diff --git a/AUTHORS.md b/AUTHORS.md index 2f756c09bc..8bacd8b169 100644 --- a/AUTHORS.md +++ b/AUTHORS.md @@ -30,6 +30,7 @@ | NHZlX | Zhao-Long Xing | | Noplz | Yuan Gao | | pakchoi | Chuan-Jiang Song | +| panyx0718 | Xin Pan | | pengli09 | Peng Li | | pkuyym | Ya-Ming Yang | | QiJune | Jun Qi | diff --git a/benchmark/fluid/machine_translation.py b/benchmark/fluid/machine_translation.py index d7a421c109..adde5f21ac 100644 --- a/benchmark/fluid/machine_translation.py +++ b/benchmark/fluid/machine_translation.py @@ -21,7 +21,7 @@ import argparse import time import distutils.util -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import paddle.fluid.core as core import paddle.fluid.framework as framework diff --git a/benchmark/fluid/mnist.py b/benchmark/fluid/mnist.py index dc10ac2ec1..1e2185dfac 100644 --- a/benchmark/fluid/mnist.py +++ b/benchmark/fluid/mnist.py @@ -20,7 +20,7 @@ import numpy as np import argparse import time -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import paddle.fluid.profiler as profiler diff --git a/benchmark/fluid/resnet.py b/benchmark/fluid/resnet.py index 1af5eaf6b4..831fa2c019 100644 --- a/benchmark/fluid/resnet.py +++ b/benchmark/fluid/resnet.py @@ -23,7 +23,7 @@ import time import cProfile, pstats, StringIO -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import paddle.fluid.core as core import paddle.fluid.profiler as profiler diff --git a/benchmark/fluid/stacked_dynamic_lstm.py b/benchmark/fluid/stacked_dynamic_lstm.py index 5fcbdd64af..73bcc47b4d 100644 --- a/benchmark/fluid/stacked_dynamic_lstm.py +++ b/benchmark/fluid/stacked_dynamic_lstm.py @@ -23,10 +23,10 @@ import random import time import numpy -import paddle.v2 as paddle -import paddle.v2.dataset.imdb as imdb +import paddle +import paddle.dataset.imdb as imdb import paddle.fluid as fluid -from paddle.v2 import batch +import paddle.batch as batch import paddle.fluid.profiler as profiler diff --git a/benchmark/fluid/vgg.py b/benchmark/fluid/vgg.py index 9d990eff62..53e34e0cbd 100644 --- a/benchmark/fluid/vgg.py +++ b/benchmark/fluid/vgg.py @@ -17,7 +17,7 @@ from __future__ import print_function import sys import time import numpy as np -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import paddle.fluid.core as core import argparse diff --git a/doc/fluid/api/data/data_reader.rst b/doc/fluid/api/data/data_reader.rst index d7c896a627..1a35d0bbc8 100644 --- a/doc/fluid/api/data/data_reader.rst +++ b/doc/fluid/api/data/data_reader.rst @@ -56,11 +56,11 @@ DataFeeder Reader ====== -.. automodule:: paddle.v2.reader +.. automodule:: paddle.reader :members: :noindex: -.. automodule:: paddle.v2.reader.creator +.. automodule:: paddle.reader.creator :members: :noindex: diff --git a/doc/fluid/api/layers.rst b/doc/fluid/api/layers.rst index 3790f09c84..ff3c9346a2 100644 --- a/doc/fluid/api/layers.rst +++ b/doc/fluid/api/layers.rst @@ -479,6 +479,13 @@ label_smooth .. autofunction:: paddle.fluid.layers.label_smooth :noindex: +roi_pool +--------- + +.. autofunction:: paddle.fluid.layers.roi_pool + :noindex: + + ops === @@ -820,3 +827,5 @@ topk .. autofunction:: paddle.fluid.layers.topk :noindex: + + diff --git a/doc/fluid/design/data_type/float16.md b/doc/fluid/design/data_type/float16.md index 1ea95ed6b5..844d2aafcf 100644 --- a/doc/fluid/design/data_type/float16.md +++ b/doc/fluid/design/data_type/float16.md @@ -3,7 +3,7 @@ ## Why float16 Half precision (float16) is a binary floating-point format that occupies 16 bits in memory. float16 is half the size of traditional 32-bit single precision format (float) and has lower precision and smaller range. -When high precision computation is not required, using float16 data type could potentially +When high precision computation is not required (which is usually the case at least in the deep learning inference stage), using float16 data type could potentially - reduce storage space, memory bandwidth, and power usages; - increase the chance of data fitting into a smaller cache of lower latency; @@ -12,7 +12,7 @@ When high precision computation is not required, using float16 data type could p ## Survey of current float16 support A brief survey of float16 support on different compilers, hardwares, and libraries can be found below. Interested readers can refer to [link1](https://github.com/PaddlePaddle/Paddle/issues/4853) and [link2](https://github.com/Xreki/Xreki.github.io/blob/master/multi_data_types_in_dl_framework/ppt/float16_and_quantized_type.md) for more info. -The goal of float16 is to serve as a key for the executor to find and run the correct version of compute method specialized for float16 in operator kernel. It should be compatible with various natively supported float16 implementations including `__half` for cuda, `float16_t` for ARM, and `Eigen::half` for Eigen to make writing customized float16 kernels easier. +The goal of float16 is to serve as a key for the executor to find and run the correct version of compute method specialized for float16 in operator kernels. It should be compatible with various natively supported float16 implementations including `__half` for cuda, `float16_t` for ARM, and `Eigen::half` for Eigen to make writing customized float16 kernels easier. ### Compiler - nvcc supports `__half` data type after CUDA 7.5. @@ -95,11 +95,89 @@ float half_to_float(float16 h); ``` which provides one-to-one conversion between float32 and float16. These twos functions will do different conversion routines based on the current hardware. CUDA/ARM instrinsics will be used when the corresonding hardware is available. If the hardware or compiler level does not support float32 to float16 conversion, software emulation will be performed to do the conversion. -## To do -After float16 class is available, some of the future items are below: +## float16 inference +In Fluid, a neural network is represented as a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), whose Python wrapper is a [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program). The basic structure of a program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program desc by executing the sequence of operators in the entrance block of the program one by one. -- Update pybind/tensor_py.h to bind c++ float16 with numpy float16. +### Operator level requirement +Each operator has many kernels for different data types, devices, and library types. The operator will select the appropriate kernel to run based on, among other things, the data type of the input variables. By default, every Fluid operator has a float data type kernel that takes float variables as input and generates float output. -- Modify `GetKernelType()` method in `framework/operator.h` to make it compatible with float16. +This means that if we provide float input to the first operator in a program, then each opeartor will use float kernel to compute float output and send it as input to the next operator to trigger the float kernel. Overall, the program will run in float mode and give us a final output of float data type. -- Create a type-casting operator that can convert the data type in tensor between float16 and other types. +The same principle applies if we want a program to run in float16 mode. We provide input variable of float16 data type to the first operator, and then one by one, each operator in the program will run the float16 kernel (provided that each operator in this program has float16 kernels registered) until we finally obtain a float16 output variable. + +So the preliminary requirement for float16 inference is to add float16 kernel to operators that are needed in a specific kind of program. For example, float16 inference on an image classification neural network like Vgg or Resnet, typically requires the following operators to have float16 kernels: convolution, pooling, multiplication, addition, batch norm, dropout, relu, and softmax. Please refer to [new_op_en](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/dev/new_op_en.md) for details of how to add new kernels to an operator. + +### Variable level requirement +Operators including convolution and multiplication (used in fully-connected layers) takes as input not only the variables generated by the preceding operators but also [parameter](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#parameter) variables, which contains the trained weights to apply to the input data. These weights are obtained in the Fluid training process and are by default of float data type. + +When these operators are running in float16 mode, the float16 kernel requires those parameter variables to contain weights of Fluid float16 data type. Thus, we need a convenient way to convert the original float weights to float16 weights. + +In Fluid, we use tensor to hold actual data for a variable on the c++ end. [Pybind](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/pybind/tensor_py.h) is used to bind c++ tensors of certain data type with numpy array of the correponding numpy data type on the Python end. Each common c++ built-in data type has a corresponding numpy data type of the same name. However, since there is no built-in float16 type in c++, we cannot directly bind numpy float16 data type with the Fluid float16 class. Since both Fluid float16 and numpy float16 use uint16 as the internal data storage type, we use c++ built-in type `uint16_t` and the corresponding numpy uint16 data type to bridge the gap via [Pybind](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/pybind/tensor_py.h). + +The following code demonstrates how to do the tensor conversion. +```Python +# var is the variable of float weights +# tensor is a numpy array of data copied from the tensor data in var +# fp16_var is the variable that will contain float16 weights converted from var +tensor = numpy.array(var.get_tensor()) +fp16_tensor = fp16_var.get_tensor() + +# After the original tensor data is converted to numpy float16 data type, +# view(numpy.uint16) is used so that the internal memory of the numpy array +# will be reinterpreted to be of uint16 data type, which is binded to +# Fluid float16 class via pybind with the help of uint16_t built-in c++ type +fp16_tensor.set(tensor.astype(numpy.float16).view(numpy.uint16), GPUPlace) +``` + +### Consistent API requirement +The basic inference in float16 mode requires users to feed input and obtain output both of float16 data type. However, in this way, the inference APIs are not consistent between float16 mode and float mode, and users may find it confusing and diffcult to use float16 inference since they need to do extra steps to provide float16 input data and convert float16 output data back to float. To have consistent API for different inference modes, we need to transpile the program desc in some way so that we can run float16 inference by feeding and fetching variables of float data type. + +This problem can be solved by introducing a type-casting operator which takes an input variable of certain data type, cast it to another specified data type, and put the casted data into the output variable. Insert cast operator where needed can make a program internally run in float16 mode. + +### float16 transpiler +Put all the above requirements in mind, we designed a float16 inference transpiler that can tranpile a float32 mode inference program desc to a float16 mode one. + +Given a float inference program and the corresponding variables of float32 weights in the [scope](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/scope.md), +this transpiler mainly does the following modifications: + +1. Insert cast operators at the beginning of the program so that the input float data will be converted to float16 data type before feeding to subsequent operators to invoke the float16 kernel. + +2. Insert cast operators at the end of the program so that the output float16 data will be converted back to float data type before users obtain the result. + +3. For each parameter variable of float weights, create in the scope a corresponding variable of float16 weights which are converted from the corresponding float weights and add this new float16 variable to the program. + +4. Update the operator information in the program so that each relevant operator use the newly created float16 variable instead of its float counterpart. + +Below is an example of usage: +```Python +# Get the float inference program +[float_inference_program, feed_target_names, + fetch_targets] = fluid.io.load_inference_model(save_dirname, exe) + +# Prepare the float input data +tensor_img = numpy.random.rand(1, 3, 32, 32).astype(numpy.float32) + +# Running inference_program in float mode +float_results = exe.run(float_inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) + +# Use float16 transpiler to speedup +float16_inference_program = float_inference_program.clone() +t = fluid.InferenceTranspiler() +t.float16_transpile(float16_inference_program, GPUPlace) + +# Running +float16_results = exe.run(float16_inference_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) +``` + +As we can see from the example above, users can simply use the `float16_transpile` method provided by the infernece transpiler class on an existing float inference program to run inference in float16 mode. + +### Speedup on GPU +Currently, Fluid inference in float16 mode is only supported on Nvidia GPU device. There is no motivation to support float16 inference on non-ARM CPUs because float16 is not natively supported there and float16 calculation will only be slower than its float counterpart. + +Nvidia started to support its native float16 data type (which has the same internal memory representation as Fluid float16 class) on CUDA 7.5. Moreover, float16 speedups on common computational intensive tasks including GEMM (general matrix-matrix multiplication) and convolution are supported since cublas 7.5 and cuDNN 5.0. + +Recently, the introduction of [tensor core](https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/) in volta architecture GPUs and the support of tensor core calculation in CUDA 9.0 and cuDNN 7.0 make float16 truly superior to float in certain deep learning applications. Please refer to this [benchmark report](https://github.com/kexinzhao/Paddle_benchmark/blob/master/float16_benchmark.md) for more details. diff --git a/doc/fluid/design/onnx/images/project_structure.png b/doc/fluid/design/onnx/images/project_structure.png new file mode 100644 index 0000000000..ab1c2ff23c Binary files /dev/null and b/doc/fluid/design/onnx/images/project_structure.png differ diff --git a/doc/fluid/design/onnx/onnx_convertor.md b/doc/fluid/design/onnx/onnx_convertor.md new file mode 100644 index 0000000000..bc1665d7c3 --- /dev/null +++ b/doc/fluid/design/onnx/onnx_convertor.md @@ -0,0 +1,131 @@ +# Background + +[ONNX (Open Neural Network Exchange)](https://github.com/onnx/onnx) bridges different deep learning frameworks by providing an open source graph format for models. The models trained in other frameworks can be converted into the ONNX format to execute inference by utilizing the built-in operators in ONNX - this is called a **frontend**. With the inverse conversion (called a **backend**), different frameworks can share any models supported by ONNX in principle. Now most mainstream frameworks have joined the ONNX community, e.g. Caffe2, PyTorch, and MXNet etc. And there is a momentum driving more and more vendors to begin supporting ONNX or even choose ONNX as the only machine learning runtime in their devices. + +Therefore, it is necessary to enable the conversion between PaddlePaddle and ONNX. This design doc is aimed at implementing a convertor, mainly for converting between **Fluid** models and ONNX (it is very likely that we may support older v2 models in the future). A complete convertor should be bidirectional - with a frontend AND a backend, but considering the importance, the we will start with the frontend i.e. Fluid models to ONNX models. + + +# How it works + +ONNX has a [working list of operators](https://github.com/onnx/onnx/blob/master/docs/Operators.md) which is versioned. + +When prioritizing implementation of a frontend over a backend, choice of coverage of Fluid -> ONNX operators comes down to choices of models to be supported (see section `Supported models`). Eventually, this will allow us to reach a really-wide coverage of all operators. + +Here are a few major considerations when it comes to converting models: + +- **Op-level conversion**: How to map the inputs, attributes, and outputs of each Paddle operator to those of the ONNX operator. In several cases, these require transformations. For each direction (frontend vs. backend), a different conversion mapping is needed. +- **Parameters (weights) initialization**: Setting initial parameters on different nodes. +- **Tensor data type mapping** (Note: Some ONNX data types are not supported in Fluid) +- **Network representation adaption**: Fluid `ProgramDesc` include nested blocks. Since ONNX is free of nesting, the `ProgramDesc` ops need to be traversed to only include ops from the global scope in the root block. The variables used as inputs and outputs should also be in this scope. +- **Model validation**: There are two kinds of validations that are necessary: + 1. We need to ensure that the inference outputs of the ops in run inside a model are the same as those when running the ONNX converted ops through an alternative ONNX backend. + 2. Checking to see if the generated nodes on the graph are validated by the internal ONNX checkers. +- **Versioning**: ONNX versions its op listing over versions. In fact, it has versioning on 3 different levels: ops, graphs, and ONNX models. This requires that we are conscious about versioning the convertor and updating tests and op convertor logic for each release. It also implies that we release pre-trained ONNX models upon each version release. + +One thing that makes this conversion more feasible in Fluid's case is the use of a static IR - the `ProgramDesc` - as opposed to a dynamic graph, as created in the cases of frameworks like PyTorch. + + +# Project structure + +

+ +

+ +The project contains four important parts: + +* **fluid**: The directory that contains wrappers for fluid related APIs. Fluid has provided some low-level APIs to parse or generate the inference model. However, directly using these low-level APIs makes the code tediously long. This module wraps low-level APIs to provide simplified interfaces. + +* **onnx**: This is a Python package provided by ONNX containing helpers for creating nodes, graphs, and eventually binary protobuf models with initializer parameters. + +* **onnx_fluid**: Contains two-way mapping (Fluid -> ONNX ops and ONNX -> Fluid ops). Called from `convert.py`, the program uses this mapping along with modifier functions to construct ONNX nodes with the help of ONNX's `make_node` helper. It also contains mapping between datatypes and tensor deprecation / amplification logic. + +* **convert.py**: The interface exposed to users. This will traverse the global program blocks/variables and construct the write-able model. + + +# Usage +The converter should be designed to very easy-to-use. Bidirectional conversion between a Fluid inference model and an ONNX binary model will be supported. Model validation will also provided to verify the correctness of converted model. + +* Convert Fluid inference model to ONNX binary model + + ``` + python convert.py --fluid_model --onnx_model validate True + ``` + +* Validate the converted model + + ``` + python validate.py --fluid_model --onnx_model + ``` + +The conversion and model validation will be completed consecutively, finally output a readable model structure description. And for the converse conversion, users only need to exchange the input and output. + + +# Challenges and mitigation + +## Cycles + +Cycles are unsupported in ONNX. In Paddle, the `while` op is the most prominent example of a cycle. + +*Resolution*: We won't support models with `while`s which can't be substituted until ONNX adds support for such ops. + +## Sequences + +Sequence processing operators like `sequence_expand`, `sequence_reshape`, `sequence_concat`, and `sequence_pool` are not supported by ONNX as well, because they do not support non-padded datatypes like LoDTensors. + +*Resolution*: Since the runtimes using our ONNX exported graphs won't be using LoDTensors in the first place, such sequence operators should be mapped to ONNX ops that will do the necessary transposing ops with the knowledge of the padding and shape of the Tensors. + +## Ops that can't easily be mapped + +There are ops that just aren't possible to map today: + +**Control flow operators** + +Paddle supports control flow ops like `If/Else` and `Switch` (if we ignore the CSP operations like `select` for now). ONNX has `If` support in the experimental phase. + +*Resolution*: Map Paddle's `If/Else` to ONNX's `If`, but ignore other control flow operators until ONNX brings support for them. + + +**Non-existent in Fluid** + +There are several ONNX operators that are not available in Fluid today, e.g. `InstanceNormalization`, `RandomUniform`, `Unsqueeze`, etc. + +*Resolution*: For the initial phase, we can choose to not support ops that our models don't care for and are subsequently not available in Fluid. However, for ops that we think might be necessary for Fluid users also, we must implement them on our side and support the ONNX conversion to them. This list is TBD. + + +**Concurrency** + +ONNX does not have any considerations for concurrency right now. + +*Resolution*: There are two ways to approach this: + +a. We choose to not support concurrent models. +b. We only support `go_op`s (basically threads) shallowly. This could mean that we enqueue `go_op` ops prior to gradient calculations OR even prior to the entire graph, and that's it - since `go_op`s do not have support for backprop anyways. One of the core target use cases of `go_op`: batch reading - can be handled through this approach. + + +**Overloaded in Fluid** + +There are ops in ONNX whose job can't be accomplished by a single corresponding Paddle operator (e.g. ), but a collection of operators. + +*Resolution*: Chain multiple Paddle operators. + + +## Lack of LoDTensors + +As stated above, ONNX only supports simple Tensor values. + +*Resolution*: Deprecate to plain old numpy-able tensors. + + +## Reconstruction from deprecated ONNX ops + +For higher-level Fluid ops, such as a few offered by the `nn` layer that do not have direct corresponding mappings but can be converted to ONNX by chaining a series of ops without cycles, it would be useful to map them back to the higher-level Fluid ops once converted back from the deprecated ONNX graphs. + +*Resolution*: Graphs that have the deprecation from Paddle -> ONNX. When converting back from ONNX, if we encounter the identical graphs by doing a forward search, we can replace the subgraphs with the matching ONNX op. + + +# Supported models + +As mentioned above, potential risks may come from the conversion of sequence-related models, including the LodTensor, ```if/else``` and ```while``` operator. So a good choice is to focus on some important feedforward models first, then implement some simple recurrent models. + +- Feedforward models: common models selected in PaddleBook, e.g. VGG, ResNet and some other models proposed by application teams. +- Recurrent models: language model, stacked LSTMs etc. diff --git a/doc/v2/api/data/data_reader.rst b/doc/v2/api/data/data_reader.rst index d7c896a627..1a35d0bbc8 100644 --- a/doc/v2/api/data/data_reader.rst +++ b/doc/v2/api/data/data_reader.rst @@ -56,11 +56,11 @@ DataFeeder Reader ====== -.. automodule:: paddle.v2.reader +.. automodule:: paddle.reader :members: :noindex: -.. automodule:: paddle.v2.reader.creator +.. automodule:: paddle.reader.creator :members: :noindex: diff --git a/doc/v2/api/data/dataset.rst b/doc/v2/api/data/dataset.rst index 02e41564b1..e7c8be4452 100644 --- a/doc/v2/api/data/dataset.rst +++ b/doc/v2/api/data/dataset.rst @@ -1,82 +1,82 @@ Dataset ======= -.. automodule:: paddle.v2.dataset +.. automodule:: paddle.dataset :members: :noindex: mnist +++++ -.. automodule:: paddle.v2.dataset.mnist +.. automodule:: paddle.dataset.mnist :members: :noindex: cifar +++++ -.. automodule:: paddle.v2.dataset.cifar +.. automodule:: paddle.dataset.cifar :members: :noindex: conll05 +++++++ -.. automodule:: paddle.v2.dataset.conll05 +.. automodule:: paddle.dataset.conll05 :members: get_dict,get_embedding,test :noindex: imdb ++++ -.. automodule:: paddle.v2.dataset.imdb +.. automodule:: paddle.dataset.imdb :members: :noindex: imikolov ++++++++ -.. automodule:: paddle.v2.dataset.imikolov +.. automodule:: paddle.dataset.imikolov :members: :noindex: movielens +++++++++ -.. automodule:: paddle.v2.dataset.movielens +.. automodule:: paddle.dataset.movielens :members: :noindex: -.. autoclass:: paddle.v2.dataset.movielens.MovieInfo +.. autoclass:: paddle.dataset.movielens.MovieInfo :noindex: - -.. autoclass:: paddle.v2.dataset.movielens.UserInfo + +.. autoclass:: paddle.dataset.movielens.UserInfo :noindex: sentiment +++++++++ -.. automodule:: paddle.v2.dataset.sentiment +.. automodule:: paddle.dataset.sentiment :members: :noindex: uci_housing +++++++++++ -.. automodule:: paddle.v2.dataset.uci_housing +.. automodule:: paddle.dataset.uci_housing :members: :noindex: wmt14 +++++ -.. automodule:: paddle.v2.dataset.wmt14 +.. automodule:: paddle.dataset.wmt14 :members: :noindex: wmt16 +++++ -.. automodule:: paddle.v2.dataset.wmt16 +.. automodule:: paddle.dataset.wmt16 :members: :noindex: diff --git a/doc/v2/howto/cluster/multi_cluster/index_en.rst b/doc/v2/howto/cluster/multi_cluster/index_en.rst index dac7aaef08..b69bd5b2db 100644 --- a/doc/v2/howto/cluster/multi_cluster/index_en.rst +++ b/doc/v2/howto/cluster/multi_cluster/index_en.rst @@ -1,19 +1,35 @@ Use different clusters ====================== -PaddlePaddle supports running jobs on several platforms including: -- `Kubernetes `_ open-source system for automating deployment, scaling, and management of containerized applications from Google. -- `OpenMPI `_ Mature high performance parallel computing framework. -- `Fabric `_ A cluster management tool. Write scripts to submit jobs or manage the cluster. +The user's cluster environment is not the same. To facilitate everyone's deployment, we provide a variety of cluster deployment methods to facilitate the submission of cluster training tasks, which will be introduced as follows: -We'll introduce cluster job management on these platforms. The examples can be found under `cluster_train_v2 `_ . +`Kubernetes `_ is a scheduling framework of Google open source container cluster, supporting a complete cluster solution for large-scale cluster production environment. The following guidelines show PaddlePaddle's support for Kubernetes: -These cluster platforms provide API or environment variables for training processes, when the job is dispatched to different nodes. Like node ID, IP or total number of nodes etc. +.. toctree:: + :maxdepth: 1 + + k8s_cn.md + k8s_distributed_cn.md + +`OpenMPI `_ is a mature high-performance parallel computing framework, which is widely used in the field of HPC. The following guide describes how to use OpenMPI to build PaddlePaddle's cluster training task: .. toctree:: :maxdepth: 1 - fabric_en.md - openmpi_en.md - k8s_en.md - k8s_aws_en.md + openmpi_cn.md + +`Fabric `_ is a convenient tool for program deployment and management. We provide a way to deploy and manage with Fabric. If you want to know more about it, please read the following guidelines: + +.. toctree:: + :maxdepth: 1 + + fabric_cn.md + +We also support the deployment of PaddlePaddle on AWS. Learn more about: + +.. toctree:: + :maxdepth: 1 + + k8s_aws_cn.md + +The examples can be found under `cluster_train_v2 `_ . \ No newline at end of file diff --git a/paddle/fluid/framework/blocking_queue.h b/paddle/fluid/framework/blocking_queue.h new file mode 100644 index 0000000000..a19558c0ae --- /dev/null +++ b/paddle/fluid/framework/blocking_queue.h @@ -0,0 +1,74 @@ +/* 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 // NOLINT +#include +#include // NOLINT +#include + +namespace paddle { +namespace framework { + +template +class BlockingQueue { + public: + void Push(const T &item) { + { + std::lock_guard g(mutex_); + q_.emplace_back(item); + } + cv_.notify_one(); + } + + template + void Extend(const U &items) { + { + std::lock_guard g(mutex_); + for (auto &item : items) { + q_.emplace_back(item); + } + } + cv_.notify_all(); + } + + std::deque PopAll(size_t ms, bool *timeout) { + auto time = + std::chrono::system_clock::now() + std::chrono::milliseconds(ms); + std::unique_lock lock(mutex_); + *timeout = !cv_.wait_until(lock, time, [this] { return !q_.empty(); }); + std::deque ret; + if (!*timeout) { + std::swap(ret, q_); + } + return ret; + } + + T Pop() { + std::unique_lock lock(mutex_); + cv_.wait(lock, [=] { return !q_.empty(); }); + T rc(std::move(q_.front())); + q_.pop_front(); + return rc; + } + + private: + std::mutex mutex_; + std::condition_variable cv_; + std::deque q_; +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/data_transform.cc b/paddle/fluid/framework/data_transform.cc index bfad9ac1e9..9c277a27da 100644 --- a/paddle/fluid/framework/data_transform.cc +++ b/paddle/fluid/framework/data_transform.cc @@ -63,16 +63,16 @@ void DataTransform(const OpKernelType& expected_kernel_type, } void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor, - Variable& out_var) { + Variable* out_var) { if (in_var.IsType()) { auto& in_lod_tensor = in_var.Get(); - auto* tran_lod_tensor = out_var.GetMutable(); + auto* tran_lod_tensor = out_var->GetMutable(); tran_lod_tensor->set_lod(in_lod_tensor.lod()); tran_lod_tensor->set_layout(in_lod_tensor.layout()); tran_lod_tensor->ShareDataWith(tensor); } else if (in_var.IsType()) { auto& in_selected_rows = in_var.Get(); - auto* trans_selected_rows = out_var.GetMutable(); + auto* trans_selected_rows = out_var->GetMutable(); trans_selected_rows->set_height(in_selected_rows.height()); trans_selected_rows->set_rows(in_selected_rows.rows()); trans_selected_rows->mutable_value()->ShareDataWith(tensor); diff --git a/paddle/fluid/framework/data_transform.h b/paddle/fluid/framework/data_transform.h index 9ec67e6f3d..dee5d8c7c1 100644 --- a/paddle/fluid/framework/data_transform.h +++ b/paddle/fluid/framework/data_transform.h @@ -35,7 +35,7 @@ void DataTransform(const OpKernelType& expected_kernel_type, const Tensor& input_tensor, Tensor* out); void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor, - Variable& out_var); + Variable* out_var); } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/broadcast_op_handle_test.cc b/paddle/fluid/framework/details/broadcast_op_handle_test.cc index 3f2dcde3e9..8f1b6d1615 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle_test.