y")) class TestLodResetOpBoth(OpTest): @@ -62,7 +62,7 @@ class TestLodResetOpBoth(OpTest): target_lod_0_in = [0, 4, 7, 10] self.inputs = { 'X': (x, lod), - 'TargetLoD': np.array(target_lod_0_in).astype('int32') + 'Y': np.array(target_lod_0_in).astype('int32') } self.attrs = {'target_lod': target_lod_0_attr} self.outputs = {'Out': (x, [target_lod_0_in])} @@ -71,7 +71,24 @@ class TestLodResetOpBoth(OpTest): self.check_output() def test_check_grad(self): - self.check_grad(["X"], "Out", no_grad_set=set("TargetLoD")) + self.check_grad(["X"], "Out", no_grad_set=set("Y")) + + +class TestLodResetOpYIsLoDTensor(OpTest): + def setUp(self): + self.op_type = "lod_reset" + x = np.random.random((10, 20)).astype("float32") + lod = [[0, 3, 5, 10]] + y = np.random.random((10, 10)).astype("float32") + target_lod_0 = [[0, 4, 7, 10]] + self.inputs = {'X': (x, lod), 'Y': (y, target_lod_0)} + self.outputs = {'Out': (x, target_lod_0)} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out", no_grad_set=set("Y")) if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_lookup_table_op.py b/python/paddle/fluid/tests/unittests/test_lookup_table_op.py index 03a5bd24a133e703855400532517c293196b64f0..ed920ad388ff0e01887404e70fe82565b4cd28fa 100644 --- a/python/paddle/fluid/tests/unittests/test_lookup_table_op.py +++ b/python/paddle/fluid/tests/unittests/test_lookup_table_op.py @@ -15,6 +15,8 @@ import unittest import numpy as np from op_test import OpTest +import paddle.fluid.core as core +from paddle.fluid.op import Operator class TestLookupTableOp(OpTest): @@ -47,5 +49,52 @@ class TestLookupTableOpWithPadding(TestLookupTableOp): pass +class TestLookupTableIdsIsSelectedRows(OpTest): + def check_with_place(self, place): + scope = core.Scope() + + # create and initialize Variable + height = 10 + rows = [0, 4, 4, 7] + row_numel = 12 + + # create and initialize W Variable + W = scope.var('W').get_tensor() + W_array = np.full((height, row_numel), 1.0).astype("float32") + for i in range(height): + W_array[i] *= i + W.set(W_array, place) + + # create and initialize Ids Variable + ids_selected_rows = scope.var('Ids').get_selected_rows() + ids_selected_rows.set_height(len(rows)) + ids_selected_rows.set_rows(rows) + np_array = np.ones((len(rows), row_numel)).astype("float32") + ids_tensor = ids_selected_rows.get_tensor() + ids_tensor.set(np_array, place) + + # create Out Variable + Out = scope.var('Out').get_selected_rows() + + # create and run lookup_table operator + concat_rows_op = Operator("lookup_table", W='W', Ids='Ids', Out='Out') + concat_rows_op.run(scope, place) + + # get result from Out + Out_tensor = Out.get_tensor() + result_array = np.array(Out_tensor) + + # all(): return True if all elements of the iterable are true (or if the iterable is empty) + for idx, row in enumerate(rows): + assert (row == result_array[idx]).all() + + def test_concat_rows(self): + places = [core.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) + for place in places: + self.check_with_place(place) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_lrn_op.py b/python/paddle/fluid/tests/unittests/test_lrn_op.py index 7f2352c5882ce36d8d681a737806f3ee0e3ace98..eaff45cbb2a58798e9d55149510bec72eea370cd 100644 --- a/python/paddle/fluid/tests/unittests/test_lrn_op.py +++ b/python/paddle/fluid/tests/unittests/test_lrn_op.py @@ -41,7 +41,7 @@ class TestLRNOp(OpTest): mid.fill(self.k) for m in range(0, self.N): for i in range(0, self.C): - for c in range(start, end + 1): + for c in range(start, end): ch = i + c if ch < 0 or ch >= self.C: continue diff --git a/python/paddle/fluid/tests/unittests/test_mul_op.py b/python/paddle/fluid/tests/unittests/test_mul_op.py index 9d1da420c7f70bd2a89d183a5f0a2b145f0ff475..40440bea1267112b84b66002a0bf921be3029265 100644 --- a/python/paddle/fluid/tests/unittests/test_mul_op.py +++ b/python/paddle/fluid/tests/unittests/test_mul_op.py @@ -14,6 +14,7 @@ import unittest import numpy as np +import paddle.fluid.core as core from op_test import OpTest @@ -69,5 +70,42 @@ class TestMulOp2(OpTest): ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) +class TestFP16MulOp1(OpTest): + def setUp(self): + self.op_type = "mul" + x = np.random.random((32, 84)).astype("float16") + y = np.random.random((84, 100)).astype("float16") + self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} + self.outputs = {'Out': np.dot(x, y)} + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-1) + + +class