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle_test.cc @@ -139,7 +139,7 @@ struct TestBroadcastOpHandle { PADDLE_ENFORCE_EQ(out_tensor.lod(), lod, "lod is not equal."); f::Tensor result_tensor; - f::TensorCopy(out_tensor, cpu_place, *(ctxs_[j]), &result_tensor); + f::TensorCopySync(out_tensor, cpu_place, &result_tensor); float* ct = result_tensor.mutable_data(cpu_place); for (int64_t i = 0; i < f::product(kDims); ++i) { @@ -185,7 +185,7 @@ struct TestBroadcastOpHandle { } f::Tensor result_tensor; - f::TensorCopy(rt, cpu_place, *(ctxs_[j]), &result_tensor); + f::TensorCopySync(rt, cpu_place, &result_tensor); float* ct = result_tensor.data(); for (int64_t i = 0; i < f::product(kDims); ++i) { diff --git a/paddle/fluid/framework/details/fetch_op_handle.cc b/paddle/fluid/framework/details/fetch_op_handle.cc index 423449abff..1e8ca20b51 100644 --- a/paddle/fluid/framework/details/fetch_op_handle.cc +++ b/paddle/fluid/framework/details/fetch_op_handle.cc @@ -66,8 +66,7 @@ void FetchOpHandle::RunImpl() { auto &t = var->Get(); if (platform::is_gpu_place(t.place())) { #ifdef PADDLE_WITH_CUDA - TensorCopy(t, cpu, *dev_ctxes_[t.place()], &tensors_[i], true); - dev_ctxes_.at(t.place())->Wait(); + TensorCopySync(t, cpu, &tensors_[i]); #endif } else { tensors_[i].ShareDataWith(t); diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc index 3413467b14..c2eb1c31b4 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc @@ -58,23 +58,20 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result, const OpDesc &op, - const platform::Place &p, - const size_t &i) const { + size_t place_id) const { + auto p = places_[place_id]; auto *op_handle = result->ops_.back().get(); op_handle->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p)); - auto var_names = op.InputArgumentNames(); - - for (auto &each_var_name : var_names) { - VarHandle *var = CreateOrGetLatestVarHandle(result, each_var_name, p, i); + for (auto &each_var_name : op.InputArgumentNames()) { + VarHandle *var = + CreateOrGetLatestVarHandle(result, each_var_name, p, place_id); op_handle->AddInput(var); } - var_names = op.OutputArgumentNames(); - - for (auto &each_var_name : var_names) { - CreateOpOutput(result, op_handle, each_var_name, p, i); + for (auto &each_var_name : op.OutputArgumentNames()) { + CreateOpOutput(result, op_handle, each_var_name, p, place_id); } } @@ -84,17 +81,18 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp(const OpDesc &op, return false; } - auto checker = [&](const std::vector opvars, - const std::vector sendvars) -> bool { - bool is_dist_train_op = false; + /** + * Check any of opvars contains `.block` and in sendvars + */ + auto checker = [](const std::vector &opvars, + const std::vector &sendvars) -> bool { for (auto &var : opvars) { if (var.find(".block") != std::string::npos && std::find(sendvars.begin(), sendvars.end(), var) != sendvars.end()) { - is_dist_train_op = true; - break; + return true; } } - return is_dist_train_op; + return false; }; if (op.Type() == "split") { @@ -117,13 +115,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( places_.size()); // Find "send" op first for split is in front of send. - OpDesc *send_op = nullptr; - for (auto *op : program.Block(0).AllOps()) { - if (op->Type() == "send") { - send_op = op; - break; - } - } + OpDesc *send_op = GetSendOpDesc(program); bool is_forwarding = true; for (auto *op : program.Block(0).AllOps()) { @@ -134,6 +126,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( } else if (IsDistTrainOp(*op, send_op)) { CreateComputationalOps(&result, *op, 1); } else if (IsScaleLossOp(*op)) { + // user can customize loss@grad if skip_scale_loss_ if (!skip_scale_loss_) { CreateScaleLossGradOp(&result); } @@ -142,10 +135,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( CreateComputationalOps(&result, *op, places_.size()); if (!is_forwarding) { // Currently, we assume that once gradient is generated, it can be - // broadcast, and each gradient is only broadcast once. But there are no - // other cases, for example, we need to adjust the gradient according to - // the input when we get the gradient, which is not considered at - // present. + // broadcast, and each gradient is only broadcast once. for (auto &og : op->OutputArgumentNames()) { if (IsParameterGradientOnce(og, &og_has_been_broadcast)) { InsertNCCLAllReduceOp(&result, og); @@ -175,6 +165,16 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( return std::unique_ptr(graph); } +OpDesc *MultiDevSSAGraphBuilder::GetSendOpDesc( + const ProgramDesc &program) const { + for (auto *op : program.Block(0).AllOps()) { + if (op->Type() == "send") { + return op; + } + } + return nullptr; +} + void MultiDevSSAGraphBuilder::InsertNCCLAllReduceOp( SSAGraph *result, const std::string &og) const { #ifdef PADDLE_WITH_CUDA @@ -243,7 +243,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(SSAGraph *result, auto p = places_[scope_idx]; auto s = local_scopes_[scope_idx]; result->ops_.emplace_back(new ComputationOpHandle(op, s, p)); - CreateOpHandleIOs(result, op, p, scope_idx); + CreateOpHandleIOs(result, op, scope_idx); } } @@ -255,7 +255,7 @@ void MultiDevSSAGraphBuilder::CreateSendOp(SSAGraph *result, result->ops_.emplace_back(new SendOpHandle(op, s, p)); // Create inputs for output on original place and no ssa output // is created for send op. - CreateOpHandleIOs(result, op, p, 0); + CreateOpHandleIOs(result, op, 0); } bool MultiDevSSAGraphBuilder::IsScaleLossOp(const OpDesc &op) const { diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.h b/paddle/fluid/framework/details/multi_devices_graph_builder.h index dc3da70eda..fa4d31bdc4 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.h +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.h @@ -48,7 +48,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { private: void CreateOpHandleIOs(SSAGraph *result, const OpDesc &op, - const platform::Place &p, const size_t &i) const; + size_t place_id) const; private: std::string loss_var_name_; @@ -65,6 +65,9 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { void CreateSendOp(SSAGraph *result, const OpDesc &op) const; + /** + * Is this operator as the end-point operator before/after send operator. + */ bool IsDistTrainOp(const OpDesc &op, OpDesc *send_op) const; void CreateComputationalOps(SSAGraph *result, const OpDesc &op, @@ -77,6 +80,12 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { std::unordered_set *og_has_been_broadcast) const; void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const; + + /** + * Get send op in the global block of program. + * nullptr if not found. + */ + OpDesc *GetSendOpDesc(const ProgramDesc &program) const; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/reduce_op_handle_test.cc b/paddle/fluid/framework/details/reduce_op_handle_test.cc index c17aabee53..ffdd7c14eb 100644 --- a/paddle/fluid/framework/details/reduce_op_handle_test.cc +++ b/paddle/fluid/framework/details/reduce_op_handle_test.cc @@ -194,7 +194,7 @@ struct TestReduceOpHandle { } f::Tensor result_tensor; - f::TensorCopy(rt, cpu_place, *(ctxs_[output_scope_idx]), &result_tensor); + f::TensorCopySync(rt, cpu_place, &result_tensor); float *ct = result_tensor.data(); for (int64_t j = 0; j < f::product(result_tensor.dims()); ++j) { @@ -239,7 +239,7 @@ struct TestReduceOpHandle { auto &rt = out_var->Get(); f::Tensor result_tensor; - f::TensorCopy(rt, cpu_place, *(ctxs_[output_scope_idx]), &result_tensor); + f::TensorCopySync(rt, cpu_place, &result_tensor); float *ct = result_tensor.data(); for (int64_t j = 0; j < f::product(result_tensor.dims()); ++j) { diff --git a/paddle/fluid/framework/details/ssa_graph.h b/paddle/fluid/framework/details/ssa_graph.h index 72684e7f97..e996a00c16 100644 --- a/paddle/fluid/framework/details/ssa_graph.h +++ b/paddle/fluid/framework/details/ssa_graph.h @@ -25,12 +25,22 @@ namespace paddle { namespace framework { namespace details { +// A SSA graph used by parallel executor. struct SSAGraph { + // all variable in each devices. + // The outside vector is the device vector. Each element of this vector is a + // map from variable name to variables. The variables, who have the same name, + // will have a different version. The offset in the + // `std::vector>` is the version of varaibles. std::vector< std::unordered_map>>> vars_; + // aux variables to represent dependency. Useful to resolve data hazard. std::unordered_set> dep_vars_; + + // all operators. NOTE that even we use a vector here, the operators is + // unordered. std::vector> ops_; }; diff --git a/paddle/fluid/framework/details/ssa_graph_builder.h b/paddle/fluid/framework/details/ssa_graph_builder.h index be1f0460e4..64e5d93081 100644 --- a/paddle/fluid/framework/details/ssa_graph_builder.h +++ b/paddle/fluid/framework/details/ssa_graph_builder.h @@ -48,6 +48,8 @@ class SSAGraphBuilder { const platform::Place &place, size_t place_offset); + // Add an output variable (each_var_name, place, place_offset) to op_handle, + // which belongs to graph static void CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle, const std::string &each_var_name, const platform::Place &place, size_t place_offset); diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index d70bbd4ef0..d089b79d91 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -22,6 +22,7 @@ #include #include "ThreadPool.h" // ThreadPool in thrird party +#include "paddle/fluid/framework/blocking_queue.h" #include "paddle/fluid/framework/details/ssa_graph_executor.h" namespace paddle { @@ -30,46 +31,6 @@ class Scope; namespace details { -template -class BlockingQueue { - public: - void Push(const T &item) { - { - std::lock_guard g(mutex_); - q_.emplace_back(item); - } - cv_.notify_one(); - } - - template - void Extend(const U &items) { - { - std::lock_guard g(mutex_); - for (auto &item : items) { - q_.emplace_back(item); - } - } - cv_.notify_all(); - } - - std::deque PopAll(size_t ms, bool *timeout) { - auto time = - std::chrono::system_clock::now() + std::chrono::milliseconds(ms); - std::unique_lock lock(mutex_); - *timeout = !cv_.wait_until(lock, time, [this] { return !q_.empty(); }); - std::deque ret; - if (!*timeout) { - std::swap(ret, q_); - } - return ret; - } - - private: - std::mutex mutex_; - std::condition_variable cv_; - std::deque q_; -}; - class ThreadedSSAGraphExecutor : public SSAGraphExecutor { public: ThreadedSSAGraphExecutor(size_t num_threads, bool use_event, diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 513e720fd0..766bf0ab0c 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -226,15 +226,15 @@ static bool has_fetch_operators( } void Executor::Run(const ProgramDesc& program, Scope* scope, - std::map& feed_targets, - std::map& fetch_targets, + std::map* feed_targets, + std::map* fetch_targets, bool create_vars, const std::string& feed_holder_name, const std::string& fetch_holder_name) { platform::RecordBlock b(kProgramId); bool has_feed_ops = - has_feed_operators(program.Block(0), feed_targets, feed_holder_name); + 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); + has_fetch_operators(program.Block(0), *fetch_targets, fetch_holder_name); ProgramDesc* copy_program = const_cast(&program); if (!has_feed_ops || !has_fetch_ops) { @@ -250,7 +250,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, feed_holder->SetPersistable(true); int i = 0; - for (auto& feed_target : feed_targets) { + for (auto& feed_target : (*feed_targets)) { std::string var_name = feed_target.first; VLOG(3) << "feed target's name: " << var_name; @@ -273,7 +273,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, fetch_holder->SetPersistable(true); int i = 0; - for (auto& fetch_target : fetch_targets) { + for (auto& fetch_target : (*fetch_targets)) { std::string var_name = fetch_target.first; VLOG(3) << "fetch target's name: " << var_name; @@ -361,16 +361,16 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, void Executor::RunPreparedContext( ExecutorPrepareContext* ctx, Scope* scope, - std::map& feed_targets, - std::map& fetch_targets, bool create_vars, + std::map* feed_targets, + std::map* fetch_targets, bool create_vars, const std::string& feed_holder_name, const std::string& fetch_holder_name) { auto& global_block = ctx->prog_.Block(ctx->block_id_); PADDLE_ENFORCE( - has_feed_operators(global_block, feed_targets, feed_holder_name), + has_feed_operators(global_block, *feed_targets, feed_holder_name), "Program in ExecutorPrepareContext should has feed_ops."); PADDLE_ENFORCE( - has_fetch_operators(global_block, fetch_targets, fetch_holder_name), + has_fetch_operators(global_block, *fetch_targets, fetch_holder_name), "Program in the prepared context should has fetch_ops."); // map the data of feed_targets to feed_holder @@ -378,8 +378,8 @@ void Executor::RunPreparedContext( if (op->Type() == kFeedOpType) { std::string feed_target_name = op->Output("Out")[0]; int idx = boost::get(op->GetAttr("col")); - SetFeedVariable(scope, *feed_targets[feed_target_name], feed_holder_name, - idx); + SetFeedVariable(scope, *(*feed_targets)[feed_target_name], + feed_holder_name, idx); } } @@ -390,7 +390,7 @@ void Executor::RunPreparedContext( if (op->Type() == kFetchOpType) { std::string fetch_target_name = op->Input("X")[0]; int idx = boost::get(op->GetAttr("col")); - *fetch_targets[fetch_target_name] = + *(*fetch_targets)[fetch_target_name] = GetFetchVariable(*scope, fetch_holder_name, idx); } } diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 43defdacf2..4a3d637e2d 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -55,8 +55,8 @@ class Executor { bool create_local_scope = true, bool create_vars = true); void Run(const ProgramDesc& program, Scope* scope, - std::map& feed_targets, - std::map& fetch_targets, + std::map* feed_targets, + std::map* fetch_targets, bool create_vars = true, const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); @@ -74,8 +74,8 @@ class Executor { bool create_vars = true); void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, - std::map& feed_targets, - std::map& fetch_targets, + std::map* feed_targets, + std::map* fetch_targets, bool create_vars = true, const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); diff --git a/paddle/fluid/framework/init.cc b/paddle/fluid/framework/init.cc index b30f276b4b..85beae775b 100644 --- a/paddle/fluid/framework/init.cc +++ b/paddle/fluid/framework/init.cc @@ -15,7 +15,6 @@ limitations under the License. */ #include #include #include -#include #include "paddle/fluid/framework/init.h" #include "paddle/fluid/framework/operator.h" @@ -31,6 +30,7 @@ std::once_flag p2p_init_flag; void InitGflags(std::vector argv) { std::call_once(gflags_init_flag, [&]() { + argv.insert(argv.begin(), "dummy"); int argc = argv.size(); char **arr = new char *[argv.size()]; std::string line; @@ -44,20 +44,23 @@ void InitGflags(std::vector argv) { }); } -void InitP2P(int count) { +void InitP2P(std::vector devices) { #ifdef PADDLE_WITH_CUDA std::call_once(p2p_init_flag, [&]() { + int count = devices.size(); for (int i = 0; i < count; ++i) { for (int j = 0; j < count; ++j) { - if (i == j) continue; + if (devices[i] == devices[j]) continue; int can_acess = -1; - PADDLE_ENFORCE(cudaDeviceCanAccessPeer(&can_acess, i, j), - "Failed to test P2P access."); + PADDLE_ENFORCE( + cudaDeviceCanAccessPeer(&can_acess, devices[i], devices[j]), + "Failed to test P2P access."); if (can_acess != 1) { - LOG(WARNING) << "Cannot enable P2P access from " << i << " to " << j; + LOG(WARNING) << "Cannot enable P2P access from " << devices[i] + << " to " << devices[j]; } else { - cudaSetDevice(i); - cudaDeviceEnablePeerAccess(j, 0); + cudaSetDevice(devices[i]); + cudaDeviceEnablePeerAccess(devices[j], 0); } } } @@ -67,11 +70,26 @@ void InitP2P(int count) { void InitDevices(bool init_p2p) { /*Init all available devices by default */ + std::vector devices; +#ifdef PADDLE_WITH_CUDA + try { + int count = platform::GetCUDADeviceCount(); + for (int i = 0; i < count; ++i) { + devices.push_back(i); + } + } catch (const std::exception &exp) { + LOG(WARNING) << "Compiled with WITH_GPU, but no GPU found in runtime."; + } +#else + LOG(WARNING) + << "'CUDA' is not supported, Please re-compile with WITH_GPU option"; +#endif + InitDevices(init_p2p, devices); +} +void InitDevices(bool init_p2p, const std::vector devices) { std::vector places; - places.emplace_back(platform::CPUPlace()); int count = 0; - #ifdef PADDLE_WITH_CUDA try { count = platform::GetCUDADeviceCount(); @@ -83,12 +101,17 @@ void InitDevices(bool init_p2p) { << "'CUDA' is not supported, Please re-compile with WITH_GPU option"; #endif - for (int i = 0; i < count; ++i) { - places.emplace_back(platform::CUDAPlace(i)); + for (size_t i = 0; i < devices.size(); ++i) { + if (devices[i] >= count || devices[i] < 0) { + LOG(WARNING) << "Invalid devices id."; + continue; + } + places.emplace_back(platform::CUDAPlace(devices[i])); } if (init_p2p) { - InitP2P(count); + InitP2P(devices); } + places.emplace_back(platform::CPUPlace()); platform::DeviceContextPool::Init(places); } diff --git a/paddle/fluid/framework/init.h b/paddle/fluid/framework/init.h index 1155ca3604..0e30594672 100644 --- a/paddle/fluid/framework/init.h +++ b/paddle/fluid/framework/init.h @@ -28,5 +28,7 @@ void InitGLOG(const std::string &prog_name); void InitDevices(bool init_p2p); +void InitDevices(bool init_p2p, const std::vector devices); + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index 46c834b38b..076c457130 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -205,8 +205,8 @@ void OpDesc::SetAttr(const std::string &name, const Attribute &v) { need_update_ = true; } -void OpDesc::SetBlockAttr(const std::string &name, BlockDesc &block) { - this->attrs_[name] = █ +void OpDesc::SetBlockAttr(const std::string &name, BlockDesc *block) { + this->attrs_[name] = block; need_update_ = true; } diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index cd6777e60a..3ee36a47c1 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include #include #include "paddle/fluid/framework/attribute.h" @@ -73,7 +74,7 @@ class OpDesc { void SetAttr(const std::string &name, const Attribute &v); - void SetBlockAttr(const std::string &name, BlockDesc &block); + void SetBlockAttr(const std::string &name, BlockDesc *block); Attribute GetAttr(const std::string &name) const; diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index f97bd08274..32576423a6 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -171,17 +171,6 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const { return ss.str(); } -void OperatorBase::Rename(const std::string& old_name, - const std::string& new_name) { - for (auto& input : inputs_) { - std::replace(input.second.begin(), input.second.end(), old_name, new_name); - } - for (auto& output : outputs_) { - std::replace(output.second.begin(), output.second.end(), old_name, - new_name); - } -} - OperatorBase::OperatorBase(const std::string& type, const VariableNameMap& inputs, const VariableNameMap& outputs, @@ -327,7 +316,6 @@ bool OpSupportGPU(const std::string& op_type) { auto it = all_kernels.find(op_type); if (it == all_kernels.end()) { // All control operator must support GPU - return true; } for (auto& kern_pair : it->second) { @@ -554,7 +542,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope, std::shared_ptr out(new Tensor); DataTransform(expected_kernel_key, kernel_type_for_var, *tensor_in, out.get()); - CopyVariableWithTensor(*var, *(out.get()), *trans_var); + CopyVariableWithTensor(*var, *(out.get()), trans_var); } } } diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index b7a7c69b4c..826cc57b72 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -79,31 +79,28 @@ class OperatorBase { virtual ~OperatorBase() {} - template - inline const T& Attr(const std::string& name) const { - PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap", - name); - return boost::get(attrs_.at(name)); - } - - /// if scope is not null, also show dimensions of arguments - virtual std::string DebugStringEx(const Scope* scope) const; - - std::string DebugString() const { return DebugStringEx(nullptr); } - - /// Net will call this interface function to Run an op. + /// Executor will call this interface function to Run an op. // The implementation should be written at RunImpl void Run(const Scope& scope, const platform::Place& place); // FIXME(typhoonzero): this is only used for recv_op to stop event_loop. virtual void Stop() {} - virtual bool IsNetOp() const { return false; } + /// if scope is not null, also show dimensions of arguments + virtual std::string DebugStringEx(const Scope* scope) const; + std::string DebugString() const { return DebugStringEx(nullptr); } virtual bool SupportGPU() const { return false; } - /// rename inputs outputs name - void Rename(const std::string& old_name, const std::string& new_name); + const std::string& Type() const { return type_; } + + template + inline const T& Attr(const std::string& name) const { + PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap", + name); + return boost::get(attrs_.at(name)); + } + const AttributeMap& Attrs() const { return attrs_; } const VariableNameMap& Inputs() const { return inputs_; } const VariableNameMap& Outputs() const { return outputs_; } @@ -112,7 +109,7 @@ class OperatorBase { std::string Input(const std::string& name) const; //! Get a input which has multiple variables. const std::vector& Inputs(const std::string& name) const; - + //! Get all inputs variable names std::vector InputVars() const; //! Get a output with argument's name described in `op_proto` @@ -120,13 +117,9 @@ class OperatorBase { //! Get an output which has multiple variables. //! TODO add a vector_view to prevent memory copy. const std::vector& Outputs(const std::string& name) const; - + //! Get all outputs variable names virtual std::vector OutputVars(bool has_intermediate) const; - const std::string& Type() const { return type_; } - void SetType(const std::string& type) { type_ = type; } - const AttributeMap& Attrs() const { return attrs_; } - // Return a new operator instance, which is as same as this. // Use unique_ptr to prevent caller forget to delete this pointer. virtual std::unique_ptr Clone() const = 0; @@ -278,20 +271,6 @@ class ExecutionContext { return res; } - void ShareLoD(const std::string& in, const std::string& out, size_t i = 0, - size_t j = 0) const { - PADDLE_ENFORCE_LT(i, InputSize(in)); - PADDLE_ENFORCE_LT(j, OutputSize(out)); - auto* in_var = MultiInputVar(in)[i]; - auto* out_var = MultiOutputVar(out)[j]; - if (!in_var->IsType()) return; - PADDLE_ENFORCE(out_var->IsType(), - "The %d-th output of Output(%s) must be LoDTensor.", j, out); - auto in_tensor = in_var->Get(); - auto* out_tensor = out_var->GetMutable(); - out_tensor->set_lod(in_tensor.lod()); - } - platform::Place GetPlace() const { return device_context_.GetPlace(); } template diff --git a/paddle/fluid/framework/program_desc.cc b/paddle/fluid/framework/program_desc.cc index 16694bcf76..64fb028f83 100644 --- a/paddle/fluid/framework/program_desc.cc +++ b/paddle/fluid/framework/program_desc.cc @@ -56,7 +56,7 @@ ProgramDesc::ProgramDesc(const ProgramDesc &o) { for (const auto &attr : op->Proto()->attrs()) { if (attr.type() == proto::AttrType::BLOCK) { size_t blk_idx = attr.block_idx(); - op->SetBlockAttr(attr.name(), *this->MutableBlock(blk_idx)); + op->SetBlockAttr(attr.name(), this->MutableBlock(blk_idx)); } } } @@ -73,7 +73,7 @@ ProgramDesc::ProgramDesc(const proto::ProgramDesc &desc) { for (const auto &attr : op->Proto()->attrs()) { if (attr.type() == proto::AttrType::BLOCK) { size_t blk_idx = attr.block_idx(); - op->SetBlockAttr(attr.name(), *this->MutableBlock(blk_idx)); + op->SetBlockAttr(attr.name(), this->MutableBlock(blk_idx)); } } } diff --git a/paddle/fluid/framework/prune.cc b/paddle/fluid/framework/prune.cc index 107c5bf8ec..57c1b822d8 100644 --- a/paddle/fluid/framework/prune.cc +++ b/paddle/fluid/framework/prune.cc @@ -14,19 +14,19 @@ limitations under the License. */ #include "paddle/fluid/framework/prune.h" +#include + #include #include #include #include #include -#include - namespace paddle { namespace framework { -const std::string kFeedOpType = "feed"; -const std::string kFetchOpType = "fetch"; +const char kFeedOpType[] = "feed"; +const char kFetchOpType[] = "fetch"; bool HasDependentVar(const proto::OpDesc& op_desc, const std::set& dependent_vars) { @@ -68,7 +68,7 @@ bool HasSubBlock(const proto::OpDesc& op_desc) { // the child block to help pruning void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output, int block_id, int parent_block_id, - std::set& dependent_vars) { + std::set* dependent_vars) { auto& block = input.blocks(block_id); auto& ops = block.ops(); @@ -90,11 +90,11 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output, std::vector should_run; for (auto op_iter = ops.rbegin(); op_iter != ops.rend(); ++op_iter) { auto& op_desc = *op_iter; - if (IsTarget(op_desc) || HasDependentVar(op_desc, dependent_vars)) { + if (IsTarget(op_desc) || HasDependentVar(op_desc, *dependent_vars)) { // insert its input to the dependency graph for (auto& var : op_desc.inputs()) { for (auto& argu : var.arguments()) { - dependent_vars.insert(argu); + dependent_vars->insert(argu); } } should_run.push_back(true); @@ -138,7 +138,7 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output, // GetSubBlockIndex(*op) is the idx of the sub_block in the input desc // output_block_id is the idx of the current block in the output desc prune_impl(input, output, GetSubBlockIndex(*op), output_block_id, - sub_block_dependent_vars); + &sub_block_dependent_vars); } } } @@ -181,7 +181,7 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output, void Prune(const proto::ProgramDesc& input, proto::ProgramDesc* output) { std::set dependent_vars; output->clear_blocks(); - prune_impl(input, output, 0, -1, dependent_vars); + prune_impl(input, output, 0, -1, &dependent_vars); } void inference_optimize_impl(proto::ProgramDesc* input, int block_id) { diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc index d2e60ab1dd..e5bc74755f 100644 --- a/paddle/fluid/framework/tensor_util.cc +++ b/paddle/fluid/framework/tensor_util.cc @@ -20,7 +20,7 @@ namespace paddle { namespace framework { void TensorCopy(const Tensor& src, const platform::Place& dst_place, - const platform::DeviceContext& ctx, Tensor* dst, bool sync) { + const platform::DeviceContext& ctx, Tensor* dst) { VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to " << dst_place; src.check_memory_size(); @@ -48,9 +48,7 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto ctx_gpu_place = boost::get(ctx_place); PADDLE_ENFORCE_EQ(src_gpu_place, ctx_gpu_place); auto stream = - sync ? nullptr - : reinterpret_cast(ctx) - .stream(); + reinterpret_cast(ctx).stream(); memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream); } else if (platform::is_cpu_place(src_place) && platform::is_gpu_place(dst_place)) { @@ -61,9 +59,7 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto ctx_gpu_place = boost::get(ctx_place); PADDLE_ENFORCE_EQ(dst_gpu_place, ctx_gpu_place); auto stream = - sync ? nullptr - : reinterpret_cast(ctx) - .stream(); + reinterpret_cast(ctx).stream(); memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, stream); } else if (platform::is_gpu_place(src_place) && platform::is_gpu_place(dst_place)) { @@ -72,9 +68,7 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto ctx_place = ctx.GetPlace(); PADDLE_ENFORCE(platform::is_gpu_place(ctx_place)); auto stream = - sync ? nullptr - : reinterpret_cast(ctx) - .stream(); + reinterpret_cast(ctx).stream(); memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream); } #endif @@ -92,6 +86,41 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, TensorCopy(src, dst_place, *dev_ctx, dst); } +void TensorCopySync(const Tensor& src, const platform::Place& dst_place, + Tensor* dst) { + VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place() + << " to " << dst_place; + src.check_memory_size(); + dst->Resize(src.dims()); + dst->set_layout(src.layout()); + auto src_place = src.place(); + auto src_ptr = src.data(); + auto dst_ptr = dst->mutable_data(dst_place, src.type()); + auto size = src.numel() * SizeOfType(src.type()); + if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) { + memory::Copy(boost::get(dst_place), dst_ptr, + boost::get(src_place), src_ptr, size); + } +#ifdef PADDLE_WITH_CUDA + else if (platform::is_gpu_place(src_place) && // NOLINT + platform::is_cpu_place(dst_place)) { + auto src_gpu_place = boost::get(src_place); + auto dst_cpu_place = boost::get(dst_place); + memory::Copy(dst_cpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr); + } else if (platform::is_cpu_place(src_place) && + platform::is_gpu_place(dst_place)) { + auto src_cpu_place = boost::get(src_place); + auto dst_gpu_place = boost::get(dst_place); + memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr); + } else if (platform::is_gpu_place(src_place) && + platform::is_gpu_place(dst_place)) { + auto src_gpu_place = boost::get(src_place); + auto dst_gpu_place = boost::get(dst_place); + memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr); + } +#endif +} + template struct AnyDTypeVisitor { Predicate predicate_; diff --git a/paddle/fluid/framework/tensor_util.h b/paddle/fluid/framework/tensor_util.h index 3af68402dc..dca279b693 100644 --- a/paddle/fluid/framework/tensor_util.h +++ b/paddle/fluid/framework/tensor_util.h @@ -24,10 +24,11 @@ namespace paddle { namespace framework { void TensorCopy(const Tensor& src, const platform::Place& dst_place, - const platform::DeviceContext& ctx, Tensor* dst, - bool sync = false); + const platform::DeviceContext& ctx, Tensor* dst); void TensorCopy(const Tensor& src, const platform::Place& dst_place, Tensor* dst); +void TensorCopySync(const Tensor& src, const platform::Place& dst_place, + Tensor* dst); template void TensorFromVector(const std::vector& src, diff --git a/paddle/fluid/inference/engine.h b/paddle/fluid/inference/engine.h new file mode 100644 index 0000000000..6b0ac92fa9 --- /dev/null +++ b/paddle/fluid/inference/engine.h @@ -0,0 +1,52 @@ +/* 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/framework/framework.pb.h" + +namespace paddle { +namespace inference { + +/* + * EngineBase is the base class of all inference engines. An inference engine + * takes a paddle program as input, and outputs the result in fluid Tensor + * format. It can be used to optimize performance of computation sub-blocks, for + * example, break down the original block into sub-blocks and execute each + * sub-blocks in different engines. + * + * For example: + * When inference, the resnet50 model can put most of the model into subgraph + * and run it on a TensorRT engine. + * + * There are several engines such as TensorRT and other frameworks, so an + * EngineBase is put forward to give an unified interface for all the + * different engine implemention. + */ +class EngineBase { + public: + using DescType = ::paddle::framework::proto::BlockDesc; + + // Build the model and do some preparation, for example, in TensorRT, run + // createInferBuilder, buildCudaEngine. + virtual void Build(const DescType& paddle_model) = 0; + + // Execute the engine, that will run the inference network. + virtual void Execute(int batch_size) = 0; + + virtual ~EngineBase() {} +}; // class EngineBase + +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/io.cc b/paddle/fluid/inference/io.cc index 78d2f16746..65db7c7b50 100644 --- a/paddle/fluid/inference/io.cc +++ b/paddle/fluid/inference/io.cc @@ -16,17 +16,29 @@ limitations under the License. */ #include #include +#include #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/feed_fetch_type.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/pybind/pybind.h" +DEFINE_string(devices, "", "The devices to be used which is joined by comma."); +DEFINE_bool(init_p2p, false, "Whether to init p2p."); + namespace paddle { namespace inference { -// Temporarily add this function for exposing framework::InitDevices() when -// linking the inference shared library. -void Init(bool init_p2p) { framework::InitDevices(init_p2p); } +void Init(const std::vector argv) { + framework::InitGflags(argv); + // init devices + std::vector devices; + std::string token; + std::istringstream tokenStream(FLAGS_devices); + while (std::getline(tokenStream, token, ',')) { + devices.push_back(std::stoi(token)); + } + framework::InitDevices(FLAGS_init_p2p, devices); +} void ReadBinaryFile(const std::string& filename, std::string* contents) { std::ifstream fin(filename, std::ios::in | std::ios::binary); diff --git a/paddle/fluid/inference/io.h b/paddle/fluid/inference/io.h index ba3e45099a..caf599b1a6 100644 --- a/paddle/fluid/inference/io.h +++ b/paddle/fluid/inference/io.h @@ -25,7 +25,7 @@ limitations under the License. */ namespace paddle { namespace inference { -void Init(bool init_p2p); +void Init(const std::vector argv); void LoadPersistables(framework::Executor* executor, framework::Scope* scope, const framework::ProgramDesc& main_program, diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt index e39c0daac7..4b5866ad5d 100644 --- a/paddle/fluid/inference/tensorrt/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt @@ -1 +1,4 @@ -nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) +if(WITH_TESTING) + nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) + nv_test(test_tensorrt_engine SRCS test_engine.cc engine.cc DEPS dynload_cuda) +endif() diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc new file mode 100644 index 0000000000..03a25f8e8b --- /dev/null +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -0,0 +1,135 @@ +/* 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/inference/tensorrt/engine.h" + +#include +#include +#include +#include +#include "paddle/fluid/inference/tensorrt/helper.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +void TensorRTEngine::Build(const DescType& paddle_model) { + PADDLE_ENFORCE(false, "not implemented"); +} + +void TensorRTEngine::Execute(int batch_size) { + infer_context_->enqueue(batch_size, buffers_.data(), *stream_, nullptr); + cudaStreamSynchronize(*stream_); +} + +TensorRTEngine::~TensorRTEngine() { + // clean buffer + for (auto& buffer : buffers_) { + if (buffer != nullptr) { + PADDLE_ENFORCE_EQ(0, cudaFree(buffer)); + buffer = nullptr; + } + } +} + +void TensorRTEngine::FreezeNetwork() { + PADDLE_ENFORCE(infer_builder_ != nullptr, + "Call InitNetwork first to initialize network."); + PADDLE_ENFORCE(infer_network_ != nullptr, + "Call InitNetwork first to initialize network."); + // build engine. + infer_builder_->setMaxBatchSize(max_batch_); + infer_builder_->setMaxWorkspaceSize(max_workspace_); + + infer_engine_.reset(infer_builder_->buildCudaEngine(*infer_network_)); + PADDLE_ENFORCE(infer_engine_ != nullptr, "build cuda engine failed!"); + + infer_context_.reset(infer_engine_->createExecutionContext()); + + // allocate GPU buffers. + buffers_.resize(buffer_sizes_.size(), nullptr); + for (auto& item : buffer_sizes_) { + if (item.second == 0) { + auto slot_offset = infer_engine_->getBindingIndex(item.first.c_str()); + item.second = kDataTypeSize[static_cast( + infer_engine_->getBindingDataType(slot_offset))] * + AccumDims(infer_engine_->getBindingDimensions(slot_offset)); + } + PADDLE_ENFORCE_EQ(0, cudaMalloc(&buffer(item.first), item.second)); + } +} + +nvinfer1::ITensor* TensorRTEngine::DeclareInput(const std::string& name, + nvinfer1::DataType dtype, + const nvinfer1::Dims& dim) { + PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate input name %s", + name); + + PADDLE_ENFORCE(infer_network_ != nullptr, "should initnetwork first"); + auto* input = infer_network_->addInput(name.c_str(), dtype, dim); + PADDLE_ENFORCE(input, "infer network add input %s failed", name); + + buffer_sizes_[name] = kDataTypeSize[static_cast(dtype)] * AccumDims(dim); + return input; +} + +void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer* layer, int offset, + const std::string& name) { + PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate output name %s", + name); + + auto* output = layer->getOutput(offset); + PADDLE_ENFORCE(output != nullptr); + output->setName(name.c_str()); + infer_network_->markOutput(*output); + // output buffers' size can only be decided latter, set zero here to mark this + // and will reset latter. + buffer_sizes_[name] = 0; +} + +void* TensorRTEngine::GetOutputInGPU(const std::string& name) { + return buffer(name); +} + +void TensorRTEngine::GetOutputInCPU(const std::string& name, void* dst, + size_t max_size) { + // determine data size + auto it = buffer_sizes_.find(name); + PADDLE_ENFORCE(it != buffer_sizes_.end()); + PADDLE_ENFORCE_GT(it->second, 0); + PADDLE_ENFORCE_GE(max_size, it->second); + + PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buffer(name), it->second, + cudaMemcpyDeviceToHost, *stream_)); +} + +void*& TensorRTEngine::buffer(const std::string& name) { + PADDLE_ENFORCE(infer_engine_ != nullptr, "call FreezeNetwork first."); + auto it = buffer_sizes_.find(name); + PADDLE_ENFORCE(it != buffer_sizes_.end()); + auto slot_offset = infer_engine_->getBindingIndex(name.c_str()); + return buffers_[slot_offset]; +} + +void TensorRTEngine::SetInputFromCPU(const std::string& name, void* data, + size_t size) { + void* buf = buffer(name); + PADDLE_ENFORCE_EQ( + 0, cudaMemcpyAsync(buf, data, size, cudaMemcpyHostToDevice, *stream_)); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h new file mode 100644 index 0000000000..82d8c3df4e --- /dev/null +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -0,0 +1,146 @@ +/* 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 +#include +#include +#include "paddle/fluid/inference/engine.h" +#include "paddle/fluid/inference/tensorrt/helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * TensorRT Engine. + * + * There are two alternative ways to use it, one is to build from a paddle + * protobuf model, another way is to manully construct the network. + */ +class TensorRTEngine : public EngineBase { + public: + // Weight is model parameter. + class Weight { + public: + Weight(nvinfer1::DataType dtype, void* value, int num_elem) { + w_.type = dtype; + w_.values = value; + w_.count = num_elem; + } + const nvinfer1::Weights& get() { return w_; } + + private: + nvinfer1::Weights w_; + }; + + TensorRTEngine(int max_batch, int max_workspace, cudaStream_t* stream, + nvinfer1::ILogger& logger = NaiveLogger::Global()) + : max_batch_(max_batch), + max_workspace_(max_workspace), + stream_(stream), + logger_(logger) {} + + virtual ~TensorRTEngine(); + + // TODO(Superjomn) implement it later when graph segmentation is supported. + void Build(const DescType& paddle_model) override; + + void Execute(int batch_size) override; + + // Initialize the inference network, so that TensorRT layers can add to this + // network. + void InitNetwork() { + infer_builder_.reset(createInferBuilder(logger_)); + infer_network_.reset(infer_builder_->createNetwork()); + } + // After finishing adding ops, freeze this network and creates the executation + // environment. + void FreezeNetwork(); + + // Add an input and set its name, data type and dimention. + nvinfer1::ITensor* DeclareInput(const std::string& name, + nvinfer1::DataType dtype, + const nvinfer1::Dims& dim); + // Set the offset-th output from a layer as the network's output, and set its + // name. + void DeclareOutput(const nvinfer1::ILayer* layer, int offset, + const std::string& name); + + // GPU memory address for an ITensor with specific name. One can operate on + // these memory directly for acceleration, for example, output the converted + // data directly to the buffer to save data copy overhead. + // NOTE this should be used after calling `FreezeNetwork`. + void*& buffer(const std::string& name); + + // Fill an input from CPU memory with name and size. + void SetInputFromCPU(const std::string& name, void* data, size_t size); + // TODO(Superjomn) is this method necessary given that buffer(xxx) can be + // accessed directly. Fill an input from GPU memory with name and size. + void SetInputFromGPU(const std::string& name, void* data, size_t size); + // Get an output called name, the output of tensorrt is in GPU, so this method + // will just return the output's GPU memory address. + void* GetOutputInGPU(const std::string& name); + // LOW EFFICENCY! Get output to CPU, this will trigger a memory copy from GPU + // to CPU. + void GetOutputInCPU(const std::string& name, void* dst, size_t max_size); + + nvinfer1::ICudaEngine* engine() { return infer_engine_.get(); } + nvinfer1::INetworkDefinition* network() { return infer_network_.get(); } + + private: + // the max batch size + int max_batch_; + // the max memory size the engine uses + int max_workspace_; + cudaStream_t* stream_; + nvinfer1::ILogger& logger_; + + std::vector buffers_; + // max data size for the buffers. + std::unordered_map buffer_sizes_; + + // TensorRT related internal members + template + struct Destroyer { + void operator()(T* x) { x->destroy(); } + }; + template + using infer_ptr = std::unique_ptr>; + infer_ptr infer_builder_; + infer_ptr infer_network_; + infer_ptr infer_engine_; + infer_ptr infer_context_; +}; // class TensorRTEngine + +// Add an layer__ into engine__ with args ARGS. +// For example: +// TRT_ENGINE_ADD_LAYER(xxx, FullyConnected, input, dim, weights, bias) +// +// Reference +// https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#charRNN_define_network +// +// will add a fully connected layer into the engine. +// TensorRT has too many layers, so that is not wise to add member functions for +// them, and an macro like this is more extensible when underlying TensorRT +// library add new layer supports. +#define TRT_ENGINE_ADD_LAYER(engine__, layer__, ARGS...) \ + engine__->network()->add##layer__(ARGS); + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/helper.h b/paddle/fluid/inference/tensorrt/helper.h new file mode 100644 index 0000000000..796283d325 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/helper.h @@ -0,0 +1,88 @@ +/* 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 +#include "paddle/fluid/platform/dynload/tensorrt.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +namespace dy = paddle::platform::dynload; + +static size_t AccumDims(nvinfer1::Dims dims) { + size_t num = dims.nbDims == 0 ? 0 : 1; + for (int i = 0; i < dims.nbDims; i++) { + PADDLE_ENFORCE_GT(dims.d[i], 0); + num *= dims.d[i]; + } + return num; +} + +// TensorRT data type to size +const int kDataTypeSize[] = { + 4, // kFLOAT + 2, // kHALF + 1, // kINT8 + 4 // kINT32 +}; + +// The following two API are implemented in TensorRT's header file, cannot load +// from the dynamic library. So create our own implementation and directly +// trigger the method from the dynamic library. +static nvinfer1::IBuilder* createInferBuilder(nvinfer1::ILogger& logger) { + return static_cast( + dy::createInferBuilder_INTERNAL(&logger, NV_TENSORRT_VERSION)); +} +static nvinfer1::IRuntime* createInferRuntime(nvinfer1::ILogger& logger) { + return static_cast( + dy::createInferRuntime_INTERNAL(&logger, NV_TENSORRT_VERSION)); +} + +// A logger for create TensorRT infer builder. +class NaiveLogger : public nvinfer1::ILogger { + public: + void log(nvinfer1::ILogger::Severity severity, const char* msg) override { + switch (severity) { + case Severity::kINFO: + LOG(INFO) << msg; + break; + case Severity::kWARNING: + LOG(WARNING) << msg; + break; + case Severity::kINTERNAL_ERROR: + case Severity::kERROR: + LOG(ERROR) << msg; + break; + default: + break; + } + } + + static nvinfer1::ILogger& Global() { + static nvinfer1::ILogger* x = new NaiveLogger; + return *x; + } + + virtual ~NaiveLogger() override {} +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/test_engine.cc b/paddle/fluid/inference/tensorrt/test_engine.cc new file mode 100644 index 0000000000..c6e1c71cdc --- /dev/null +++ b/paddle/fluid/inference/tensorrt/test_engine.cc @@ -0,0 +1,83 @@ +/* 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 +#include +#include + +#include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +class TensorRTEngineTest : public ::testing::Test { + protected: + void SetUp() override { + ASSERT_EQ(0, cudaStreamCreate(&stream_)); + engine_ = new TensorRTEngine(1, 1 << 10, &stream_); + engine_->InitNetwork(); + } + + void TearDown() override { + delete engine_; + cudaStreamDestroy(stream_); + } + + protected: + TensorRTEngine* engine_; + cudaStream_t stream_; +}; + +TEST_F(TensorRTEngineTest, add_layer) { + const int size = 1; + + float raw_weight[size] = {2.}; // Weight in CPU memory. + float raw_bias[size] = {3.}; + + LOG(INFO) << "create weights"; + TensorRTEngine::Weight weight(nvinfer1::DataType::kFLOAT, raw_weight, size); + TensorRTEngine::Weight bias(nvinfer1::DataType::kFLOAT, raw_bias, size); + auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT, + nvinfer1::DimsCHW{1, 1, 1}); + auto* fc_layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, *x, size, + weight.get(), bias.get()); + PADDLE_ENFORCE(fc_layer != nullptr); + + engine_->DeclareOutput(fc_layer, 0, "y"); + LOG(INFO) << "freeze network"; + engine_->FreezeNetwork(); + ASSERT_EQ(engine_->engine()->getNbBindings(), 2); + + // fill in real data + float x_v = 1234; + engine_->SetInputFromCPU("x", reinterpret_cast(&x_v), + 1 * sizeof(float)); + LOG(INFO) << "to execute"; + engine_->Execute(1); + + LOG(INFO) << "to get output"; + // void* y_v = + float y_cpu; + engine_->GetOutputInCPU("y", &y_cpu, sizeof(float)); + + LOG(INFO) << "to checkout output"; + ASSERT_EQ(y_cpu, x_v * 2 + 3); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/test_tensorrt.cc b/paddle/fluid/inference/tensorrt/test_tensorrt.cc index a81a708e7a..aed5b5e1a2 100644 --- a/paddle/fluid/inference/tensorrt/test_tensorrt.cc +++ b/paddle/fluid/inference/tensorrt/test_tensorrt.cc @@ -1,16 +1,16 @@ /* 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 +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 +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. */ +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 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 1e6555bb02..1a685b9e2e 100644 --- a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc +++ b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc @@ -62,5 +62,21 @@ TEST(inference, image_classification) { LOG(INFO) << output2.dims(); CheckError(output1, output2); + + // float16 inference requires cuda GPUs with >= 5.3 compute capability + if (paddle::platform::GetCUDAComputeCapability(0) >= 53) { + paddle::framework::LoDTensor output3; + std::vector cpu_fetchs3; + cpu_fetchs3.push_back(&output3); + + LOG(INFO) << "--- GPU Runs in float16 mode: ---"; + std::string fp16_dirname = dirname; + fp16_dirname.replace(fp16_dirname.find("book/"), + std::string("book/").size(), "book/float16_"); + TestInference( + fp16_dirname, cpu_feeds, cpu_fetchs3, FLAGS_repeat); + + CheckError(output2, output3); + } #endif } diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 117472599f..af2a7a5620 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -178,10 +178,10 @@ void TestInference(const std::string& dirname, std::unique_ptr ctx; if (PrepareContext) { ctx = executor.Prepare(*inference_program, 0); - executor.RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets, - CreateVars); + executor.RunPreparedContext(ctx.get(), scope, &feed_targets, + &fetch_targets, CreateVars); } else { - executor.Run(*inference_program, scope, feed_targets, fetch_targets, + executor.Run(*inference_program, scope, &feed_targets, &fetch_targets, CreateVars); } @@ -197,10 +197,10 @@ void TestInference(const std::string& dirname, if (PrepareContext) { // Note: if you change the inference_program, you need to call // executor.Prepare() again to get a new ExecutorPrepareContext. - executor.RunPreparedContext(ctx.get(), scope, feed_targets, - fetch_targets, CreateVars); + executor.RunPreparedContext(ctx.get(), scope, &feed_targets, + &fetch_targets, CreateVars); } else { - executor.Run(*inference_program, scope, feed_targets, fetch_targets, + executor.Run(*inference_program, scope, &feed_targets, &fetch_targets, CreateVars); } } diff --git a/paddle/fluid/operators/adam_op.h b/paddle/fluid/operators/adam_op.h index b332b67163..f82ff47b52 100644 --- a/paddle/fluid/operators/adam_op.h +++ b/paddle/fluid/operators/adam_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once #include // for sqrt in CPU and CUDA +#include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" @@ -24,8 +25,14 @@ namespace operators { namespace scatter = paddle::operators::math::scatter; +struct GPUAdam; +struct CPUAdam; + +template +struct AdamFunctor; + template -struct AdamFunctor { +struct AdamFunctor { T beta1_; T beta2_; T epsilon_; @@ -71,6 +78,7 @@ struct AdamFunctor { // Calculation lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow); + mom1 = beta1_ * mom1 + (1 - beta1_) * g; mom2 = beta2_ * mom2 + (1 - beta2_) * g * g; p -= lr * (mom1 / (sqrt(mom2) + epsilon_)); @@ -82,6 +90,71 @@ struct AdamFunctor { } }; +template +struct AdamFunctor { + T beta1_; + T beta2_; + T epsilon_; + + const T* beta1_pow_; + const T* beta2_pow_; + const T* moment1_; + T* moment1_out_; + const T* moment2_; + T* moment2_out_; + const T* lr_; + const T* grad_; + const T* param_; + T* param_out_; + + AdamFunctor(T beta1, T beta2, T epsilon, const T* beta1_pow, + const T* beta2_pow, const T* mom1, T* mom1_out, const T* mom2, + T* mom2_out, const T* lr, const T* grad, const T* param, + T* param_out) + : beta1_(beta1), + beta2_(beta2), + epsilon_(epsilon), + beta1_pow_(beta1_pow), + beta2_pow_(beta2_pow), + moment1_(mom1), + moment1_out_(mom1_out), + moment2_(mom2), + moment2_out_(mom2_out), + lr_(lr), + grad_(grad), + param_(param), + param_out_(param_out) {} + + void operator()(size_t numel) const { + Eigen::Map> g{ + grad_, static_cast(numel)}; + Eigen::Map> mom1{ + moment1_, static_cast(numel)}; + Eigen::Map> mom2{ + moment2_, static_cast(numel)}; + Eigen::Map> param{ + param_, static_cast(numel)}; + + Eigen::Map> param_out{ + param_out_, static_cast(numel)}; + Eigen::Map> moment1_out{ + moment1_out_, static_cast(numel)}; + Eigen::Map> moment2_out{ + moment2_out_, static_cast(numel)}; + + T lr = *lr_; + T beta1_pow = *beta1_pow_; + T beta2_pow = *beta2_pow_; + + // Calculation + lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow); + + moment1_out = beta1_ * mom1 + (1 - beta1_) * g; + moment2_out = beta2_ * mom2 + (1 - beta2_) * g * g; + param_out = param - lr * (moment1_out / (moment2_out.sqrt() + epsilon_)); + } +}; + template struct SparseAdamFunctor { T beta1_; @@ -134,6 +207,7 @@ struct SparseAdamFunctor { T p = param_[rows_[i] * row_numel_ + j]; lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow); + mom1 = beta1_ * mom1 + (1 - beta1_) * g; mom2 = beta2_ * mom2 + (1 - beta2_) * g * g; p -= lr * (mom1 / (sqrt(mom2) + epsilon_)); @@ -177,19 +251,34 @@ class AdamOpKernel : public framework::OpKernel { if (grad_var->IsType()) { auto& grad = Ref(ctx.Input("Grad"), "Must set Grad"); - AdamFunctor functor( - beta1, beta2, epsilon, beta1_pow.template data(), - beta2_pow.template data(), mom1.template data(), - mom1_out.template mutable_data(ctx.GetPlace()), - mom2.template data(), - mom2_out.template mutable_data(ctx.GetPlace()), - lr.template data(), grad.template data(), - param.template data(), - param_out.template mutable_data(ctx.GetPlace())); - platform::ForRange for_range( - static_cast(ctx.device_context()), - param.numel()); - for_range(functor); + + if (platform::is_cpu_place(ctx.GetPlace())) { + AdamFunctor functor( + beta1, beta2, epsilon, beta1_pow.template data(), + beta2_pow.template data(), mom1.template data(), + mom1_out.template mutable_data(ctx.GetPlace()), + mom2.template data(), + mom2_out.template mutable_data(ctx.GetPlace()), + lr.template data(), grad.template data(), + param.template data(), + param_out.template mutable_data(ctx.GetPlace())); + functor(param.numel()); + } else if (platform::is_gpu_place(ctx.GetPlace())) { + AdamFunctor functor( + beta1, beta2, epsilon, beta1_pow.template data(), + beta2_pow.template data(), mom1.template data(), + mom1_out.template mutable_data(ctx.GetPlace()), + mom2.template data(), + mom2_out.template mutable_data(ctx.GetPlace()), + lr.template data(), grad.template data(), + param.template data(), + param_out.template mutable_data(ctx.GetPlace())); + + platform::ForRange for_range( + static_cast(ctx.device_context()), + param.numel()); + for_range(functor); + } } else if (grad_var->IsType()) { auto& grad = Ref(ctx.Input("Grad"), "Must set Grad"); diff --git a/paddle/fluid/operators/beam_search_decode_op.h b/paddle/fluid/operators/beam_search_decode_op.h index 4cb0457d92..3c01f81c83 100644 --- a/paddle/fluid/operators/beam_search_decode_op.h +++ b/paddle/fluid/operators/beam_search_decode_op.h @@ -223,8 +223,9 @@ void BeamSearchDecoder::ConvertSentenceVectorToLodTensor( sentence_vector_list[src_idx].size()); } - auto cpu_place = new paddle::platform::CPUPlace(); - paddle::platform::CPUDeviceContext cpu_ctx(*cpu_place); + auto cpu_place = std::unique_ptr( + new paddle::platform::CPUPlace()); + paddle::platform::CPUDeviceContext cpu_ctx(*cpu_place.get()); framework::LoD lod; lod.push_back(source_level_lod); diff --git a/paddle/fluid/operators/bilinear_interp_op.cc b/paddle/fluid/operators/bilinear_interp_op.cc new file mode 100644 index 0000000000..69f79bf93b --- /dev/null +++ b/paddle/fluid/operators/bilinear_interp_op.cc @@ -0,0 +1,94 @@ +/* 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/operators/bilinear_interp_op.h" +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +class BilinearInterpOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of BilinearInterOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of BilinearInterOp should not be null."); + + auto dim_x = ctx->GetInputDim("X"); // NCHW format + int out_h = ctx->Attrs().Get("out_h"); + int out_w = ctx->Attrs().Get("out_w"); + PADDLE_ENFORCE_EQ(dim_x.size(), 4, "X's dimension must be 4"); + + std::vector dim_out({dim_x[0], dim_x[1], out_h, out_w}); + ctx->SetOutputDim("Out", framework::make_ddim(dim_out)); + } +}; + +class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker { + public: + BilinearInterpOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", + "(Tensor) The input tensor of bilinear interpolation, " + "This is a 4-D tensor with shape of (N x C x h x w)"); + AddOutput("Out", + "(Tensor) The dimension of output is (N x C x out_h x out_w]"); + + AddAttr("out_h", "(int) output height of bilinear interpolation op."); + AddAttr("out_w", "(int) output width of bilinear interpolation op."); + AddComment(R"DOC( + Bilinear interpolation is an extension of linear interpolation for + interpolating functions of two variables (e.g. H-direction and + W-direction in this op) on a rectilinear 2D grid. + + The key idea is to perform linear interpolation first in one + direction, and then again in the other direction. + + For details, please refer to Wikipedia: + https://en.wikipedia.org/wiki/Bilinear_interpolation + )DOC"); + } +}; + +class BilinearInterpOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null"); + auto dim_x = ctx->GetInputDim("X"); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), dim_x); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(bilinear_interp, ops::BilinearInterpOp, + ops::BilinearInterpOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(bilinear_interp_grad, ops::BilinearInterpOpGrad); +REGISTER_OP_CPU_KERNEL(bilinear_interp, ops::BilinearInterpKernel); +REGISTER_OP_CPU_KERNEL(bilinear_interp_grad, + ops::BilinearInterpGradKernel); diff --git a/paddle/fluid/operators/bilinear_interp_op.cu b/paddle/fluid/operators/bilinear_interp_op.cu new file mode 100644 index 0000000000..82eb9e83bd --- /dev/null +++ b/paddle/fluid/operators/bilinear_interp_op.cu @@ -0,0 +1,186 @@ +/* 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/operators/bilinear_interp_op.h" +#include "paddle/fluid/platform/cuda_helper.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +template +__global__ void KeBilinearInterpFw( + const T* in, const size_t in_img_h, const size_t in_img_w, + const size_t input_h, const size_t input_w, T* out, const size_t out_img_h, + const size_t out_img_w, const size_t output_h, const size_t output_w, + const size_t num_channels, const T ratio_h, const T ratioW) { + int nthreads = output_h * output_w; + int tid = blockIdx.x * blockDim.x + threadIdx.x; + if (tid < nthreads) { + int out_id_h = tid / output_w; + int out_id_w = tid % output_w; + int in_img_size = input_w / num_channels; + int out_img_size = output_w / num_channels; + int channel_id = out_id_w / out_img_size; + + int out_img_idy = (out_id_w % out_img_size) / out_img_w; + int in_img_idy = ratio_h * out_img_idy; + int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0; + T h1lambda = ratio_h * out_img_idy - in_img_idy; + T h2lambda = 1.f - h1lambda; + + int out_img_idx = tid % out_img_w; + int in_img_idx = ratioW * out_img_idx; + int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0; + T w1lambda = ratioW * out_img_idx - in_img_idx; + T w2lambda = 1.f - w1lambda; + + const T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + + in_img_idy * in_img_w + in_img_idx]; + + // bilinear interpolation + out[out_id_h * output_w + out_id_w] = + h2lambda * (w2lambda * in_pos[0] + w1lambda * in_pos[w_id]) + + h1lambda * (w2lambda * in_pos[h_id * in_img_w] + + w1lambda * in_pos[h_id * in_img_w + w_id]); + } +} + +template +__global__ void KeBilinearInterpBw( + T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h, + const size_t input_w, const T* out, const size_t out_img_h, + const size_t out_img_w, const size_t output_h, const size_t output_w, + const size_t num_channels, const T ratio_h, const T ratioW) { + int nthreads = output_h * output_w; + int tid = blockIdx.x * blockDim.x + threadIdx.x; + if (tid < nthreads) { + int out_id_h = tid / output_w; + int out_id_w = tid % output_w; + int in_img_size = input_w / num_channels; + int out_img_size = output_w / num_channels; + int channel_id = out_id_w / out_img_size; + + int out_img_idy = (out_id_w % out_img_size) / out_img_w; + int in_img_idy = ratio_h * out_img_idy; + int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0; + T h1lambda = ratio_h * out_img_idy - in_img_idy; + T h2lambda = 1.f - h1lambda; + + int out_img_idx = tid % out_img_w; + int in_img_idx = ratioW * out_img_idx; + int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0; + T w1lambda = ratioW * out_img_idx - in_img_idx; + T w2lambda = 1.f - w1lambda; + + T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + + in_img_idy * in_img_w + in_img_idx]; + const T* out_pos = &out[out_id_h * output_w + out_id_w]; + atomicAdd(&in_pos[0], h2lambda * w2lambda * out_pos[0]); + atomicAdd(&in_pos[w_id], h2lambda * w1lambda * out_pos[0]); + atomicAdd(&in_pos[h_id * in_img_w], h1lambda * w2lambda * out_pos[0]); + atomicAdd(&in_pos[h_id * in_img_w + w_id], + h1lambda * w1lambda * out_pos[0]); + } +} + +template +class BilinearInterpOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + auto* input_t = ctx.Input("X"); // float tensor + auto* output_t = ctx.Output("Out"); // float tensor + auto* input = input_t->data(); + auto* output = output_t->mutable_data(ctx.GetPlace()); + + int out_h = ctx.Attr("out_h"); + int out_w = ctx.Attr("out_w"); + int batch_size = input_t->dims()[0]; + int channels = input_t->dims()[1]; + int in_h = input_t->dims()[2]; + int in_w = input_t->dims()[3]; + + int in_hw = in_h * in_w; + int out_hw = out_h * out_w; + int in_chw = channels * in_hw; + int out_chw = channels * out_hw; + + T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + + if (in_h == out_h && in_w == out_w) { + memcpy(output, input, input_t->numel() * sizeof(T)); + } else { + int threadNum = batch_size * out_chw; + int blocks = (threadNum + 1024 - 1) / 1024; + + KeBilinearInterpFw< + T><<>>( + input, in_h, in_w, batch_size, in_chw, output, out_h, out_w, + batch_size, out_chw, channels, ratio_h, ratio_w); + } + } +}; + +template +class BilinearInterpGradOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* d_input_t = ctx.Output(framework::GradVarName("X")); + auto* d_output_t = ctx.Input(framework::GradVarName("Out")); + auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); + auto* d_output = d_output_t->data(); + + auto& device_ctx = + ctx.template device_context(); + math::SetConstant zero; + zero(device_ctx, d_input_t, static_cast(0.0)); + + int out_h = ctx.Attr("out_h"); + int out_w = ctx.Attr("out_w"); + int batch_size = d_input_t->dims()[0]; + int channels = d_input_t->dims()[1]; + int in_h = d_input_t->dims()[2]; + int in_w = d_input_t->dims()[3]; + + int in_hw = in_h * in_w; + int out_hw = out_h * out_w; + int in_chw = channels * in_hw; + int out_chw = channels * out_hw; + + T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + + if (in_h == out_h && in_w == out_w) { + memcpy(d_input, d_output, d_input_t->numel() * sizeof(T)); + } else { + int threadNum = batch_size * out_chw; + int blocks = (threadNum + 1024 - 1) / 1024; + + KeBilinearInterpBw< + T><<>>( + d_input, in_h, in_w, batch_size, in_chw, d_output, out_h, out_w, + batch_size, out_chw, channels, ratio_h, ratio_w); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL(bilinear_interp, + ops::BilinearInterpOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(bilinear_interp_grad, + ops::BilinearInterpGradOpCUDAKernel); diff --git a/paddle/fluid/operators/bilinear_interp_op.h b/paddle/fluid/operators/bilinear_interp_op.h new file mode 100644 index 0000000000..f6cd77e4d4 --- /dev/null +++ b/paddle/fluid/operators/bilinear_interp_op.h @@ -0,0 +1,143 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class BilinearInterpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input_t = ctx.Input("X"); // float tensor + auto* output_t = ctx.Output("Out"); // float tensor + auto* input = input_t->data(); + auto* output = output_t->mutable_data(ctx.GetPlace()); + + int out_h = ctx.Attr("out_h"); + int out_w = ctx.Attr("out_w"); + int batch_size = input_t->dims()[0]; + int channels = input_t->dims()[1]; + int in_h = input_t->dims()[2]; + int in_w = input_t->dims()[3]; + + int in_hw = in_h * in_w; + int out_hw = out_h * out_w; + int in_chw = channels * in_hw; + int out_chw = channels * out_hw; + + T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + + if (in_h == out_h && in_w == out_w) { + memcpy(output, input, input_t->numel() * sizeof(T)); + } else { + for (int k = 0; k < batch_size; ++k) { // loop for batches + for (int i = 0; i < out_h; ++i) { // loop for images + int h = ratio_h * i; + int hid = (h < in_h - 1) ? 1 : 0; + T h1lambda = ratio_h * i - h; + T h2lambda = 1 - h1lambda; + + for (int j = 0; j < out_w; ++j) { + int w = ratio_w * j; + int wid = (w < in_w - 1) ? 1 : 0; + T w1lambda = ratio_w * j - w; + T w2lambda = 1 - w1lambda; + // calculate four position for bilinear interpolation + const T* in_pos = &input[k * in_chw + h * in_w + w]; + T* out_pos = &output[k * out_chw + i * out_w + j]; + + for (int c = 0; c < channels; ++c) { // loop for channels + // bilinear interpolation + out_pos[0] = + h2lambda * (w2lambda * in_pos[0] + w1lambda * in_pos[wid]) + + h1lambda * (w2lambda * in_pos[hid * in_w] + + w1lambda * in_pos[hid * in_w + wid]); + in_pos += in_hw; + out_pos += out_hw; + } + } + } + } + } + } +}; + +template +class BilinearInterpGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* d_input_t = ctx.Output(framework::GradVarName("X")); + auto* d_output_t = ctx.Input(framework::GradVarName("Out")); + auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); + auto* d_output = d_output_t->data(); + + auto& device_ctx = + ctx.template device_context(); + math::SetConstant zero; + zero(device_ctx, d_input_t, static_cast(0.0)); + + int out_h = ctx.Attr("out_h"); + int out_w = ctx.Attr("out_w"); + int batch_size = d_input_t->dims()[0]; + int channels = d_input_t->dims()[1]; + int in_h = d_input_t->dims()[2]; + int in_w = d_input_t->dims()[3]; + + int in_hw = in_h * in_w; + int out_hw = out_h * out_w; + int in_chw = channels * in_hw; + int out_chw = channels * out_hw; + + T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + + if (in_h == out_h && in_w == out_w) { + memcpy(d_input, d_output, d_input_t->numel() * sizeof(T)); + } else { + for (int k = 0; k < batch_size; ++k) { // loop for batches + for (int i = 0; i < out_h; ++i) { // loop for images + int h = ratio_h * i; + int hid = (h < in_h - 1) ? 