TestFP16MulOp2(OpTest): + def setUp(self): + self.op_type = "mul" + x = np.random.random((15, 4, 12, 10)).astype("float16") + y = np.random.random((4, 30, 8, 2, 9)).astype("float16") + self.inputs = {'X': x.view(np.uint16), 'Y': y.view(np.uint16)} + self.attrs = { + 'x_num_col_dims': 2, + 'y_num_col_dims': 2, + } + result = np.dot( + x.reshape(15 * 4, 12 * 10), y.reshape(4 * 30, 8 * 2 * 9)) + result = result.reshape(15, 4, 8, 2, 9) + self.outputs = {'Out': result} + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-1) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_optimizer.py b/python/paddle/fluid/tests/unittests/test_optimizer.py index 9d87f4daa98da42fcc33aba4b51a4528343fb137..e775db1d10f4561b6fb90631757a25c9f74cb777 100644 --- a/python/paddle/fluid/tests/unittests/test_optimizer.py +++ b/python/paddle/fluid/tests/unittests/test_optimizer.py @@ -21,31 +21,43 @@ from paddle.fluid.backward import append_backward class TestOptimizer(unittest.TestCase): def test_sgd_optimizer(self): - init_program = framework.Program() - program = framework.Program() - block = program.global_block() - mul_x = block.create_parameter( - dtype="float32", shape=[5, 10], lod_level=0, name="mul.x") - mul_y = block.create_var( - dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") - mul_out = block.create_var( - dtype="float32", shape=[5, 8], lod_level=0, name="mul.out") - mean_out = block.create_var( - dtype="float32", shape=[1], lod_level=0, name="mean.out") - block.append_op( - type="mul", - inputs={"X": mul_x, - "Y": mul_y}, - outputs={"Out": mul_out}, - attrs={"x_num_col_dims": 1}) - block.append_op( - type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out}) - sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.01) - opts, _ = sgd_optimizer.minimize(mean_out, init_program) + def check_sgd_optimizer(optimizer_attr): + init_program = framework.Program() + program = framework.Program() + block = program.global_block() + mul_x = block.create_parameter( + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + optimize_attr=optimizer_attr) + mul_y = block.create_var( + dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") + mul_out = block.create_var( + dtype="float32", shape=[5, 8], lod_level=0, name="mul.out") + mean_out = block.create_var( + dtype="float32", shape=[1], lod_level=0, name="mean.out") + block.append_op( + type="mul", + inputs={"X": mul_x, + "Y": mul_y}, + outputs={"Out": mul_out}, + attrs={"x_num_col_dims": 1}) + block.append_op( + type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out}) + sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.01) + opts, _ = sgd_optimizer.minimize(mean_out, init_program) + return opts + + opts = check_sgd_optimizer({'learning_rate': 1.1}) self.assertEqual(len(opts), 3) self.assertEqual([op.type for op in opts], ["fill_constant", "elementwise_mul", "sgd"]) + opts = check_sgd_optimizer({'learning_rate': 1.0}) + self.assertEqual(len(opts), 1) + self.assertEqual([op.type for op in opts], ["sgd"]) + class TestMomentumOptimizer(unittest.TestCase): class MockMomentum(optimizer.MomentumOptimizer): @@ -60,7 +72,11 @@ class TestMomentumOptimizer(unittest.TestCase): program = framework.Program() block = program.global_block() mul_x = block.create_parameter( - dtype="float32", shape=[5, 10], lod_level=0, name="mul.x") + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + optimize_attr={'learning_rate': 1.1}) mul_y = block.create_var( dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") mul_out = block.create_var( @@ -110,7 +126,11 @@ class TestMomentumOptimizer(unittest.TestCase): program = framework.Program() block = program.global_block() mul_x = block.create_parameter( - dtype="float32", shape=[5, 10], lod_level=0, name="mul.x") + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + optimize_attr={'learning_rate': 1.1}) mul_y = block.create_var( dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") mul_out = block.create_var( @@ -169,7 +189,11 @@ class TestAdagradOptimizer(unittest.TestCase): program = framework.Program() block = program.global_block() mul_x = block.create_parameter( - dtype="float32", shape=[5, 10], lod_level=0, name="mul.x") + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + optimize_attr={'learning_rate': 1.1}) mul_y = block.create_var( dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") mul_out = block.create_var( @@ -229,7 +253,11 @@ class TestAdamOptimizer(unittest.TestCase): program = framework.Program() block = program.global_block() mul_x = block.create_parameter( - dtype="float32", shape=[5, 10], lod_level=0, name="mul.x") + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + optimize_attr={'learning_rate': 1.1}) mul_y = block.create_var( dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") mul_out = block.create_var( @@ -292,7 +320,11 @@ class TestAdamaxOptimizer(unittest.TestCase): program = framework.Program() block = program.global_block() mul_x = block.create_parameter( - dtype="float32", shape=[5, 10], lod_level=0, name="mul.x") + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + optimize_attr={'learning_rate': 1.1}) mul_y = block.create_var( dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") mul_out = block.create_var( @@ -352,7 +384,11 @@ class TestDecayedAdagradOptimizer(unittest.TestCase): program = framework.Program() block = program.global_block() mul_x = block.create_parameter( - dtype="float32", shape=[5, 10], lod_level=0, name="mul.x") + dtype="float32", + shape=[5, 10], + lod_level=0, + name="mul.x", + optimize_attr={'learning_rate': 1.1}) mul_y = block.create_var( dtype="float32", shape=[10, 8], lod_level=0, name="mul.y") mul_out = block.create_var( diff --git a/python/paddle/fluid/tests/unittests/test_parallel_op.py b/python/paddle/fluid/tests/unittests/test_parallel_op.py index 1a7551c57b26f576ab286e7b18177b9120261623..79bea148f9398152a02d70946cdc5fff1f47ba6b 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_op.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_op.py @@ -15,6 +15,7 @@ import unittest import paddle.fluid as fluid +import paddle.fluid.profiler as profiler import numpy @@ -60,20 +61,23 @@ class BaseParallelForTest(unittest.TestCase): feed=feed, fetch=fetch, place=gpu, - use_parallel=False) + use_parallel=False, + use_gpu=True) result_gpu_parallel = self._run_test_impl_( callback=callback, feed=feed, fetch=fetch, place=gpu, - use_parallel=True) + use_parallel=True, + use_gpu=True) result_gpu_nccl = self._run_test_impl_( callback=callback, feed=feed, fetch=fetch, place=gpu, use_parallel=True, - use_nccl=True) + use_nccl=True, + use_gpu=True) self._assert_same_(fetch, result_cpu, result_cpu_parallel, result_gpu, result_gpu_parallel, result_gpu_nccl) else: @@ -85,7 +89,8 @@ class BaseParallelForTest(unittest.TestCase): fetch, place, use_parallel=False, - use_nccl=False): + use_nccl=False, + use_gpu=False): """ Run a single test, returns the fetch values Args: @@ -132,7 +137,12 @@ class BaseParallelForTest(unittest.TestCase): exe = fluid.Executor(place) exe.run(startup) - return exe.run(main, feed=feed, fetch_list=fetch) + if use_gpu: + profile_type = 'GPU' + else: + profile_type = 'CPU' + with profiler.profiler(profile_type, 'total', '/tmp/profiler'): + return exe.run(main, feed=feed, fetch_list=fetch) def _assert_same_(self, fetch, *args): """ diff --git a/python/paddle/fluid/tests/unittests/test_pool2d_op.py b/python/paddle/fluid/tests/unittests/test_pool2d_op.py index d2107fb4796a588b87d09f3d67e08566c12ffefb..764fa575fba1615de3171e848890b3836e640849 100644 --- a/python/paddle/fluid/tests/unittests/test_pool2d_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool2d_op.py @@ -78,19 +78,22 @@ def avg_pool2D_forward_naive(x, class TestPool2d_Op(OpTest): def setUp(self): + self.op_type = "pool2d" self.use_cudnn = False + self.use_mkldnn = False + self.dtype = np.float32 self.init_test_case() self.init_global_pool() - self.init_op_type() + self.init_kernel_type() self.init_pool_type() self.init_ceil_mode() if self.global_pool: self.paddings = [0 for _ in range(len(self.paddings))] - input = np.random.random(self.shape).astype("float32") + input = np.random.random(self.shape).astype(self.dtype) output = self.pool2D_forward_naive(input, self.ksize, self.strides, self.paddings, self.global_pool, - self.ceil_mode).astype("float32") - self.inputs = {'X': input} + self.ceil_mode).astype(self.dtype) + self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(input)} self.attrs = { 'strides': self.strides, @@ -99,11 +102,12 @@ class TestPool2d_Op(OpTest): 'pooling_type': self.pool_type, 'global_pooling': self.global_pool, 'use_cudnn': self.use_cudnn, + 'use_mkldnn': self.use_mkldnn, 'ceil_mode': self.ceil_mode, 'data_format': 'AnyLayout' # TODO(dzhwinter) : should be fix latter } - self.outputs = {'Out': output.astype('float32')} + self.outputs = {'Out': output} def test_check_output(self): if self.use_cudnn: @@ -113,6 +117,8 @@ class TestPool2d_Op(OpTest): self.check_output() def test_check_grad(self): + if self.dtype == np.float16: + return if self.use_cudnn and self.pool_type != "max": place = core.CUDAPlace(0) self.check_grad_with_place( @@ -126,8 +132,8 @@ class TestPool2d_Op(OpTest): self.strides = [1, 1] self.paddings = [0, 0] - def init_op_type(self): - self.op_type = "pool2d" + def init_kernel_type(self): + pass def init_pool_type(self): self.pool_type = "avg" @@ -147,9 +153,6 @@ class TestCase1(TestPool2d_Op): self.strides = [1, 1] self.paddings = [0, 0] - def init_op_type(self): - self.op_type = "pool2d" - def init_pool_type(self): self.pool_type = "avg" self.pool2D_forward_naive = avg_pool2D_forward_naive @@ -165,9 +168,6 @@ class TestCase2(TestPool2d_Op): self.strides = [1, 1] self.paddings = [1, 1] - def init_op_type(self): - self.op_type = "pool2d" - def init_pool_type(self): self.pool_type = "avg" self.pool2D_forward_naive = avg_pool2D_forward_naive @@ -177,27 +177,18 @@ class TestCase2(TestPool2d_Op): class TestCase3(TestPool2d_Op): - def init_op_type(self): - self.op_type = "pool2d" - def init_pool_type(self): self.pool_type = "max" self.pool2D_forward_naive = max_pool2D_forward_naive class TestCase4(TestCase1): - def init_op_type(self): - self.op_type = "pool2d" - def init_pool_type(self): self.pool_type = "max" self.pool2D_forward_naive = max_pool2D_forward_naive class TestCase5(TestCase2): - def init_op_type(self): - self.op_type = "pool2d" - def init_pool_type(self): self.pool_type = "max" self.pool2D_forward_naive = max_pool2D_forward_naive @@ -205,39 +196,105 @@ class TestCase5(TestCase2): #--------------------test pool2d-------------------- class TestCUDNNCase1(TestPool2d_Op): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool2d" + + +class TestFP16CUDNNCase1(TestPool2d_Op): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase2(TestCase1): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool2d" + + +class TestFP16CUDNNCase2(TestCase1): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase3(TestCase2): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool2d" + + +class TestFP16CUDNNCase3(TestCase2): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase4(TestCase3): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool2d" + + +class TestFP16CUDNNCase4(TestCase3): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase5(TestCase4): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool2d" + + +class TestFP16CUDNNCase5(TestCase4): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCUDNNCase6(TestCase5): - def init_op_type(self): + def init_kernel_type(self): self.use_cudnn = True - self.op_type = "pool2d" + + +class TestFP16CUDNNCase6(TestCase5): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) class TestCeilModeCase1(TestCUDNNCase1): @@ -260,5 +317,36 @@ class TestCeilModeCase4(TestCase2): self.ceil_mode = True +#--------------------test pool2d MKLDNN-------------------- +class TestMKLDNNCase1(TestPool2d_Op): + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNCase2(TestCase1): + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNCase3(TestCase2): + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNCase4(TestCase3): + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNCase5(TestCase4): + def init_kernel_type(self): + self.use_mkldnn = True + + +class TestMKLDNNCase6(TestCase5): + def init_kernel_type(self): + self.use_mkldnn = True + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_profiler.py b/python/paddle/fluid/tests/unittests/test_profiler.py index 1da6b94eea30e65913ce713e0e5e355507534161..cf6fe14a86aa1ab6ea3f60ad15f33d708e9b803a 100644 --- a/python/paddle/fluid/tests/unittests/test_profiler.py +++ b/python/paddle/fluid/tests/unittests/test_profiler.py @@ -31,8 +31,22 @@ class TestProfiler(unittest.TestCase): with fluid.program_guard(main_program, startup_program): image = fluid.layers.data(name='x', shape=[784], dtype='float32') - hidden1 = fluid.layers.fc(input=image, size=128, act='relu') - hidden2 = fluid.layers.fc(input=hidden1, size=64, act='relu') + hidden1 = fluid.layers.fc(input=image, size=64, act='relu') + i = layers.zeros(shape=[1], dtype='int64') + counter = fluid.layers.zeros( + shape=[1], dtype='int64', force_cpu=True) + until = layers.fill_constant([1], dtype='int64', value=10) + data_arr = layers.array_write(hidden1, i) + cond = fluid.layers.less_than(x=counter, y=until) + while_op = fluid.layers.While(cond=cond) + with