1 : 0; + T h1lambda = ratio_h * i - h; + T h2lambda = 1 - h1lambda; + + for (int j = 0; j < out_w; ++j) { + int w = ratio_w * j; + int wid = (w < in_w - 1) ? 1 : 0; + T w1lambda = ratio_w * j - w; + T w2lambda = 1 - w1lambda; + T* in_pos = &d_input[k * in_chw + h * in_w + w]; + const T* out_pos = &d_output[k * out_chw + i * out_w + j]; + + for (int c = 0; c < channels; ++c) { // loop for channels + in_pos[0] += h2lambda * w2lambda * out_pos[0]; + in_pos[wid] += h2lambda * w1lambda * out_pos[0]; + in_pos[hid * in_w] += h1lambda * w2lambda * out_pos[0]; + in_pos[hid * in_w + wid] += h1lambda * w1lambda * out_pos[0]; + in_pos += in_hw; + out_pos += out_hw; + } + } + } + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/concurrency/channel_util.cc b/paddle/fluid/operators/concurrency/channel_util.cc index 246c99489c..fba4abf189 100644 --- a/paddle/fluid/operators/concurrency/channel_util.cc +++ b/paddle/fluid/operators/concurrency/channel_util.cc @@ -12,7 +12,7 @@ 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/operators/concurrency/channel_util.h" #include "paddle/fluid/framework/var_type.h" namespace poc = paddle::operators::concurrency; diff --git a/paddle/fluid/operators/conditional_block_op.cc b/paddle/fluid/operators/conditional_block_op.cc index 137fee99e8..27f74a789b 100644 --- a/paddle/fluid/operators/conditional_block_op.cc +++ b/paddle/fluid/operators/conditional_block_op.cc @@ -227,7 +227,7 @@ class ConditionalBlockGradMaker : public framework::SingleGradOpDescMaker { grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X", false)); grad_op->SetOutput(framework::GradVarName("Params"), InputGrad("Params", false)); - grad_op->SetBlockAttr("sub_block", *this->grad_block_[0]); + grad_op->SetBlockAttr("sub_block", this->grad_block_[0]); grad_op->SetAttr("is_scalar_condition", GetAttr("is_scalar_condition")); return std::unique_ptr(grad_op); } diff --git a/paddle/fluid/operators/detail/grpc_client.h b/paddle/fluid/operators/detail/grpc_client.h index 4425b19328..f6229b71bc 100644 --- a/paddle/fluid/operators/detail/grpc_client.h +++ b/paddle/fluid/operators/detail/grpc_client.h @@ -29,12 +29,12 @@ limitations under the License. */ #include "grpc++/support/byte_buffer.h" #include "grpc++/support/slice.h" #include "grpc/support/log.h" +#include "paddle/fluid/framework/blocking_queue.h" #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/operators/detail/sendrecvop_utils.h" -#include "paddle/fluid/operators/detail/simple_block_queue.h" namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc index 119e146e07..95f4738b4f 100644 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ b/paddle/fluid/operators/detail/grpc_server.cc @@ -30,9 +30,13 @@ enum CallStatus { PROCESS = 0, FINISH }; class RequestBase { public: explicit RequestBase(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, + ::grpc::ServerCompletionQueue* cq, bool sync_mode, const platform::DeviceContext* dev_ctx) - : service_(service), cq_(cq), status_(PROCESS), dev_ctx_(dev_ctx) { + : service_(service), + cq_(cq), + sync_mode_(sync_mode), + status_(PROCESS), + dev_ctx_(dev_ctx) { PADDLE_ENFORCE(cq_); } virtual ~RequestBase() {} @@ -49,6 +53,7 @@ class RequestBase { ::grpc::ServerContext ctx_; GrpcService::AsyncService* service_; ::grpc::ServerCompletionQueue* cq_; + const bool sync_mode_; CallStatus status_; const platform::DeviceContext* dev_ctx_; }; @@ -56,11 +61,17 @@ class RequestBase { class RequestSend final : public RequestBase { public: explicit RequestSend(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, + ::grpc::ServerCompletionQueue* cq, bool sync_mode, framework::Scope* scope, ReceivedQueue* queue, const platform::DeviceContext* dev_ctx) - : RequestBase(service, cq, dev_ctx), queue_(queue), responder_(&ctx_) { - request_.reset(new VariableResponse(scope, dev_ctx_)); + : RequestBase(service, cq, sync_mode, dev_ctx), + queue_(queue), + responder_(&ctx_) { + if (sync_mode_) { + request_.reset(new VariableResponse(scope, dev_ctx_, false)); + } else { + request_.reset(new VariableResponse(scope, dev_ctx_, true)); + } int method_id = static_cast(detail::GrpcMethod::kSendVariable); service_->RequestAsyncUnary(method_id, &ctx_, request_.get(), &responder_, cq_, cq_, this); @@ -87,11 +98,11 @@ class RequestSend final : public RequestBase { class RequestGet final : public RequestBase { public: explicit RequestGet(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, + ::grpc::ServerCompletionQueue* cq, bool sync_mode, framework::Scope* scope, const platform::DeviceContext* dev_ctx, - SimpleBlockQueue* queue) - : RequestBase(service, cq, dev_ctx), + framework::BlockingQueue* queue) + : RequestBase(service, cq, sync_mode, dev_ctx), responder_(&ctx_), scope_(scope), queue_(queue) { @@ -128,25 +139,29 @@ class RequestGet final : public RequestBase { sendrecv::VariableMessage request_; ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; framework::Scope* scope_; - SimpleBlockQueue* queue_; + framework::BlockingQueue* queue_; }; class RequestPrefetch final : public RequestBase { public: explicit RequestPrefetch(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, + ::grpc::ServerCompletionQueue* cq, bool sync_mode, framework::Scope* scope, const platform::DeviceContext* dev_ctx, framework::Executor* executor, framework::ProgramDesc* program, framework::ExecutorPrepareContext* prefetch_ctx) - : RequestBase(service, cq, dev_ctx), + : RequestBase(service, cq, sync_mode, dev_ctx), responder_(&ctx_), scope_(scope), executor_(executor), program_(program), prefetch_ctx_(prefetch_ctx) { - request_.reset(new VariableResponse(scope, dev_ctx_)); + if (sync_mode_) { + request_.reset(new VariableResponse(scope, dev_ctx_, false)); + } else { + request_.reset(new VariableResponse(scope, dev_ctx_, true)); + } int method_id = static_cast(detail::GrpcMethod::kPrefetchVariable); service_->RequestAsyncUnary(method_id, &ctx_, request_.get(), &responder_, cq_, cq_, this); @@ -181,7 +196,6 @@ class RequestPrefetch final : public RequestBase { framework::Executor* executor_; framework::ProgramDesc* program_; framework::ExecutorPrepareContext* prefetch_ctx_; - int blkid_; }; void AsyncGRPCServer::WaitClientGet(int count) { @@ -254,8 +268,8 @@ void AsyncGRPCServer::TryToRegisterNewSendOne() { VLOG(3) << "shutdown, do not TryToRegisterNewSendOne"; return; } - RequestSend* send = new RequestSend(&service_, cq_send_.get(), scope_, - &var_recv_queue_, dev_ctx_); + RequestSend* send = new RequestSend(&service_, cq_send_.get(), sync_mode_, + scope_, &var_recv_queue_, dev_ctx_); VLOG(4) << "Create RequestSend status:" << send->Status(); } @@ -265,8 +279,8 @@ void AsyncGRPCServer::TryToRegisterNewGetOne() { VLOG(3) << "shutdown, do not TryToRegisterNewGetOne"; return; } - RequestGet* get = new RequestGet(&service_, cq_get_.get(), scope_, dev_ctx_, - &var_get_queue_); + RequestGet* get = new RequestGet(&service_, cq_get_.get(), sync_mode_, scope_, + dev_ctx_, &var_get_queue_); VLOG(4) << "Create RequestGet status:" << get->Status(); } @@ -277,8 +291,8 @@ void AsyncGRPCServer::TryToRegisterNewPrefetchOne() { return; } RequestPrefetch* prefetch = - new RequestPrefetch(&service_, cq_prefetch_.get(), scope_, dev_ctx_, - executor_, program_, prefetch_ctx_); + new RequestPrefetch(&service_, cq_prefetch_.get(), sync_mode_, scope_, + dev_ctx_, executor_, program_, prefetch_ctx_); VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status(); } @@ -301,9 +315,11 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq, VLOG(3) << "HandleRequest for " << cq_name << " while after Next"; PADDLE_ENFORCE(tag); - // FIXME(typhoonzero): de-couple the barriers with recv_op - if (!is_shut_down_ && cq_name == "cq_get") WaitCond(1); - if (!is_shut_down_ && cq_name == "cq_send") WaitCond(0); + if (sync_mode_) { + // FIXME(typhoonzero): de-couple the barriers with recv_op + if (!is_shut_down_ && cq_name == "cq_get") WaitCond(1); + if (!is_shut_down_ && cq_name == "cq_send") WaitCond(0); + } RequestBase* base = reinterpret_cast(tag); // reference: @@ -320,13 +336,13 @@ void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq, switch (base->Status()) { case PROCESS: { - VLOG(4) << cq_name << " status:" << base->Status(); + VLOG(4) << cq_name << " PROCESS status:" << base->Status(); TryToRegisterNewOne(); base->Process(); break; } case FINISH: { - VLOG(4) << cq_name << " status:" << base->Status(); + VLOG(4) << cq_name << " FINISH status:" << base->Status(); delete base; break; } diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h index 452ff5e967..99b87b8c6c 100644 --- a/paddle/fluid/operators/detail/grpc_server.h +++ b/paddle/fluid/operators/detail/grpc_server.h @@ -19,6 +19,7 @@ limitations under the License. */ #include #include "grpc++/grpc++.h" +#include "paddle/fluid/framework/blocking_queue.h" #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/program_desc.h" @@ -29,7 +30,6 @@ limitations under the License. */ #include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" #include "paddle/fluid/operators/detail/send_recv.pb.h" #include "paddle/fluid/operators/detail/sendrecvop_utils.h" -#include "paddle/fluid/operators/detail/simple_block_queue.h" namespace paddle { namespace operators { @@ -37,14 +37,15 @@ namespace detail { typedef std::pair> ReceivedMessage; -typedef SimpleBlockQueue ReceivedQueue; +typedef framework::BlockingQueue ReceivedQueue; typedef std::pair MessageWithName; class RequestBase; class AsyncGRPCServer final { public: - explicit AsyncGRPCServer(const std::string &address) : address_(address) {} + explicit AsyncGRPCServer(const std::string &address, bool sync_mode) + : address_(address), sync_mode_(sync_mode) {} void RunSyncUpdate(); @@ -95,11 +96,12 @@ class AsyncGRPCServer final { std::unique_ptr<::grpc::Server> server_; std::string address_; + const bool sync_mode_; framework::Scope *scope_; const platform::DeviceContext *dev_ctx_; // received variable from RPC, operators fetch variable from this queue. - SimpleBlockQueue var_get_queue_; + framework::BlockingQueue var_get_queue_; // client send variable to this queue. ReceivedQueue var_recv_queue_; diff --git a/paddle/fluid/operators/detail/grpc_server_test.cc b/paddle/fluid/operators/detail/grpc_server_test.cc index c51933718f..25b95d608d 100644 --- a/paddle/fluid/operators/detail/grpc_server_test.cc +++ b/paddle/fluid/operators/detail/grpc_server_test.cc @@ -89,7 +89,7 @@ void InitTensorsOnServer(framework::Scope* scope, platform::CPUPlace* place, } void StartServer(const std::string& endpoint) { - rpc_service_.reset(new detail::AsyncGRPCServer(endpoint)); + rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, true)); framework::ProgramDesc program; framework::Scope scope; platform::CPUPlace place; diff --git a/paddle/fluid/operators/detail/simple_block_queue.h b/paddle/fluid/operators/detail/simple_block_queue.h deleted file mode 100644 index 69773e05df..0000000000 --- a/paddle/fluid/operators/detail/simple_block_queue.h +++ /dev/null @@ -1,52 +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 // NOLINT -#include -#include // NOLINT - -namespace paddle { -namespace operators { -namespace detail { - -template -class SimpleBlockQueue { - private: - std::mutex mutex_; - std::condition_variable condition_; - std::deque queue_; - - public: - void Push(T const& value) { - { - std::unique_lock lock(this->mutex_); - queue_.push_front(value); - } - this->condition_.notify_one(); - } - - T Pop() { - std::unique_lock lock(this->mutex_); - this->condition_.wait(lock, [=] { return !this->queue_.empty(); }); - T rc(std::move(this->queue_.back())); - this->queue_.pop_back(); - return rc; - } -}; - -} // namespace detail -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/detail/variable_response.h b/paddle/fluid/operators/detail/variable_response.h index 3018a5c4af..bf624da2a6 100644 --- a/paddle/fluid/operators/detail/variable_response.h +++ b/paddle/fluid/operators/detail/variable_response.h @@ -46,7 +46,9 @@ class VariableResponse { } virtual ~VariableResponse() { - if (create_scope_) scope_->DeleteScope(local_scope_); + if (create_scope_) { + scope_->DeleteScope(local_scope_); + } } // return: @@ -63,6 +65,8 @@ class VariableResponse { const framework::Scope& GetLocalScope() const { return *local_scope_; } + framework::Scope* GetMutableLocalScope() const { return local_scope_; } + inline std::string Varname() { return meta_.varname(); } inline std::string OutVarname() { return meta_.out_varname(); } diff --git a/paddle/fluid/operators/fetch_op.cc b/paddle/fluid/operators/fetch_op.cc index 7c7f3e9059..18deec5813 100644 --- a/paddle/fluid/operators/fetch_op.cc +++ b/paddle/fluid/operators/fetch_op.cc @@ -57,10 +57,7 @@ class FetchOp : public framework::OperatorBase { // FIXME(yuyang18): Should we assume the fetch operator always generate // CPU outputs? - auto &dev_ctx = *pool.Get(src_item.place()); - - TensorCopy(src_item, platform::CPUPlace(), dev_ctx, &dst_item); - dev_ctx.Wait(); + TensorCopySync(src_item, platform::CPUPlace(), &dst_item); dst_item.set_lod(src_item.lod()); VLOG(3) << "Fetch variable " << fetch_var_name << " to " << out_name; diff --git a/paddle/fluid/operators/gru_op.h b/paddle/fluid/operators/gru_op.h index 1d5c291495..3b0d93e54b 100644 --- a/paddle/fluid/operators/gru_op.h +++ b/paddle/fluid/operators/gru_op.h @@ -34,7 +34,7 @@ inline void ReorderInitState(const DeviceContext& ctx, framework::Tensor* dst, bool indexed_src) { math::CopyMatrixRowsFunctor row_shuffle; dst->mutable_data(src.dims(), ctx.GetPlace()); - row_shuffle(ctx, src, index_lod, *dst, indexed_src); + row_shuffle(ctx, src, index_lod, dst, indexed_src); } template @@ -56,14 +56,12 @@ class GRUKernel : public framework::OpKernel { auto* hidden = context.Output("Hidden"); hidden->mutable_data(context.GetPlace()); - context.ShareLoD("Input", "Hidden"); - auto hidden_dims = hidden->dims(); bool is_reverse = context.Attr("is_reverse"); math::LoDTensor2BatchFunctor to_batch; auto& dev_ctx = context.template device_context(); - to_batch(dev_ctx, *input, *batch_gate, true, is_reverse); + to_batch(dev_ctx, *input, batch_gate, true, is_reverse); if (bias) { math::RowwiseAdd add_bias; @@ -115,7 +113,7 @@ class GRUKernel : public framework::OpKernel { math::Batch2LoDTensorFunctor to_seq; batch_hidden->set_lod(batch_gate->lod()); - to_seq(dev_ctx, *batch_hidden, *hidden); + to_seq(dev_ctx, *batch_hidden, hidden); } void Compute(const framework::ExecutionContext& context) const override { @@ -176,7 +174,7 @@ class GRUGradKernel : public framework::OpKernel { bool is_reverse = context.Attr("is_reverse"); batch_hidden_grad.set_lod(batch_hidden->lod()); - to_batch(dev_ctx, *hidden_grad, batch_hidden_grad, false, is_reverse); + to_batch(dev_ctx, *hidden_grad, &batch_hidden_grad, false, is_reverse); math::GRUMetaValue gru_value; gru_value.gate_weight = const_cast(weight_data); @@ -238,7 +236,7 @@ class GRUGradKernel : public framework::OpKernel { input_grad->mutable_data(context.GetPlace()); math::Batch2LoDTensorFunctor to_seq; batch_gate_grad.set_lod(batch_gate->lod()); - to_seq(dev_ctx, batch_gate_grad, *input_grad); + to_seq(dev_ctx, batch_gate_grad, input_grad); } if (bias_grad) { bias_grad->mutable_data(context.GetPlace()); diff --git a/paddle/fluid/operators/iou_similarity_op.h b/paddle/fluid/operators/iou_similarity_op.h index c76448c736..9f193ebc59 100644 --- a/paddle/fluid/operators/iou_similarity_op.h +++ b/paddle/fluid/operators/iou_similarity_op.h @@ -41,22 +41,24 @@ struct IOUSimilarityFunctor { IOUSimilarityFunctor(const T* x, const T* y, T* z, int cols) : x_(x), y_(y), z_(z), cols_(static_cast(cols)) {} - inline HOSTDEVICE void operator()(size_t row_id) const { + inline HOSTDEVICE void operator()(size_t tid) const { + size_t row_id = tid / cols_; + size_t col_id = tid % cols_; + T x_min1 = x_[row_id * 4]; T y_min1 = x_[row_id * 4 + 1]; T x_max1 = x_[row_id * 4 + 2]; T y_max1 = x_[row_id * 4 + 3]; - for (size_t i = 0; i < cols_; ++i) { - T x_min2 = y_[i * 4]; - T y_min2 = y_[i * 4 + 1]; - T x_max2 = y_[i * 4 + 2]; - T y_max2 = y_[i * 4 + 3]; - T sim = IOUSimilarity(x_min1, y_min1, x_max1, y_max1, x_min2, y_min2, - x_max2, y_max2); + T x_min2 = y_[col_id * 4]; + T y_min2 = y_[col_id * 4 + 1]; + T x_max2 = y_[col_id * 4 + 2]; + T y_max2 = y_[col_id * 4 + 3]; + + T sim = IOUSimilarity(x_min1, y_min1, x_max1, y_max1, x_min2, y_min2, + x_max2, y_max2); - z_[row_id * cols_ + i] = sim; - } + z_[row_id * cols_ + col_id] = sim; } const T* x_; const T* y_; @@ -81,7 +83,7 @@ class IOUSimilarityKernel : public framework::OpKernel { out->mutable_data(ctx.GetPlace()), y_n); platform::ForRange for_range( - static_cast(ctx.device_context()), x_n); + static_cast(ctx.device_context()), x_n * y_n); for_range(functor); } }; // namespace operators diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index af235fb6a0..57cff680ab 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -27,6 +27,38 @@ void RunServer(std::shared_ptr service) { VLOG(4) << "RunServer thread end"; } +static void split(const std::string &str, char sep, + std::vector *pieces) { + pieces->clear(); + if (str.empty()) { + return; + } + size_t pos = 0; + size_t next = str.find(sep, pos); + while (next != std::string::npos) { + pieces->push_back(str.substr(pos, next - pos)); + pos = next + 1; + next = str.find(sep, pos); + } + if (!str.substr(pos).empty()) { + pieces->push_back(str.substr(pos)); + } +} + +static void AsyncExecuteBlock(framework::Executor *executor, + framework::ExecutorPrepareContext *prepared, + framework::Scope *scope) { + std::future future = framework::Async([&executor, &prepared, &scope]() { + try { + executor->RunPreparedContext(prepared, scope, false, false); + } catch (std::exception &e) { + LOG(ERROR) << "run sub program error " << e.what(); + } + }); + // TODO(qiao) maybe we can remove this + future.wait(); +} + static void ParallelExecuteBlocks( const std::vector ¶llel_blkids, framework::Executor *executor, const std::vector> @@ -169,15 +201,82 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor, } // while(true) } +void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, + framework::ProgramDesc *program, + framework::Scope *recv_scope, + framework::BlockDesc *prefetch_block) const { + VLOG(3) << "RunAsyncLoop in"; + // grad name to block id + std::unordered_map grad_to_block_id; + std::unordered_map id_to_grad; + + auto grad_to_block_id_str = + Attr>("grad_to_block_id"); + for (auto &grad_and_id : grad_to_block_id_str) { + std::vector pieces; + split(grad_and_id, ':', &pieces); + VLOG(3) << "after split, grad = " << pieces[0] << ", id=" << pieces[1]; + PADDLE_ENFORCE_EQ(pieces.size(), 2); + PADDLE_ENFORCE_EQ(grad_to_block_id.count(pieces[0]), 0); + int block_id = std::stoi(pieces[1]); + grad_to_block_id[pieces[0]] = block_id; + id_to_grad[block_id] = pieces[0]; + } + size_t num_blocks = program->Size(); + PADDLE_ENFORCE_GE(num_blocks, 2, + "server program should have at least 2 blocks"); + + std::vector block_list; + for (size_t blkid = 1; blkid < num_blocks; ++blkid) { + block_list.push_back(blkid); + } + auto optimize_prepared = executor->Prepare(*program, block_list); + std::unordered_map> + grad_to_prepared_ctx; + for (size_t i = 0; i < block_list.size(); ++i) { + grad_to_prepared_ctx[id_to_grad[block_list[i]]] = optimize_prepared[i]; + } + + VLOG(3) << "RunAsyncLoop into while"; + bool exit_flag = false; + while (!exit_flag) { + const detail::ReceivedMessage v = rpc_service_->Get(); + auto recv_var_name = v.first; + if (recv_var_name == LISTEN_TERMINATE_MESSAGE) { + LOG(INFO) << "received terminate message and exit"; + exit_flag = true; + break; + } else { + VLOG(3) << "received grad: " << recv_var_name; + auto var = v.second->GetVar(); + if (var == nullptr) { + LOG(ERROR) << "Can not find server side var: " << recv_var_name; + PADDLE_THROW("Can not find server side var"); + } + AsyncExecuteBlock(executor, grad_to_prepared_ctx[recv_var_name].get(), + v.second->GetMutableLocalScope()); + } + + if (exit_flag) { + rpc_service_->ShutDown(); + break; + } + } // while(true) +} + void ListenAndServOp::RunImpl(const framework::Scope &scope, const platform::Place &dev_place) const { platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); auto &dev_ctx = *pool.Get(dev_place); framework::Scope &recv_scope = scope.NewScope(); + bool sync_mode = Attr("sync_mode"); + PADDLE_ENFORCE(!rpc_service_); std::string endpoint = Attr("endpoint"); - rpc_service_.reset(new detail::AsyncGRPCServer(endpoint)); + + rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, sync_mode)); auto *optimize_block = Attr(kOptimizeBlock); auto *prefetch_block = Attr(kPrefetchBlock); @@ -202,7 +301,11 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, sleep(5); // Write to a file of server selected port for python use. SavePort(rpc_service_); - RunSyncLoop(&executor, program, &recv_scope, prefetch_block); + if (sync_mode) { + RunSyncLoop(&executor, program, &recv_scope, prefetch_block); + } else { + RunAsyncLoop(&executor, program, &recv_scope, prefetch_block); + } } class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { @@ -221,6 +324,12 @@ from send_op and send back variables to recv_op. "IP address to listen on.") .SetDefault("127.0.0.1:6164") .AddCustomChecker([](const std::string &ip) { return !ip.empty(); }); + AddAttr>( + "grad_to_block_id", + "['param1@GRAD.block0:1', 'param2@GRAD.blockn:2'] " + "a map from grad name to it's optimize block id") + .SetDefault({}); + AddAttr("sync_mode", "if works at sync_mode or not").SetDefault(true); AddAttr(kOptimizeBlock, "BlockID to run on server side."); AddAttr(kPrefetchBlock, diff --git a/paddle/fluid/operators/listen_and_serv_op.h b/paddle/fluid/operators/listen_and_serv_op.h index dfb7c77c8e..3cc0f30477 100644 --- a/paddle/fluid/operators/listen_and_serv_op.h +++ b/paddle/fluid/operators/listen_and_serv_op.h @@ -46,6 +46,11 @@ class ListenAndServOp : public framework::OperatorBase { framework::Scope* recv_scope, framework::BlockDesc* prefetch_block) const; + void RunAsyncLoop(framework::Executor* executor, + framework::ProgramDesc* program, + framework::Scope* recv_scope, + framework::BlockDesc* prefetch_block) const; + void Stop() override; void RunImpl(const framework::Scope& scope, diff --git a/paddle/fluid/operators/lstm_op.h b/paddle/fluid/operators/lstm_op.h index a1ef0eb278..0707aded8c 100644 --- a/paddle/fluid/operators/lstm_op.h +++ b/paddle/fluid/operators/lstm_op.h @@ -33,7 +33,7 @@ inline void ReorderInitState(const DeviceContext& ctx, framework::Tensor* dst, bool indexed_src) { math::CopyMatrixRowsFunctor row_shuffle; dst->mutable_data(src.dims(), ctx.GetPlace()); - row_shuffle(ctx, src, index_lod, *dst, indexed_src); + row_shuffle(ctx, src, index_lod, dst, indexed_src); } template @@ -57,7 +57,7 @@ class LSTMKernel : public framework::OpKernel { bool is_reverse = ctx.Attr("is_reverse"); math::LoDTensor2BatchFunctor to_batch; auto& device_ctx = ctx.template device_context(); - to_batch(device_ctx, *input, *batch_gate, true, is_reverse); + to_batch(device_ctx, *input, batch_gate, true, is_reverse); auto in_dims = input->dims(); int frame_size = static_cast(in_dims[1] / 4); @@ -161,11 +161,11 @@ class LSTMKernel : public framework::OpKernel { math::Batch2LoDTensorFunctor to_seq; batch_hidden.set_lod(batch_gate->lod()); // restore the output hidden in LoDTensor from the batch hidden - to_seq(device_ctx, batch_hidden, *hidden_out); + to_seq(device_ctx, batch_hidden, hidden_out); batch_cell.set_lod(batch_gate->lod()); // restore the output cell state in LoDTensor from the batch cell - to_seq(device_ctx, batch_cell, *cell_out); + to_seq(device_ctx, batch_cell, cell_out); } }; @@ -257,7 +257,7 @@ class LSTMGradKernel : public framework::OpKernel { const framework::DDim& dims, framework::LoDTensor& dst) { dst.mutable_data(dims, ctx.GetPlace()); dst.set_lod(batch_gate->lod()); - to_batch(ctx, src, dst, false); + to_batch(ctx, src, &dst, false); }; LoDTensor batch_hidden, batch_hidden_g, batch_cell; @@ -351,7 +351,7 @@ class LSTMGradKernel : public framework::OpKernel { if (in_g) { /* backward data */ in_g->mutable_data(ctx.GetPlace()); - to_seq(device_ctx, batch_gate_g, *in_g); + to_seq(device_ctx, batch_gate_g, in_g); } if (bias && bias_g) { /* backward bias */ diff --git a/paddle/fluid/operators/lstmp_op.h b/paddle/fluid/operators/lstmp_op.h index 172db54896..628936a310 100644 --- a/paddle/fluid/operators/lstmp_op.h +++ b/paddle/fluid/operators/lstmp_op.h @@ -40,7 +40,7 @@ inline void ReorderInitState(const DeviceContext& ctx, framework::Tensor* dst, bool indexed_src) { math::CopyMatrixRowsFunctor row_shuffle; dst->mutable_data(src.dims(), ctx.GetPlace()); - row_shuffle(ctx, src, index, *dst, indexed_src); + row_shuffle(ctx, src, index, dst, indexed_src); } template @@ -81,7 +81,7 @@ class LSTMPKernel : public framework::OpKernel { bool is_reverse = ctx.Attr("is_reverse"); math::LoDTensor2BatchFunctor to_batch; auto& device_ctx = ctx.template device_context(); - to_batch(device_ctx, *input, *batch_gate, true, is_reverse); + to_batch(device_ctx, *input, batch_gate, true, is_reverse); auto in_dims = input->dims(); int frame_size = static_cast(in_dims[1] / 4); @@ -208,11 +208,11 @@ class LSTMPKernel : public framework::OpKernel { math::Batch2LoDTensorFunctor to_seq; batch_proj.set_lod(batch_gate->lod()); // restore the output hidden in LoDTensor from the batch hidden - to_seq(device_ctx, batch_proj, *proj_out); + to_seq(device_ctx, batch_proj, proj_out); batch_cell.set_lod(batch_gate->lod()); // restore the output cell state in LoDTensor from the batch cell - to_seq(device_ctx, batch_cell, *cell_out); + to_seq(device_ctx, batch_cell, cell_out); } }; @@ -332,7 +332,7 @@ class LSTMPGradKernel : public framework::OpKernel { const framework::DDim& dims, framework::LoDTensor& dst) { dst.mutable_data(dims, ctx.GetPlace()); dst.set_lod(batch_gate->lod()); - to_batch(ctx, src, dst, false); + to_batch(ctx, src, &dst, false); }; LoDTensor batch_hidden_g, batch_proj, batch_proj_g, batch_cell; @@ -471,7 +471,7 @@ class LSTMPGradKernel : public framework::OpKernel { if (in_g) { /* backward data */ in_g->mutable_data(ctx.GetPlace()); - to_seq(device_ctx, batch_gate_g, *in_g); + to_seq(device_ctx, batch_gate_g, in_g); } if (bias && bias_g) { /* backward bias */ diff --git a/paddle/fluid/operators/math/concat_test.cc b/paddle/fluid/operators/math/concat_test.cc index 1741af8148..f0847aafae 100644 --- a/paddle/fluid/operators/math/concat_test.cc +++ b/paddle/fluid/operators/math/concat_test.cc @@ -17,17 +17,14 @@ limitations under the License. */ #include #include "paddle/fluid/framework/tensor_util.h" -using namespace paddle::framework; -using namespace paddle::platform; - template void testConcat() { - Tensor input_a_cpu; - Tensor input_b_cpu; - Tensor out_cpu; - Tensor input_a; - Tensor input_b; - Tensor out; + paddle::framework::Tensor input_a_cpu; + paddle::framework::Tensor input_b_cpu; + paddle::framework::Tensor out_cpu; + paddle::framework::Tensor input_a; + paddle::framework::Tensor input_b; + paddle::framework::Tensor out; DeviceContext* context = new DeviceContext(Place()); // DeviceContext context(Place()); @@ -40,18 +37,18 @@ void testConcat() { * output: * out.shape: [5, 3, 4] */ - auto dim_a = make_ddim({2, 3, 4}); - auto dim_b = make_ddim({3, 3, 4}); - auto dim_out = make_ddim({5, 3, 4}); + auto dim_a = paddle::framework::make_ddim({2, 3, 4}); + auto dim_b = paddle::framework::make_ddim({3, 3, 4}); + auto dim_out = paddle::framework::make_ddim({5, 3, 4}); input_a.mutable_data(dim_a, Place()); input_b.mutable_data(dim_b, Place()); out.mutable_data(dim_out, Place()); if (paddle::platform::is_gpu_place(Place())) { - input_a_cpu.mutable_data(dim_a, CPUPlace()); - input_b_cpu.mutable_data(dim_b, CPUPlace()); - out_cpu.mutable_data(dim_out, CPUPlace()); + input_a_cpu.mutable_data(dim_a, paddle::platform::CPUPlace()); + input_b_cpu.mutable_data(dim_b, paddle::platform::CPUPlace()); + out_cpu.mutable_data(dim_out, paddle::platform::CPUPlace()); } int* a_ptr; @@ -72,11 +69,11 @@ void testConcat() { } if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(input_a_cpu, Place(), *context, &input_a); - TensorCopy(input_b_cpu, Place(), *context, &input_b); + paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a); + paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b); } - std::vector input; + std::vector input; input.push_back(input_a); input.push_back(input_b); @@ -89,7 +86,8 @@ void testConcat() { int* out_ptr; if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(out, CPUPlace(), *context, &out_cpu); + paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context, + &out_cpu); out_ptr = out_cpu.data(); } else { out_ptr = out.data(); @@ -115,9 +113,9 @@ void testConcat() { * output: * out.shape: [2, 7, 4] */ - dim_a = make_ddim({2, 3, 4}); - dim_b = make_ddim({2, 4, 4}); - dim_out = make_ddim({2, 7, 4}); + dim_a = paddle::framework::make_ddim({2, 3, 4}); + dim_b = paddle::framework::make_ddim({2, 4, 4}); + dim_out = paddle::framework::make_ddim({2, 7, 4}); input_a.Resize(dim_a); input_b.Resize(dim_b); @@ -144,8 +142,8 @@ void testConcat() { } if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(input_a_cpu, Place(), *context, &input_a); - TensorCopy(input_b_cpu, Place(), *context, &input_b); + paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a); + paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b); } input.clear(); @@ -159,7 +157,8 @@ void testConcat() { PADDLE_ENFORCE_EQ(input_b.dims(), dim_b); if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(out, CPUPlace(), *context, &out_cpu); + paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context, + &out_cpu); out_ptr = out_cpu.data(); } else { out_ptr = out.data(); @@ -187,9 +186,9 @@ void testConcat() { * output: * out.shape: [2, 3, 9] */ - dim_a = make_ddim({2, 3, 4}); - dim_b = make_ddim({2, 3, 5}); - dim_out = make_ddim({2, 3, 9}); + dim_a = paddle::framework::make_ddim({2, 3, 4}); + dim_b = paddle::framework::make_ddim({2, 3, 5}); + dim_out = paddle::framework::make_ddim({2, 3, 9}); input_a.Resize(dim_a); input_b.Resize(dim_b); @@ -216,8 +215,8 @@ void testConcat() { } if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(input_a_cpu, Place(), *context, &input_a); - TensorCopy(input_b_cpu, Place(), *context, &input_b); + paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a); + paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b); } input.clear(); @@ -231,7 +230,8 @@ void testConcat() { PADDLE_ENFORCE_EQ(input_b.dims(), dim_b); if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(out, CPUPlace(), *context, &out_cpu); + paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context, + &out_cpu); out_ptr = out_cpu.data(); } else { out_ptr = out.data(); @@ -261,9 +261,9 @@ void testConcat() { * output: * out.shape: [2, 6, 4] */ - dim_a = make_ddim({2, 3, 4}); - dim_b = make_ddim({2, 3, 4}); - dim_out = make_ddim({2, 6, 4}); + dim_a = paddle::framework::make_ddim({2, 3, 4}); + dim_b = paddle::framework::make_ddim({2, 3, 4}); + dim_out = paddle::framework::make_ddim({2, 6, 4}); input_a.Resize(dim_a); input_b.Resize(dim_b); @@ -290,8 +290,8 @@ void testConcat() { } if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(input_a_cpu, Place(), *context, &input_a); - TensorCopy(input_b_cpu, Place(), *context, &input_b); + paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a); + paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b); } input.clear(); @@ -305,7 +305,8 @@ void testConcat() { PADDLE_ENFORCE_EQ(input_b.dims(), dim_b); if (paddle::platform::is_gpu_place(Place())) { - TensorCopy(out, CPUPlace(), *context, &out_cpu); + paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context, + &out_cpu); out_ptr = out_cpu.data(); } else { out_ptr = out.data(); diff --git a/paddle/fluid/operators/math/context_project.h b/paddle/fluid/operators/math/context_project.h index 4da94383af..027a019a28 100644 --- a/paddle/fluid/operators/math/context_project.h +++ b/paddle/fluid/operators/math/context_project.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once +#include +#include #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/operators/math/im2col.h" #include "paddle/fluid/operators/math/math_function.h" diff --git a/paddle/fluid/operators/math/cross_entropy.cu b/paddle/fluid/operators/math/cross_entropy.cu index f4935c2813..da73f575f3 100644 --- a/paddle/fluid/operators/math/cross_entropy.cu +++ b/paddle/fluid/operators/math/cross_entropy.cu @@ -108,7 +108,9 @@ class CrossEntropyFunctor { if (softLabel) { const T* label_data = labels->data(); - int block = class_num > 512 ? 512 : pow(2, int(std::log2(class_num))); + int block = class_num > 512 + ? 512 + : pow(2, static_cast(std::log2(class_num))); SoftCrossEntropyKernel<<< batch_size, block, block * sizeof(T), diff --git a/paddle/fluid/operators/math/depthwise_conv.cu b/paddle/fluid/operators/math/depthwise_conv.cu index a5e6e4031b..d360728484 100644 --- a/paddle/fluid/operators/math/depthwise_conv.cu +++ b/paddle/fluid/operators/math/depthwise_conv.cu @@ -12,6 +12,7 @@ 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/operators/math/depthwise_conv.h" #include "paddle/fluid/platform/cuda_helper.h" diff --git a/paddle/fluid/operators/math/depthwise_conv.h b/paddle/fluid/operators/math/depthwise_conv.h index 081bda891d..97aec40188 100644 --- a/paddle/fluid/operators/math/depthwise_conv.h +++ b/paddle/fluid/operators/math/depthwise_conv.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/hostdevice.h" diff --git a/paddle/fluid/operators/math/detail/activation_functions.h b/paddle/fluid/operators/math/detail/activation_functions.h index d205ebf210..b127fbe8c8 100644 --- a/paddle/fluid/operators/math/detail/activation_functions.h +++ b/paddle/fluid/operators/math/detail/activation_functions.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once #include +#include #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/hostdevice.h" diff --git a/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h b/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h index ee7b16da41..0b1034a080 100644 --- a/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h +++ b/paddle/fluid/operators/math/detail/lstm_gpu_kernel.h @@ -13,13 +13,13 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include + #include "paddle/fluid/operators/math/detail/activation_functions.h" #include "paddle/fluid/operators/math/lstm_compute.h" #include "paddle/fluid/platform/cuda_helper.h" #include "paddle/fluid/platform/device_context.h" -#include - namespace paddle { namespace operators { namespace math { diff --git a/paddle/fluid/operators/math/im2col.cc b/paddle/fluid/operators/math/im2col.cc index 123e10586f..336d6febc2 100644 --- a/paddle/fluid/operators/math/im2col.cc +++ b/paddle/fluid/operators/math/im2col.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/im2col.h" +#include namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/im2col.cu b/paddle/fluid/operators/math/im2col.cu index f41c78140f..1268e21e06 100644 --- a/paddle/fluid/operators/math/im2col.cu +++ b/paddle/fluid/operators/math/im2col.cu @@ -12,6 +12,8 @@ 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 "paddle/fluid/operators/math/im2col.h" #include "paddle/fluid/platform/cuda_helper.h" diff --git a/paddle/fluid/operators/math/im2col.h b/paddle/fluid/operators/math/im2col.h index 451ec9d534..26d94e0f2e 100644 --- a/paddle/fluid/operators/math/im2col.h +++ b/paddle/fluid/operators/math/im2col.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/platform/device_context.h" diff --git a/paddle/fluid/operators/math/im2col_test.cc b/paddle/fluid/operators/math/im2col_test.cc index b3978536bc..8e3f0f2868 100644 --- a/paddle/fluid/operators/math/im2col_test.cc +++ b/paddle/fluid/operators/math/im2col_test.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/im2col.h" #include +#include template void testIm2col() { @@ -62,7 +63,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - TensorCopy(input_tmp, *place, *context, &input); + TensorCopySync(input_tmp, *place, &input); } output_cfo.mutable_data( {1, filter_size, filter_size, output_height, output_width}, *place); @@ -87,7 +88,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { out_cfo_ptr = output_cfo.data(); } else { - TensorCopy(output_cfo, paddle::platform::CPUPlace(), *context, &output_tmp); + TensorCopySync(output_cfo, paddle::platform::CPUPlace(), &output_tmp); out_cfo_ptr = output_tmp.data(); } for (int i = 0; i < 6; ++i) { @@ -98,7 +99,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { out_ocf_ptr = output_ocf.data(); } else { - TensorCopy(output_ocf, paddle::platform::CPUPlace(), *context, &output_tmp); + TensorCopySync(output_ocf, paddle::platform::CPUPlace(), &output_tmp); out_ocf_ptr = output_tmp.data(); } @@ -119,7 +120,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - TensorCopy(input_tmp, *place, *context, &input); + TensorCopySync(input_tmp, *place, &input); } col2im(*context, output_cfo, dilation, stride, padding, &input); @@ -128,7 +129,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); } else { - TensorCopy(input, paddle::platform::CPUPlace(), *context, &input_tmp); + TensorCopySync(input, paddle::platform::CPUPlace(), &input_tmp); in_ptr = input_tmp.data(); } for (int i = 0; i < 6; ++i) { @@ -140,7 +141,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - TensorCopy(input_tmp, *place, *context, &input); + TensorCopySync(input_tmp, *place, &input); } col2im_ocf(*context, output_ocf, dilation, stride, padding, &input); @@ -148,7 +149,7 @@ void testIm2col() { if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); } else { - TensorCopy(input, paddle::platform::CPUPlace(), *context, &input_tmp); + TensorCopySync(input, paddle::platform::CPUPlace(), &input_tmp); in_ptr = input_tmp.data(); } for (int i = 0; i < 6; ++i) { diff --git a/paddle/fluid/operators/math/math_function.cc b/paddle/fluid/operators/math/math_function.cc index 44fd739fb1..b5ae41c8f9 100644 --- a/paddle/fluid/operators/math/math_function.cc +++ b/paddle/fluid/operators/math/math_function.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/math_function.h" +#include #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/operators/math/math_function_impl.h" #include "paddle/fluid/platform/float16.h" @@ -161,7 +162,8 @@ 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) { + float16* C, const int batchCount, const int64_t strideA, + const int64_t strideB) { PADDLE_THROW("float16 batched_gemm not supported on CPU"); } @@ -172,7 +174,8 @@ 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 float alpha, const float* A, const float* B, const float beta, - float* C, const int batchCount, const int strideA, const int strideB) { + float* C, const int batchCount, const int64_t strideA, + const int64_t strideB) { int lda = (transA == CblasNoTrans) ? K : M; int ldb = (transB == CblasNoTrans) ? N : K; int ldc = N; @@ -194,7 +197,8 @@ 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 double alpha, const double* A, const double* B, const double beta, - double* C, const int batchCount, const int strideA, const int strideB) { + double* C, const int batchCount, const int64_t strideA, + const int64_t strideB) { int lda = (transA == CblasNoTrans) ? K : M; int ldb = (transB == CblasNoTrans) ? N : K; int ldc = N; @@ -220,7 +224,8 @@ 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 float alpha, const float* A, const float* B, const float beta, - float* C, const int batchCount, const int strideA, const int strideB) { + float* C, const int batchCount, const int64_t strideA, + const int64_t strideB) { for (int k = 0; k < batchCount; ++k) { const float* Ak = &A[k * strideA]; const float* Bk = &B[k * strideB]; @@ -235,7 +240,8 @@ 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 double alpha, const double* A, const double* B, const double beta, - double* C, const int batchCount, const int strideA, const int strideB) { + double* C, const int batchCount, const int64_t strideA, + const int64_t strideB) { for (int k = 0; k < batchCount; ++k) { const double* Ak = &A[k * strideA]; const double* Bk = &B[k * strideB]; diff --git a/paddle/fluid/operators/math/math_function.cu b/paddle/fluid/operators/math/math_function.cu index 9badf26c9b..2aa819625e 100644 --- a/paddle/fluid/operators/math/math_function.cu +++ b/paddle/fluid/operators/math/math_function.cu @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #define EIGEN_USE_GPU +#include #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/math_function_impl.h" @@ -267,7 +268,8 @@ 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) { + float16* C, const int batchCount, const int64_t strideA, + const int64_t strideB) { #if CUDA_VERSION >= 8000 // Note that cublas follows fortran order, so the order is different from // the cblas convention. @@ -278,7 +280,7 @@ void batched_gemm( (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 int64_t strideC = M * N; const half h_alpha = static_cast(alpha); const half h_beta = static_cast(beta); @@ -303,7 +305,8 @@ 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 float alpha, const float* A, const float* B, const float beta, - float* C, const int batchCount, const int strideA, const int strideB) { + float* C, const int batchCount, const int64_t strideA, + const int64_t strideB) { #if CUDA_VERSION >= 8000 // Note that cublas follows fortran order, so the order is different from // the cblas convention. @@ -314,7 +317,7 @@ void batched_gemm( (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 int64_t strideC = M * N; PADDLE_ENFORCE(platform::dynload::cublasSgemmStridedBatched( context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, @@ -329,7 +332,8 @@ 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 double alpha, const double* A, const double* B, const double beta, - double* C, const int batchCount, const int strideA, const int strideB) { + double* C, const int batchCount, const int64_t strideA, + const int64_t strideB) { #if CUDA_VERSION >= 8000 // Note that cublas follows fortran order, so the order is different from // the cblas convention. @@ -340,7 +344,7 @@ void batched_gemm( (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 int64_t strideC = M * N; PADDLE_ENFORCE(platform::dynload::cublasDgemmStridedBatched( context.cublas_handle(), cuTransB, cuTransA, N, M, K, &alpha, B, ldb, diff --git a/paddle/fluid/operators/math/math_function.h b/paddle/fluid/operators/math/math_function.h index cdbc7bfb37..cdd0297472 100644 --- a/paddle/fluid/operators/math/math_function.h +++ b/paddle/fluid/operators/math/math_function.h @@ -26,7 +26,7 @@ limitations under the License. */ #ifndef LAPACK_FOUND extern "C" { -#include +#include // NOLINT int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda, int* ipiv); int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda, @@ -39,6 +39,7 @@ int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda, #endif #include +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/tensor.h" @@ -78,8 +79,8 @@ template void batched_gemm(const DeviceContext& context, const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB, const int M, const int N, const int K, const T alpha, const T* A, const T* B, - const T beta, T* C, const int batchCount, const int strideA, - const int strideB); + const T beta, T* C, const int batchCount, + const int64_t strideA, const int64_t strideB); template void gemv(const DeviceContext& context, const bool trans_a, const int M, diff --git a/paddle/fluid/operators/math/math_function_impl.h b/paddle/fluid/operators/math/math_function_impl.h index f9d4e45324..b9bd49d77d 100644 --- a/paddle/fluid/operators/math/math_function_impl.h +++ b/paddle/fluid/operators/math/math_function_impl.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/operators/math/math_function.h" diff --git a/paddle/fluid/operators/math/math_function_test.cu b/paddle/fluid/operators/math/math_function_test.cu index 8982d9d066..7986326e96 100644 --- a/paddle/fluid/operators/math/math_function_test.cu +++ b/paddle/fluid/operators/math/math_function_test.cu @@ -40,15 +40,15 @@ TEST(math_function, notrans_mul_trans_fp32) { float arr[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr, 6 * sizeof(float)); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input1, gpu_place, context, &input2_gpu); + TensorCopySync(input1, gpu_place, &input1_gpu); + TensorCopySync(input1, gpu_place, &input2_gpu); out_gpu.mutable_data({2, 2}, gpu_place); paddle::operators::math::matmul( context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0); - TensorCopy(out_gpu, cpu_place, context, &out); + TensorCopySync(out_gpu, cpu_place, &out); float* out_ptr = out.data(); context.Wait(); @@ -80,8 +80,8 @@ TEST(math_function, notrans_mul_trans_fp16) { 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); + TensorCopySync(input1, gpu_place, &input1_gpu); + TensorCopySync(input1, gpu_place, &input2_gpu); out_gpu.mutable_data({2, 2}, gpu_place); @@ -89,7 +89,7 @@ TEST(math_function, notrans_mul_trans_fp16) { context, input1_gpu, false, input2_gpu, true, float16(1), &out_gpu, float16(0)); - TensorCopy(out_gpu, cpu_place, context, &out); + TensorCopySync(out_gpu, cpu_place, &out); float16* out_ptr = out.data(); context.Wait(); @@ -117,15 +117,15 @@ TEST(math_function, trans_mul_notrans_fp32) { float arr[6] = {0, 1, 2, 3, 4, 5}; memcpy(input1_ptr, arr, 6 * sizeof(float)); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input1, gpu_place, context, &input2_gpu); + TensorCopySync(input1, gpu_place, &input1_gpu); + TensorCopySync(input1, gpu_place, &input2_gpu); out_gpu.mutable_data({3, 3}, gpu_place); paddle::operators::math::matmul( context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0); - TensorCopy(out_gpu, cpu_place, context, &out); + TensorCopySync(out_gpu, cpu_place, &out); float* out_ptr = out.data(); context.Wait(); @@ -162,8 +162,8 @@ TEST(math_function, trans_mul_notrans_fp16) { 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); + TensorCopySync(input1, gpu_place, &input1_gpu); + TensorCopySync(input1, gpu_place, &input2_gpu); out_gpu.mutable_data({3, 3}, gpu_place); @@ -171,7 +171,7 @@ TEST(math_function, trans_mul_notrans_fp16) { context, input1_gpu, true, input2_gpu, false, float16(1), &out_gpu, float16(0)); - TensorCopy(out_gpu, cpu_place, context, &out); + TensorCopySync(out_gpu, cpu_place, &out); float16* out_ptr = out.data(); context.Wait(); @@ -214,9 +214,9 @@ TEST(math_function, gemm_notrans_cublas_fp32) { float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; memcpy(input3_ptr, arr3, 8 * sizeof(float)); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input2, gpu_place, context, &input2_gpu); - TensorCopy(input3, gpu_place, context, &input3_gpu); + TensorCopySync(input1, gpu_place, &input1_gpu); + TensorCopySync(input2, gpu_place, &input2_gpu); + TensorCopySync(input3, gpu_place, &input3_gpu); float* a = input1_gpu.data(); float* b = input2_gpu.data(); float* c = input3_gpu.mutable_data(gpu_place); @@ -224,7 +224,7 @@ TEST(math_function, gemm_notrans_cublas_fp32) { paddle::operators::math::gemm( context, false, false, m, n, k, 1, a, 3, b + 1, 4, 1, c + 1, 4); - TensorCopy(input3_gpu, cpu_place, context, &input3); + TensorCopySync(input3_gpu, cpu_place, &input3); // numpy code: // a = np.arange(6).reshape(2, 3) @@ -274,9 +274,9 @@ TEST(math_function, gemm_notrans_cublas_fp16) { 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); + TensorCopySync(input1, gpu_place, &input1_gpu); + TensorCopySync(input2, gpu_place, &input2_gpu); + TensorCopySync(input3, gpu_place, &input3_gpu); float16* a = input1_gpu.data(); float16* b = input2_gpu.data(); float16* c = input3_gpu.mutable_data(gpu_place); @@ -285,7 +285,7 @@ TEST(math_function, gemm_notrans_cublas_fp16) { 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); + TensorCopySync(input3_gpu, cpu_place, &input3); // numpy code: // a = np.arange(6).reshape(2, 3) @@ -332,9 +332,9 @@ TEST(math_function, gemm_trans_cublas_fp32) { float arr3[8] = {0, 1, 2, 3, 4, 5, 6, 7}; memcpy(input3_ptr, arr3, 8 * sizeof(float)); - TensorCopy(input1, gpu_place, context, &input1_gpu); - TensorCopy(input2, gpu_place, context, &input2_gpu); - TensorCopy(input3, gpu_place, context, &input3_gpu); + TensorCopySync(input1, gpu_place, &input1_gpu); + TensorCopySync(input2, gpu_place, &input2_gpu); + TensorCopySync(input3, gpu_place, &input3_gpu); float* a = input1_gpu.data(); float* b = input2_gpu.data(); float* c = input3_gpu.mutable_data(gpu_place); @@ -342,7 +342,7 @@ TEST(math_function, gemm_trans_cublas_fp32) { paddle::operators::math::gemm( context, false, true, m, n, k, 1, a, 3, b + 3, 3, 1, c + 1, 4); - TensorCopy(input3_gpu, cpu_place, context, &input3); + TensorCopySync(input3_gpu, cpu_place, &input3); context.Wait(); EXPECT_EQ(input3_ptr[0], 0); @@ -386,9 +386,9 @@ TEST(math_function, gemm_trans_cublas_fp16) { 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); + TensorCopySync(input1, gpu_place, &input1_gpu); + TensorCopySync(input2, gpu_place, &input2_gpu); + TensorCopySync(input3, gpu_place, &input3_gpu); float16* a = input1_gpu.data(); float16* b = input2_gpu.data(); float16* c = input3_gpu.mutable_data(gpu_place); @@ -397,7 +397,7 @@ TEST(math_function, gemm_trans_cublas_fp16) { 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); + TensorCopySync(input3_gpu, cpu_place, &input3); context.Wait(); EXPECT_EQ(static_cast(input3_ptr[0]), 0); @@ -441,14 +441,14 @@ void GemvTest(int m, int n, bool trans) { data_b[i] = static_cast(i); } - TensorCopy(mat_a, gpu_place, context, &g_mat_a); - TensorCopy(vec_b, gpu_place, context, &g_vec_b); + TensorCopySync(mat_a, gpu_place, &g_mat_a); + TensorCopySync(vec_b, gpu_place, &g_vec_b); paddle::operators::math::gemv( context, trans, static_cast(m), static_cast(n), 1., g_data_a, g_data_b, 0., g_data_c); - TensorCopy(g_vec_c, cpu_place, context, &vec_c); + TensorCopySync(g_vec_c, cpu_place, &vec_c); if (!trans) { for (int i = 0; i < m; ++i) { diff --git a/paddle/fluid/operators/math/matmul.h b/paddle/fluid/operators/math/matmul.h index 6e2d35cd0f..0006c5062f 100644 --- a/paddle/fluid/operators/math/matmul.h +++ b/paddle/fluid/operators/math/matmul.h @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include +#include #include "paddle/fluid/operators/math/math_function.h" namespace paddle { diff --git a/paddle/fluid/operators/math/sampler.cc b/paddle/fluid/operators/math/sampler.cc index 3ec6538d7f..3066dc0ba2 100644 --- a/paddle/fluid/operators/math/sampler.cc +++ b/paddle/fluid/operators/math/sampler.cc @@ -12,7 +12,7 @@ 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 "sampler.h" +#include "paddle/fluid/operators/math/sampler.h" namespace paddle { namespace random { diff --git a/paddle/fluid/operators/math/sampler.h b/paddle/fluid/operators/math/sampler.h index 9d6a6c28c4..b82691f269 100644 --- a/paddle/fluid/operators/math/sampler.h +++ b/paddle/fluid/operators/math/sampler.h @@ -13,9 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include #include -typedef long int64; namespace paddle { namespace operators { namespace math { @@ -27,25 +27,25 @@ namespace math { */ class Sampler { public: - explicit Sampler(int64 range) : range_(range) { + explicit Sampler(int64_t range) : range_(range) { PADDLE_ENFORCE_GT(range, 0); std::random_device r; seed_ = r(); } - explicit Sampler(int64 range, unsigned int seed) + explicit Sampler(int64_t range, unsigned int seed) : range_(range), seed_(seed) { PADDLE_ENFORCE_GT(range, 0); } virtual ~Sampler(); // Sample a single value - virtual int64 Sample() const = 0; + virtual int64_t Sample() const = 0; // The probability that a single call to Sample() returns the given value. - virtual float Probability(int64 value) const = 0; + virtual float Probability(int64_t value) const = 0; - int64 range() { return range_; }; + int64 range() { return range_; } protected: - const int64 range_; + const int64_t range_; unsigned int seed_; }; @@ -56,15 +56,15 @@ class Sampler { */ class UniformSampler : public Sampler { public: - explicit UniformSampler(int64 range); + explicit UniformSampler(int64_t range); - explicit UniformSampler(int64 range, unsigned int seed); + explicit UniformSampler(int64_t range, unsigned int seed); ~UniformSampler() override {} int64 Sample() const override; - float Probability(int64 value) const override; + float Probability(int64_t value) const override; private: const float inv_range_; @@ -79,15 +79,15 @@ class UniformSampler : public Sampler { */ class LogUniformSampler : public Sampler { public: - explicit LogUniformSampler(int64 range); + explicit LogUniformSampler(int64_t range); - explicit LogUniformSampler(int64 range, unsigned int seed); + explicit LogUniformSampler(int64_t range, unsigned int seed); ~LogUniformSampler() override {} int64 Sample() const override; - float Probability(int64 value) const override; + float Probability(int64_t value) const override; private: const float log_range_; @@ -95,6 +95,6 @@ class LogUniformSampler : public Sampler { std::shared_ptr> dist_; }; -} // math +} // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/math/selected_rows_functor.cc b/paddle/fluid/operators/math/selected_rows_functor.cc index 5da3d15277..a830dc5250 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cc +++ b/paddle/fluid/operators/math/selected_rows_functor.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include +#include #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" diff --git a/paddle/fluid/operators/math/selected_rows_functor.cu b/paddle/fluid/operators/math/selected_rows_functor.cu index 5d78fd9d21..7b31ee8e38 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cu +++ b/paddle/fluid/operators/math/selected_rows_functor.cu @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include +#include #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" diff --git a/paddle/fluid/operators/math/selected_rows_functor_test.cc b/paddle/fluid/operators/math/selected_rows_functor_test.cc index 679b6568ad..70bed820ee 100644 --- a/paddle/fluid/operators/math/selected_rows_functor_test.cc +++ b/paddle/fluid/operators/math/selected_rows_functor_test.cc @@ -13,41 +13,50 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/selected_rows_functor.h" +#include #include "gtest/gtest.h" #include "paddle/fluid/operators/math/math_function.h" TEST(selected_rows_functor, cpu_add) { - using namespace paddle::framework; - using namespace paddle::platform; - using namespace paddle::operators::math; - - CPUPlace cpu_place; - CPUDeviceContext ctx(cpu_place); - SetConstant functor; + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + functor; int64_t height = 10; int64_t row_numel = 10; std::vector rows1{0, 4, 7}; - std::unique_ptr selected_rows1{new SelectedRows(rows1, height)}; + std::unique_ptr selected_rows1{ + new paddle::framework::SelectedRows(rows1, height)}; auto* in1_value = selected_rows1->mutable_value(); in1_value->mutable_data( - make_ddim({static_cast(rows1.size()), row_numel}), cpu_place); + paddle::framework::make_ddim( + {static_cast(rows1.size()), row_numel}), + cpu_place); functor(ctx, in1_value, 1.0); std::vector rows2{0, 5, 7, 9}; - std::unique_ptr selected_rows2{new SelectedRows(rows2, height)}; + std::unique_ptr selected_rows2{ + new paddle::framework::SelectedRows(rows2, height)}; auto* in2_value = selected_rows2->mutable_value(); in2_value->mutable_data( - make_ddim({static_cast(rows2.size()), row_numel}), cpu_place); + paddle::framework::make_ddim( + {static_cast(rows2.size()), row_numel}), + cpu_place); functor(ctx, in2_value, 2.0); - std::unique_ptr output{new SelectedRows()}; + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; auto* out_value = output->mutable_value(); // simplely concat two SelectedRows - out_value->mutable_data(make_ddim({7, 10}), cpu_place); + out_value->mutable_data(paddle::framework::make_ddim({7, 10}), + cpu_place); - SelectedRowsAdd add_functor; + paddle::operators::math::SelectedRowsAdd + add_functor; add_functor(ctx, *selected_rows1, *selected_rows2, output.get()); auto out_height = output->height(); @@ -78,14 +87,20 @@ TEST(selected_rows_functor, cpu_add) { EXPECT_EQ(out_data[5 * row_numel + 7], 2.0); EXPECT_EQ(out_data[6 * row_numel + 9], 2.0); - std::unique_ptr tensor1{new Tensor()}; - tensor1->mutable_data(make_ddim({height, row_numel}), cpu_place); + std::unique_ptr tensor1{ + new paddle::framework::Tensor()}; + tensor1->mutable_data( + paddle::framework::make_ddim({height, row_numel}), cpu_place); functor(ctx, tensor1.get(), 3.0); - std::unique_ptr tensor2{new Tensor()}; - tensor2->mutable_data(make_ddim({height, row_numel}), cpu_place); + std::unique_ptr tensor2{ + new paddle::framework::Tensor()}; + tensor2->mutable_data( + paddle::framework::make_ddim({height, row_numel}), cpu_place); - SelectedRowsAddTensor add_tensor_functor; + paddle::operators::math::SelectedRowsAddTensor< + paddle::platform::CPUDeviceContext, float> + add_tensor_functor; add_tensor_functor(ctx, *output, *tensor1, tensor2.get()); auto* tensor2_data = tensor2->data(); @@ -106,38 +121,46 @@ TEST(selected_rows_functor, cpu_add) { } TEST(selected_rows_functor, cpu_add_to) { - using namespace paddle::framework; - using namespace paddle::platform; - using namespace paddle::operators::math; - - CPUPlace cpu_place; - CPUDeviceContext ctx(cpu_place); - SetConstant functor; + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + functor; int64_t height = 10; int64_t row_numel = 10; std::vector rows1{0, 4, 7}; - std::unique_ptr selected_rows1{new SelectedRows(rows1, height)}; + std::unique_ptr selected_rows1{ + new paddle::framework::SelectedRows(rows1, height)}; auto* in1_value = selected_rows1->mutable_value(); in1_value->mutable_data( - make_ddim({static_cast(rows1.size()), row_numel}), cpu_place); + paddle::framework::make_ddim( + {static_cast(rows1.size()), row_numel}), + cpu_place); functor(ctx, in1_value, 1.0); std::vector rows2{0, 5, 7, 9}; - std::unique_ptr selected_rows2{new SelectedRows(rows2, height)}; + std::unique_ptr selected_rows2{ + new paddle::framework::SelectedRows(rows2, height)}; auto* in2_value = selected_rows2->mutable_value(); in2_value->mutable_data( - make_ddim({static_cast(rows2.size()), row_numel}), cpu_place); + paddle::framework::make_ddim( + {static_cast(rows2.size()), row_numel}), + cpu_place); functor(ctx, in2_value, 2.0); - std::unique_ptr output{new SelectedRows()}; + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; output->set_height(height); auto* out_value = output->mutable_value(); // simplely concat two SelectedRows - out_value->mutable_data(make_ddim({7, 10}), cpu_place); + out_value->mutable_data(paddle::framework::make_ddim({7, 10}), + cpu_place); - SelectedRowsAddTo add_to_functor; + paddle::operators::math::SelectedRowsAddTo + add_to_functor; add_to_functor(ctx, *selected_rows1, 0, output.get()); add_to_functor(ctx, *selected_rows2, in1_value->numel(), output.get()); @@ -169,11 +192,15 @@ TEST(selected_rows_functor, cpu_add_to) { EXPECT_EQ(out_data[5 * row_numel + 7], 2.0); EXPECT_EQ(out_data[6 * row_numel + 9], 2.0); - std::unique_ptr tensor1{new Tensor()}; - tensor1->mutable_data(make_ddim({height, row_numel}), cpu_place); + std::unique_ptr tensor1{ + new paddle::framework::Tensor()}; + tensor1->mutable_data( + paddle::framework::make_ddim({height, row_numel}), cpu_place); functor(ctx, tensor1.get(), 3.0); - SelectedRowsAddToTensor add_to_tensor_functor; + paddle::operators::math::SelectedRowsAddToTensor< + paddle::platform::CPUDeviceContext, float> + add_to_tensor_functor; add_to_tensor_functor(ctx, *output, tensor1.get()); auto* tensor1_data = tensor1->data(); diff --git a/paddle/fluid/operators/math/sequence2batch.cc b/paddle/fluid/operators/math/sequence2batch.cc index 8899abff36..b546b87282 100644 --- a/paddle/fluid/operators/math/sequence2batch.cc +++ b/paddle/fluid/operators/math/sequence2batch.cc @@ -23,11 +23,11 @@ class CopyMatrixRowsFunctor { public: void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& src, - framework::Vector index_lod, framework::Tensor& dst, + framework::Vector index_lod, framework::Tensor* dst, bool is_src_index) { size_t* index = index_lod.data(); auto src_dims = src.dims(); - auto dst_dims = dst.dims(); + auto dst_dims = dst->dims(); PADDLE_ENFORCE_EQ(src_dims.size(), 2UL, "The src must be matrix with rank 2."); PADDLE_ENFORCE_EQ(dst_dims.size(), 2UL, @@ -37,7 +37,7 @@ class CopyMatrixRowsFunctor { auto height = dst_dims[0]; auto width = dst_dims[1]; auto* src_data = src.data(); - auto* dst_data = dst.data(); + auto* dst_data = dst->data(); for (int i = 0; i < height; ++i) { if (is_src_index) { memcpy(dst_data + i * width, src_data + index[i] * width, diff --git a/paddle/fluid/operators/math/sequence2batch.cu b/paddle/fluid/operators/math/sequence2batch.cu index 3185f10d41..be73adfc0c 100644 --- a/paddle/fluid/operators/math/sequence2batch.cu +++ b/paddle/fluid/operators/math/sequence2batch.cu @@ -43,10 +43,10 @@ class CopyMatrixRowsFunctor { public: void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& src, - framework::Vector index_lod, framework::Tensor& dst, + framework::Vector index_lod, framework::Tensor* dst, bool is_src_index) { auto src_dims = src.dims(); - auto dst_dims = dst.dims(); + auto dst_dims = dst->dims(); PADDLE_ENFORCE_EQ(src_dims.size(), 2, "The src must be matrix with rank 2."); PADDLE_ENFORCE_EQ(dst_dims.size(), 2, @@ -56,7 +56,7 @@ class CopyMatrixRowsFunctor { auto height = dst_dims[0]; auto width = dst_dims[1]; auto* src_data = src.data(); - auto* dst_data = dst.data(); + auto* dst_data = dst->data(); dim3 threads(128, 8); dim3 grid(8, 1); diff --git a/paddle/fluid/operators/math/sequence2batch.h b/paddle/fluid/operators/math/sequence2batch.h index e78aafd37d..0abda999a5 100644 --- a/paddle/fluid/operators/math/sequence2batch.h +++ b/paddle/fluid/operators/math/sequence2batch.h @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/tensor.h" @@ -35,7 +37,7 @@ class CopyMatrixRowsFunctor { // copy the input src to the indexed rows of output dst. // The indexed rows are based on the input index. void operator()(const DeviceContext& context, const framework::Tensor& src, - framework::Vector index_lod, framework::Tensor& dst, + framework::Vector index_lod, framework::Tensor* dst, bool is_src_index); }; @@ -58,10 +60,10 @@ class LoDTensor2BatchFunctor { public: void operator()(const DeviceContext& context, const framework::LoDTensor& lod_tensor, - framework::LoDTensor& batch, bool is_cal_batch_lod, + framework::LoDTensor* batch, bool is_cal_batch_lod, bool is_reverse = false) const { if (!is_cal_batch_lod) { - auto lods = batch.lod(); + auto lods = batch->lod(); PADDLE_ENFORCE_GT(lods.size(), 2UL); PADDLE_ENFORCE_EQ(lods[1].size(), static_cast(lod_tensor.dims()[0])); @@ -141,7 +143,7 @@ class LoDTensor2BatchFunctor { for (size_t i = 0; i < seq_info.size(); ++i) { seq_order[i] = seq_info[i].seq_idx; } - batch.set_lod(batch_lods); + batch->set_lod(batch_lods); CopyMatrixRowsFunctor to_batch; to_batch(context, lod_tensor, batch_lods[1], batch, true); @@ -153,11 +155,11 @@ class Batch2LoDTensorFunctor { public: void operator()(const DeviceContext& context, const framework::LoDTensor& batch, - framework::LoDTensor& lod_tensor) const { + framework::LoDTensor* lod_tensor) const { auto in_lod = batch.lod(); PADDLE_ENFORCE_GT(in_lod.size(), 2UL); PADDLE_ENFORCE_EQ(in_lod[1].size(), - static_cast(lod_tensor.dims()[0])); + static_cast(lod_tensor->dims()[0])); CopyMatrixRowsFunctor to_seq; to_seq(context, batch, in_lod[1], lod_tensor, false); } diff --git a/paddle/fluid/operators/math/sequence_padding_test.cc b/paddle/fluid/operators/math/sequence_padding_test.cc index bece46e753..e3d6214485 100644 --- a/paddle/fluid/operators/math/sequence_padding_test.cc +++ b/paddle/fluid/operators/math/sequence_padding_test.cc @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/sequence_padding.h" #include +#include template void TestSequencePadding(const paddle::framework::LoD& lod, @@ -75,7 +76,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod, delete place; delete context; -}; +} TEST(Seq2BatchPadding, CPU) { paddle::framework::LoD lod1; diff --git a/paddle/fluid/operators/math/sequence_pooling.cc b/paddle/fluid/operators/math/sequence_pooling.cc index 5ae42ab973..f25d3d3f1e 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cc +++ b/paddle/fluid/operators/math/sequence_pooling.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/sequence_pooling.h" +#include #include "paddle/fluid/operators/math/math_function.h" namespace paddle { diff --git a/paddle/fluid/operators/math/sequence_pooling.cu b/paddle/fluid/operators/math/sequence_pooling.cu index 1935364da3..36f6402396 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cu +++ b/paddle/fluid/operators/math/sequence_pooling.cu @@ -12,6 +12,7 @@ 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/operators/math/math_function.h" #include "paddle/fluid/operators/math/sequence_pooling.h" #include "paddle/fluid/platform/cuda_helper.h" diff --git a/paddle/fluid/operators/math/sequence_pooling.h b/paddle/fluid/operators/math/sequence_pooling.h index 38e7802229..8dcbee65d0 100644 --- a/paddle/fluid/operators/math/sequence_pooling.h +++ b/paddle/fluid/operators/math/sequence_pooling.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/device_context.h" diff --git a/paddle/fluid/operators/math/sequence_scale.cc b/paddle/fluid/operators/math/sequence_scale.cc index 2c46d4183b..ee5b22ca85 100644 --- a/paddle/fluid/operators/math/sequence_scale.cc +++ b/paddle/fluid/operators/math/sequence_scale.cc @@ -21,15 +21,15 @@ namespace math { template class ScaleLoDTensorFunctor { public: - void operator()(const platform::CPUDeviceContext& context, - framework::LoDTensor& seq, const T* scales) { + void operator()(const platform::CPUDeviceContext& context, const T* scales, + framework::LoDTensor* seq) { const size_t level = 0; - auto lod = seq.lod(); + auto lod = seq->lod(); const size_t num_seq = lod[level].size() - 1; - size_t seq_width = seq.dims()[1]; + size_t seq_width = seq->dims()[1]; framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - T* seq_data = seq.mutable_data(context.GetPlace()); + T* seq_data = seq->mutable_data(context.GetPlace()); for (size_t i = 0; i < num_seq; ++i) { for (size_t j = lod[level][i] * seq_width; j < lod[level][i + 1] * seq_width; ++j) { diff --git a/paddle/fluid/operators/math/sequence_scale.cu b/paddle/fluid/operators/math/sequence_scale.cu index 74085153c6..430bf13c3f 100644 --- a/paddle/fluid/operators/math/sequence_scale.cu +++ b/paddle/fluid/operators/math/sequence_scale.cu @@ -35,14 +35,14 @@ __global__ void SequenceScaleKernel(T* seq, size_t* lod, const T* scales, template class ScaleLoDTensorFunctor { public: - void operator()(const platform::CUDADeviceContext& context, - framework::LoDTensor& seq, const T* scales) { + void operator()(const platform::CUDADeviceContext& context, const T* scales, + framework::LoDTensor* seq) { const size_t level = 0; - auto lod = seq.lod(); + auto lod = seq->lod(); const size_t num_seq = lod[level].size() - 1; - const size_t seq_width = seq.numel() / seq.dims()[0]; + const size_t seq_width = seq->numel() / seq->dims()[0]; framework::LoD abs_offset_lod = framework::ToAbsOffset(lod); - T* seq_data = seq.mutable_data(context.GetPlace()); + T* seq_data = seq->mutable_data(context.GetPlace()); SequenceScaleKernel<<< num_seq, PADDLE_CUDA_NUM_THREADS, 0, context.stream()>>>( diff --git a/paddle/fluid/operators/math/sequence_scale.h b/paddle/fluid/operators/math/sequence_scale.h index 6cdcbe21cb..202243985c 100644 --- a/paddle/fluid/operators/math/sequence_scale.h +++ b/paddle/fluid/operators/math/sequence_scale.h @@ -46,8 +46,8 @@ namespace math { template class ScaleLoDTensorFunctor { public: - void operator()(const DeviceContext& context, framework::LoDTensor& seq, - const T* scales); + void operator()(const DeviceContext& context, const T* scales, + framework::LoDTensor* seq); }; } // namespace math diff --git a/paddle/fluid/operators/math/vol2col.cc b/paddle/fluid/operators/math/vol2col.cc index 09e9f85cca..e92adc09ba 100644 --- a/paddle/fluid/operators/math/vol2col.cc +++ b/paddle/fluid/operators/math/vol2col.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/math/vol2col.h" +#include namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/math/vol2col.cu b/paddle/fluid/operators/math/vol2col.cu index 619730d394..e0f3ef3687 100644 --- a/paddle/fluid/operators/math/vol2col.cu +++ b/paddle/fluid/operators/math/vol2col.cu @@ -12,6 +12,8 @@ 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 "paddle/fluid/operators/math/vol2col.h" #include "paddle/fluid/platform/cuda_helper.h" diff --git a/paddle/fluid/operators/math/vol2col.h b/paddle/fluid/operators/math/vol2col.h index dbc2ed7a69..5f59de8f02 100644 --- a/paddle/fluid/operators/math/vol2col.h +++ b/paddle/fluid/operators/math/vol2col.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/platform/device_context.h" diff --git a/paddle/fluid/operators/math/vol2col_test.cc b/paddle/fluid/operators/math/vol2col_test.cc index eb91f862e3..aa979c4f10 100644 --- a/paddle/fluid/operators/math/vol2col_test.cc +++ b/paddle/fluid/operators/math/vol2col_test.cc @@ -15,6 +15,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/vol2col.h" #include #include +#include template void testVol2col() { @@ -71,7 +72,7 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - paddle::framework::TensorCopy(input_tmp, *place, *context, &input); + paddle::framework::TensorCopySync(input_tmp, *place, &input); } output.mutable_data({1, filter_size, filter_size, filter_size, output_depth, output_height, output_width}, @@ -85,7 +86,7 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { out_cfo_ptr = output.data(); } else { - TensorCopy(output, paddle::platform::CPUPlace(), *context, &output_tmp); + TensorCopySync(output, paddle::platform::CPUPlace(), &output_tmp); out_cfo_ptr = output_tmp.data(); } @@ -99,7 +100,7 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { input = input_tmp; } else { - TensorCopy(input_tmp, *place, *context, &input); + TensorCopySync(input_tmp, *place, &input); } paddle::operators::math::Col2VolFunctor col2vol; @@ -109,7 +110,7 @@ void testVol2col() { if (paddle::platform::is_cpu_place(*place)) { in_ptr = input.data(); } else { - TensorCopy(input, paddle::platform::CPUPlace(), *context, &input_tmp); + TensorCopySync(input, paddle::platform::CPUPlace(), &input_tmp); in_ptr = input_tmp.data(); } diff --git a/paddle/fluid/operators/nccl/nccl_gpu_common.h b/paddle/fluid/operators/nccl/nccl_gpu_common.h index 113f93e346..558ff4cc09 100644 --- a/paddle/fluid/operators/nccl/nccl_gpu_common.h +++ b/paddle/fluid/operators/nccl/nccl_gpu_common.h @@ -15,9 +15,9 @@ limitations under the License. */ #pragma once #include -#include +#include // NOLINT #include -#include +#include // NOLINT #include #include #include diff --git a/paddle/fluid/operators/nccl_op_test.cu.cc b/paddle/fluid/operators/nccl_op_test.cu.cc index 20b8a5c98a..ef54d79fdf 100644 --- a/paddle/fluid/operators/nccl_op_test.cu.cc +++ b/paddle/fluid/operators/nccl_op_test.cu.cc @@ -228,10 +228,8 @@ TEST_F(NCCLTester, ncclReduceOp) { result_tensor->Resize(kDims); auto *ct = result_tensor->mutable_data(cpu_place); - paddle::memory::Copy( - cpu_place, ct, p::CUDAPlace(gpu_list_[kRoot]), rt, - recv_tensor.numel() * sizeof(float), - static_cast(dev_ctxs_[kRoot])->stream()); + paddle::memory::Copy(cpu_place, ct, p::CUDAPlace(gpu_list_[kRoot]), rt, + recv_tensor.numel() * sizeof(float), nullptr); for (int64_t j = 0; j < f::product(kDims); ++j) { ASSERT_NEAR(ct[j], expected_result, 1e-5); diff --git a/paddle/fluid/operators/parallel_do_op.cc b/paddle/fluid/operators/parallel_do_op.cc index b28c16b13f..ae34fe2184 100644 --- a/paddle/fluid/operators/parallel_do_op.cc +++ b/paddle/fluid/operators/parallel_do_op.cc @@ -364,7 +364,7 @@ class ParallelDoGradOpDescMaker : public framework::SingleGradOpDescMaker { } } grad->SetAttrMap(this->Attrs()); - grad->SetBlockAttr(kParallelBlock, *grad_block_[0]); + grad->SetBlockAttr(kParallelBlock, grad_block_[0]); return std::unique_ptr(grad); } diff --git a/paddle/fluid/operators/reader/CMakeLists.txt b/paddle/fluid/operators/reader/CMakeLists.txt index 845528860f..3106978eb0 100644 --- a/paddle/fluid/operators/reader/CMakeLists.txt +++ b/paddle/fluid/operators/reader/CMakeLists.txt @@ -23,5 +23,7 @@ reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_o reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc) reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc) reader_library(create_threaded_reader_op SRCS create_threaded_reader_op.cc) + +cc_test(reader_blocking_queue_test SRCS reader_blocking_queue_test.cc) # Export local libraries to parent set(READER_LIBRARY ${LOCAL_READER_LIBS} PARENT_SCOPE) diff --git a/paddle/fluid/operators/reader/blocking_queue.h b/paddle/fluid/operators/reader/blocking_queue.h new file mode 100644 index 0000000000..71684b1417 --- /dev/null +++ b/paddle/fluid/operators/reader/blocking_queue.h @@ -0,0 +1,112 @@ +// 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 // NOLINT +#include + +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace reader { + +template +class BlockingQueue { + // BlockingQueue is for buffered reading and is supposed to use only the + // reader package. It is true that we could and we should have been using + // framework::Channel, but which has currently a deadlock bug. BlockingQueue + // is a workaround and a simplified version of framework::Channel as it + // doesn't support GPU and it implements on buffered blocking queue. + public: + explicit BlockingQueue(size_t capacity) + : capacity_(capacity), closed_(false) { + PADDLE_ENFORCE_GT( + capacity_, 0, + "The capacity of a reader::BlockingQueue must be greater than 0."); + } + + bool Send(const T& elem) { + std::unique_lock lock(mutex_); + send_cv_.wait(lock, [&] { return queue_.size() < capacity_ || closed_; }); + if (closed_) { + VLOG(5) + << "WARNING: Sending an element to a closed reader::BlokcingQueue."; + return false; + } + PADDLE_ENFORCE_LT(queue_.size(), capacity_); + queue_.push_back(elem); + receive_cv_.notify_one(); + return true; + } + + bool Send(T&& elem) { + std::unique_lock lock(mutex_); + send_cv_.wait(lock, [&] { return queue_.size() < capacity_ || closed_; }); + if (closed_) { + VLOG(5) + << "WARNING: Sending an element to a closed reader::BlokcingQueue."; + return false; + } + PADDLE_ENFORCE_LT(queue_.size(), capacity_); + queue_.emplace_back(std::move(elem)); + receive_cv_.notify_one(); + return true; + } + + bool Receive(T* elem) { + std::unique_lock lock(mutex_); + receive_cv_.wait(lock, [&] { return !queue_.empty() || closed_; }); + if (!queue_.empty()) { + PADDLE_ENFORCE_NOT_NULL(elem); + *elem = queue_.front(); + queue_.pop_front(); + send_cv_.notify_one(); + return true; + } else { + PADDLE_ENFORCE(closed_); + return false; + } + } + + void Close() { + std::lock_guard lock(mutex_); + closed_ = true; + send_cv_.notify_all(); + receive_cv_.notify_all(); + } + + bool IsClosed() { + std::lock_guard lock(mutex_); + return closed_; + } + + size_t Cap() { + std::lock_guard lock(mutex_); + return capacity_; + } + + private: + size_t capacity_; + bool closed_; + std::deque queue_; + + std::mutex mutex_; + std::condition_variable receive_cv_; + std::condition_variable send_cv_; +}; +} // namespace reader +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc index 4372f23fc1..e5efac4615 100644 --- a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc +++ b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc @@ -14,7 +14,7 @@ #include // NOLINT -#include "paddle/fluid/framework/channel.h" +#include "paddle/fluid/operators/reader/blocking_queue.h" #include "paddle/fluid/operators/reader/reader_op_registry.h" namespace paddle { @@ -23,13 +23,13 @@ namespace reader { // 'Double buffer' means we shall maintain two batches of input data at the same // time. So the kCacheSize shoul be at least 2. -static constexpr size_t kCacheSize = 2; +static constexpr size_t kCacheSize = 3; // There will be two bacthes out of the channel during training: // 1. the one waiting to be sent to the channel // 2. the one just be received from the channel, which is also being used by // subsequent operators. // So the channel size should be kChacheSize - 2 -static constexpr size_t kChannelSize = 0; // kCacheSize - 2 +static constexpr size_t kChannelSize = 1; // kCacheSize - 2 class DoubleBufferReader : public framework::DecoratedReader { public: @@ -55,10 +55,8 @@ class DoubleBufferReader : public framework::DecoratedReader { ~DoubleBufferReader() { EndPrefetcher(); } private: - bool HasNext() const; - void StartPrefetcher() { - channel_ = framework::MakeChannel(kChannelSize); + channel_ = new reader::BlockingQueue(kChannelSize); prefetcher_ = std::thread([this] { PrefetchThreadFunc(); }); } @@ -74,7 +72,7 @@ class DoubleBufferReader : public framework::DecoratedReader { void PrefetchThreadFunc(); std::thread prefetcher_; - framework::Channel* channel_; + reader::BlockingQueue* channel_; platform::Place place_; std::vector> cpu_tensor_cache_; std::vector> gpu_tensor_cache_; @@ -139,17 +137,16 @@ class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase { }; void DoubleBufferReader::ReadNext(std::vector* out) { - out->clear(); - if (HasNext()) { - size_t cached_tensor_id; - channel_->Receive(&cached_tensor_id); + size_t cached_tensor_id; + if (channel_->Receive(&cached_tensor_id)) { if (platform::is_gpu_place(place_)) { *out = gpu_tensor_cache_[cached_tensor_id]; - ctxs_[cached_tensor_id]->Wait(); } else { // CPU place *out = cpu_tensor_cache_[cached_tensor_id]; } + } else { + out->clear(); } } @@ -159,12 +156,6 @@ void DoubleBufferReader::ReInit() { StartPrefetcher(); } -bool DoubleBufferReader::HasNext() const { - while (!channel_->IsClosed() && !channel_->CanReceive()) { - } - return channel_->CanReceive(); -} - void DoubleBufferReader::PrefetchThreadFunc() { VLOG(5) << "A new prefetch thread starts."; size_t cached_tensor_id = 0; @@ -177,18 +168,14 @@ void DoubleBufferReader::PrefetchThreadFunc() { } if (platform::is_gpu_place(place_)) { auto& gpu_batch = gpu_tensor_cache_[cached_tensor_id]; - auto* gpu_ctx = ctxs_[cached_tensor_id].get(); gpu_batch.resize(cpu_batch.size()); for (size_t i = 0; i < cpu_batch.size(); ++i) { - framework::TensorCopy(cpu_batch[i], place_, *gpu_ctx, &gpu_batch[i], - true); + // TODO(fengjiayi): Use asynchronous TensorCopy instead + framework::TensorCopySync(cpu_batch[i], place_, &gpu_batch[i]); gpu_batch[i].set_lod(cpu_batch[i].lod()); } } - try { - size_t tmp = cached_tensor_id; - channel_->Send(&tmp); - } catch (paddle::platform::EnforceNotMet e) { + if (!channel_->Send(cached_tensor_id)) { VLOG(5) << "WARNING: The double buffer channel has been closed. The " "prefetch thread will terminate."; break; diff --git a/paddle/fluid/operators/reader/open_files_op.cc b/paddle/fluid/operators/reader/open_files_op.cc index 779dc8a6a0..91ad7d5658 100644 --- a/paddle/fluid/operators/reader/open_files_op.cc +++ b/paddle/fluid/operators/reader/open_files_op.cc @@ -14,7 +14,7 @@ #include // NOLINT -#include "paddle/fluid/framework/channel.h" +#include "paddle/fluid/operators/reader/blocking_queue.h" #include "paddle/fluid/operators/reader/reader_op_registry.h" namespace paddle { @@ -37,7 +37,6 @@ class MultiFileReader : public framework::ReaderBase { ~MultiFileReader() { EndScheduler(); } private: - bool HasNext(); void StartNewScheduler(); void EndScheduler(); void ScheduleThreadFunc(); @@ -48,15 +47,14 @@ class MultiFileReader : public framework::ReaderBase { std::thread scheduler_; std::vector prefetchers_; size_t buffer_size_; - framework::Channel* waiting_file_idx_; - framework::Channel* available_thread_idx_; - framework::Channel>* buffer_; + reader::BlockingQueue* waiting_file_idx_; + reader::BlockingQueue* available_thread_idx_; + reader::BlockingQueue>* buffer_; }; void MultiFileReader::ReadNext(std::vector* out) { - out->clear(); - if (HasNext()) { - buffer_->Receive(out); + if (!buffer_->Receive(out)) { + out->clear(); } } @@ -65,25 +63,19 @@ void MultiFileReader::ReInit() { StartNewScheduler(); } -bool MultiFileReader::HasNext() { - while (!buffer_->IsClosed() && !buffer_->CanReceive()) { - } - return buffer_->CanReceive(); -} - void MultiFileReader::StartNewScheduler() { size_t thread_num = prefetchers_.size(); - waiting_file_idx_ = framework::MakeChannel(file_names_.size()); - available_thread_idx_ = framework::MakeChannel(thread_num); - buffer_ = - framework::MakeChannel>(buffer_size_); + waiting_file_idx_ = new reader::BlockingQueue(file_names_.size()); + available_thread_idx_ = new reader::BlockingQueue(thread_num); + buffer_ = new reader::BlockingQueue>( + buffer_size_); for (size_t i = 0; i < file_names_.size(); ++i) { - waiting_file_idx_->Send(&i); + waiting_file_idx_->Send(i); } waiting_file_idx_->Close(); for (size_t i = 0; i < thread_num; ++i) { - available_thread_idx_->Send(&i); + available_thread_idx_->Send(i); } scheduler_ = std::thread([this] { ScheduleThreadFunc(); }); @@ -149,7 +141,7 @@ void MultiFileReader::PrefetchThreadFunc(std::string file_name, break; } try { - buffer_->Send(&ins); + buffer_->Send(std::move(ins)); } catch (paddle::platform::EnforceNotMet e) { VLOG(5) << "WARNING: The buffer channel has been closed. The prefetch " "thread of file '" @@ -158,9 +150,7 @@ void MultiFileReader::PrefetchThreadFunc(std::string file_name, } } - try { - available_thread_idx_->Send(&thread_idx); - } catch (paddle::platform::EnforceNotMet e) { + if (!available_thread_idx_->Send(thread_idx)) { VLOG(5) << "WARNING: The available_thread_idx_ channel has been closed. " "Fail to send thread_idx."; } diff --git a/paddle/fluid/operators/reader/reader_blocking_queue_test.cc b/paddle/fluid/operators/reader/reader_blocking_queue_test.cc new file mode 100644 index 0000000000..7d1b381d56 --- /dev/null +++ b/paddle/fluid/operators/reader/reader_blocking_queue_test.cc @@ -0,0 +1,219 @@ +// 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 // NOLINT +#include +#include // NOLINT +#include +#include "gtest/gtest.h" + +#include "paddle/fluid/operators/reader/blocking_queue.h" + +using paddle::operators::reader::BlockingQueue; + +TEST(BlockingQueue, CapacityTest) { + size_t cap = 10; + BlockingQueue q(cap); + EXPECT_EQ(q.Cap(), cap); +} + +void FirstInFirstOut(size_t queue_cap, size_t elem_num, size_t send_time_gap, + size_t receive_time_gap) { + BlockingQueue q(queue_cap); + std::thread sender([&]() { + for (size_t i = 0; i < elem_num; ++i) { + std::this_thread::sleep_for(std::chrono::milliseconds(send_time_gap)); + EXPECT_TRUE(q.Send(i)); + } + q.Close(); + }); + size_t count = 0; + while (true) { + std::this_thread::sleep_for(std::chrono::milliseconds(receive_time_gap)); + size_t elem; + if (!q.Receive(&elem)) { + break; + } + EXPECT_EQ(elem, count++); + } + sender.join(); + EXPECT_EQ(count, elem_num); + EXPECT_TRUE(q.IsClosed()); +} + +TEST(BlockingQueue, FirstInFirstOutTest) { + FirstInFirstOut(2, 5, 2, 50); + FirstInFirstOut(2, 5, 50, 2); + FirstInFirstOut(10, 3, 50, 2); + FirstInFirstOut(10, 3, 2, 50); +} + +TEST(BlockingQueue, SenderBlockingTest) { + const size_t queue_cap = 2; + BlockingQueue q(queue_cap); + size_t send_count = 0; + std::thread sender([&]() { + for (size_t i = 0; i < 5; ++i) { + if (!q.Send(i)) { + break; + } + ++send_count; + } + }); + std::this_thread::sleep_for(std::chrono::milliseconds(200)); + q.Close(); + sender.join(); + EXPECT_EQ(send_count, queue_cap); + std::vector res; + while (true) { + size_t elem; + if (!q.Receive(&elem)) { + break; + } + res.push_back(elem); + } + EXPECT_EQ(res.size(), queue_cap); + for (size_t i = 0; i < res.size(); ++i) { + EXPECT_EQ(res[i], i); + } +} + +TEST(BlockingQueue, ReceiverBlockingTest) { + const size_t queue_cap = 5; + BlockingQueue q(queue_cap); + std::vector receive_res; + std::thread receiver([&]() { + size_t elem; + while (true) { + if (!q.Receive(&elem)) { + break; + } + receive_res.push_back(elem); + } + }); + std::vector to_send{2, 1, 7}; + for (auto e : to_send) { + q.Send(e); + } + q.Close(); + receiver.join(); + EXPECT_EQ(receive_res.size(), to_send.size()); + for (size_t i = 0; i < to_send.size(); ++i) { + EXPECT_EQ(receive_res[i], to_send[i]); + } +} + +void CheckIsUnorderedSame(const std::vector>& v1, + const std::vector>& v2) { + std::set s1; + std::set s2; + for (auto vec : v1) { + for (size_t elem : vec) { + s1.insert(elem); + } + } + for (auto vec : v2) { + for (size_t elem : vec) { + s2.insert(elem); + } + } + EXPECT_EQ(s1.size(), s2.size()); + auto it1 = s1.begin(); + auto it2 = s2.begin(); + while (it1 != s1.end()) { + EXPECT_EQ(*it1, *it2); + ++it1; + ++it2; + } +} + +void MultiSenderMultiReceiver(const size_t queue_cap, + const std::vector>& to_send, + size_t receiver_num, size_t send_time_gap, + size_t receive_time_gap) { + BlockingQueue q(queue_cap); + size_t sender_num = to_send.size(); + std::vector senders; + for (size_t s_idx = 0; s_idx < sender_num; ++s_idx) { + senders.emplace_back(std::thread([&, s_idx] { + for (size_t elem : to_send[s_idx]) { + std::this_thread::sleep_for(std::chrono::milliseconds(send_time_gap)); + EXPECT_TRUE(q.Send(elem)); + } + })); + } + std::vector receivers; + std::mutex mu; + std::vector> res; + for (size_t r_idx = 0; r_idx < receiver_num; ++r_idx) { + receivers.emplace_back(std::thread([&] { + std::vector receiver_res; + while (true) { + std::this_thread::sleep_for( + std::chrono::milliseconds(receive_time_gap)); + size_t elem; + if (!q.Receive(&elem)) { + break; + } + receiver_res.push_back(elem); + } + std::lock_guard lock(mu); + res.push_back(receiver_res); + })); + } + for (auto& t : senders) { + t.join(); + } + q.Close(); + for (auto& t : receivers) { + t.join(); + } + CheckIsUnorderedSame(to_send, res); +} + +TEST(BlockingQueue, MultiSenderMultiReaderTest) { + std::vector> to_send_1{{2, 3, 4}, {9}, {0, 7, 15, 6}}; + MultiSenderMultiReceiver(2, to_send_1, 2, 0, 0); + MultiSenderMultiReceiver(10, to_send_1, 2, 0, 0); + MultiSenderMultiReceiver(2, to_send_1, 20, 0, 0); + MultiSenderMultiReceiver(2, to_send_1, 2, 50, 0); + MultiSenderMultiReceiver(2, to_send_1, 2, 0, 50); + + std::vector> to_send_2{ + {2, 3, 4}, {}, {0, 7, 15, 6, 9, 32}}; + MultiSenderMultiReceiver(2, to_send_2, 3, 0, 0); + MultiSenderMultiReceiver(20, to_send_2, 3, 0, 0); + MultiSenderMultiReceiver(2, to_send_2, 30, 0, 0); + MultiSenderMultiReceiver(2, to_send_2, 3, 50, 0); + MultiSenderMultiReceiver(2, to_send_2, 3, 0, 50); +} + +struct MyClass { + MyClass() : val_(0) {} + explicit MyClass(int val) : val_(val) {} + MyClass(const MyClass& b) { val_ = b.val_; } + MyClass(MyClass&& b) { val_ = b.val_; } + void operator=(const MyClass& b) { val_ = b.val_; } + + int val_; +}; + +TEST(BlockingQueue, MyClassTest) { + BlockingQueue q(2); + MyClass a(200); + q.Send(std::move(a)); + MyClass b; + q.Receive(&b); + EXPECT_EQ(a.val_, b.val_); +} diff --git a/paddle/fluid/operators/reader/reader_op_registry.cc b/paddle/fluid/operators/reader/reader_op_registry.cc index fc8dc747ff..3ff4536819 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.cc +++ b/paddle/fluid/operators/reader/reader_op_registry.cc @@ -12,7 +12,9 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "reader_op_registry.h" +#include "paddle/fluid/operators/reader/reader_op_registry.h" +#include +#include namespace paddle { namespace operators { diff --git a/paddle/fluid/operators/reader/reader_op_registry.h b/paddle/fluid/operators/reader/reader_op_registry.h index 929d32ad8b..ec25f55ef5 100644 --- a/paddle/fluid/operators/reader/reader_op_registry.h +++ b/paddle/fluid/operators/reader/reader_op_registry.h @@ -14,6 +14,8 @@ #pragma once +#include +#include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" diff --git a/paddle/fluid/operators/recurrent_op.cc b/paddle/fluid/operators/recurrent_op.cc index 00241e7682..72c2905872 100644 --- a/paddle/fluid/operators/recurrent_op.cc +++ b/paddle/fluid/operators/recurrent_op.cc @@ -596,7 +596,7 @@ class RecurrentGradOpDescMaker : public framework::SingleGradOpDescMaker { } } grad->SetAttrMap(this->Attrs()); - grad->SetBlockAttr(kStepBlock, *grad_block_[0]); + grad->SetBlockAttr(kStepBlock, grad_block_[0]); return std::unique_ptr(grad); } diff --git a/paddle/fluid/operators/reshape_op.h b/paddle/fluid/operators/reshape_op.h index 8320c257c9..ccd7063fe6 100644 --- a/paddle/fluid/operators/reshape_op.h +++ b/paddle/fluid/operators/reshape_op.h @@ -93,8 +93,14 @@ class ReshapeOp : public framework::OperatorWithKernel { if (unk_dim_idx != -1) { output_shape[unk_dim_idx] = -in_size / capacity; - PADDLE_ENFORCE_EQ(output_shape[unk_dim_idx] * capacity, -in_size, - "Invalid shape is given."); + // in_size < 0 and is un-determinate in compile time, skip the check, + // for example, in_dims = [-1, 8, 1, 1], shape = [-1, 3, 8], + // capacity = -24, in_size = -8, output_shape[0] = 0 + // the following check will fail. + if (in_size > 0) { + PADDLE_ENFORCE_EQ(output_shape[unk_dim_idx] * capacity, -in_size, + "Invalid shape is given."); + } } else { PADDLE_ENFORCE_EQ(capacity, in_size, "Invalid shape is given."); } @@ -124,10 +130,8 @@ class ReshapeKernel : public framework::OpKernel { auto *shape_data = shape_tensor->data(); framework::Tensor cpu_shape_tensor; if (platform::is_gpu_place(ctx.GetPlace())) { - TensorCopy(*shape_tensor, platform::CPUPlace(), ctx.device_context(), - &cpu_shape_tensor); + TensorCopySync(*shape_tensor, platform::CPUPlace(), &cpu_shape_tensor); shape_data = cpu_shape_tensor.data(); - ctx.device_context().Wait(); } auto shape = std::vector(shape_data, shape_data + shape_tensor->numel()); @@ -146,9 +150,7 @@ class ReshapeKernel : public framework::OpKernel { out->Resize(out_dims); if (!inplace) { out->mutable_data(ctx.GetPlace()); - framework::TensorCopy(*in, ctx.GetPlace(), ctx.device_context(), out); - ctx.device_context().Wait(); - // TensorCopy will resize to in_dims. + framework::TensorCopySync(*in, ctx.GetPlace(), out); out->Resize(out_dims); } else { out->ShareDataWith(*in); diff --git a/paddle/fluid/operators/roi_pool_op.cc b/paddle/fluid/operators/roi_pool_op.cc index 224ec93d28..397e49ef20 100644 --- a/paddle/fluid/operators/roi_pool_op.cc +++ b/paddle/fluid/operators/roi_pool_op.cc @@ -18,8 +18,7 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; - -static constexpr int kROISize = 5; +using LoDTensor = framework::LoDTensor; class ROIPoolOp : public framework::OperatorWithKernel { public: @@ -40,11 +39,11 @@ class ROIPoolOp : public framework::OperatorWithKernel { PADDLE_ENFORCE(input_dims.size() == 4, "The format of input tensor is NCHW."); PADDLE_ENFORCE(rois_dims.size() == 2, - "ROIs should be a 2-D tensor of shape (num_rois, 5)" - "given as [[batch_id, x1, y1, x2, y2], …]."); + "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]."); PADDLE_ENFORCE(rois_dims[1] == kROISize, - "ROIs should be a 2-D tensor of shape (num_rois, 5)" - "given as [[batch_id, x1, y1, x2, y2], …]."); + "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]."); int pooled_height = ctx->Attrs().Get("pooled_height"); int pooled_width = ctx->Attrs().Get("pooled_width"); @@ -109,10 +108,10 @@ class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker { "H is the height of the feature, and " "W is the width of the feature."); AddInput("ROIs", - "(Tensor), " + "(LoDTensor), " "ROIs (Regions of Interest) to pool over. " - "should be a 2-D tensor of shape (num_rois, 5)" - "given as [[batch_id, x1, y1, x2, y2], …]. " + "should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]. " "Where batch_id is the id of the data, " "(x1, y1) is the top left coordinates, and " "(x2, y2) is the bottom right coordinates."); diff --git a/paddle/fluid/operators/roi_pool_op.cu b/paddle/fluid/operators/roi_pool_op.cu index 1931629d13..0bdfee0434 100644 --- a/paddle/fluid/operators/roi_pool_op.cu +++ b/paddle/fluid/operators/roi_pool_op.cu @@ -19,10 +19,10 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; static constexpr int kNumCUDAThreads = 512; static constexpr int kNumMaxinumNumBlocks = 4096; -static constexpr int kROISize = 5; static inline int NumBlocks(const int N) { return std::min((N + kNumCUDAThreads - 1) / kNumCUDAThreads, @@ -30,13 +30,11 @@ static inline int NumBlocks(const int N) { } template -__global__ void GPUROIPoolForward(const int nthreads, const T* input_data, - const int64_t* input_rois, - const float spatial_scale, const int channels, - const int height, const int width, - const int pooled_height, - const int pooled_width, T* output_data, - int64_t* argmax_data) { +__global__ void GPUROIPoolForward( + const int nthreads, const T* input_data, const int64_t* input_rois, + const float spatial_scale, const int channels, const int height, + const int width, const int pooled_height, const int pooled_width, + int* roi_batch_id_data, T* output_data, int64_t* argmax_data) { int index = blockIdx.x * blockDim.x + threadIdx.x; int offset = blockDim.x * gridDim.x; for (size_t i = index; i < nthreads; i += offset) { @@ -46,11 +44,11 @@ __global__ void GPUROIPoolForward(const int nthreads, const T* input_data, int n = index / pooled_width / pooled_height / channels; const int64_t* offset_input_rois = input_rois + n * kROISize; - int roi_batch_ind = offset_input_rois[0]; - int roi_start_w = round(offset_input_rois[1] * spatial_scale); - int roi_start_h = round(offset_input_rois[2] * spatial_scale); - int roi_end_w = round(offset_input_rois[3] * spatial_scale); - int roi_end_h = round(offset_input_rois[4] * spatial_scale); + int roi_batch_ind = roi_batch_id_data[n]; + int roi_start_w = round(offset_input_rois[0] * spatial_scale); + int roi_start_h = round(offset_input_rois[1] * spatial_scale); + int roi_end_w = round(offset_input_rois[2] * spatial_scale); + int roi_end_h = round(offset_input_rois[3] * spatial_scale); int roi_width = max(roi_end_w - roi_start_w + 1, 1); int roi_height = max(roi_end_h - roi_start_h + 1, 1); @@ -93,7 +91,8 @@ __global__ void GPUROIPoolBackward( const int nthreads, const int64_t* input_rois, const T* output_grad, const int64_t* argmax_data, const int num_rois, const float spatial_scale, const int channels, const int height, const int width, - const int pooled_height, const int pooled_width, T* input_grad) { + const int pooled_height, const int pooled_width, int* roi_batch_id_data, + T* input_grad) { int index = blockIdx.x * blockDim.x + threadIdx.x; int offset = blockDim.x * gridDim.x; for (int i = index; i < nthreads; i += offset) { @@ -102,8 +101,7 @@ __global__ void GPUROIPoolBackward( int c = (index / pooled_width / pooled_height) % channels; int n = index / pooled_width / pooled_height / channels; - const int64_t* offset_input_rois = input_rois + n * kROISize; - int roi_batch_ind = offset_input_rois[0]; + int roi_batch_ind = roi_batch_id_data[n]; int input_offset = (roi_batch_ind * channels + c) * height * width; int output_offset = (n * channels + c) * pooled_height * pooled_width; const T* offset_output_grad = output_grad + output_offset; @@ -124,7 +122,7 @@ class GPUROIPoolOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); - auto* rois = ctx.Input("ROIs"); + auto* rois = ctx.Input("ROIs"); auto* out = ctx.Output("Out"); auto* argmax = ctx.Output("Argmax"); @@ -133,23 +131,46 @@ class GPUROIPoolOpKernel : public framework::OpKernel { auto spatial_scale = ctx.Attr("spatial_scale"); auto in_dims = in->dims(); + int batch_size = in_dims[0]; auto in_stride = framework::stride(in_dims); int channels = in_dims[1]; int height = in_dims[2]; int width = in_dims[3]; - size_t rois_num = rois->dims()[0]; + int rois_num = rois->dims()[0]; if (rois_num == 0) return; int output_size = out->numel(); int blocks = NumBlocks(output_size); int threads = kNumCUDAThreads; + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(platform::CPUPlace()); + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + PADDLE_ENFORCE_EQ( + rois_batch_size, batch_size, + "The rois_batch_size and imgs batch_size must be the same."); + int rois_num_with_lod = rois_lod[rois_batch_size]; + PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod, + "The rois_num from input and lod must be the same."); + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + + framework::Tensor roi_batch_id_list_gpu; + framework::TensorCopy(roi_batch_id_list, ctx.GetPlace(), + ctx.device_context(), &roi_batch_id_list_gpu); + GPUROIPoolForward< T><<>>( output_size, in->data(), rois->data(), spatial_scale, channels, height, width, pooled_height, pooled_width, - out->mutable_data(ctx.GetPlace()), + roi_batch_id_list_gpu.data(), out->mutable_data(ctx.GetPlace()), argmax->mutable_data(ctx.GetPlace())); } }; @@ -159,7 +180,7 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); - auto* rois = ctx.Input("ROIs"); + auto* rois = ctx.Input("ROIs"); auto* argmax = ctx.Input("Argmax"); auto* out_grad = ctx.Input(framework::GradVarName("Out")); @@ -169,12 +190,27 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel { auto pooled_width = ctx.Attr("pooled_width"); auto spatial_scale = ctx.Attr("spatial_scale"); - size_t rois_num = rois->dims()[0]; + int rois_num = rois->dims()[0]; int channels = in->dims()[1]; int height = in->dims()[2]; int width = in->dims()[3]; if (x_grad) { + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(platform::CPUPlace()); + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + framework::Tensor roi_batch_id_list_gpu; + framework::TensorCopy(roi_batch_id_list, ctx.GetPlace(), + ctx.device_context(), &roi_batch_id_list_gpu); + x_grad->mutable_data(ctx.GetPlace()); math::SetConstant set_zero; set_zero(ctx.cuda_device_context(), x_grad, static_cast(0)); @@ -189,6 +225,7 @@ class GPUROIPoolGradOpKernel : public framework::OpKernel { output_grad_size, rois->data(), out_grad->data(), argmax->data(), rois_num, spatial_scale, channels, height, width, pooled_height, pooled_width, + roi_batch_id_list_gpu.data(), x_grad->mutable_data(ctx.GetPlace())); } } diff --git a/paddle/fluid/operators/roi_pool_op.h b/paddle/fluid/operators/roi_pool_op.h index 54e0749031..c4f739b2c6 100644 --- a/paddle/fluid/operators/roi_pool_op.h +++ b/paddle/fluid/operators/roi_pool_op.h @@ -21,12 +21,14 @@ limitations under the License. */ namespace paddle { namespace operators { +static constexpr int kROISize = 4; + template class CPUROIPoolOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); - auto* rois = ctx.Input("ROIs"); + auto* rois = ctx.Input("ROIs"); auto* out = ctx.Output("Out"); auto* argmax = ctx.Output("Argmax"); @@ -47,24 +49,36 @@ class CPUROIPoolOpKernel : public framework::OpKernel { auto out_stride = framework::stride(out->dims()); const T* input_data = in->data(); - const int64_t* rois_data = rois->data(); - T* output_data = out->mutable_data(ctx.GetPlace()); - int64_t* argmax_data = argmax->mutable_data(ctx.GetPlace()); - for (int n = 0; n < rois_num; ++n) { - int roi_batch_id = rois_data[0]; - PADDLE_ENFORCE_GE(roi_batch_id, 0); - PADDLE_ENFORCE_LT(roi_batch_id, batch_size); - rois_data += roi_stride[0]; + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(ctx.GetPlace()); + + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + PADDLE_ENFORCE_EQ( + rois_batch_size, batch_size, + "The rois_batch_size and imgs batch_size must be the same."); + int rois_num_with_lod = rois_lod[rois_batch_size]; + PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod, + "The rois_num from input and lod must be the same."); + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } } - rois_data = rois->data(); + T* output_data = out->mutable_data(ctx.GetPlace()); + int64_t* argmax_data = argmax->mutable_data(ctx.GetPlace()); + + const int64_t* rois_data = rois->data(); for (int n = 0; n < rois_num; ++n) { - int roi_batch_id = rois_data[0]; - int roi_start_w = round(rois_data[1] * spatial_scale); - int roi_start_h = round(rois_data[2] * spatial_scale); - int roi_end_w = round(rois_data[3] * spatial_scale); - int roi_end_h = round(rois_data[4] * spatial_scale); + int roi_batch_id = roi_batch_id_data[n]; + int roi_start_w = round(rois_data[0] * spatial_scale); + int roi_start_h = round(rois_data[1] * spatial_scale); + int roi_end_w = round(rois_data[2] * spatial_scale); + int roi_end_h = round(rois_data[3] * spatial_scale); // Force malformed ROIs to be 1x1 int roi_height = std::max(roi_end_h - roi_start_h + 1, 1); @@ -133,7 +147,7 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); - auto* rois = ctx.Input("ROIs"); + auto* rois = ctx.Input("ROIs"); auto* argmax = ctx.Input("Argmax"); auto* out_grad = ctx.Input(framework::GradVarName("Out")); @@ -143,6 +157,20 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel { auto pooled_width = ctx.Attr("pooled_width"); if (in_grad) { + int rois_num = rois->dims()[0]; + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(ctx.GetPlace()); + + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + const int64_t* rois_data = rois->data(); const T* out_grad_data = out_grad->data(); const int64_t* argmax_data = argmax->data(); @@ -156,11 +184,10 @@ class CPUROIPoolGradOpKernel : public framework::OpKernel { auto roi_stride = framework::stride(rois->dims()); auto out_stride = framework::stride(out_grad->dims()); - int rois_num = rois->dims()[0]; int channels = in->dims()[1]; for (int n = 0; n < rois_num; ++n) { - int roi_batch_idx = rois_data[0]; + int roi_batch_idx = roi_batch_id_data[n]; T* batch_grad_data = in_grad_data + roi_batch_idx * in_stride[0]; for (int c = 0; c < channels; ++c) { for (int ph = 0; ph < pooled_height; ++ph) { diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index 82ff087d0a..e4386b640a 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -41,6 +41,8 @@ class SendOp : public framework::OperatorBase { std::vector endpoints = Attr>("endpoints"); + bool sync_mode = Attr("sync_mode"); + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); @@ -64,11 +66,13 @@ class SendOp : public framework::OperatorBase { } PADDLE_ENFORCE(rpc_client->Wait()); - for (auto& ep : endpoints) { - VLOG(3) << "batch barrier, ep: " << ep; - rpc_client->AsyncSendBatchBarrier(ep); + if (sync_mode) { + for (auto& ep : endpoints) { + VLOG(3) << "batch barrier, ep: " << ep; + rpc_client->AsyncSendBatchBarrier(ep); + } + PADDLE_ENFORCE(rpc_client->Wait()); } - PADDLE_ENFORCE(rpc_client->Wait()); if (outs.size() > 0) { for (size_t i = 0; i < outs.size(); i++) { @@ -112,6 +116,7 @@ This operator will send tensor to recv_op at the parameter server. "Server endpoints in the order of input " "variables for mapping") .SetDefault({}); + AddAttr("sync_mode", "work in sync_mode or not").SetDefault(true); } }; diff --git a/paddle/fluid/operators/send_recv_op_test.cc b/paddle/fluid/operators/send_recv_op_test.cc index 81350fee38..d2e1f3cb2f 100644 --- a/paddle/fluid/operators/send_recv_op_test.cc +++ b/paddle/fluid/operators/send_recv_op_test.cc @@ -137,6 +137,8 @@ void StartServerNet(bool is_sparse) { attrs.insert({"GradList", std::vector({"x1"})}); attrs.insert({"OptimizeBlock", optimize_block}); attrs.insert({"PrefetchBlock", prefetch_block}); + attrs.insert({"grad_to_block_id", std::vector({""})}); + attrs.insert({"sync_mode", true}); listen_and_serv_op = f::OpRegistry::CreateOp("listen_and_serv", {{"X", {"x1"}}}, {}, attrs); listen_and_serv_op->Run(scope, place); diff --git a/paddle/fluid/operators/sequence_conv_op.h b/paddle/fluid/operators/sequence_conv_op.h index b59504bb98..3916cdbb6a 100644 --- a/paddle/fluid/operators/sequence_conv_op.h +++ b/paddle/fluid/operators/sequence_conv_op.h @@ -33,7 +33,6 @@ class SequenceConvKernel : public framework::OpKernel { auto filter = *context.Input("Filter"); out->mutable_data(context.GetPlace()); - context.ShareLoD("X", "Out"); int context_start = context.Attr("contextStart"); int context_length = context.Attr("contextLength"); diff --git a/paddle/fluid/operators/softmax_mkldnn_op.cc b/paddle/fluid/operators/softmax_mkldnn_op.cc index d00bd1447e..71b541d98f 100644 --- a/paddle/fluid/operators/softmax_mkldnn_op.cc +++ b/paddle/fluid/operators/softmax_mkldnn_op.cc @@ -77,7 +77,7 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel { const bool is_test = ctx.Attr("is_test"); if (!is_test) { T threshold = exp(-64); - for (size_t i = 0; i < dst_tz[0] * dst_tz[1]; ++i) { + for (int i = 0; i < dst_tz[0] * dst_tz[1]; ++i) { output_data[i] = output_data[i] < threshold ? threshold : output_data[i]; } diff --git a/paddle/fluid/operators/warpctc_op.h b/paddle/fluid/operators/warpctc_op.h index afbfe69973..85131d0025 100644 --- a/paddle/fluid/operators/warpctc_op.h +++ b/paddle/fluid/operators/warpctc_op.h @@ -222,8 +222,8 @@ class WarpCTCGradKernel : public framework::OpKernel { const T* loss_grad_data = loss_grad->data(); math::ScaleLoDTensorFunctor()( - ctx.template device_context(), *logits_grad, - loss_grad_data); + ctx.template device_context(), loss_grad_data, + logits_grad); } }; diff --git a/paddle/fluid/operators/while_op.cc b/paddle/fluid/operators/while_op.cc index 8b62b242cf..710cc9fc2e 100644 --- a/paddle/fluid/operators/while_op.cc +++ b/paddle/fluid/operators/while_op.cc @@ -288,7 +288,7 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker { while_grad->SetInput(framework::GradVarName(kOutputs), output_grads_list); while_grad->SetAttrMap(this->Attrs()); - while_grad->SetBlockAttr(kStepBlock, *grad_block); + while_grad->SetBlockAttr(kStepBlock, grad_block); // record the original output gradient names, since the gradient name of // while operator could be renamed. while_grad->SetAttr("original_output_grad", output_grads_list); diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt index 917bdc64ab..598fd4d419 100644 --- a/paddle/fluid/platform/CMakeLists.txt +++ b/paddle/fluid/platform/CMakeLists.txt @@ -12,7 +12,7 @@ add_custom_command(TARGET profiler_py_proto POST_BUILD WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) if(WITH_GPU) - cc_library(enforce SRCS enforce.cc DEPS) + nv_library(enforce SRCS enforce.cc) else() cc_library(enforce SRCS enforce.cc) endif() diff --git a/paddle/fluid/platform/dynload/cublas.h b/paddle/fluid/platform/dynload/cublas.h index 1ab55d6b9b..81acaff87d 100644 --- a/paddle/fluid/platform/dynload/cublas.h +++ b/paddle/fluid/platform/dynload/cublas.h @@ -14,10 +14,12 @@ #pragma once +#include #include #include #include #include // NOLINT +#include #include "paddle/fluid/platform/dynload/dynamic_loader.h" namespace paddle { @@ -37,14 +39,14 @@ extern void *cublas_dso_handle; #ifdef PADDLE_USE_DSO #define DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP(__name) \ struct DynLoad__##__name { \ + using FUNC_TYPE = decltype(&::__name); \ template \ inline cublasStatus_t operator()(Args... args) { \ - typedef cublasStatus_t (*cublasFunc)(Args...); \ std::call_once(cublas_dso_flag, []() { \ cublas_dso_handle = paddle::platform::dynload::GetCublasDsoHandle(); \ }); \ void *p_##__name = dlsym(cublas_dso_handle, #__name); \ - return reinterpret_cast(p_##__name)(args...); \ + return reinterpret_cast(p_##__name)(args...); \ } \ }; \ extern DynLoad__##__name __name @@ -71,8 +73,8 @@ extern void *cublas_dso_handle; __macro(cublasDgemm_v2); \ __macro(cublasHgemm); \ __macro(cublasSgemmEx); \ - __macro(cublasSgeam_v2); \ - __macro(cublasDgeam_v2); \ + __macro(cublasSgeam); \ + __macro(cublasDgeam); \ __macro(cublasCreate_v2); \ __macro(cublasDestroy_v2); \ __macro(cublasSetStream_v2); \ diff --git a/paddle/fluid/platform/dynload/cudnn.h b/paddle/fluid/platform/dynload/cudnn.h index 24475b62ca..34d83e3956 100644 --- a/paddle/fluid/platform/dynload/cudnn.h +++ b/paddle/fluid/platform/dynload/cudnn.h @@ -34,7 +34,7 @@ extern void EnforceCUDNNLoaded(const char* fn_name); struct DynLoad__##__name { \ template \ auto operator()(Args... args) -> decltype(__name(args...)) { \ - using cudnn_func = decltype(__name(args...)) (*)(Args...); \ + using cudnn_func = decltype(&::__name); \ std::call_once(cudnn_dso_flag, []() { \ cudnn_dso_handle = paddle::platform::dynload::GetCUDNNDsoHandle(); \ }); \ diff --git a/paddle/fluid/platform/dynload/cupti.h b/paddle/fluid/platform/dynload/cupti.h index d0d676b9d8..e64de7c20f 100644 --- a/paddle/fluid/platform/dynload/cupti.h +++ b/paddle/fluid/platform/dynload/cupti.h @@ -41,7 +41,7 @@ extern void *cupti_dso_handle; struct DynLoad__##__name { \ template \ inline CUptiResult CUPTIAPI operator()(Args... args) { \ - typedef CUptiResult CUPTIAPI (*cuptiFunc)(Args...); \ + using cuptiFunc = decltype(&::__name); \ std::call_once(cupti_dso_flag, []() { \ cupti_dso_handle = paddle::platform::dynload::GetCUPTIDsoHandle(); \ }); \ diff --git a/paddle/fluid/platform/dynload/curand.h b/paddle/fluid/platform/dynload/curand.h index 4697fb6cd9..46ad4379d5 100644 --- a/paddle/fluid/platform/dynload/curand.h +++ b/paddle/fluid/platform/dynload/curand.h @@ -30,7 +30,7 @@ extern void *curand_dso_handle; struct DynLoad__##__name { \ template \ curandStatus_t operator()(Args... args) { \ - typedef curandStatus_t (*curandFunc)(Args...); \ + using curandFunc = decltype(&::__name); \ std::call_once(curand_dso_flag, []() { \ curand_dso_handle = paddle::platform::dynload::GetCurandDsoHandle(); \ }); \ diff --git a/paddle/fluid/platform/dynload/nccl.h b/paddle/fluid/platform/dynload/nccl.h index c5a10a78a4..37902ae20c 100644 --- a/paddle/fluid/platform/dynload/nccl.h +++ b/paddle/fluid/platform/dynload/nccl.h @@ -33,7 +33,7 @@ extern void* nccl_dso_handle; struct DynLoad__##__name { \ template \ auto operator()(Args... args) -> decltype(__name(args...)) { \ - using nccl_func = decltype(__name(args...)) (*)(Args...); \ + using nccl_func = decltype(&::__name); \ std::call_once(nccl_dso_flag, []() { \ nccl_dso_handle = paddle::platform::dynload::GetNCCLDsoHandle(); \ }); \ diff --git a/paddle/fluid/platform/dynload/warpctc.h b/paddle/fluid/platform/dynload/warpctc.h index 7fa4683704..7c70649d21 100644 --- a/paddle/fluid/platform/dynload/warpctc.h +++ b/paddle/fluid/platform/dynload/warpctc.h @@ -36,7 +36,7 @@ extern void* warpctc_dso_handle; struct DynLoad__##__name { \ template \ auto operator()(Args... args) -> decltype(__name(args...)) { \ - using warpctcFunc = decltype(__name(args...)) (*)(Args...); \ + using warpctcFunc = decltype(&::__name); \ std::call_once(warpctc_dso_flag, []() { \ warpctc_dso_handle = paddle::platform::dynload::GetWarpCTCDsoHandle(); \ }); \ diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h index 159d1d5f4e..dcd711a33f 100644 --- a/paddle/fluid/pybind/tensor_py.h +++ b/paddle/fluid/pybind/tensor_py.h @@ -63,15 +63,9 @@ struct CastToPyBufferImpl { auto *dst_ptr = static_cast(dst_tensor.mutable_data( tensor.dims(), platform::CPUPlace())); - platform::DeviceContextPool &pool = - platform::DeviceContextPool::Instance(); - auto dev_ctx = static_cast( - pool.Get(tensor.place())); - - paddle::platform::GpuMemcpyAsync( - dst_ptr, src_ptr, sizeof(CUR_TYPE) * tensor.numel(), - cudaMemcpyDeviceToHost, dev_ctx->stream()); - dev_ctx->Wait(); + paddle::platform::GpuMemcpySync(dst_ptr, src_ptr, + sizeof(CUR_TYPE) * tensor.numel(), + cudaMemcpyDeviceToHost); #else PADDLE_THROW("'CUDAPlace' is not supported in CPU only device."); #endif @@ -184,17 +178,8 @@ void PyCUDATensorSetFromArray( 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(T) * array.size(), - cudaMemcpyHostToDevice, dev_ctx->stream()); - // NOTE: For safety, here wait the copy complete. - // It because the CPU array.data() could be destroyed after this method. - // If we make this method async, it could be copied data from a memory buffer - // that has been freed. - dev_ctx->Wait(); + paddle::platform::GpuMemcpySync(dst, array.data(), sizeof(T) * array.size(), + cudaMemcpyHostToDevice); } template <> @@ -214,18 +199,9 @@ void PyCUDATensorSetFromArray( 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()); - // NOTE: For safety, here wait the copy complete. - // It because the CPU array.data() could be destroyed after this method. - // If we make this method async, it could be copied data from a memory buffer - // that has been freed. - dev_ctx->Wait(); + paddle::platform::GpuMemcpySync(dst, array.data(), + sizeof(uint16_t) * array.size(), + cudaMemcpyHostToDevice); } template diff --git a/paddle/scripts/docker/README.md b/paddle/scripts/README.md similarity index 66% rename from paddle/scripts/docker/README.md rename to paddle/scripts/README.md index 78c0cc3782..9e8b135c1b 100644 --- a/paddle/scripts/docker/README.md +++ b/paddle/scripts/README.md @@ -13,40 +13,49 @@ We want to make the building procedures: 1. Build docker images with PaddlePaddle pre-installed, so that we can run PaddlePaddle applications directly in docker or on Kubernetes clusters. -To achieve this, we created a repo: https://github.com/PaddlePaddle/buildtools -which gives several docker images that are `manylinux1` sufficient. Then we -can build PaddlePaddle using these images to generate corresponding `whl` -binaries. +To achieve this, we maintain a dockerhub repo:https://hub.docker.com/r/paddlepaddle/paddle +which provides pre-built environment images to build PaddlePaddle and generate corresponding `whl` +binaries.(**We strongly recommend building paddlepaddle in our pre-specified Docker environment.**) -## Run The Build +## Development Workflow + +Here we describe how the workflow goes on. We start from considering our daily development environment. + +Developers work on a computer, which is usually a laptop or desktop: + + + +or, they might rely on a more sophisticated box (like with GPUs): + + + +A principle here is that source code lies on the development computer (host) so that editors like Eclipse can parse the source code to support auto-completion. + +## Build With Docker ### Build Environments -The pre-built build environment images are: +The lastest pre-built build environment images are: | Image | Tag | | ----- | --- | -| paddlepaddle/paddle_manylinux_devel | cuda7.5_cudnn5 | -| paddlepaddle/paddle_manylinux_devel | cuda8.0_cudnn5 | -| paddlepaddle/paddle_manylinux_devel | cuda7.5_cudnn7 | -| paddlepaddle/paddle_manylinux_devel | cuda9.0_cudnn7 | +| paddlepaddle/paddle | latest-dev | +| paddlepaddle/paddle | latest-dev-android | ### Start Build -Choose one docker image that suit your environment and run the following -command to start a build: - ```bash git clone https://github.com/PaddlePaddle/Paddle.git cd Paddle -docker run --rm -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=OFF" -e "RUN_TEST=OFF" -e "PYTHON_ABI=cp27-cp27mu" paddlepaddle/paddle_manylinux_devel /paddle/paddle/scripts/docker/build.sh +./paddle/scripts/paddle_docker_build.sh build ``` After the build finishes, you can get output `whl` package under `build/python/dist`. -This command mounts the source directory on the host into `/paddle` in the container, then run the build script `/paddle/paddle/scripts/docker/build.sh` -in the container. When it writes to `/paddle/build` in the container, it writes to `$PWD/build` on the host indeed. +This command will download the most recent dev image from docker hub, start a container in the backend and then run the build script `/paddle/paddle/scripts/paddle_build.sh build` in the container. +The container mounts the source directory on the host into `/paddle`. +When it writes to `/paddle/build` in the container, it writes to `$PWD/build` on the host indeed. ### Build Options @@ -68,7 +77,6 @@ Users can specify the following Docker build arguments with either "ON" or "OFF" | `WITH_DOC` | OFF | Build docs after build binaries. | | `WOBOQ` | OFF | Generate WOBOQ code viewer under `build/woboq_out` | - ## Docker Images You can get the latest PaddlePaddle docker images by @@ -144,59 +152,37 @@ docker push kubectl ... ``` -## Docker Images for Developers - -We have a special docker image for developers: -`paddlepaddle/paddle:-dev`. This image is also generated from -https://github.com/PaddlePaddle/buildtools - -This a development image contains only the -development tools and standardizes the building procedure. Users include: - -- developers -- no longer need to install development tools on the host, and can build their current work on the host (development computer). -- release engineers -- use this to build the official release from certain branch/tag on Github.com. -- document writers / Website developers -- Our documents are in the source repo in the form of .md/.rst files and comments in source code. We need tools to extract the information, typeset, and generate Web pages. - -Of course, developers can install building tools on their development computers. But different versions of PaddlePaddle might require different set or version of building tools. Also, it makes collaborative debugging easier if all developers use a unified development environment. - -The development image contains the following tools: - - - gcc/clang - - nvcc - - Python - - sphinx - - woboq - - sshd - -Many developers work on a remote computer with GPU; they could ssh into the computer and `docker exec` into the development container. However, running `sshd` in the container allows developers to ssh into the container directly. - - -### Development Workflow - -Here we describe how the workflow goes on. We start from considering our daily development environment. +### Reading source code with woboq codebrowser -Developers work on a computer, which is usually a laptop or desktop: +For developers who are interested in the C++ source code, you can build C++ source code into HTML pages using [Woboq codebrowser](https://github.com/woboq/woboq_codebrowser). - +- The following command builds PaddlePaddle, generates HTML pages from C++ source code, and writes HTML pages into `$HOME/woboq_out` on the host: -or, they might rely on a more sophisticated box (like with GPUs): +```bash +./paddle/scripts/paddle_docker_build.sh html +``` - +- You can open the generated HTML files in your Web browser. Or, if you want to run a Nginx container to serve them for a wider audience, you can run: -A principle here is that source code lies on the development computer (host) so that editors like Eclipse can parse the source code to support auto-completion. +``` +docker run -v $HOME/woboq_out:/usr/share/nginx/html -d -p 8080:80 nginx +``` -### Reading source code with woboq codebrowser +## More Options -For developers who are interested in the C++ source code, please use -e "WOBOQ=ON" to enable the building of C++ source code into HTML pages using [Woboq codebrowser](https://github.com/woboq/woboq_codebrowser). +### Build Without Docker -- The following command builds PaddlePaddle, generates HTML pages from C++ source code, and writes HTML pages into `$HOME/woboq_out` on the host: +Follow the *Dockerfile* in the paddlepaddle repo to set up your local dev environment and run: ```bash -docker run -v $PWD:/paddle -v $HOME/woboq_out:/woboq_out -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=ON" -e "WOBOQ=ON" paddlepaddle/paddle:latest-dev +./paddle/scripts/paddle_build.sh build ``` -- You can open the generated HTML files in your Web browser. Or, if you want to run a Nginx container to serve them for a wider audience, you can run: +### Additional Tasks -``` -docker run -v $HOME/woboq_out:/usr/share/nginx/html -d -p 8080:80 nginx +You can get the help menu for the build scripts by running with no options: + +```bash +./paddle/scripts/paddle_build.sh +or ./paddle/scripts/paddle_docker_build.sh ``` diff --git a/paddle/scripts/docker/doc/paddle-development-environment-gpu.graffle b/paddle/scripts/doc/paddle-development-environment-gpu.graffle similarity index 100% rename from paddle/scripts/docker/doc/paddle-development-environment-gpu.graffle rename to paddle/scripts/doc/paddle-development-environment-gpu.graffle diff --git a/paddle/scripts/docker/doc/paddle-development-environment-gpu.png b/paddle/scripts/doc/paddle-development-environment-gpu.png similarity index 100% rename from paddle/scripts/docker/doc/paddle-development-environment-gpu.png rename to paddle/scripts/doc/paddle-development-environment-gpu.png diff --git a/paddle/scripts/docker/doc/paddle-development-environment.graffle b/paddle/scripts/doc/paddle-development-environment.graffle similarity index 100% rename from paddle/scripts/docker/doc/paddle-development-environment.graffle rename to paddle/scripts/doc/paddle-development-environment.graffle diff --git a/paddle/scripts/docker/doc/paddle-development-environment.png b/paddle/scripts/doc/paddle-development-environment.png similarity index 100% rename from paddle/scripts/docker/doc/paddle-development-environment.png rename to paddle/scripts/doc/paddle-development-environment.png diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh new file mode 100755 index 0000000000..654c8272a1 --- /dev/null +++ b/paddle/scripts/paddle_build.sh @@ -0,0 +1,508 @@ +#!/usr/bin/env bash + +# 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. + + +#================================================= +# Utils +#================================================= + +function print_usage() { + RED='\033[0;31m' + BLUE='\033[0;34m' + BOLD='\033[1m' + NONE='\033[0m' + + echo -e "\n${RED}Usage${NONE}: + ${BOLD}$0${NONE} [OPTION]" + + echo -e "\n${RED}Options${NONE}: + ${BLUE}build${NONE}: run build for x86 platform + ${BLUE}build_android${NONE}: run build for android platform + ${BLUE}build_ios${NONE}: run build for ios platform + ${BLUE}test${NONE}: run all unit tests + ${BLUE}bind_test${NONE}: parallel tests bind to different GPU + ${BLUE}doc${NONE}: generate paddle documents + ${BLUE}html${NONE}: convert C++ source code into HTML + ${BLUE}dockerfile${NONE}: generate paddle release dockerfile + ${BLUE}capi${NONE}: generate paddle CAPI package + ${BLUE}fluid_inference_lib${NONE}: deploy fluid inference library + ${BLUE}check_style${NONE}: run code style check + " +} + +function init() { + PADDLE_ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}")/../../" && pwd )" +} + +function cmake_gen() { + mkdir -p ${PADDLE_ROOT}/build + cd ${PADDLE_ROOT}/build + + # build script will not fail if *.deb does not exist + rm *.deb 2>/dev/null || true + # delete previous built whl packages + rm -rf python/dist 2>/dev/null || true + + # Support build for all python versions, currently + # including cp27-cp27m and cp27-cp27mu. + PYTHON_FLAGS="" + if [ "$1" != "" ]; then + echo "using python abi: $1" + if [ "$1" == "cp27-cp27m" ]; then + export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:} + export PATH=/opt/python/cp27-cp27m/bin/:${PATH} + PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python + -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7 + -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so" + elif [ "$1" == "cp27-cp27mu" ]; then + export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:} + export PATH=/opt/python/cp27-cp27mu/bin/:${PATH} + PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python + -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7 + -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so" + fi + fi + + cat <&2 + echo "Please use pre-commit to check what is wrong." 1>&2 + exit 1 +} + +function check_style() { + trap 'abort' 0 + set -e + + # install glide + curl https://glide.sh/get | bash + eval "$(GIMME_GO_VERSION=1.8.3 gimme)" + + # set up go environment for running gometalinter + mkdir -p $GOPATH/src/github.com/PaddlePaddle/ + ln -sf ${PADDLE_ROOT} $GOPATH/src/github.com/PaddlePaddle/Paddle + cd $GOPATH/src/github.com/PaddlePaddle/Paddle/go; glide install; cd - + + go get github.com/alecthomas/gometalinter + gometalinter --install + + cd ${PADDLE_ROOT} + export PATH=/usr/bin:$PATH + pre-commit install + clang-format --version + + if ! pre-commit run -a ; then + git diff + exit 1 + fi + + trap : 0 +} + +#================================================= +# Build +#================================================= + +function build() { + mkdir -p ${PADDLE_ROOT}/build + cd ${PADDLE_ROOT}/build + cat <= 21." + ANDROID_API=21 + fi + else # armeabi, armeabi-v7a + ANDROID_ARCH=arm + fi + + ANDROID_STANDALONE_TOOLCHAIN=$ANDROID_TOOLCHAINS_DIR/$ANDROID_ARCH-android-$ANDROID_API + + cat < ${PADDLE_ROOT}/build/Dockerfile < + ENV HOME /root +EOF + + if [[ ${WITH_GPU} == "ON" ]]; then + NCCL_DEPS="apt-get install -y libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 &&" + else + NCCL_DEPS="" + fi + + if [[ ${WITH_FLUID_ONLY:-OFF} == "OFF" ]]; then + PADDLE_VERSION="paddle version" + CMD='"paddle", "version"' + else + PADDLE_VERSION="true" + CMD='"true"' + fi + + cat >> /paddle/build/Dockerfile < /dev/null + return $? +} + +function start_build_docker() { + docker pull $IMG + + if container_running "${CONTAINER_ID}"; then + docker stop "${CONTAINER_ID}" 1>/dev/null + docker rm -f "${CONTAINER_ID}" 1>/dev/null + fi + + DOCKER_ENV=$(cat < - -#include -#include +#include // NOLINT +#include // NOLINT #include "paddle/utils/CustomStackTrace.h" #include "paddle/utils/Locks.h" @@ -39,14 +37,10 @@ void testNormalImpl( threads.reserve(FLAGS_test_thread_num); for (int32_t i = 0; i < FLAGS_test_thread_num; ++i) { - threads.emplace_back(new std::thread([&tracer, - &countDown, - &layerSize, - &startBarrier, - &doneBarrier, - &callback] { - callback(tracer, countDown, layerSize, startBarrier, doneBarrier); - })); + threads.emplace_back( + new std::thread([&tracer, &startBarrier, &doneBarrier, &callback] { + callback(tracer, countDown, layerSize, startBarrier, doneBarrier); + })); } size_t cntDown = countDown; while (cntDown-- > 0) { diff --git a/python/paddle/fluid/distribute_transpiler.py b/python/paddle/fluid/distribute_transpiler.py index d07e0f696e..44542749f8 100644 --- a/python/paddle/fluid/distribute_transpiler.py +++ b/python/paddle/fluid/distribute_transpiler.py @@ -143,7 +143,8 @@ class DistributeTranspiler: program=None, pservers="127.0.0.1:6174", trainers=1, - split_method=splitter.round_robin): + split_method=splitter.round_robin, + sync_mode=True): """ Transpile the program to distributed data-parallelism programs. The main_program will be transformed to use a remote parameter server @@ -184,6 +185,9 @@ class DistributeTranspiler: :param split_method: A function to determin how to split variables to different servers equally. :type split_method: function + :type sync_mode: boolean default True + :param sync_mode: if sync_mode is set True, it means that dist transpiler + will transpile the program into sync_mode pserver and trainer program. """ assert (callable(split_method)) if program is None: @@ -191,6 +195,7 @@ class DistributeTranspiler: self.origin_program = program self.trainer_num = trainers self.optimize_ops = optimize_ops + self.sync_mode = sync_mode # TODO(typhoonzero): currently trainer_id is fetched from cluster system # like Kubernetes, we should port this to use etcd later when developing # fluid distributed training with fault-tolerance. @@ -295,8 +300,11 @@ class DistributeTranspiler: inputs={"X": send_inputs}, outputs={"Out": send_outputs, "RPCClient": rpc_client_var}, - attrs={"endpoints": pserver_endpoints, - "epmap": eplist}) + attrs={ + "endpoints": pserver_endpoints, + "epmap": eplist, + "sync_mode": self.sync_mode + }) # step4: Concat the parameters splits together after recv. for varname, splited_var in param_var_mapping.iteritems(): if len(splited_var) <= 1: @@ -356,7 +364,7 @@ class DistributeTranspiler: type=v.type, dtype=v.dtype, shape=v.shape) - if self.trainer_num > 1: + if self.sync_mode and self.trainer_num > 1: for trainer_id in xrange(self.trainer_num): var = pserver_program.global_block().create_var( name="%s.trainer_%d" % (orig_var_name, trainer_id), @@ -402,13 +410,13 @@ class DistributeTranspiler: for op in self.optimize_ops: if op.type == "scale": for in_name in op.input_arg_names: - if in_name.startswith("beta1_pow_acc") or\ - in_name.startswith("beta2_pow_acc"): + if in_name.startswith("beta1_pow_acc") or \ + in_name.startswith("beta2_pow_acc"): global_ops.append(op) - def __append_optimize_op__(op, block): + def __append_optimize_op__(op, block, grad_to_block_id): if self._is_opt_op(op): - self._append_pserver_ops(block, op, endpoint, + self._append_pserver_ops(block, op, endpoint, grad_to_block_id, default_main_program()) else: self._append_pserver_non_opt_ops(block, op) @@ -422,21 +430,22 @@ class DistributeTranspiler: self._append_pserver_non_opt_ops(lr_decay_block, op) # append op to the current block + grad_to_block_id = [] pre_block_idx = pserver_program.num_blocks - 1 for idx, opt_op in enumerate(opt_op_on_pserver): per_opt_block = pserver_program.create_block(pre_block_idx) for _, op in enumerate(self.optimize_ops): # optimizer is connected to itself if ufind.is_connected(op, opt_op) and op not in global_ops: - __append_optimize_op__(op, per_opt_block) + __append_optimize_op__(op, per_opt_block, grad_to_block_id) # append global ops - opt_state_block = None if global_ops: opt_state_block = pserver_program.create_block( pserver_program.num_blocks - 1) for glb_op in global_ops: - __append_optimize_op__(glb_op, opt_state_block) + __append_optimize_op__(glb_op, opt_state_block, + grad_to_block_id) # NOT USED: single block version: # @@ -472,7 +481,9 @@ class DistributeTranspiler: "OptimizeBlock": pserver_program.block(1), "endpoint": endpoint, "Fanin": self.trainer_num, - "PrefetchBlock": prefetch_block + "PrefetchBlock": prefetch_block, + "sync_mode": self.sync_mode, + "grad_to_block_id": grad_to_block_id }) pserver_program.sync_with_cpp() @@ -683,17 +694,6 @@ class DistributeTranspiler: self.table_name)], persistable=False) - # create grad vars in pserver program - table_grad_var = self.table_param_grad[1] - table_grad_list = [ - pserver_program.global_block().create_var( - name="%s.trainer_%d.pserver_%d" % - (table_grad_var.name, index, pserver_index), - type=table_grad_var.type, - shape=table_grad_var.shape, - dtype=table_grad_var.dtype) for index in range(self.trainer_num) - ] - # create table optimize block in pserver program table_opt_op = [ op for op in self.optimize_ops @@ -703,11 +703,24 @@ class DistributeTranspiler: # only support sgd now assert table_opt_op.type == "sgd" - # append sum op for table_grad_list - table_opt_block.append_op( - type="sum", - inputs={"X": table_grad_list}, - outputs={"Out": [grad_var]}) + if self.sync_mode: + # create grad vars in pserver program + table_grad_var = self.table_param_grad[1] + table_grad_list = [ + pserver_program.global_block().create_var( + name="%s.trainer_%d.pserver_%d" % + (table_grad_var.name, index, pserver_index), + type=table_grad_var.type, + shape=table_grad_var.shape, + dtype=table_grad_var.dtype) + for index in range(self.trainer_num) + ] + + # append sum op for table_grad_list + table_opt_block.append_op( + type="sum", + inputs={"X": table_grad_list}, + outputs={"Out": [grad_var]}) lr_var = pserver_program.global_block().vars[table_opt_op.input( "LearningRate")[0]] @@ -746,7 +759,7 @@ class DistributeTranspiler: for varname, splited in block_map.iteritems(): orig_var = program.global_block().var(varname) if len(splited) == 1: - if add_trainer_suffix: + if self.sync_mode and add_trainer_suffix: new_var_name = "%s.trainer_%d" % \ (orig_var.name, self.trainer_id) program.global_block().rename_var(varname, new_var_name) @@ -770,7 +783,7 @@ class DistributeTranspiler: if len(orig_shape) >= 2: splited_shape.extend(orig_shape[1:]) new_var_name = "" - if add_trainer_suffix: + if self.sync_mode and add_trainer_suffix: new_var_name = "%s.block%d.trainer_%d" % \ (varname, i, self.trainer_id) else: @@ -879,7 +892,7 @@ class DistributeTranspiler: return orig_var_name def _append_pserver_ops(self, optimize_block, opt_op, endpoint, - origin_program): + grad_to_block_id, origin_program): program = optimize_block.program pserver_block = program.global_block() new_inputs = dict() @@ -900,7 +913,9 @@ class DistributeTranspiler: return merged_var = \ pserver_block.vars[self._orig_varname(grad_block.name)] - if self.trainer_num > 1: + grad_to_block_id.append(merged_var.name + ":" + str( + optimize_block.idx)) + if self.sync_mode and self.trainer_num > 1: vars2merge = [] for i in xrange(self.trainer_num): per_trainer_name = "%s.trainer_%d" % \ @@ -918,6 +933,7 @@ class DistributeTranspiler: inputs={"X": merged_var}, outputs={"Out": merged_var}, attrs={"scale": 1.0 / float(self.trainer_num)}) + new_inputs[key] = merged_var elif key == "Param": # param is already created on global program diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index 340882ea9e..53486ecffc 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -1070,16 +1070,25 @@ class Program(object): for t in targets: if not isinstance(t, Operator): if isinstance(t, Variable): - if t.op is None: - global_block = self.global_block() - for op in global_block.ops: - if t.name in op.output_arg_names: - t.op = op - break + # After transpiler processing, the op that output this + # variable maybe has been changed, so t.op is not reliable + # and we need to find the current op that generate this + # variable here. + t.op = None + global_block = self.global_block() + for idx, op in enumerate(global_block.ops): + if t.name in op.output_arg_names: + t.op = op + break + t = t.op + if t is None: + raise ValueError( + "The target variable must have an " + "associated operator that generates it.") else: - raise ValueError(("All targets of prune() can only be " - "Variable or Operator.")) + raise ValueError("All targets of prune() can only be " + "Variable or Operator.") targets_idx.append([t.block.idx, t.idx]) res = Program() diff --git a/python/paddle/fluid/inference_transpiler.py b/python/paddle/fluid/inference_transpiler.py index 39b01610f9..f4ad717b9e 100644 --- a/python/paddle/fluid/inference_transpiler.py +++ b/python/paddle/fluid/inference_transpiler.py @@ -121,7 +121,60 @@ class InferenceTranspiler: # And a better solution will be considered later. program = program.clone() + def float16_transpile(self, program, place, scope=None): + ''' + Transpile the program desc and cast the weights to float16 data type to + enable float16 inference. + + Since the operator in a program desc will automatically choose the + right compute kernel to run based on the data type of the input tensor. + We actually don't need to change the program desc to run in float16 mode. + + However, in this way, users who are used to feeding and fetching tensors + of float32 data type when running typical inference may find it confusing + and difficult to run inference in float16 mode as they need to convert + input data to float16 dtype and then convert the results back to float32 + dtype to match the rest of code. + + So this function appends cast ops to the program desc where necessary so + that users are able to run inference in float16 mode while providing input + tensor (feed_holder) of float data type and obtaining output tensor + (fetch_holder) of float data type. + + Moreover, it is desired that when we have the scope and program desc to run + inference in float32 mode, we can use a single API to do the necessary + modification and then user can run float16 inference on the fly. To make + this happen, this function also create new parameters in the scope to have the + converted float16 weights and change the operators in program desc to use + these new parameters. + + :param program: program to transpile + :type program: Program + :param place: inference place + :type place: Place + :param scope: inference scope + :type scope: Scope + ''' + if scope is None: + scope = global_scope() + + self.scope = scope + self.place = place + self.block = program.block(0) + self.input_map = {} # store the input names should be adjusted + + self._modify_feed_fetch() + self._convert_param_to_float16() + self._adjust_input(skip=True) + self._remove_unused_var() + + # TODO(luotao): use clone() method to flush the program.desc in force, + # since some large program.desc will not be flushed immediately. + # And a better solution will be considered later. + program = program.clone() + # ====================== private transpiler functions ===================== + def _insert_bias_op(self, index, current_op, bn_op): ''' Construct elementwise_add operator for adding bias @@ -216,9 +269,27 @@ class InferenceTranspiler: # collect the renamed input self.input_map[bn_op.output("Y")[0]] = bias_op.output("Out")[0] - def _adjust_input(self): + def _adjust_input(self, skip=False): + ''' + Change the input variable name in operators. + + When we are in the process of modifying a program desc, we usually + replace some variables with some other variables, where we create + a dictionary input_map to record the one-to-one correspondence + between each old variable and the new one. + + After that, this function will search all the operators that use the + old variables and change the info in op to use the new variables. There + maybe some exceptions to this rule when we are using the float16 transpiler + and insert cast ops to cast float32 variable to float16 one. After we + insert the cast op to cast var_1 to var_1_fp16, we don't want to change + the input of cast op to var_1_fp16 after using this function. + ''' + skip_ops = {"cast"} for i in range(len(self.block.ops)): current_op = self.block.ops[i] + if skip and current_op.type in skip_ops: + continue for input_arg in current_op.input_arg_names: if input_arg in self.input_map: current_op.rename_input(input_arg, @@ -238,3 +309,138 @@ class InferenceTranspiler: for var in self.block.vars.keys(): if var not in args: self.block.remove_var(var) + + def _modify_feed_fetch(self): + ''' + Modify feed fetch op/vars for float16 inference. + + For each feed op: + feed_op->feed_target_var + + Change it to: + feed_op->feed_target_var->cast_op(from other dtype to float16)->tmp_var + + For each fetch op: + fetch_target_var->fetch_op + + Change it to: + tmp_var->cast_op(from float16 to other dtype)->fetch_target_var->fetch_op + + :return: None + ''' + + def find_op(var): + # It is possible that var.op is not up to date after some + # modifications to program desc. Here we force to make it up to date. + var.op = None + for op in self.block.ops: + if var.name in op.output_arg_names: + var.op = op + break + + if var.op is None: + raise ValueError("The target variable must have an " + "associated operator that generates it.") + + i = 0 + while i < len(self.block.ops): + cur_op = self.block.ops[i] + if cur_op.type == "feed": + var_name = cur_op.output("Out")[0] + tmp_var_name = var_name + ".fp16" + var = self.block.vars[var_name] + tmp_var = self.block.create_var( + name=tmp_var_name.encode('ascii'), + type=var.type, + dtype=core.VarDesc.VarType.FP16, + shape=var.shape, + persistable=var.persistable) + self.block.insert_op( + i + 1, + type="cast", + inputs={"X": var}, + outputs={"Out": tmp_var}, + attrs={ + 'in_dtype': int(var.dtype), + 'out_dtype': int(tmp_var.dtype) + }) + self.input_map[var_name] = tmp_var_name + i = i + 1 + elif cur_op.type == "fetch": + var_name = cur_op.input("X")[0] + tmp_var_name = var_name + ".fp16" + var = self.block.vars[var_name] + tmp_var = self.block.create_var( + name=tmp_var_name.encode('ascii'), + type=var.type, + dtype=core.VarDesc.VarType.FP16, + shape=var.shape, + persistable=var.persistable) + find_op(var) + var.op.rename_output(var_name, tmp_var_name) + self.block.insert_op( + i, + type="cast", + inputs={"X": tmp_var}, + outputs={"Out": var}, + attrs={ + 'in_dtype': int(tmp_var.dtype), + 'out_dtype': int(var.dtype) + }) + i = i + 1 + i = i + 1 + + def _convert_param_to_float16(self): + def _get_no_fp16_conversion_var_names(): + ''' + Get the set of input variable names that shouldn't be converted to float16. + + When we want to run inference in float16 mode, most parameters need to be + firstly converted to float16. However, there are some parameters that + shouldn't be converted to float16 because the corresponding operator + requires float32 parameters even in float16 mode (when the input data is + of float16 data type). Currently, the only operator that has this exclusion + is the batch norm op. + + :return: set of input variable names + :type var_names: set + ''' + op_names = {'batch_norm'} + var_names = [] + for op in self.block.ops: + if op.type in op_names: + var_names += op.input_arg_names + return set(var_names) + + def _should_be_converted(var): + return var.persistable and \ + var.name not in self.no_conversion_vars and \ + var.type != core.VarDesc.VarType.FEED_MINIBATCH and \ + var.type != core.VarDesc.VarType.FETCH_LIST + + self.no_conversion_vars = _get_no_fp16_conversion_var_names() + conversion_var_list = filter(_should_be_converted, + self.block.vars.values()) + for var in conversion_var_list: + fp16_var_name = var.name + ".fp16" + fp16_var = self.block.create_parameter( + name=fp16_var_name.encode('ascii'), + type=var.type, + dtype=core.VarDesc.VarType.FP16, + shape=var.shape) + + # cast the data in the tensor of the original var to float16 + # data type and store it in the tensor of the new float16 var + self.scope.var(fp16_var_name) + fp16_tensor = self.scope.find_var(fp16_var_name).get_tensor() + tensor = np.array(self.scope.find_var(var.name).get_tensor()) + # After the old tensor data is converted to np.float16, view(np.uint16) + # is used so that the internal memory of the numpy array will be + # reinterpreted to be of np.uint16 data type, which is binded to fluid + # float16 data type via the help of pybind in tensor_py.h. + fp16_tensor.set( + tensor.astype(np.float16).view(np.uint16), self.place) + + # old var will be replaced by the fp16 var in program desc + self.input_map[var.name] = fp16_var_name + self.block.remove_var(var.name) diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index f7f1ca2598..08b8a878b6 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -336,7 +336,7 @@ def save_inference_model(dirname, if main_program is None: main_program = default_main_program() - copy_program = main_program + copy_program = main_program.clone() if not os.path.isdir(dirname): os.makedirs(dirname) diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 9a0c328033..7f16bf2a0c 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -79,6 +79,7 @@ __all__ = [ 'lrn', 'pad', 'label_smooth', + 'roi_pool', ] @@ -3759,3 +3760,53 @@ def label_smooth(label, outputs={"Out": smooth_label}, attrs={"epsilon": float(epsilon)}) return smooth_label + + +def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): + """ + Region of interest pooling (also known as RoI pooling) is to perform + is to perform max pooling on inputs of nonuniform sizes to obtain + fixed-size feature maps (e.g. 7*7). + The operator has three steps: + 1. Dividing each region proposal into equal-sized sections with + the pooled_width and pooled_height + 2. Finding the largest value in each section + 3. Copying these max values to the output buffer + + Args: + input (Variable): The input for ROI pooling. + rois (Variable): ROIs (Regions of Interest) to pool over. It should + be a 2-D one level LoTensor of shape [num_rois, 4]. + The layout is [x1, y1, x2, y2], where (x1, y1) + is the top left coordinates, and (x2, y2) is the + bottom right coordinates. The num_rois is the + total number of ROIs in this batch data. + pooled_height (integer): The pooled output height. Default: 1 + pooled_width (integer): The pooled output width. Default: 1 + spatial_scale (float): Multiplicative spatial scale factor. To + translate ROI coords from their input scale + to the scale used when pooling. Default: 1.0 + + Returns: + pool_out (Variable): The output is a 4-D tensor of the shape + (num_rois, channels, pooled_h, pooled_w). + + Examples: + pool_out = fluid.layers.roi_pool(input=x, rois=rois, 7, 7, 1.0) + """ + helper = LayerHelper('roi_pool', **locals()) + dtype = helper.input_dtype() + pool_out = helper.create_tmp_variable(dtype) + argmaxes = helper.create_tmp_variable(dtype='int32') + helper.append_op( + type="roi_pool", + inputs={"X": input, + "ROIs": rois}, + outputs={"Out": pool_out, + "Argmax": argmaxes}, + attrs={ + "pooled_height": pooled_height, + "pooled_width": pooled_width, + "spatial_scale": spatial_scale + }) + return pool_out diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index da066c34bd..4be0dc6a6b 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -193,10 +193,7 @@ def assign(input, output): helper = LayerHelper('assign', **locals()) if isinstance(input, Variable): helper.append_op( - type='scale', - inputs={'X': [input]}, - outputs={'Out': [output]}, - attrs={'scale': 1.0}) + type='assign', inputs={'X': [input]}, outputs={'Out': [output]}) elif isinstance(input, numpy.ndarray): dtype = convert_np_dtype_to_dtype_(input.dtype) if dtype == VarDesc.VarType.FP32: diff --git a/python/paddle/fluid/tests/book/test_image_classification.py b/python/paddle/fluid/tests/book/test_image_classification.py index d3c14b83fa..09f994c370 100644 --- a/python/paddle/fluid/tests/book/test_image_classification.py +++ b/python/paddle/fluid/tests/book/test_image_classification.py @@ -244,7 +244,7 @@ def infer(use_cuda, save_dirname=None): assert len(results[0]) == len(transpiler_results[0]) for i in range(len(results[0])): np.testing.assert_almost_equal( - results[0][i], transpiler_results[0][i], decimal=6) + results[0][i], transpiler_results[0][i], decimal=5) print("infer results: ", results[0]) @@ -252,6 +252,26 @@ def infer(use_cuda, save_dirname=None): fetch_targets, exe, inference_transpiler_program) + if use_cuda and fluid.core.is_float16_supported(place): + # Use float16_transpiler to speedup + fp16_transpiler_program = inference_transpiler_program.clone() + t.float16_transpile(fp16_transpiler_program, place) + + fp16_results = exe.run(fp16_transpiler_program, + feed={feed_target_names[0]: tensor_img}, + fetch_list=fetch_targets) + + assert len(results[0]) == len(fp16_results[0]) + for i in range(len(results[0])): + np.testing.assert_almost_equal( + results[0][i], fp16_results[0][i], decimal=2) + + print("float16 infer results: ", fp16_results[0]) + + fluid.io.save_inference_model("float16_" + save_dirname, + feed_target_names, fetch_targets, exe, + fp16_transpiler_program) + def main(net_type, use_cuda, is_local=True): if use_cuda and not fluid.core.is_compiled_with_cuda(): diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py new file mode 100644 index 0000000000..bffb4f3b66 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py @@ -0,0 +1,95 @@ +# 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 + + +def bilinear_interp_np(input, out_h, out_w): + batch_size, channel, in_h, in_w = input.shape + if out_h > 1: + ratio_h = (in_h - 1.0) / (out_h - 1.0) + else: + ratio_h = 0.0 + if out_w > 1: + ratio_w = (in_w - 1.0) / (out_w - 1.0) + else: + ratio_w = 0.0 + + out = np.zeros((batch_size, channel, out_h, out_w)) + for i in range(out_h): + h = int(ratio_h * i) + hid = 1 if h < in_h - 1 else 0 + h1lambda = ratio_h * i - h + h2lambda = 1.0 - h1lambda + for j in range(out_w): + w = int(ratio_w * j) + wid = 1 if w < in_w - 1 else 0 + w1lambda = ratio_w * j - w + w2lambda = 1.0 - w1lambda + + out[:, :, i, j] = h2lambda*(w2lambda*input[:, :, h, w] + + w1lambda*input[:, :, h, w+wid]) + \ + h1lambda*(w2lambda*input[:, :, h+hid, w] + + w1lambda*input[:, :, h+hid, w+wid]) + return out.astype("float32") + + +class TestBilinearInterpOp(OpTest): + def setUp(self): + self.init_test_case() + self.op_type = "bilinear_interp" + input_np = np.random.random(self.input_shape).astype("float32") + output_np = bilinear_interp_np(input_np, self.out_h, self.out_w) + + self.inputs = {'X': input_np} + self.attrs = {'out_h': self.out_h, 'out_w': self.out_w} + self.outputs = {'Out': output_np} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out', in_place=True) + + def init_test_case(self): + self.input_shape = [2, 3, 4, 4] + self.out_h = 2 + self.out_w = 2 + + +class TestCase1(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [4, 1, 7, 8] + self.out_h = 1 + self.out_w = 1 + + +class TestCase2(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [3, 3, 9, 6] + self.out_h = 12 + self.out_w = 12 + + +class TestCase3(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [1, 1, 128, 64] + self.out_h = 64 + self.out_w = 128 + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py b/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py index e33436b63c..8f62ac20a5 100644 --- a/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py +++ b/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py @@ -14,6 +14,7 @@ import unittest import numpy as np +import numpy.random as random import sys import math from op_test import OpTest @@ -25,14 +26,27 @@ class TestIOUSimilarityOp(OpTest): def setUp(self): self.op_type = "iou_similarity" - self.boxes1 = np.array( - [[4.0, 3.0, 7.0, 5.0], [5.0, 6.0, 10.0, 7.0]]).astype('float32') - self.boxes2 = np.array([[3.0, 4.0, 6.0, 8.0], [14.0, 14.0, 15.0, 15.0], - [0.0, 0.0, 20.0, 20.0]]).astype('float32') - self.output = np.array( - [[2.0 / 16.0, 0, 6.0 / 400.0], - [1.0 / 16.0, 0.0, 5.0 / 400.0]]).astype('float32') - + self.boxes1 = random.rand(2, 4).astype('float32') + self.boxes2 = random.rand(3, 4).astype('float32') + self.output = random.rand(2, 3).astype('float32') + for row in range(self.boxes1.shape[0]): + for col in range(self.boxes2.shape[0]): + xmin1, ymin1, xmax1, ymax1 = self.boxes1[row] + xmin2, ymin2, xmax2, ymax2 = self.boxes2[col] + area1 = (ymax1 - ymin1) * (xmax1 - xmin1) + area2 = (ymax2 - ymin2) * (xmax2 - xmin2) + inter_xmax = min(xmax1, xmax2) + inter_ymax = min(ymax1, ymax2) + inter_xmin = max(xmin1, xmin2) + inter_ymin = max(ymin1, ymin2) + inter_height = inter_ymax - inter_ymin + inter_width = inter_xmax - inter_xmin + inter_height = max(inter_height, 0) + inter_width = max(inter_width, 0) + inter_area = inter_width * inter_height + union_area = area1 + area2 - inter_area + sim_score = inter_area / union_area + self.output[row, col] = sim_score self.inputs = {'X': self.boxes1, 'Y': self.boxes2} self.outputs = {'Out': self.output} diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 17d6afdee1..c5414abf0f 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -359,6 +359,16 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(indices) print(str(program)) + def test_roi_pool(self): + program = Program() + with program_guard(program): + x = layers.data(name="x", shape=[256, 30, 30], dtype="float32") + rois = layers.data( + name="rois", shape=[4], dtype="float32", lod_level=1) + output = layers.roi_pool(x, rois, 7, 7, 0.6) + self.assertIsNotNone(output) + print(str(program)) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_roi_pool_op.py b/python/paddle/fluid/tests/unittests/test_roi_pool_op.py index e556d51b02..3d754aff3a 100644 --- a/python/paddle/fluid/tests/unittests/test_roi_pool_op.py +++ b/python/paddle/fluid/tests/unittests/test_roi_pool_op.py @@ -25,7 +25,7 @@ class TestROIPoolOp(OpTest): self.make_rois() self.calc_roi_pool() - self.inputs = {'X': self.x, 'ROIs': self.rois} + self.inputs = {'X': self.x, 'ROIs': (self.rois[:, 1:5], self.rois_lod)} self.attrs = { 'spatial_scale': self.spatial_scale, @@ -36,7 +36,7 @@ class TestROIPoolOp(OpTest): self.outputs = {'Out': self.outs, 'Argmax': self.argmaxes} def init_test_case(self): - self.batch_size = 5 + self.batch_size = 3 self.channels = 3 self.height = 6 self.width = 4 @@ -47,7 +47,6 @@ class TestROIPoolOp(OpTest): self.spatial_scale = 1.0 / 4.0 self.pooled_height = 2 self.pooled_width = 2 - self.rois_num = 2 self.x = np.random.random(self.x_dim).astype('float32') @@ -106,20 +105,24 @@ class TestROIPoolOp(OpTest): def make_rois(self): rois = [] - batch_ids = np.random.randint(0, self.batch_size, size=self.rois_num) - for i in range(self.rois_num): - x1 = np.random.random_integers( - 0, self.width / self.spatial_scale - self.pooled_width) - y1 = np.random.random_integers( - 0, self.height / self.spatial_scale - self.pooled_height) - - x2 = np.random.random_integers(x1 + self.pooled_width, - self.width / self.spatial_scale) - y2 = np.random.random_integers(y1 + self.pooled_height, - self.height / self.spatial_scale) - - roi = [batch_ids[i], x1, y1, x2, y2] - rois.append(roi) + self.rois_lod = [[]] + for bno in range(self.batch_size): + self.rois_lod[0].append(len(rois)) + for i in range(bno + 1): + x1 = np.random.random_integers( + 0, self.width / self.spatial_scale - self.pooled_width) + y1 = np.random.random_integers( + 0, self.height / self.spatial_scale - self.pooled_height) + + x2 = np.random.random_integers(x1 + self.pooled_width, + self.width / self.spatial_scale) + y2 = np.random.random_integers(y1 + self.pooled_height, + self.height / self.spatial_scale) + + roi = [bno, x1, y1, x2, y2] + rois.append(roi) + self.rois_lod[0].append(len(rois)) + self.rois_num = len(rois) self.rois = np.array(rois).astype("int64") def setUp(self): diff --git a/tools/aws_benchmarking/README.md b/tools/aws_benchmarking/README.md index 22a468466a..4fdd4b0de4 100644 --- a/tools/aws_benchmarking/README.md +++ b/tools/aws_benchmarking/README.md @@ -77,10 +77,10 @@ Training nodes will run your `ENTRYPOINT` script with the following environment Now let's start the training process: ```bash -docker run -i -v $HOME/.aws:/root/.aws -v :/root/.pem \ +docker run -i -v $HOME/.aws:/root/.aws -v :/root/.pem \ putcn/paddle_aws_client \ --action create \ ---key_name \ +--key_name \ --security_group_id \ --docker_image myreponame/paddle_benchmark \ --pserver_count 2 \ @@ -154,8 +154,31 @@ Master exposes 4 major services: ### Parameters -TBD, please refer to client/cluster_launcher.py for now + - key_name: required, aws key pair name + - security_group_id: required, the security group id associated with your VPC + - vpc_id: The VPC in which you wish to run test, if not provided, this tool will use your default VPC. + - subnet_id: The Subnet_id in which you wish to run test, if not provided, this tool will create a new sub net to run test. + - pserver_instance_type: your pserver instance type, c5.2xlarge by default, which is a memory optimized machine. + - trainer_instance_type: your trainer instance type, p2.8xlarge by default, which is a GPU machine with 8 cards. + - task_name: the name you want to identify your job, if not provided, this tool will generate one for you. + - pserver_image_id: ami id for system image. Please note, although the default one has nvidia-docker installed, pserver is always launched with `docker` instead of `nvidia-docker`, please DO NOT init your training program with GPU place. + - pserver_command: pserver start command, format example: python,vgg.py,batch_size:128,is_local:no, which will be translated as `python vgg.py --batch_size 128 --is_local no` when trying to start the training in pserver. "--device CPU" is passed as default. + - trainer_image_id: ami id for system image, default one has nvidia-docker ready. + - trainer_command: trainer start command. Format is the same as pserver's, "--device GPU" is passed as default. + - availability_zone: aws zone id to place ec2 instances, us-east-2a by default. + - trainer_count: Trainer count, 1 by default. + - pserver_count: Pserver count, 1 by default. + - action: create|cleanup|status, "create" by default. + - pserver_port: the port for pserver to open service, 5436 by default. + - docker_image: the training docker image id. + - master_service_port: the port for master to open service, 5436 by default. + - master_server_public_ip: the master service ip, this is required when action is not "create" + - master_docker_image: master's docker image id, "putcn/paddle_aws_master:latest" by default + - no_clean_up: no instance termination when training is finished or failed when this value is set "yes". This is for debug purpose, so that you can inspect into the instances when the process is finished. + ### Trouble shooting -TBD + 1. How to check logs + + Master log is served at `http://:/status`, and you can list all the log files from `http://:/logs`, and access either one of them by `http://:/log/` diff --git a/tools/aws_benchmarking/server/cluster_master.py b/tools/aws_benchmarking/server/cluster_master.py index 7952e61159..1333a942bf 100644 --- a/tools/aws_benchmarking/server/cluster_master.py +++ b/tools/aws_benchmarking/server/cluster_master.py @@ -640,6 +640,7 @@ def start_server(args): elif request_path == "/cleanup": self._set_headers() logging.info("Received request to cleanup cluster") + args.no_clean_up = False cleanup(args.task_name) self.wfile.write("cleanup in progress") diff --git a/tools/manylinux1/build_scripts/install_nccl2.sh b/tools/manylinux1/build_scripts/install_nccl2.sh index 7efc1fe865..282c5c290d 100644 --- a/tools/manylinux1/build_scripts/install_nccl2.sh +++ b/tools/manylinux1/build_scripts/install_nccl2.sh @@ -1,11 +1,18 @@ #!/bin/bash -DEB="nccl-repo-ubuntu1604-2.1.4-ga-cuda8.0_1-1_amd64.deb" +VERSION=$(nvcc --version | grep release | grep -oEi "release ([0-9]+)\.([0-9])"| sed "s/release //") +if [ "$VERSION" == "9.0" ]; then + DEB="nccl-repo-ubuntu1604-2.1.15-ga-cuda9.0_1-1_amd64.deb" + URL="http://nccl2-deb.gz.bcebos.com/nccl-repo-ubuntu1604-2.1.15-ga-cuda9.0_1-1_amd64.deb" +else + DEB="nccl-repo-ubuntu1604-2.1.15-ga-cuda8.0_1-1_amd64.deb" + URL="http://nccl2-deb.gz.bcebos.com/nccl-repo-ubuntu1604-2.1.15-ga-cuda8.0_1-1_amd64.deb" +fi + 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" +wget -O $DIR/$DEB $URL cd $DIR && ar x $DEB && tar xf data.tar.xz DEBS=$(find ./var/ -name "*.deb")