while_op.block(): + hidden_n = fluid.layers.fc(input=hidden1, size=64, act='relu') + layers.array_write(hidden_n, i, data_arr) + fluid.layers.increment(x=counter, value=1, in_place=True) + layers.less_than(x=counter, y=until, cond=cond) + + hidden_n = layers.array_read(data_arr, i) + hidden2 = fluid.layers.fc(input=hidden_n, size=64, act='relu') predict = fluid.layers.fc(input=hidden2, size=10, act='softmax') label = fluid.layers.data(name='y', shape=[1], dtype='int64') cost = fluid.layers.cross_entropy(input=predict, label=label) diff --git a/python/paddle/fluid/tests/unittests/test_recordio_reader.py b/python/paddle/fluid/tests/unittests/test_recordio_reader.py new file mode 100644 index 0000000000000000000000000000000000000000..24a0074d9b9621d902d12eb8cb29d9b65be22ed3 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_recordio_reader.py @@ -0,0 +1,81 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# 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. + +import unittest + +import paddle.fluid as fluid +import paddle.v2 as paddle +import paddle.v2.dataset.mnist as mnist + + +class TestRecordIO(unittest.TestCase): + def setUp(self): + # Convert mnist to recordio file + with fluid.program_guard(fluid.Program(), fluid.Program()): + reader = paddle.batch(mnist.train(), batch_size=32) + feeder = fluid.DataFeeder( + feed_list=[ # order is image and label + fluid.layers.data( + name='image', shape=[784]), + fluid.layers.data( + name='label', shape=[1], dtype='int64'), + ], + place=fluid.CPUPlace()) + self.num_batches = fluid.recordio_writer.convert_reader_to_recordio_file( + './mnist.recordio', reader, feeder) + + def test_main(self, decorator_callback=None): + # use new program + with fluid.program_guard(fluid.Program(), fluid.Program()): + data_file = fluid.layers.open_recordio_file( + './mnist.recordio', + shapes=[[-1, 784], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64']) + if decorator_callback is not None: + data_file = decorator_callback(data_file) + img, label = fluid.layers.read_file(data_file) + + hidden = fluid.layers.fc(input=img, size=100, act='tanh') + prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + avg_loss = fluid.layers.mean(loss) + + fluid.optimizer.Adam(learning_rate=1e-3).minimize(avg_loss) + + if fluid.core.is_compiled_with_cuda(): + place = fluid.CUDAPlace(0) + else: + place = fluid.CPUPlace() + + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + avg_loss_np = [] + + # train a pass + batch_id = 0 + while not data_file.eof(): + tmp, = exe.run(fetch_list=[avg_loss]) + avg_loss_np.append(tmp) + batch_id += 1 + data_file.reset() + self.assertEqual(batch_id, self.num_batches) + self.assertLess(avg_loss_np[-1], avg_loss_np[0]) + + def test_shuffle_reader(self): + self.test_main(decorator_callback=lambda reader: fluid.layers.create_shuffle_reader(reader, buffer_size=200)) + + def test_double_buffer_reader(self): + self.test_main(decorator_callback=lambda reader: fluid.layers.create_double_buffer_reader(reader, + place='cuda:0' if fluid.core.is_compiled_with_cuda() else 'cpu')) diff --git a/python/paddle/fluid/tests/unittests/test_reduce_op.py b/python/paddle/fluid/tests/unittests/test_reduce_op.py index 5e656bddb7927b6e7935255c120e5e858505125a..9b0cc3534dc551e7fdf7ef8111cad1c172f8bfa4 100644 --- a/python/paddle/fluid/tests/unittests/test_reduce_op.py +++ b/python/paddle/fluid/tests/unittests/test_reduce_op.py @@ -70,6 +70,19 @@ class TestMinOp(OpTest): self.check_output() +class TestProdOp(OpTest): + def setUp(self): + self.op_type = "reduce_prod" + self.inputs = {'X': np.random.random((5, 6, 10)).astype("float64")} + self.outputs = {'Out': self.inputs['X'].prod(axis=0)} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') + + class TestKeepDimReduce(OpTest): def setUp(self): self.op_type = "reduce_sum" diff --git a/python/paddle/fluid/tests/unittests/test_scatter_op.py b/python/paddle/fluid/tests/unittests/test_scatter_op.py index bb02a40d449860cf6c0576662e79a5e36e6e0635..fb1728743630b3ea908ae835444eff7fd71b72c8 100644 --- a/python/paddle/fluid/tests/unittests/test_scatter_op.py +++ b/python/paddle/fluid/tests/unittests/test_scatter_op.py @@ -25,7 +25,7 @@ class TestScatterOp(OpTest): updates_np = np.random.random((2, 3)).astype("float32") output_np = np.copy(ref_np) output_np[index_np] = updates_np - self.inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np} + self.inputs = {'X': ref_np, 'Ids': index_np, 'Updates': updates_np} self.outputs = {'Out': output_np} def test_check_output(self): diff --git a/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py b/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py index 9e5c1e7a3d0bdf514de11e797d7139f577002c52..d6dc99bb3106feee33daa52bffb386f07cc16de5 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py @@ -16,11 +16,15 @@ import unittest import numpy as np from op_test import OpTest from test_softmax_op import stable_softmax +import paddle.fluid.core as core class TestSequenceSoftmaxOp(OpTest): def setUp(self): self.op_type = "sequence_softmax" + self.use_cudnn = False + self.init_op_type() + x = np.random.uniform(0.1, 1, (11, 1)).astype("float32") lod = [[0, 4, 5, 8, 11]] @@ -34,12 +38,31 @@ class TestSequenceSoftmaxOp(OpTest): self.inputs = {"X": (x, lod)} self.outputs = {"Out": out} + self.attrs = {'use_cudnn': self.use_cudnn, } + + def init_op_type(self): + pass def test_check_output(self): - self.check_output() + if self.use_cudnn: + place = core.CUDAPlace(0) + self.check_output_with_place(place, atol=1e-5) + else: + self.check_output() def test_check_grad(self): - self.check_grad(["X"], "Out", max_relative_error=0.01) + if self.use_cudnn: + place = core.CUDAPlace(0) + self.check_grad_with_place( + place, ["X"], "Out", max_relative_error=0.01) + else: + self.check_grad(["X"], "Out", max_relative_error=0.01) + + +# ----------------cudnn Sequencesoftmax---------------- +class TestSequenceSoftmaxCUDNNOp(TestSequenceSoftmaxOp): + def init_op_type(self): + self.use_cudnn = True if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_softmax_op.py b/python/paddle/fluid/tests/unittests/test_softmax_op.py index 8f8312edca7e2d98eb4e881f671c6afdda01c57a..4f20da2b926823db9e7ec92c95178b6d3d1feec9 100644 --- a/python/paddle/fluid/tests/unittests/test_softmax_op.py +++ b/python/paddle/fluid/tests/unittests/test_softmax_op.py @@ -15,6 +15,7 @@ import unittest import numpy as np from op_test import OpTest +import paddle.fluid.core as core def stable_softmax(x): @@ -27,18 +28,37 @@ def stable_softmax(x): class TestSoftmaxOp(OpTest): def setUp(self): self.op_type = "softmax" + self.use_cudnn = False self.inputs = { 'X': np.random.uniform(0.1, 1, [10, 10]).astype("float32") } self.outputs = { 'Out': np.apply_along_axis(stable_softmax, 1, self.inputs['X']) } + self.attrs = {'use_cudnn': self.use_cudnn, } + + def init_op_type(self): + pass def test_check_output(self): - self.check_output() + if self.use_cudnn: + place = core.CUDAPlace(0) + self.check_output_with_place(place, atol=1e-5) + else: + self.check_output() def test_check_grad(self): - self.check_grad(['X'], 'Out') + if self.use_cudnn: + place = core.CUDAPlace(0) + self.check_grad_with_place( + place, ["X"], "Out", max_relative_error=0.01) + else: + self.check_grad(["X"], "Out", max_relative_error=0.01) + + +class TestSoftmaxCUDNNOp(TestSoftmaxOp): + def init_op_type(self): + self.use_cudnn = True if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py b/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py index 889fea2ce66e64d439b51498722e571f48cd1f96..c0d9fc8f22a7c4f791d80a9cad87d003b5d54299 100644 --- a/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py +++ b/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py @@ -26,7 +26,7 @@ class TestSoftmaxWithCrossEntropyOp(OpTest): def setUp(self): self.op_type = "softmax_with_cross_entropy" - batch_size = 2 + batch_size = 41 class_num = 37 logits = np.random.uniform(0.1, 1.0, @@ -59,7 +59,7 @@ class TestSoftmaxWithCrossEntropyOp2(OpTest): def setUp(self): self.op_type = "softmax_with_cross_entropy" - batch_size = 2 + batch_size = 41 class_num = 37 logits = np.random.uniform(0.1, 1.0, diff --git a/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py b/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py index 286d305a777a4683d42a4d3d2d5d5f0c5b6ac12a..61040a39ced6dc57d05a10bf0605c80011db45c3 100644 --- a/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py @@ -60,8 +60,8 @@ class TestSpliteSelectedRows(unittest.TestCase): # expected output selected rows expected_out0_rows = [0, 4] - expected_out1_rows = [5, 7] - expected_out4_rows = [20] + expected_out1_rows = [0, 2] + expected_out4_rows = [0] op = Operator( "split_selected_rows", @@ -101,7 +101,7 @@ class TestSpliteSelectedRows(unittest.TestCase): out0_grad_tensor.set(np_array, place) out1_grad = scope.var("out1@GRAD").get_selected_rows() - rows1 = [7, 5] + rows1 = [2, 0] out1_grad.set_rows(rows1) out1_grad.set_height(height) out1_grad_tensor = out1_grad.get_tensor() diff --git a/python/setup.py.in b/python/setup.py.in index f830039a3af581d593d510326f15139377cb25f1..4cb5409524457b7bc5a99c88a0dbbfc8834923fa 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -62,20 +62,22 @@ write_version_py(filename='@PADDLE_SOURCE_DIR@/python/paddle/version.py') packages=['paddle', - 'paddle.proto', - 'paddle.trainer', - 'paddle.trainer_config_helpers', 'paddle.utils', - 'paddle.v2', - 'paddle.v2.dataset', - 'paddle.v2.reader', - 'paddle.v2.master', - 'paddle.v2.plot', 'paddle.fluid', 'paddle.fluid.proto', 'paddle.fluid.proto.profiler', - 'paddle.fluid.layers', - 'py_paddle'] + 'paddle.fluid.layers'] + +if '${WITH_FLUID}'== 'OFF': + packages+=['paddle.proto', + 'paddle.trainer', + 'paddle.trainer_config_helpers', + 'paddle.v2', + 'paddle.v2.dataset', + 'paddle.v2.reader', + 'paddle.v2.master', + 'paddle.v2.plot', + 'py_paddle'] with open('@PADDLE_SOURCE_DIR@/python/requirements.txt') as f: setup_requires = f.read().splitlines() @@ -84,11 +86,29 @@ if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']: setup_requires+=['opencv-python'] # the prefix is sys.prefix which should always be usr -paddle_bin_dir = 'opt/paddle/bin' -paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer', - '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model', - '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main', - '${PADDLE_BINARY_DIR}/paddle/scripts/paddle'] +paddle_bins = '' +if '${WITH_FLUID}'== 'OFF': + paddle_bin_dir = 'opt/paddle/bin' + paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer', + '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model', + '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main', + '${PADDLE_BINARY_DIR}/paddle/scripts/paddle'] + +package_data={'paddle.fluid': ['core.so']} +if '${WITH_FLUID}'== 'OFF': + package_data['paddle.v2.master']=['libpaddle_master.so'] + package_data['py_paddle']=['*.py','_swig_paddle.so'] + +package_dir={ + '': '${CMAKE_CURRENT_SOURCE_DIR}', + # The paddle.fluid.proto will be generated while compiling. + # So that package points to other directory. + 'paddle.fluid.proto.profiler': '${PADDLE_BINARY_DIR}/paddle/fluid/platform', + 'paddle.fluid.proto': '${PADDLE_BINARY_DIR}/paddle/fluid/framework', +} +if '${WITH_FLUID}'== 'OFF': + package_dir['py_paddle']='${PADDLE_SOURCE_DIR}/paddle/py_paddle' + paddle_rt_lib_dir = 'lib' paddle_rt_libs = ['${WARPCTC_LIBRARIES}'] @@ -101,19 +121,8 @@ setup(name='${PACKAGE_NAME}', install_requires=setup_requires, packages=packages, ext_modules=[Extension('_foo', ['stub.cc'])], - package_data={ - 'paddle.v2.master': ['libpaddle_master.so'], - 'paddle.fluid': ['core.so'], - 'py_paddle':['*.py','_swig_paddle.so'] - }, - package_dir={ - '': '${CMAKE_CURRENT_SOURCE_DIR}', - # The paddle.fluid.proto will be generated while compiling. - # So that package points to other directory. - 'paddle.fluid.proto.profiler': '${PADDLE_BINARY_DIR}/paddle/fluid/platform', - 'paddle.fluid.proto': '${PADDLE_BINARY_DIR}/paddle/fluid/framework', - 'py_paddle': '${PADDLE_SOURCE_DIR}/paddle/py_paddle' - }, + package_data=package_data, + package_dir=package_dir, scripts=paddle_bins, data_files=[(paddle_rt_lib_dir, paddle_rt_libs)] ) diff --git a/tools/manylinux1/Dockerfile.x64 b/tools/manylinux1/Dockerfile.x64 index 93cab692e363cde43bdd4dd9ad20f4a2c06be121..bca0b77ad71a3f65dda15191e5f540bfc2e043d1 100644 --- a/tools/manylinux1/Dockerfile.x64 +++ b/tools/manylinux1/Dockerfile.x64 @@ -13,8 +13,10 @@ ENV PATH /opt/rh/devtoolset-2/root/usr/bin:$PATH ENV LD_LIBRARY_PATH /opt/rh/devtoolset-2/root/usr/lib64:/opt/rh/devtoolset-2/root/usr/lib:/usr/local/lib64:/usr/local/lib:${LD_LIBRARY_PATH} ENV PKG_CONFIG_PATH=/usr/local/lib/pkgconfig +RUN yum install -y sqlite-devel zlib-devel openssl-devel pcre-devel vim tk-devel tkinter libtool xz COPY build_scripts /build_scripts -RUN bash build_scripts/build.sh && rm -r build_scripts +RUN bash build_scripts/build.sh && \ + bash build_scripts/install_nccl2.sh && rm -r build_scripts ENV SSL_CERT_FILE=/opt/_internal/certs.pem @@ -34,9 +36,6 @@ RUN cd /opt && wget -q --no-check-certificate https://github.com/google/protobuf tar xzf protobuf-cpp-3.1.0.tar.gz && \ cd protobuf-3.1.0 && ./configure && make -j4 && make install && cd .. && rm -f protobuf-cpp-3.1.0.tar.gz - -RUN yum install -y sqlite-devel zlib-devel openssl-devel pcre-devel vim tk-devel tkinter libtool - RUN wget -O /root/requirements.txt https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/python/requirements.txt RUN LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH} /opt/python/cp27-cp27mu/bin/pip install -r /root/requirements.txt && \ @@ -47,10 +46,7 @@ RUN LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH} /o RUN LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH} /opt/python/cp27-cp27mu/bin/pip install pre-commit 'ipython==5.3.0' opencv-python && \ LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH} /opt/python/cp27-cp27m/bin/pip install pre-commit 'ipython==5.3.0' opencv-python -RUN wget -O /opt/swig-2.0.12.tar.gz https://sourceforge.net/projects/swig/files/swig/swig-2.0.12/swig-2.0.12.tar.gz/download && \ +RUN wget -O /opt/swig-2.0.12.tar.gz https://cytranet.dl.sourceforge.net/project/swig/swig/swig-2.0.12/swig-2.0.12.tar.gz && \ cd /opt && tar xzf swig-2.0.12.tar.gz && cd /opt/swig-2.0.12 && ./configure && make && make install && cd /opt && rm swig-2.0.12.tar.gz -RUN mkdir -p /src && cd /src && git clone https://github.com/NVIDIA/nccl.git nccl && cd nccl &&\ - make -j `nproc` install && cd .. && rm -rf nccl - CMD ["bash", "/paddle/paddle/scripts/docker/build.sh"] diff --git a/tools/manylinux1/build_scripts/install_nccl2.sh b/tools/manylinux1/build_scripts/install_nccl2.sh new file mode 100644 index 0000000000000000000000000000000000000000..7efc1fe8651a02684c92fa4b924a4f4416503710 --- /dev/null +++ b/tools/manylinux1/build_scripts/install_nccl2.sh @@ -0,0 +1,18 @@ +#!/bin/bash +DEB="nccl-repo-ubuntu1604-2.1.4-ga-cuda8.0_1-1_amd64.deb" +DIR="/nccl2" +mkdir -p $DIR +# we cached the nccl2 deb package in BOS, so we can download it with wget +# install nccl2: http://docs.nvidia.com/deeplearning/sdk/nccl-install-guide/index.html#down +wget -O $DIR/$DEB \ + "http://nccl2-deb.gz.bcebos.com/nccl-repo-ubuntu1604-2.1.4-ga-cuda8.0_1-1_amd64.deb?responseContentDisposition=attachment" + +cd $DIR && ar x $DEB && tar xf data.tar.xz +DEBS=$(find ./var/ -name "*.deb") +for sub_deb in $DEBS; do + echo $sub_deb + ar x $sub_deb && tar xf data.tar.xz +done +mv -f usr/include/nccl.h /usr/local/include/ +mv -f usr/lib/libnccl* /usr/local/lib/ +rm -rf $DIR diff --git a/tools/timeline.py b/tools/timeline.py index d1d1dae2bd85e85b069dc41bdd3e3e54579f9c49..f4083c824e7333a74661d096d4954609f767c83e 100644 --- a/tools/timeline.py +++ b/tools/timeline.py @@ -121,27 +121,34 @@ class Timeline(object): def _allocate_pids(self): for event in self._profile_pb.events: - if event.device_id not in self._devices: - pid = self._allocate_pid() - self._devices[event.device_id] = pid - if event.device_id >= 0: - self._chrome_trace.emit_pid("gpu:%s:stream:%d" % - (pid, event.stream_id), pid) - elif event.device_id == -1: - self._chrome_trace.emit_pid("cpu:thread_hash:%d" % - event.stream_id, pid) + if event.type == profiler_pb2.Event.CPU: + if (event.device_id, "CPU") not in self._devices: + pid = self._allocate_pid() + self._devices[(event.device_id, "CPU")] = pid + self._chrome_trace.emit_pid("cpu:block:%d" % + (event.device_id), pid) + elif event.type == profiler_pb2.Event.GPUKernel: + if (event.device_id, "GPUKernel") not in self._devices: + pid = self._allocate_pid() + self._devices[(event.device_id, "GPUKernel")] = pid + self._chrome_trace.emit_pid("gpu:%d" % (event.device_id), + pid) def _allocate_events(self): for event in self._profile_pb.events: - pid = self._devices[event.device_id] + if event.type == profiler_pb2.Event.CPU: + type = "CPU" + elif event.type == profiler_pb2.Event.GPUKernel: + type = "GPUKernel" + pid = self._devices[(event.device_id, type)] args = {'name': event.name} if event.memcopy.bytes > 0: args = {'mem_bytes': event.memcopy.bytes} # TODO(panyx0718): Chrome tracing only handles ms. However, some # ops takes micro-seconds. Hence, we keep the ns here. - self._chrome_trace.emit_region(event.start_ns, - (event.end_ns - event.start_ns) / - 1.0, pid, 0, 'Op', event.name, args) + self._chrome_trace.emit_region( + event.start_ns, (event.end_ns - event.start_ns) / 1.0, pid, + event.sub_device_id, 'Op', event.name, args) def generate_chrome_trace(self): self._allocate_pids() @@ -159,7 +166,7 @@ if args.timeline_path: with open(profile_path, 'r') as f: profile_s = f.read() profile_pb = profiler_pb2.Profile() - text_format.Merge(profile_s, profile_pb) + profile_pb.ParseFromString(profile_s) tl = Timeline(profile_pb) with open(timeline_path, 'w') as f: