diff --git a/AUTHORS.md b/AUTHORS.md index 4ee05420982d13f686cf13e8957ce41dfcdd2cb8..11f227be7148d8d6e055538347a8c31679406c84 100644 --- a/AUTHORS.md +++ b/AUTHORS.md @@ -4,6 +4,7 @@ | backyes | Yan-Fei Wang | | baiyfbupt | Yi-Fan Bai | | beckett1124 | Bin Qi | +| ChengduoZH | Cheng-Duo Zhao| | chengxiaohua1105 | Xiao-Hua Cheng | | cxwangyi, yiwangbaidu, wangkuiyi | Yi Wang | | cxysteven | Xing-Yi Cheng | diff --git a/Dockerfile b/Dockerfile index 80a96983ec1ca6b9ec440f7e95de6c328eb1ed40..4d6165b79a1d94b8f27d7f3ee1b6e2cee5992d31 100644 --- a/Dockerfile +++ b/Dockerfile @@ -29,7 +29,7 @@ RUN apt-get update && \ wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \ curl sed grep graphviz libjpeg-dev zlib1g-dev \ python-matplotlib gcc-4.8 g++-4.8 \ - automake locales clang-format swig doxygen cmake \ + automake locales clang-format swig cmake \ liblapack-dev liblapacke-dev \ clang-3.8 llvm-3.8 libclang-3.8-dev \ net-tools libtool ccache && \ diff --git a/benchmark/fluid/Dockerfile b/benchmark/fluid/Dockerfile new file mode 100644 index 0000000000000000000000000000000000000000..46140a9d1be01a50cd74dab2799e3731e8d3debd --- /dev/null +++ b/benchmark/fluid/Dockerfile @@ -0,0 +1,22 @@ +FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04 +RUN apt-get update && apt-get install -y python python-pip iputils-ping libgtk2.0-dev wget vim net-tools iftop +RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/lib/libcudnn.so && ln -s /usr/lib/x86_64-linux-gnu/libnccl.so.2 /usr/lib/libnccl.so +RUN pip install -U pip +RUN pip install -U kubernetes opencv-python paddlepaddle + +# IMPORTANT: +# Add "ENV http_proxy=http://ip:port" if your download is slow, and don't forget to unset it at runtime. + +RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()\npaddle.dataset.flowers.fetch()" | python' +RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.mnist.train()\npaddle.dataset.mnist.test()\npaddle.dataset.imdb.fetch()" | python' +RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.imikolov.fetch()" | python' +RUN pip uninstall -y paddlepaddle && mkdir /workspace + +ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin +ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/k8s_tools.py /root + +ADD *.whl / +RUN pip install /*.whl && rm -f /*.whl && chmod +x /usr/bin/paddle_k8s + +ENV LD_LIBRARY_PATH=/usr/local/lib +ADD fluid_benchmark.py dataset.py models/ /workspace/ diff --git a/benchmark/fluid/README.md b/benchmark/fluid/README.md index 7071e9fdcd394a5a4db4d0d599610a72d98c0a3c..1b0c7dce8bd6faab0c4c59caa1cbe337483cbd16 100644 --- a/benchmark/fluid/README.md +++ b/benchmark/fluid/README.md @@ -44,11 +44,25 @@ Currently supported `--model` argument include: ## Run Distributed Benchmark on Kubernetes Cluster +You may need to build a Docker image before submitting a cluster job onto Kubernetes, or you will +have to start all those processes mannually on each node, which is not recommended. + +To build the Docker image, you need to choose a paddle "whl" package to run with, you may either +download it from +http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_en.html or +build it by your own. Once you've got the "whl" package, put it under the current directory and run: + +```bash +docker build -t [your docker image name]:[your docker image tag] . +``` + +Then push the image to a Docker registry that your Kubernetes cluster can reach. + We provide a script `kube_gen_job.py` to generate Kubernetes yaml files to submit distributed benchmark jobs to your cluster. To generate a job yaml, just run: ```bash -python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --parallel 1 --device GPU --update_method pserver " --disttype pserver +python kube_gen_job.py --jobname myjob --pscpu 4 --cpu 8 --gpu 8 --psmemory 20 --memory 40 --pservers 4 --trainers 4 --entry "python fluid_benchmark.py --model mnist --gpus 8 --device GPU --update_method pserver " --disttype pserver ``` Then the yaml files are generated under directory `myjob`, you can run: diff --git a/benchmark/fluid/kube_gen_job.py b/benchmark/fluid/kube_gen_job.py index 39ba207fd96f71563504017e77dc0e87c249b3f8..9da8a69af1d7b671b2648b1b3702776c1c0650b0 100644 --- a/benchmark/fluid/kube_gen_job.py +++ b/benchmark/fluid/kube_gen_job.py @@ -49,7 +49,7 @@ def parse_args(): parser.add_argument( '--fluid', default=1, type=int, help='whether is fluid job') parser.add_argument( - '--rdma', action='store_ture', help='whether mount rdma libs') + '--rdma', action='store_true', help='whether mount rdma libs') parser.add_argument( '--disttype', default="pserver", diff --git a/benchmark/fluid/run.sh b/benchmark/fluid/run.sh index f6dfd20bf2ee0b668b6d4238d4511253b2233035..afaab5f4de43fa7e94feeed4a1de991351c04b76 100644 --- a/benchmark/fluid/run.sh +++ b/benchmark/fluid/run.sh @@ -37,7 +37,8 @@ nohup stdbuf -oL nvidia-smi \ -l 1 & # mnist # mnist gpu mnist 128 -FLAGS_benchmark=true stdbuf -oL python fluid/mnist.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=mnist \ --device=GPU \ --batch_size=128 \ --skip_batch_num=5 \ @@ -46,7 +47,8 @@ FLAGS_benchmark=true stdbuf -oL python fluid/mnist.py \ # vgg16 # gpu cifar10 128 -FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=vgg16 \ --device=GPU \ --batch_size=128 \ --skip_batch_num=5 \ @@ -54,7 +56,8 @@ FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \ 2>&1 | tee -a vgg16_gpu_128.log # flowers gpu 128 -FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=vgg16 \ --device=GPU \ --batch_size=32 \ --data_set=flowers \ @@ -64,40 +67,39 @@ FLAGS_benchmark=true stdbuf -oL python fluid/vgg16.py \ # resnet50 # resnet50 gpu cifar10 128 -FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=resnet50 \ --device=GPU \ --batch_size=128 \ --data_set=cifar10 \ - --model=resnet_cifar10 \ --skip_batch_num=5 \ --iterations=30 \ 2>&1 | tee -a resnet50_gpu_128.log # resnet50 gpu flowers 64 -FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=resnet50 \ --device=GPU \ --batch_size=64 \ --data_set=flowers \ - --model=resnet_imagenet \ --skip_batch_num=5 \ --iterations=30 \ 2>&1 | tee -a resnet50_gpu_flowers_64.log # lstm # lstm gpu imdb 32 # tensorflow only support batch=32 -FLAGS_benchmark=true stdbuf -oL python fluid/stacked_dynamic_lstm.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=stacked_dynamic_lstm \ --device=GPU \ --batch_size=32 \ --skip_batch_num=5 \ --iterations=30 \ - --hidden_dim=512 \ - --emb_dim=512 \ - --crop_size=1500 \ 2>&1 | tee -a lstm_gpu_32.log # seq2seq # seq2seq gpu wmb 128 -FLAGS_benchmark=true stdbuf -oL python fluid/machine_translation.py \ +FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \ + --model=machine_translation \ --device=GPU \ --batch_size=128 \ --skip_batch_num=5 \ diff --git a/doc/fluid/api/layers.rst b/doc/fluid/api/layers.rst index f53da4d194f8d2428b4121fa1bb31f3fc95a9f64..dbb99d3c03f39f650b2cb0dbe8ee49cd413db6e3 100644 --- a/doc/fluid/api/layers.rst +++ b/doc/fluid/api/layers.rst @@ -1009,3 +1009,9 @@ ____ .. autofunction:: paddle.fluid.layers.upsampling_bilinear2d :noindex: +gather +____ + +.. autofunction:: paddle.fluid.layers.gather + :noindex: + diff --git a/doc/fluid/getstarted/Developer's_Guide_to_Paddle_Fluid.md b/doc/fluid/getstarted/Developer's_Guide_to_Paddle_Fluid.md index 0c0156c8e46378e7bbeea8072938b8ccfb9ab6d7..79df6c59578e2acf495a3453ab61f069c3f09a49 100644 --- a/doc/fluid/getstarted/Developer's_Guide_to_Paddle_Fluid.md +++ b/doc/fluid/getstarted/Developer's_Guide_to_Paddle_Fluid.md @@ -86,7 +86,7 @@

- +

--- @@ -123,12 +123,12 @@ - 在科学计算领域，计算图是一种描述计算的经典方式。下图展示了从前向计算图（蓝色）开始，通过添加反向（红色）和优化算法相关（绿色）操作，构建出整个计算图的过程： -- +-

- + - Fluid ==使用`Program`而不是计算图==来描述模型和优化过程。`Program`由`Block`、`Operator`和`Variable`构成，相关概念会在后文详细展开。 - 编译时 Fluid 接受前向计算（这里可以先简单的理解为是一段有序的计算流）`Program`，为这段前向计算按照：前向 -> 反向 -> 梯度 clip -> 正则 -> 优化 的顺序，添加相关 `Operator`和`Variable`到`Program`到完整的计算。 @@ -328,7 +328,7 @@ ---- +--- ### 编译时概念 ：==**[Transpiler](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/motivation/fluid_compiler.md)**== @@ -402,7 +402,7 @@ - `Scope` - 计算相关 - - `Block` + - `Block` - `Kernel`、`OpWithKernel`、`OpWithoutKernel` @@ -439,7 +439,7 @@
-- 执行相关 ：`Executor` +- 执行相关 ：`Executor` @@ -798,7 +798,7 @@ class GPUAllocator : public SystemAllocator { - step 1：添加Place类型，由用户实现添加到框架 - 可以将Place类型理解为一个整数加上一个枚举型，包括：设备号 + 设备类型 - +

@@ -824,7 +824,7 @@ class GPUAllocator : public SystemAllocator { 1. DataType 执行数据类型 FP32/FP64/INT32/INT64 1. Memory layout： 运行时 Tensor 在内存中的排布格式 NCHW、 NHWC 1. 使用的库 - + 来区分Kernel，为同一个operator注册多个 Kernel。 ```cpp @@ -876,7 +876,7 @@ step 3: 运行时的 KernelType 推断和Kernel切换， --- @@ -1107,7 +1107,7 @@ void Run(const framework::Scope &scope,

-

+

@@ -1127,13 +1127,13 @@ void Run(const framework::Scope &scope, - 设计概览 - - 重构概览 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/refactorization.md) - - fluid [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/fluid.md) + - 重构概览 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/refactorization.md) + - fluid [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/fluid.md) - fluid_compiler [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/motivation/fluid_compiler.md) - 核心概念 - variable 描述 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/var_desc.md) - Tensor [->](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.md) - - LoDTensor [->](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) + - LoDTensor [->](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md) - TensorArray [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/tensor_array.md) - Program [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md) - Block [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md) @@ -1152,7 +1152,7 @@ void Run(const framework::Scope &scope, - 支持新设硬件设备库 [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/support_new_device.md) - 添加新的Operator [->](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_cn.md) - 添加新的Kernel [->]( -https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_kernel_en.md) +https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_kernel_en.md) @@ -1167,10 +1167,10 @@ https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_kernel_ Docker编译PaddlePaddle源码: [->](http://www.paddlepaddle.org/docs/develop/documentation/fluid/zh/build_and_install/docker_install_cn.html) - + PaddlePaddle 在 Dockerhub 地址：[->]( https://hub.docker.com/r/paddlepaddle/paddle/tags/) - + 1. 获取PaddlePaddle的Docker镜像 ```bash docker pull paddlepaddle/paddle:latest-dev @@ -1183,7 +1183,7 @@ PaddlePaddle 在 Dockerhub 地址：[->]( ``` 1. 进入docker container后，从源码编译，请参考文档 [->]( http://www.paddlepaddle.org/docs/develop/documentation/fluid/zh/build_and_install/build_from_source_cn.html) - + --- @@ -1196,7 +1196,7 @@ PaddlePaddle 在 Dockerhub 地址：[->]( 1. 开发推荐使用tag为`latest-dev`的镜像，其中打包了所有编译依赖。`latest`及`lastest-gpu`是production镜像，主要用于运行PaddlePaddle程序。 2. 在Docker中运行GPU程序，推荐使用nvidia-docker，[否则需要将CUDA库和设备挂载到Docker容器内](http://www.paddlepaddle.org/docs/develop/documentation/fluid/zh/build_and_install/docker_install_cn.html)。 - + ```bash nvidia-docker run -it -v $PWD/Paddle:/paddle paddlepaddle/paddle:latest-dev /bin/bash ``` @@ -1353,9 +1353,9 @@ Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，CUDA实 } }; ``` - + - + --- ###### 实现带Kernel的Operator step2: 定义Operator类 @@ -1420,11 +1420,11 @@ class ClipOp : public framework::OperatorWithKernel { 2. override InferShape函数（参考 [clip_op](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/clip_op.cc#L24)） 1. 什么是`functor` ? - + - 类或结构体仅重载了`()`，一般是可被多个kernel复用的计算函数。 - + ```cpp template class CrossEntropyFunctor { @@ -1438,9 +1438,9 @@ class ClipOp : public framework::OperatorWithKernel { }; ``` - + - 在 clip_op 内也会看到将一段计算函数抽象为functor的使用法： [->](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/clip_op.h#L27)。 - + --- @@ -1504,7 +1504,7 @@ class ClipKernel : public framework::OpKernel { - 需要注意，Fluid中，不区分Cost Op和中间层Op，所有Op都必须正确处理接收到的梯度 2. 反向Op的输出 - 对可学习参数的求导结果 - - 对所有输入的求导结果 + - 对所有输入的求导结果 @@ -1520,7 +1520,7 @@ class ClipKernel : public framework::OpKernel { 1. 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。 - + ```cpp namespace ops = paddle::operators; REGISTER_OP(clip, ops::ClipOp, ops::ClipOpMaker, clip_grad, @@ -1530,13 +1530,13 @@ class ClipKernel : public framework::OpKernel { REGISTER_OP_CPU_KERNEL( clip_grad, ops::ClipGradKernel); ``` - + - 在上面的代码片段中： 1. `REGISTER_OP` ： 注册`ops::ClipOp`类，类型名为`clip`，该类的`ProtoMaker`为`ops::ClipOpMaker`，注册`ops::ClipOpGrad`，类型名为`clip_grad` 1. `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op，例如：优化算法相关的Op 1. `REGISTER_OP_CPU_KERNEL` ：注册`ops::ClipKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::ClipGradKernel`类 - + 1. 按照同样方法，在`.cu`文件中注册GPU Kernel - 如果CUDA Kernel的实现基于Eigen，需在 `.cu`的开始加上宏定义 `#define EIGEN_USE_GPU` @@ -1593,7 +1593,7 @@ class ClipKernel : public framework::OpKernel { ```bash make test ARGS="-R test_mul_op -V" ``` - + 或者: ``` @@ -1613,7 +1613,7 @@ class ClipKernel : public framework::OpKernel { - 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。 - + --- ### ==10.== 使用相关问题 @@ -1735,7 +1735,7 @@ class ClipKernel : public framework::OpKernel { y_data = np.random.randint(0, 8, [1]).astype("int32") y_tensor = core.Tensor() y_tensor.set(y_data, place) - + x_data = np.random.uniform(0.1, 1, [11, 8]).astype("float32") x_tensor = core.Tensor() x_tensor.set(x_data, place) diff --git a/doc/fluid/getstarted/index_cn.rst b/doc/fluid/getstarted/index_cn.rst index 75af7354be93a6eeabfa9ccf86903505402a7ca6..3daea71d0933a2774227ff2b5e744392ca6b1765 100644 --- a/doc/fluid/getstarted/index_cn.rst +++ b/doc/fluid/getstarted/index_cn.rst @@ -17,3 +17,4 @@ :maxdepth: 1 concepts/use_concepts_cn.rst + developer's_guide_to_paddle_fluid.md diff --git a/doc/fluid/getstarted/index_en.rst b/doc/fluid/getstarted/index_en.rst index 75a43f4af87c34830ec940068196e6ca72640501..fb20bb4f245281c3acf67c417979dc63c144fef3 100644 --- a/doc/fluid/getstarted/index_en.rst +++ b/doc/fluid/getstarted/index_en.rst @@ -16,3 +16,4 @@ Here is an example of linear regression. It introduces workflow of PaddlePaddle, :maxdepth: 1 concepts/index_en.rst + developer's_guide_to_paddle_fluid.md diff --git a/doc/fluid/getstarted/quickstart_cn.rst b/doc/fluid/getstarted/quickstart_cn.rst index 135beb75d0330f39d062753aa2aa83a077f36bb1..6a964d4f8561f30aa10936d2399698c51583442c 100644 --- a/doc/fluid/getstarted/quickstart_cn.rst +++ b/doc/fluid/getstarted/quickstart_cn.rst @@ -11,7 +11,7 @@ PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14. pip install paddlepaddle -如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行： +如果需要安装支持GPU的版本（cuda8.0_cudnn5_avx_openblas），需要执行： .. code-block:: bash @@ -28,18 +28,18 @@ PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14. import paddle.dataset.uci_housing as uci_housing import paddle.fluid as fluid - + with fluid.scope_guard(fluid.core.Scope()): # initialize executor with cpu exe = fluid.Executor(place=fluid.CPUPlace()) - # load inference model + # load inference model [inference_program, feed_target_names,fetch_targets] = \ fluid.io.load_inference_model(uci_housing.fluid_model(), exe) # run inference - result = exe.run(inference_program, - feed={feed_target_names[0]: uci_housing.predict_reader()}, + result = exe.run(inference_program, + feed={feed_target_names[0]: uci_housing.predict_reader()}, fetch_list=fetch_targets) - # print predicted price is $12,273.97 + # print predicted price is $12,273.97 print 'Predicted price: ${:,.2f}'.format(result[0][0][0] * 1000) 执行 :code:`python housing.py` 瞧！ 它应该打印出预测住房数据的清单。 diff --git a/doc/fluid/getstarted/quickstart_en.rst b/doc/fluid/getstarted/quickstart_en.rst index df6619cfd039fc1fdca8cde57db9cc6aebf8f029..680122f25893a5a48fac103266bda4788f891f6d 100644 --- a/doc/fluid/getstarted/quickstart_en.rst +++ b/doc/fluid/getstarted/quickstart_en.rst @@ -12,7 +12,7 @@ Simply run the following command to install, the version is cpu_avx_openblas: pip install paddlepaddle -If you need to install GPU version (cuda7.5_cudnn5_avx_openblas), run: +If you need to install GPU version (cuda8.0_cudnn5_avx_openblas), run: .. code-block:: bash @@ -31,18 +31,18 @@ code: import paddle.dataset.uci_housing as uci_housing import paddle.fluid as fluid - + with fluid.scope_guard(fluid.core.Scope()): # initialize executor with cpu exe = fluid.Executor(place=fluid.CPUPlace()) - # load inference model + # load inference model [inference_program, feed_target_names,fetch_targets] = \ fluid.io.load_inference_model(uci_housing.fluid_model(), exe) # run inference - result = exe.run(inference_program, - feed={feed_target_names[0]: uci_housing.predict_reader()}, + result = exe.run(inference_program, + feed={feed_target_names[0]: uci_housing.predict_reader()}, fetch_list=fetch_targets) - # print predicted price is $12,273.97 + # print predicted price is $12,273.97 print 'Predicted price: ${:,.2f}'.format(result[0][0][0] * 1000) Run :code:`python housing.py` and voila! It should print out a list of predictions diff --git a/doc/fluid/howto/index_cn.rst b/doc/fluid/howto/index_cn.rst index b7c620179724ebe97a0a47b75a57b376b21ccf90..b57af64f44da82926c4862578f3072960ca5aa92 100644 --- a/doc/fluid/howto/index_cn.rst +++ b/doc/fluid/howto/index_cn.rst @@ -4,5 +4,5 @@ .. toctree:: :maxdepth: 1 + inference/index_cn.rst optimization/index_cn.rst - inference/inference_support_in_fluid.md diff --git a/doc/fluid/howto/index_en.rst b/doc/fluid/howto/index_en.rst index f3ca41cdbf1d40ec8afaf045233a38755d8a777a..fd21e167ce3a46da167db1e9d7013804f730e047 100644 --- a/doc/fluid/howto/index_en.rst +++ b/doc/fluid/howto/index_en.rst @@ -5,4 +5,3 @@ HOW TO :maxdepth: 1 optimization/index_en.rst - inference/inference_support_in_fluid.md diff --git a/doc/fluid/howto/inference/build_and_install_lib_cn.rst b/doc/fluid/howto/inference/build_and_install_lib_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..c8d9992fcc92c25f8c14f71c79bde9f79fd92b1f --- /dev/null +++ b/doc/fluid/howto/inference/build_and_install_lib_cn.rst @@ -0,0 +1,96 @@ +安装与编译C++预测库 +=========================== + +直接下载安装 +------------- + +====================== ======================================== +版本说明 C++预测库 +====================== ======================================== +cpu_avx_mkl `fluid.tgz `_ +cpu_avx_openblas `fluid.tgz `_ +cpu_noavx_openblas `fluid.tgz `_ +cuda7.5_cudnn5_avx_mkl `fluid.tgz `_ +cuda8.0_cudnn5_avx_mkl `fluid.tgz `_ +cuda8.0_cudnn7_avx_mkl `fluid.tgz `_ +====================== ======================================== + +从源码编译 +---------- +用户也可以从 PaddlePaddle 核心代码编译C++预测库，只需在编译时配制下面这些编译选项： + +================= ========= +选项 值 +================= ========= +CMAKE_BUILD_TYPE Release +FLUID_INSTALL_DIR 安装路径 +WITH_FLUID_ONLY ON（推荐） +WITH_SWIG_PY OFF（推荐 +WITH_PYTHON OFF（推荐） +WITH_GPU ON/OFF +WITH_MKL ON/OFF +================= ========= + +建议按照推荐值设置，以避免链接不必要的库。其它可选编译选项按需进行设定。 + +下面的代码片段从github拉取最新代码，配制编译选项（需要将PADDLE_ROOT替换为PaddlePaddle预测库的安装路径）： + + .. code-block:: bash + + pip install paddlepaddle-gpu + PADDLE_ROOT=/path/of/capi + git clone https://github.com/PaddlePaddle/Paddle.git + cd Paddle + mkdir build + cd build + cmake -DFLUID_INSTALL_DIR=$PADDLE_ROOT \ + -DCMAKE_BUILD_TYPE=Release \ + -DWITH_FLUID_ONLY=ON \ + -DWITH_SWIG_PY=OFF \ + -DWITH_PYTHON=OFF \ + -DWITH_MKL=OFF \ + -DWITH_GPU=OFF \ + .. + make + make inference_lib_dist + +成功编译后，使用C++预测库所需的依赖（包括：（1）编译出的PaddlePaddle预测库和头文件；（2）第三方链接库和头文件；（3）版本信息与编译选项信息） +均会存放于PADDLE_ROOT目录中。目录结构如下： + + .. code-block:: text + + PaddleRoot/ + ├── CMakeCache.txt + ├── paddle + │   └── fluid + │   ├── framework + │   ├── inference + │   ├── memory + │   ├── platform + │   ├── pybind + │   └── string + ├── third_party + │   ├── boost + │   │   └── boost + │   ├── eigen3 + │   │   ├── Eigen + │   │   └── unsupported + │   └── install + │   ├── gflags + │   ├── glog + │   ├── mklml + │   ├── protobuf + │   ├── snappy + │   ├── snappystream + │   └── zlib + └── version.txt + +version.txt 中记录了该预测库的版本信息，包括Git Commit ID、使用OpenBlas或MKL数学库、CUDA/CUDNN版本号，如： + + .. code-block:: text + + GIT COMMIT ID: c95cd4742f02bb009e651a00b07b21c979637dc8 + WITH_MKL: ON + WITH_GPU: ON + CUDA version: 8.0 + CUDNN version: v5 diff --git a/doc/fluid/howto/inference/index_cn.rst b/doc/fluid/howto/inference/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..a903423548decd0992bf19772fb2cb143f6a12b5 --- /dev/null +++ b/doc/fluid/howto/inference/index_cn.rst @@ -0,0 +1,8 @@ +预测库 +------------ + +.. toctree:: + :maxdepth: 1 + + build_and_install_lib_cn.rst + inference_support_in_fluid_cn.md diff --git a/doc/fluid/howto/inference/inference_support_in_fluid.md b/doc/fluid/howto/inference/inference_support_in_fluid_cn.md similarity index 90% rename from doc/fluid/howto/inference/inference_support_in_fluid.md rename to doc/fluid/howto/inference/inference_support_in_fluid_cn.md index d272cd3e3bdac49b9ed1a21531de1b0be03d881e..309b17fccd5c461c9c22beb64eb4c6792b7e4a7a 100644 --- a/doc/fluid/howto/inference/inference_support_in_fluid.md +++ b/doc/fluid/howto/inference/inference_support_in_fluid_cn.md @@ -1,9 +1,8 @@ -# Fluid Inference使用指南 +# 使用指南 ## 目录： - Python Inference API -- 编译Fluid Inference库 - Inference C++ API - Inference实例 - Inference计算优化 @@ -55,62 +54,6 @@ return [program, feed_target_names, fetch_targets] ``` - -## 编译Fluid Inference库 - - - **不需要额外的CMake选项** - - 1、 配置CMake命令，更多配置请参考[源码编译PaddlePaddle](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/build_from_source_cn.html) - ```bash - $ git clone https://github.com/PaddlePaddle/Paddle.git - $ cd Paddle - $ mkdir build - $ cd build - $ cmake -DCMAKE_INSTALL_PREFIX=your/path/to/paddle_inference_lib \ - -DCMAKE_BUILD_TYPE=Release \ - -DWITH_PYTHON=ON \ - -DWITH_MKL=OFF \ - -DWITH_GPU=OFF \ - .. - ``` - - - 2、 编译PaddlePaddle - ```bash - $ make - ``` - - - 3、 部署。执行如下命令将PaddlePaddle Fluid Inference库部署到`your/path/to/paddle_inference_lib`目录。 - ```bash - $ make inference_lib_dist - ``` - -- 目录结构 - - ```bash - $ cd your/path/to/paddle_inference_lib - $ tree - . - |-- paddle - | `-- fluid - | |-- framework - | |-- inference - | | |-- io.h - | | `-- libpaddle_fluid.so - | |-- memory - | |-- platform - | `-- string - |-- third_party - | |-- eigen3 - | `-- install - | |-- gflags - | |-- glog - | `-- protobuf - `-- ... - ``` - - 假设`PADDLE_ROOT=your/path/to/paddle_inference_lib`。 - - - ## 链接Fluid Inference库 - 示例项目([链接](https://github.com/luotao1/fluid_inference_example.git)) diff --git a/paddle/contrib/inference/CMakeLists.txt b/paddle/contrib/inference/CMakeLists.txt index 9c55f189bcc5cbf0ce84f11e9653fa20b84a51f7..6847f7db7fc0f6b41ced1260d409ca6eba9b53eb 100644 --- a/paddle/contrib/inference/CMakeLists.txt +++ b/paddle/contrib/inference/CMakeLists.txt @@ -17,46 +17,33 @@ if(APPLE) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") endif(APPLE) -function(inference_api_test TARGET_NAME TEST_SRC) +function(inference_api_test TARGET_NAME) set(options "") set(oneValueArgs "") set(multiValueArgs ARGS) cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) - set(arg_list "") + cc_test(test_paddle_inference_${TARGET_NAME} + SRCS test_paddle_inference_${TARGET_NAME}.cc + DEPS paddle_fluid_api paddle_inference_api + ARGS --dirname=${PYTHON_TESTS_DIR}/book/) if(inference_test_ARGS) - foreach(arg ${inference_test_ARGS}) - list(APPEND arg_list "_${arg}") - endforeach() - else() - list(APPEND arg_list "_") + set_tests_properties(test_paddle_inference_${TARGET_NAME} + PROPERTIES DEPENDS "${inference_test_ARGS}") endif() - foreach(arg ${arg_list}) - string(REGEX REPLACE "^_$" "" arg "${arg}") - cc_test(${TARGET_NAME} - SRCS ${TEST_SRC} - DEPS paddle_fluid_api paddle_inference_api paddle_inference_api_impl - ARGS --dirname=${PYTHON_TESTS_DIR}/book/) - # TODO(panyx0178): Figure out how to add word2vec and image_classification - # as deps. - # set_tests_properties(${TARGET_NAME} - # PROPERTIES DEPENDS ${DEP_TEST}) - endforeach() endfunction(inference_api_test) cc_library(paddle_inference_api - SRCS paddle_inference_api.cc + SRCS paddle_inference_api.cc paddle_inference_api_impl.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) -cc_library(paddle_inference_api_impl - SRCS paddle_inference_api_impl.cc - DEPS paddle_inference_api paddle_fluid_api) +if(WITH_TESTING) + cc_test(test_paddle_inference_api + SRCS test_paddle_inference_api.cc + DEPS paddle_inference_api) -cc_test(test_paddle_inference_api - SRCS test_paddle_inference_api.cc - DEPS paddle_inference_api) - -inference_api_test(test_paddle_inference_api_impl - test_paddle_inference_api_impl.cc) + inference_api_test(api_impl + ARGS test_word2vec test_image_classification) +endif() diff --git a/paddle/contrib/inference/paddle_inference_api.h b/paddle/contrib/inference/paddle_inference_api.h index f804d9b28697a6703d63d9a640c4ec337effaba6..5fe8399762bba69bc99ed9ae694db32f532ed953 100644 --- a/paddle/contrib/inference/paddle_inference_api.h +++ b/paddle/contrib/inference/paddle_inference_api.h @@ -40,15 +40,24 @@ struct PaddleBuf { struct PaddleTensor { std::string name; // variable name. std::vector shape; + // TODO(Superjomn) for LoD support, add a vector> field if needed. PaddleBuf data; // blob of data. PaddleDType dtype; }; +enum class PaddleEngineKind { + kNative = 0, // Use the native Fluid facility. + // TODO(Superjomn) support following engines latter. + // kAnakin, // Use Anakin for inference. + // kTensorRT, // Use TensorRT for inference. + // kAutoMixedAnakin, // Automatically mix Fluid with Anakin. + // kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT. +}; + /* -* A simple Inference API for Paddle. Currently this API might just be used by -* non-sequence scenerios. -* TODO(Superjomn) Prepare another API for NLP-related usages. -*/ + * A simple Inference API for Paddle. Currently this API can be used by + * non-sequence scenerios. + */ class PaddlePredictor { public: struct Config; @@ -66,34 +75,35 @@ class PaddlePredictor { // be thread-safe. virtual std::unique_ptr Clone() = 0; - virtual bool InitShared() { return false; } // Destroy the Predictor. virtual ~PaddlePredictor() {} - friend std::unique_ptr CreatePaddlePredictor( - const PaddlePredictor::Config& config); - // The common configs for all the predictors. struct Config { - enum class EngineKind; - std::string model_dir; // path to the model directory. bool enable_engine{false}; // Enable to execute (part of) the model on - // third-party engines. - EngineKind engine_kind{Config::EngineKind::kNone}; - - enum class EngineKind { - kNone = -1, // Use the native Fluid facility. - kAnakin, // Use Anakin for inference. - kTensorRT, // Use TensorRT for inference. - kAutoMixedAnakin, // Automatically mix Fluid with Anakin. - kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT. - }; }; }; -// A factory to help create difference predictor. -template +struct NativeConfig : public PaddlePredictor::Config { + // GPU related fields. + bool use_gpu{false}; + int device{0}; + float fraction_of_gpu_memory{-1.f}; // Negative to notify initialization. + + std::string prog_file; + std::string param_file; +}; + +// A factory to help create different predictors. +// +// FOR EXTENSION DEVELOPER: +// Different predictors are designated by config type and engine kind. Similar +// configs can be merged, but there shouldn't be a huge config containing +// different fields for more than one kind of predictors. +// +// Similarly, each engine kind should map to a unique predictor implementation. +template std::unique_ptr CreatePaddlePredictor(const ConfigT& config); } // namespace paddle diff --git a/paddle/contrib/inference/paddle_inference_api_impl.cc b/paddle/contrib/inference/paddle_inference_api_impl.cc index ebe4c3291802707009f30616463705d966e244d6..99a64662d4d04e3cf9dfdafe5b5ab9e5dac0af8a 100644 --- a/paddle/contrib/inference/paddle_inference_api_impl.cc +++ b/paddle/contrib/inference/paddle_inference_api_impl.cc @@ -54,11 +54,10 @@ std::string num2str(T a) { } } // namespace -bool PaddlePredictorImpl::Init() { +bool NativePaddlePredictor::Init() { VLOG(3) << "Predictor::init()"; - // TODO(panyx0718): Should CPU vs GPU device be decided by id? - if (config_.device >= 0) { + if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); } else { place_ = paddle::platform::CPUPlace(); @@ -85,19 +84,21 @@ bool PaddlePredictorImpl::Init() { } ctx_ = executor_->Prepare(*inference_program_, 0); - // Create variables - // TODO(panyx0718): Why need to test share_variables here? - if (config_.share_variables) { - executor_->CreateVariables(*inference_program_, scope_.get(), 0); - } + // Create temporary variables first, so that the first batch do not need to + // create variables in the runtime. This is the logics of the old inference + // API. + // TODO(Superjomn) this should be modified when `Clone` is valid for + // multi-thread application. + executor_->CreateVariables(*inference_program_, scope_.get(), 0); + // Get the feed_target_names and fetch_target_names feed_target_names_ = inference_program_->GetFeedTargetNames(); fetch_target_names_ = inference_program_->GetFetchTargetNames(); return true; } -bool PaddlePredictorImpl::Run(const std::vector &inputs, - std::vector *output_data) { +bool NativePaddlePredictor::Run(const std::vector &inputs, + std::vector *output_data) { VLOG(3) << "Predictor::predict"; Timer timer; timer.tic(); @@ -124,7 +125,7 @@ bool PaddlePredictorImpl::Run(const std::vector &inputs, scope_.get(), &feed_targets, &fetch_targets, - !config_.share_variables); + false /* don't create variable eatch time */); if (!GetFetch(fetchs, output_data)) { LOG(ERROR) << "fail to get fetchs"; return false; @@ -133,59 +134,20 @@ bool PaddlePredictorImpl::Run(const std::vector &inputs, return true; } -std::unique_ptr PaddlePredictorImpl::Clone() { +std::unique_ptr NativePaddlePredictor::Clone() { VLOG(3) << "Predictor::clone"; - std::unique_ptr cls(new PaddlePredictorImpl(config_)); - if (!cls->InitShared()) { - LOG(ERROR) << "fail to call InitShared"; + std::unique_ptr cls(new NativePaddlePredictor(config_)); + + if (!dynamic_cast(cls.get())->Init()) { + LOG(ERROR) << "fail to call Init"; return nullptr; } // fix manylinux compile error. return std::move(cls); } -// TODO(panyx0718): Consider merge with Init()? -bool PaddlePredictorImpl::InitShared() { - VLOG(3) << "Predictor::init_shared"; - // 1. Define place, executor, scope - if (this->config_.device >= 0) { - place_ = platform::CUDAPlace(); - } else { - place_ = platform::CPUPlace(); - } - this->executor_.reset(new framework::Executor(this->place_)); - this->scope_.reset(new framework::Scope()); - // Initialize the inference program - if (!this->config_.model_dir.empty()) { - // Parameters are saved in separate files sited in - // the specified `dirname`. - this->inference_program_ = inference::Load( - this->executor_.get(), this->scope_.get(), this->config_.model_dir); - } else if (!this->config_.prog_file.empty() && - !this->config_.param_file.empty()) { - // All parameters are saved in a single file. - // The file names should be consistent with that used - // in Python API `fluid.io.save_inference_model`. - this->inference_program_ = inference::Load(this->executor_.get(), - this->scope_.get(), - this->config_.prog_file, - this->config_.param_file); - } - this->ctx_ = this->executor_->Prepare(*this->inference_program_, 0); - // 3. create variables - // TODO(panyx0718): why test share_variables. - if (config_.share_variables) { - this->executor_->CreateVariables( - *this->inference_program_, this->scope_.get(), 0); - } - // 4. Get the feed_target_names and fetch_target_names - this->feed_target_names_ = this->inference_program_->GetFeedTargetNames(); - this->fetch_target_names_ = this->inference_program_->GetFetchTargetNames(); - return true; -} - -bool PaddlePredictorImpl::SetFeed(const std::vector &inputs, - std::vector *feeds) { +bool NativePaddlePredictor::SetFeed(const std::vector &inputs, + std::vector *feeds) { VLOG(3) << "Predictor::set_feed"; if (inputs.size() != feed_target_names_.size()) { LOG(ERROR) << "wrong feed input size."; @@ -213,7 +175,7 @@ bool PaddlePredictorImpl::SetFeed(const std::vector &inputs, return true; } -bool PaddlePredictorImpl::GetFetch( +bool NativePaddlePredictor::GetFetch( const std::vector &fetchs, std::vector *outputs) { VLOG(3) << "Predictor::get_fetch"; @@ -280,23 +242,29 @@ bool PaddlePredictorImpl::GetFetch( } template <> -std::unique_ptr CreatePaddlePredictor( - const ConfigImpl &config) { - VLOG(3) << "create PaddlePredictorImpl"; - // 1. GPU memeroy - std::vector flags; - if (config.fraction_of_gpu_memory >= 0.0f || - config.fraction_of_gpu_memory <= 0.95f) { - flags.push_back("dummpy"); - std::string flag = "--fraction_of_gpu_memory_to_use=" + - num2str(config.fraction_of_gpu_memory); - flags.push_back(flag); - VLOG(3) << "set flag: " << flag; - framework::InitGflags(flags); +std::unique_ptr +CreatePaddlePredictor( + const NativeConfig &config) { + VLOG(3) << "create NativePaddlePredictor"; + if (config.use_gpu) { + // 1. GPU memeroy + PADDLE_ENFORCE( + config.fraction_of_gpu_memory > 0.f, + "fraction_of_gpu_memory in the config should be set to range (0., 1.]"); + std::vector flags; + if (config.fraction_of_gpu_memory >= 0.0f || + config.fraction_of_gpu_memory <= 0.95f) { + flags.push_back("dummpy"); + std::string flag = "--fraction_of_gpu_memory_to_use=" + + num2str(config.fraction_of_gpu_memory); + flags.push_back(flag); + VLOG(3) << "set flag: " << flag; + framework::InitGflags(flags); + } } - std::unique_ptr predictor(new PaddlePredictorImpl(config)); - if (!dynamic_cast(predictor.get())->Init()) { + std::unique_ptr predictor(new NativePaddlePredictor(config)); + if (!dynamic_cast(predictor.get())->Init()) { return nullptr; } return std::move(predictor); diff --git a/paddle/contrib/inference/paddle_inference_api_impl.h b/paddle/contrib/inference/paddle_inference_api_impl.h index c545461680723b429b2253392060ea36b84ce708..84707e223d7aa3d1ebca933923e932b3973613ae 100644 --- a/paddle/contrib/inference/paddle_inference_api_impl.h +++ b/paddle/contrib/inference/paddle_inference_api_impl.h @@ -29,17 +29,10 @@ namespace paddle { -struct ConfigImpl : public PaddlePredictor::Config { - int device; - float fraction_of_gpu_memory; - std::string prog_file; - std::string param_file; - bool share_variables; -}; - -class PaddlePredictorImpl : public PaddlePredictor { +class NativePaddlePredictor : public PaddlePredictor { public: - explicit PaddlePredictorImpl(const ConfigImpl &config) : config_(config) {} + explicit NativePaddlePredictor(const NativeConfig &config) + : config_(config) {} bool Init(); @@ -48,16 +41,15 @@ class PaddlePredictorImpl : public PaddlePredictor { std::unique_ptr Clone() override; - ~PaddlePredictorImpl() override{}; + ~NativePaddlePredictor() override{}; private: - bool InitShared() override; bool SetFeed(const std::vector &input_datas, std::vector *feeds); bool GetFetch(const std::vector &fetchs, std::vector *output_data); - ConfigImpl config_; + NativeConfig config_; platform::Place place_; std::unique_ptr executor_; std::unique_ptr scope_; diff --git a/paddle/contrib/inference/test_paddle_inference_api_impl.cc b/paddle/contrib/inference/test_paddle_inference_api_impl.cc index 096293a4e25df0c78150d85dc091d7ca6539bf40..07b17acd484b13af2ab4019aafa4a08c6b9f59d4 100644 --- a/paddle/contrib/inference/test_paddle_inference_api_impl.cc +++ b/paddle/contrib/inference/test_paddle_inference_api_impl.cc @@ -40,19 +40,19 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { return pt; } -ConfigImpl GetConfig() { - ConfigImpl config; +NativeConfig GetConfig() { + NativeConfig config; config.model_dir = FLAGS_dirname + "word2vec.inference.model"; LOG(INFO) << "dirname " << config.model_dir; config.fraction_of_gpu_memory = 0.15; + config.use_gpu = true; config.device = 0; - config.share_variables = true; return config; } TEST(paddle_inference_api_impl, word2vec) { - ConfigImpl config = GetConfig(); - std::unique_ptr predictor = CreatePaddlePredictor(config); + NativeConfig config = GetConfig(); + auto predictor = CreatePaddlePredictor(config); framework::LoDTensor first_word, second_word, third_word, fourth_word; framework::LoD lod{{0, 1}}; @@ -104,7 +104,7 @@ TEST(paddle_inference_api_impl, image_classification) { int batch_size = 2; bool use_mkldnn = false; bool repeat = false; - ConfigImpl config = GetConfig(); + NativeConfig config = GetConfig(); config.model_dir = FLAGS_dirname + "image_classification_resnet.inference.model"; @@ -133,7 +133,7 @@ TEST(paddle_inference_api_impl, image_classification) { is_combined, use_mkldnn); - std::unique_ptr predictor = CreatePaddlePredictor(config); + auto predictor = CreatePaddlePredictor(config); std::vector paddle_tensor_feeds; paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&input)); @@ -144,8 +144,7 @@ TEST(paddle_inference_api_impl, image_classification) { float* data = static_cast(outputs[0].data.data); float* lod_data = output1.data(); for (size_t j = 0; j < len / sizeof(float); ++j) { - EXPECT_LT(lod_data[j] - data[j], 1e-10); - EXPECT_GT(lod_data[j] - data[j], -1e-10); + EXPECT_NEAR(lod_data[j], data[j], 1e-3); } free(data); } diff --git a/paddle/fluid/framework/block_desc.cc b/paddle/fluid/framework/block_desc.cc index fd409ed4c0f7a504686765909e9c71692aab8824..e7842e9b8130d35e511e02dfb1dc27f307d17f38 100644 --- a/paddle/fluid/framework/block_desc.cc +++ b/paddle/fluid/framework/block_desc.cc @@ -200,7 +200,7 @@ BlockDesc::BlockDesc(ProgramDesc *prog, proto::BlockDesc *desc) vars_[var_desc.name()].reset(new VarDesc(var_desc)); } for (const proto::OpDesc &op_desc : desc_->ops()) { - ops_.emplace_back(new OpDesc(op_desc, prog, this)); + ops_.emplace_back(new OpDesc(op_desc, this)); } } @@ -209,7 +209,7 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, : prog_(prog), desc_(desc) { need_update_ = true; for (auto &op : other.ops_) { - ops_.emplace_back(new OpDesc(*op->Proto(), prog, this)); + ops_.emplace_back(new OpDesc(*op, this)); } for (auto &it : other.vars_) { auto *var = new VarDesc(*it.second); diff --git a/paddle/fluid/framework/block_desc.h b/paddle/fluid/framework/block_desc.h index 600601669c5d56a3ffc2fb9c804ffad5fde58f0b..189dd6c52f85b5bf623b98c64c07c0c7269505d4 100644 --- a/paddle/fluid/framework/block_desc.h +++ b/paddle/fluid/framework/block_desc.h @@ -105,7 +105,7 @@ class BlockDesc { size_t OpSize() const { return ops_.size(); } - OpDesc *Op(int idx) { return ops_.at(idx).get(); } + OpDesc *Op(int idx) const { return ops_.at(idx).get(); } void Flush(); diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc index d8e711994c5dba15ce0a1c237558b121888902e3..17baacd13eecac8f410631fe9e94788da4fff848 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc @@ -11,11 +11,15 @@ // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" +#include #include +#include #include +#include + #include "paddle/fluid/framework/details/broadcast_op_handle.h" #include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" #include "paddle/fluid/framework/details/reduce_op_handle.h" #include "paddle/fluid/framework/details/rpc_op_handle.h" #include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h" @@ -26,9 +30,6 @@ #include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h" #endif -#include -#include - DEFINE_string(ssa_graph_path, "/tmp/ssa_graph.dot", "the ssa graph path only print with GLOG_v=10," "default /tmp/graph.dot"); @@ -148,9 +149,9 @@ bool MultiDevSSAGraphBuilder::IsDistTrainOp( std::unique_ptr MultiDevSSAGraphBuilder::Build( const ProgramDesc &program) const { - std::unordered_map var_types; + std::unordered_map all_vars; for (auto *var : program.Block(0).AllVars()) { - var_types[var->Name()] = var->GetType(); + all_vars[var->Name()] = var; } auto graph = new SSAGraph(); @@ -167,12 +168,28 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( auto send_vars = FindDistTrainSendVars(program); auto recv_vars = FindDistTrainRecvVars(program); - size_t cur_device_id = 0; std::vector> var_name_on_devices; std::vector> bcast_var_name_set; var_name_on_devices.resize(places_.size()); bcast_var_name_set.resize(places_.size()); + size_t cur_device_id = 0; + std::vector balance_grads(places_.size(), 0); + + auto get_appropriate_dev = [&](std::string &g_name) -> size_t { + auto var_desc = all_vars.at(g_name); + PADDLE_ENFORCE_NOT_NULL(var_desc); + auto dim = framework::make_ddim(var_desc->GetShape()); + int64_t numel = framework::product(dim); + PADDLE_ENFORCE_GE(numel, 0); + auto smallest = + std::min_element(std::begin(balance_grads), std::end(balance_grads)); + size_t dev_id = + static_cast(std::distance(std::begin(balance_grads), smallest)); + balance_grads[dev_id] += numel; + return dev_id; + }; + bool is_forwarding = true; for (auto *op : program.Block(0).AllOps()) { if (boost::get( @@ -220,13 +237,13 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( switch (strategy_.reduce_) { case BuildStrategy::ReduceStrategy::kReduce: + cur_device_id = get_appropriate_dev(g_name); CreateReduceOp(&result, g_name, cur_device_id); var_name_on_devices[cur_device_id].emplace(g_name); bcast_var_name_set[cur_device_id].emplace(p_name); - cur_device_id = (cur_device_id + 1) % places_.size(); break; case BuildStrategy::ReduceStrategy::kAllReduce: - if (IsSparseGradient(var_types, g_name)) { + if (IsSparseGradient(all_vars, g_name)) { CreateReduceOp(&result, g_name, 0); CreateBroadcastOp(&result, g_name, 0); } else { @@ -269,10 +286,10 @@ std::unique_ptr MultiDevSSAGraphBuilder::Build( } bool MultiDevSSAGraphBuilder::IsSparseGradient( - const std::unordered_map &var_types, + const std::unordered_map &all_vars, const std::string &og) const { - PADDLE_ENFORCE(var_types.count(og) != 0); - if (var_types.at(og) == proto::VarType::SELECTED_ROWS) { + PADDLE_ENFORCE(all_vars.count(og) != 0); + if (all_vars.at(og)->GetType() == proto::VarType::SELECTED_ROWS) { return true; } return false; diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.h b/paddle/fluid/framework/details/multi_devices_graph_builder.h index e07597dbd80889c366babe79455beb12c9eb80d9..544cbe585c7423b5f3eb98ee698ca5668376f1ca 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_builder.h +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.h @@ -106,7 +106,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder { size_t src_dev_id) const; bool IsSparseGradient( - const std::unordered_map &var_types, + const std::unordered_map &all_vars, const std::string &og) const; private: diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index 09b67e5a1741c68c5f5487340e8fc86ff31e00a4..f92769192c218eb7cdc2350ff6e4721b45005806 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -103,7 +103,7 @@ void OpDesc::CopyFrom(const OpDesc &op_desc) { need_update_ = true; } -OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block) +OpDesc::OpDesc(const proto::OpDesc &desc, BlockDesc *block) : desc_(desc), need_update_(false) { // restore inputs_ int input_size = desc_.inputs_size(); diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index 1a330db7cc5555a939950043ac90a321573b292d..a02d3e269129596f65a2fb346e76c1af7fbead95 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -33,13 +33,14 @@ class OpDesc { OpDesc(const std::string &type, const VariableNameMap &inputs, const VariableNameMap &outputs, const AttributeMap &attrs); - OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block); + OpDesc(const proto::OpDesc &desc, BlockDesc *block); explicit OpDesc(BlockDesc *block) : block_(block) {} OpDesc(const OpDesc &other, BlockDesc *block) { *this = other; block_ = block; + need_update_ = true; } void CopyFrom(const OpDesc &op_desc); diff --git a/paddle/fluid/framework/program_desc.cc b/paddle/fluid/framework/program_desc.cc index 64fb028f83a539d17885186d5d8ee6ef26f095e9..1e01a6e900404990e16674755367d2fc6d832725 100644 --- a/paddle/fluid/framework/program_desc.cc +++ b/paddle/fluid/framework/program_desc.cc @@ -51,12 +51,15 @@ ProgramDesc::ProgramDesc(const ProgramDesc &o) { auto *block = desc_.mutable_blocks(i); blocks_.emplace_back(new BlockDesc(*o.blocks_[i], block, this)); } - for (auto &block : blocks_) { - for (auto *op : block->AllOps()) { - 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)); + for (size_t block_id = 0; block_id < blocks_.size(); ++block_id) { + auto all_ops = blocks_[block_id]->AllOps(); + for (size_t op_id = 0; op_id < all_ops.size(); ++op_id) { + auto &op = all_ops[op_id]; + for (const std::string &attr_name : op->AttrNames()) { + if (op->GetAttrType(attr_name) == proto::AttrType::BLOCK) { + int sub_block_id = + o.Block(block_id).Op(op_id)->GetBlockAttr(attr_name); + op->SetBlockAttr(attr_name, MutableBlock(sub_block_id)); } } } @@ -86,6 +89,16 @@ ProgramDesc::ProgramDesc(const std::string &binary_str) { for (auto &block_desc : *desc_.mutable_blocks()) { blocks_.emplace_back(new BlockDesc(this, &block_desc)); } + for (auto &block : blocks_) { + for (auto *op : block->AllOps()) { + 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)); + } + } + } + } } const std::vector ProgramDesc::GetFeedTargetNames() { diff --git a/paddle/fluid/framework/reader.cc b/paddle/fluid/framework/reader.cc index 76126f3dc64d71770d13f9d66bb30f176c112629..0b36f1116d15004b355e854e101abb9ad3297836 100644 --- a/paddle/fluid/framework/reader.cc +++ b/paddle/fluid/framework/reader.cc @@ -25,8 +25,10 @@ void FileReader::ReadNext(std::vector *out) { if (out->empty()) { return; } + + PADDLE_ENFORCE_EQ(out->size(), dims_.size()); for (size_t i = 0; i < dims_.size(); ++i) { - auto &actual = out->at(i).dims(); + auto &actual = (*out)[i].dims(); auto &expect = dims_[i]; PADDLE_ENFORCE_EQ(actual.size(), expect.size()); diff --git a/paddle/fluid/framework/tensor_impl.h b/paddle/fluid/framework/tensor_impl.h index 0a1db7758bd9ec0dac133efcbf495de1d690021d..2f19ec0f0a9338e2b96d1f64eac45387bae4d1eb 100644 --- a/paddle/fluid/framework/tensor_impl.h +++ b/paddle/fluid/framework/tensor_impl.h @@ -39,7 +39,7 @@ template inline const T* Tensor::data() const { check_memory_size(); PADDLE_ENFORCE(std::is_same::value || - holder_->type().hash_code() == typeid(T).hash_code(), + holder_->type() == std::type_index(typeid(T)), "Tensor holds the wrong type, it holds %s", this->holder_->type().name()); @@ -53,7 +53,7 @@ template inline T* Tensor::data() { check_memory_size(); PADDLE_ENFORCE(std::is_same::value || - holder_->type().hash_code() == typeid(T).hash_code(), + holder_->type() == std::type_index(typeid(T)), "Tensor holds the wrong type, it holds %s", this->holder_->type().name()); return reinterpret_cast(reinterpret_cast(holder_->ptr()) + diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index cc4a725dfb3b3e7723a3a3a4008b20acdb53899d..ec16a1c600a3bafc1c4cbbd920360253c106e3a1 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -5,14 +5,19 @@ cc_library(paddle_fluid_api SRCS io.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB}) -# Create static library get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES) -cc_library(paddle_fluid DEPS ${fluid_modules}) +if(WITH_CONTRIB) + set(fluid_modules "${fluid_modules}" paddle_inference_api) +endif() + +# Create static library +cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api) # Create shared library cc_library(paddle_fluid_shared SHARED SRCS io.cc - DEPS ${fluid_modules}) + DEPS ${fluid_modules} paddle_fluid_api) + set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid) if(NOT APPLE) # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac. diff --git a/paddle/fluid/inference/analysis/data_flow_graph.h b/paddle/fluid/inference/analysis/data_flow_graph.h index 9f6ce40ede25248a4f779b379c132806a4ec06ba..913e344d371ddf3ea05a53c216e5b3bea8f11c7b 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph.h +++ b/paddle/fluid/inference/analysis/data_flow_graph.h @@ -21,7 +21,10 @@ limitations under the License. */ #include #include +#include #include +#include +#include #include "paddle/fluid/inference/analysis/graph_traits.h" #include "paddle/fluid/inference/analysis/node.h" diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc index 60f159da9140516284449a0274906df004b23ac5..dcee75cee50ede1d2b660e88e06544440bd5ef77 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc @@ -44,6 +44,6 @@ TEST_F(DFG_Tester, Test) { LOG(INFO) << graph.nodes.size(); } -} // analysis -} // inference -} // paddle +}; // namespace analysis +}; // namespace inference +}; // namespace paddle diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc index f848a7d1add79c3032da7defc34a406dccf29d2e..9f67c989cca4a936cd320b73efaae277263fb3e2 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc @@ -12,9 +12,11 @@ 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/analysis/fluid_to_data_flow_graph_pass.h" +#include #include +#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" + namespace paddle { namespace inference { namespace analysis { diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h index cd0d4fabaafe844bcc5bb8bfc2586971197d9167..33517e57becdffc0416f204247eac5feadb7ed82 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h @@ -19,6 +19,8 @@ #pragma once +#include + #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/inference/analysis/data_flow_graph.h" #include "paddle/fluid/inference/analysis/pass.h" diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc index 851c98bef305fa9e20dced5f7c26e9d1b6ddf4f2..817d32c92cdbdc234eef9ed5156891c2b11ced4c 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc @@ -32,6 +32,6 @@ TEST_F(DFG_Tester, Init) { LOG(INFO) << '\n' << graph.DotString(); } -} // analysis -} // inference -} // paddle +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/helper.h b/paddle/fluid/inference/analysis/helper.h index 24ea9a4bae7132eb1692b0ffb02f8ab5e02b21a9..153dca576bd6734d62f00c4a7cb9b503506b33e2 100644 --- a/paddle/fluid/inference/analysis/helper.h +++ b/paddle/fluid/inference/analysis/helper.h @@ -50,7 +50,7 @@ struct DataTypeNamer { return dic_.at(x); } - const std::string &repr(size_t &hash) const { + const std::string &repr(size_t &hash) const { // NOLINT PADDLE_ENFORCE(dic_.count(hash), "unknown type for representation"); return dic_.at(hash); } @@ -62,7 +62,9 @@ struct DataTypeNamer { SET_TYPE(float); } - std::unordered_map dic_; + std::unordered_map + dic_; }; #undef SET_TYPE diff --git a/paddle/fluid/inference/analysis/pass.h b/paddle/fluid/inference/analysis/pass.h index 5c89b1304d84abc9a4942f12da46b4bfe76f44f5..aa0e8667b5e4a9e6156c25fcad03bb8eee3287f6 100644 --- a/paddle/fluid/inference/analysis/pass.h +++ b/paddle/fluid/inference/analysis/pass.h @@ -16,6 +16,7 @@ limitations under the License. */ #include #include +#include #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/inference/analysis/data_flow_graph.h" diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.h b/paddle/fluid/inference/analysis/subgraph_splitter.h index ed90a0dcf31e154c4d82be08ce35e2f11d11c139..a31afbe6933da8d3c7a88142cc12d63b98b55796 100644 --- a/paddle/fluid/inference/analysis/subgraph_splitter.h +++ b/paddle/fluid/inference/analysis/subgraph_splitter.h @@ -18,6 +18,8 @@ limitations under the License. */ #pragma once +#include + #include "paddle/fluid/inference/analysis/data_flow_graph.h" #include "paddle/fluid/inference/analysis/node.h" diff --git a/paddle/fluid/inference/analysis/ut_helper.h b/paddle/fluid/inference/analysis/ut_helper.h index c86083d12153921672e15c172b874f77a8b46cde..722fa99a48a5f2b0e778904de0c35977d0ee3cc0 100644 --- a/paddle/fluid/inference/analysis/ut_helper.h +++ b/paddle/fluid/inference/analysis/ut_helper.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include #include +#include #include "paddle/fluid/framework/executor.h" #include "paddle/fluid/inference/analysis/data_flow_graph.h" #include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index 5ada1d631269209e912e2d4817382ea2c6c67353..23ca8bfac84f35ebdca2e2a1a8538d366358ca8b 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -8,3 +8,5 @@ nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor) nv_test(test_trt_mul_op SRCS test_mul_op.cc mul_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL) +nv_test(test_trt_fc_op SRCS test_fc_op.cc fc_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine mul_op SERIAL) diff --git a/paddle/fluid/inference/tensorrt/convert/activation_op.cc b/paddle/fluid/inference/tensorrt/convert/activation_op.cc index 6297051e5a30f1daa512d25d5aa3ab3b2f79f1d1..79d01b640a214ed5eb86173a36d5e85a6626066f 100644 --- a/paddle/fluid/inference/tensorrt/convert/activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/activation_op.cc @@ -24,7 +24,7 @@ class ReluOpConverter : public OpConverter { void operator()(const framework::proto::OpDesc& op) override { // Here the two nullptr looks strange, that's because the // framework::OpDesc's constructor is strange. - framework::OpDesc op_desc(op, nullptr, nullptr); + framework::OpDesc op_desc(op, nullptr); LOG(INFO) << "convert a fluid relu op to tensorrt activation layer whose " "type is Relu"; const nvinfer1::ITensor* input_tensor = diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc index 209936c3bafb0d31546856dc36c1b48053a0634b..668d344f1bba1c012dcb42c71b996209b4703d78 100644 --- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -21,7 +21,8 @@ namespace tensorrt { class Conv2dOpConverter : public OpConverter { public: Conv2dOpConverter() {} - void operator()(const framework::proto::OpDesc& op) override { + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope) override { LOG(INFO) << "convert a fluid conv2d op to tensorrt conv layer without bias"; } diff --git a/paddle/fluid/inference/tensorrt/convert/fc_op.cc b/paddle/fluid/inference/tensorrt/convert/fc_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..bd05608d7620ee4d917b30f919fba70f6aeff17f --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/fc_op.cc @@ -0,0 +1,119 @@ +/* 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/framework/eigen.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/engine.h" +#include "paddle/fluid/platform/place.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +// Reorder the elements from istrides to ostrides, borrowed from TRT convert in +// tensorflow. +// https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/tensorrt/convert/convert_nodes.cc#L318 +template +void Reorder2(nvinfer1::DimsHW shape, const T* idata, nvinfer1::DimsHW istrides, + T* odata, nvinfer1::DimsHW ostrides) { + for (int h = 0; h < shape.h(); ++h) { + for (int w = 0; w < shape.w(); ++w) { + odata[h * ostrides.h() + w * ostrides.w()] = + idata[h * ostrides.h() + w * ostrides.w()]; + } + } +} + +// Reorder the data layout from CK to KC. +void ReorderCKtoKC(TensorRTEngine::Weight& iweights, + TensorRTEngine::Weight* oweights) { + int c = iweights.dims[0]; + int k = iweights.dims[1]; + oweights->dims.assign({k, c}); + nvinfer1::DimsHW istrides = {1, k}; + nvinfer1::DimsHW ostrides = {c, 1}; + Reorder2({k, c}, static_cast(iweights.get().values), istrides, + static_cast(const_cast(oweights->get().values)), + ostrides); +} + +/* + * FC converter convert a MUL op in Fluid to a FC layer in TRT. + */ +class FcOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope) override { + VLOG(4) << "convert a fluid fc op to tensorrt fc layer without bias"; + + framework::OpDesc op_desc(op, nullptr, nullptr); + PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); + PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1); // Y is a weight + PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1); + + // Declare inputs + auto* X = engine_->GetITensor(op_desc.Input("X").front()); + + // Declare weights + auto* Y_v = scope.FindVar(op_desc.Input("Y").front()); + PADDLE_ENFORCE_NOT_NULL(Y_v); + auto* Y_t = Y_v->GetMutable(); + // This may trigger a GPU->CPU copy, because TRT's weight can only be + // assigned from CPU memory, that can't be avoided. + auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); + PADDLE_ENFORCE_EQ(Y_t->dims().size(), 2UL); // a matrix + size_t n_output = Y_t->dims()[1]; + + framework::LoDTensor tmp; + tmp.Resize(Y_t->dims()); + memcpy(tmp.mutable_data(platform::CPUPlace()), Y_t->data(), + Y_t->dims()[0] * Y_t->dims()[1]); + + TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT, + static_cast(weight_data), + Y_t->memory_size() / sizeof(float)}; + TensorRTEngine::Weight tmp_weight(nvinfer1::DataType::kFLOAT, + static_cast(tmp.data()), + Y_t->memory_size() / sizeof(float)); + weight.dims.assign({Y_t->dims()[0], Y_t->dims()[1]}); + tmp_weight.dims = weight.dims; + + // The data layout of TRT FC layer's weight is different from fluid's FC, + // need to reorder the elements. + ReorderCKtoKC(tmp_weight, &weight); + + // Currently, the framework can only handle one fluid op -> one TRT layer, + // but fc fuses `mul` and `bias` (2 fluid ops), so here is a trick, just + // handle `mul`, leave `add` as another layer. + // DEBUG + TensorRTEngine::Weight bias{nvinfer1::DataType::kFLOAT, nullptr, 0}; + + auto* layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, + *const_cast(X), + n_output, weight.get(), bias.get()); + + auto output_name = op_desc.Output("Out").front(); + engine_->DeclareOutput(layer, 0, output_name); + } +}; + +REGISTER_TRT_OP_CONVERTER(fc, FcOpConverter); + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +USE_OP(mul); diff --git a/paddle/fluid/inference/tensorrt/convert/mul_op.cc b/paddle/fluid/inference/tensorrt/convert/mul_op.cc index ed09f54bde00d12aaec829ba90cc08ebfef57e92..6bb07709c7ee1c6b29c46425849a4f472d3df59d 100644 --- a/paddle/fluid/inference/tensorrt/convert/mul_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/mul_op.cc @@ -24,10 +24,11 @@ namespace tensorrt { class MulOpConverter : public OpConverter { public: MulOpConverter() {} - void operator()(const framework::proto::OpDesc& op) override { - VLOG(4) << "convert a fluid mul op to tensorrt fc layer without bias"; + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope) override { + VLOG(4) << "convert a fluid mul op to tensorrt mul layer without bias"; - framework::OpDesc op_desc(op, nullptr, nullptr); + framework::OpDesc op_desc(op, nullptr); // Declare inputs auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); auto* input2 = engine_->GetITensor(op_desc.Input("Y")[0]); diff --git a/paddle/fluid/inference/tensorrt/convert/op_converter.h b/paddle/fluid/inference/tensorrt/convert/op_converter.h index 1cd3ed9a00acead2599420f88499bd0d74c2974b..4d21e241c0fe0abd9d454aa4f5f5ffeda747bed9 100644 --- a/paddle/fluid/inference/tensorrt/convert/op_converter.h +++ b/paddle/fluid/inference/tensorrt/convert/op_converter.h @@ -31,27 +31,42 @@ namespace tensorrt { class OpConverter { public: OpConverter() {} - virtual void operator()(const framework::proto::OpDesc& op) {} - void Run(const framework::proto::OpDesc& op, TensorRTEngine* engine) { - std::string type = op.type(); - auto* it = Registry::Lookup(type); - PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", type); - it->SetEngine(engine); - (*it)(op); - } + // Converter logic for an op. + virtual void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope) {} + + // Convert a single fluid operaotr and add the corresponding layer to TRT. + void ConvertOp(const framework::proto::OpDesc& op, + const std::unordered_set& parameters, + const framework::Scope& scope, TensorRTEngine* engine) { + framework::OpDesc op_desc(op, nullptr, nullptr); + + OpConverter* it{nullptr}; - // convert fluid op to tensorrt layer - void ConvertOp(const framework::proto::OpDesc& op, TensorRTEngine* engine) { - OpConverter::Run(op, engine); + if (op_desc.Type() == "mul") { + PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1UL); + std::string Y = op_desc.Input("Y")[0]; + if (parameters.count(Y)) { + it = Registry::Lookup("fc"); + } + } + if (!it) { + it = Registry::Lookup(op_desc.Type()); + } + PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", + op_desc.Type()); + it->SetEngine(engine); + (*it)(op, scope); } // convert fluid block to tensorrt network void ConvertBlock(const framework::proto::BlockDesc& block, - TensorRTEngine* engine) { + const std::unordered_set& parameters, + const framework::Scope& scope, TensorRTEngine* engine) { for (int i = 0; i < block.ops_size(); i++) { const auto& op = block.ops(i); - OpConverter::Run(op, engine); + ConvertOp(op, parameters, scope, engine); } } diff --git a/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc b/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a30253072ac581ceca85ca10151a176f87a7cb39 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_fc_op.cc @@ -0,0 +1,46 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(fc_op, test) { + std::unordered_set parameters({"mul-Y"}); + framework::Scope scope; + TRTConvertValidation validator(20, parameters, scope, 1000); + + validator.DeclInputVar("mul-X", nvinfer1::Dims4(8, 3, 1, 1)); + validator.DeclParamVar("mul-Y", nvinfer1::Dims2(3, 2)); + validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(8, 2)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("mul"); + desc.SetInput("X", {"mul-X"}); + desc.SetInput("Y", {"mul-Y"}); + desc.SetOutput("Out", {"mul-Out"}); + + validator.SetOp(*desc.Proto()); + + validator.Execute(10); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc b/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc index d8b61d5f08ffd071c112b4677fcb6f6f50784bbc..1ce1130e5d660d717a1262a1fbdb4b620462c0b3 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_mul_op.cc @@ -21,7 +21,9 @@ namespace inference { namespace tensorrt { TEST(MulOpConverter, main) { - TRTConvertValidation validator(10, 1000); + framework::Scope scope; + std::unordered_set parameters; + TRTConvertValidation validator(10, parameters, scope, 1000); validator.DeclInputVar("mul-X", nvinfer1::Dims2(10, 6)); validator.DeclInputVar("mul-Y", nvinfer1::Dims2(6, 10)); validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(10, 10)); diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc index 9ae7de9cbfa656fbcbb48557bd4b548115897c6d..1d3f5eabb2f839b2acfa9da6527589df1ec3767f 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -12,9 +12,10 @@ 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/convert/op_converter.h" + #include #include "paddle/fluid/framework/program_desc.h" -#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" namespace paddle { namespace inference { @@ -27,7 +28,9 @@ TEST(OpConverter, ConvertBlock) { conv2d_op->SetType("conv2d"); OpConverter converter; - converter.ConvertBlock(*block->Proto(), nullptr /*TensorRTEngine*/); + framework::Scope scope; + converter.ConvertBlock(*block->Proto(), {}, scope, + nullptr /*TensorRTEngine*/); } } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/convert/ut_helper.h b/paddle/fluid/inference/tensorrt/convert/ut_helper.h index 37fcb5c50309db0ad0924a057a6b481750665531..d7e05dd5b5b235b7b166b22c5b094dc364e28dfc 100644 --- a/paddle/fluid/inference/tensorrt/convert/ut_helper.h +++ b/paddle/fluid/inference/tensorrt/convert/ut_helper.h @@ -19,6 +19,9 @@ limitations under the License. */ #pragma once +#include +#include + #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/analysis/helper.h" @@ -58,7 +61,10 @@ class TRTConvertValidation { public: TRTConvertValidation() = delete; - TRTConvertValidation(int batch_size, int workspace_size = 1 << 10) { + TRTConvertValidation(int batch_size, + const std::unordered_set& parameters, + framework::Scope& scope, int workspace_size = 1 << 10) + : parameters_(parameters), scope_(scope) { // create engine. engine_.reset(new TensorRTEngine(10, 1 << 10, &stream_)); engine_->InitNetwork(); @@ -73,19 +79,22 @@ class TRTConvertValidation { engine_->DeclareInput(name, nvinfer1::DataType::kFLOAT, dims); } + // Declare a parameter varaible in the scope. + void DeclParamVar(const std::string& name, const nvinfer1::Dims& dims) { + DeclVar(name, dims); + } + void DeclOutputVar(const std::string& name, const nvinfer1::Dims& dims) { DeclVar(name, dims); } + // Declare a variable in a fluid Scope. void DeclVar(const std::string& name, const nvinfer1::Dims& dims) { platform::CPUPlace place; platform::CPUDeviceContext ctx(place); // Init Fluid tensor. - std::vector dim_vec(dims.nbDims); - for (int i = 0; i < dims.nbDims; i++) { - dim_vec[i] = dims.d[i]; - } + std::vector dim_vec(dims.d, dims.d + dims.nbDims); auto* x = scope_.Var(name); auto* x_tensor = x->GetMutable(); x_tensor->Resize(framework::make_ddim(dim_vec)); @@ -96,20 +105,22 @@ class TRTConvertValidation { op_ = framework::OpRegistry::CreateOp(desc); OpConverter op_converter; - op_converter.ConvertOp(desc, engine_.get()); + op_converter.ConvertOp(desc, parameters_, scope_, engine_.get()); engine_->FreezeNetwork(); // Declare outputs. - op_desc_.reset(new framework::OpDesc(desc, nullptr, nullptr)); + op_desc_.reset(new framework::OpDesc(desc, nullptr)); // Set Inputs. for (const auto& input : op_desc_->InputArgumentNames()) { + if (parameters_.count(input)) continue; auto* var = scope_.FindVar(input); PADDLE_ENFORCE(var); auto tensor = var->GetMutable(); + engine_->SetInputFromCPU( - input, static_cast(tensor->data()), + input, static_cast(tensor->data()), sizeof(float) * analysis::AccuDims(tensor->dims(), tensor->dims().size())); } @@ -117,18 +128,21 @@ class TRTConvertValidation { void Execute(int batch_size) { // Execute Fluid Op - // Execute TRT platform::CPUPlace place; platform::CPUDeviceContext ctx(place); - engine_->Execute(batch_size); - op_->Run(scope_, place); + // Execute TRT. + engine_->Execute(batch_size); + cudaStreamSynchronize(*engine_->stream()); ASSERT_FALSE(op_desc_->OutputArgumentNames().empty()); + const size_t output_space_size = 200; for (const auto& output : op_desc_->OutputArgumentNames()) { std::vector fluid_out; - std::vector trt_out(200); - engine_->GetOutputInCPU(output, &trt_out[0], 200 * sizeof(float)); + std::vector trt_out(output_space_size); + engine_->GetOutputInCPU(output, &trt_out[0], + output_space_size * sizeof(float)); + cudaStreamSynchronize(*engine_->stream()); auto* var = scope_.FindVar(output); auto tensor = var->GetMutable(); @@ -136,7 +150,7 @@ class TRTConvertValidation { // Compare two output ASSERT_FALSE(fluid_out.empty()); for (size_t i = 0; i < fluid_out.size(); i++) { - EXPECT_LT(std::abs(fluid_out[i] - trt_out[i]), 0.001); + EXPECT_LT(std::abs(fluid_out[i] - trt_out[i]), 1e-6); } } } @@ -146,9 +160,10 @@ class TRTConvertValidation { private: std::unique_ptr engine_; cudaStream_t stream_; - framework::Scope scope_; std::unique_ptr op_; std::unique_ptr op_desc_; + const std::unordered_set& parameters_; + framework::Scope& scope_; }; } // namespace tensorrt diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index a88236ae98e1816fc43796ead596c432b798d7de..3d75fefc1a735168131a6c67ac073e80aba32945 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -106,6 +106,7 @@ void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer* layer, int offset, name); auto* output = layer->getOutput(offset); + SetITensor(name, output); PADDLE_ENFORCE(output != nullptr); output->setName(name.c_str()); infer_network_->markOutput(*output); diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index d9d3163b66d4c4c302d12edcc42f00e1cdfa5a30..fabcfd9e80cc0ef2637201a1499ebbe2d6adfd8c 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -37,13 +37,15 @@ class TensorRTEngine : public EngineBase { // Weight is model parameter. class Weight { public: - Weight(nvinfer1::DataType dtype, void* value, int num_elem) { + Weight(nvinfer1::DataType dtype, void* value, size_t num_elem) { w_.type = dtype; w_.values = value; w_.count = num_elem; } const nvinfer1::Weights& get() { return w_; } + std::vector dims; + private: nvinfer1::Weights w_; }; diff --git a/paddle/fluid/operators/bilinear_interp_op.cc b/paddle/fluid/operators/bilinear_interp_op.cc index d46fda54e7a9d5bc737a7ec2116daca33ffa015f..3321adf2743c28f6eeca8b5cc91ef89beed6b97c 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cc +++ b/paddle/fluid/operators/bilinear_interp_op.cc @@ -34,9 +34,22 @@ class BilinearInterpOp : public framework::OperatorWithKernel { int out_w = ctx->Attrs().Get("out_w"); PADDLE_ENFORCE_EQ(dim_x.size(), 4, "X's dimension must be 4"); + if (ctx->HasInput("OutSize")) { + auto out_size_dim = ctx->GetInputDim("OutSize"); + PADDLE_ENFORCE_EQ(out_size_dim.size(), 1, + "OutSize's dimension size must be 1"); + PADDLE_ENFORCE_EQ(out_size_dim[0], 2, "OutSize's dim[0] must be 2"); + } std::vector dim_out({dim_x[0], dim_x[1], out_h, out_w}); ctx->SetOutputDim("Out", framework::make_ddim(dim_out)); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace()); + } }; class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker { @@ -45,6 +58,10 @@ class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker { 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)"); + AddInput("OutSize", + "(Tensor) This is a 1-D tensor with two number. " + "The first number is height and the second number is width.") + .AsDispensable(); AddOutput("Out", "(Tensor) The dimension of output is (N x C x out_h x out_w]"); @@ -78,6 +95,12 @@ class BilinearInterpOpGrad : public framework::OperatorWithKernel { ctx->SetOutputDim(framework::GradVarName("X"), dim_x); } } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace()); + } }; } // namespace operators diff --git a/paddle/fluid/operators/bilinear_interp_op.cu b/paddle/fluid/operators/bilinear_interp_op.cu index 510190f1aaf02960284216a1bedd409011088499..4c1971538495c6f111e9db18f4014786f6f0dd58 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cu +++ b/paddle/fluid/operators/bilinear_interp_op.cu @@ -102,10 +102,21 @@ class BilinearInterpOpCUDAKernel : public framework::OpKernel { 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"); + auto out_dims = output_t->dims(); + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + Tensor sizes; + framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes); + auto size_data = sizes.data(); + out_h = size_data[0]; + out_w = size_data[1]; + } + auto* output = output_t->mutable_data( + {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace()); + int batch_size = input_t->dims()[0]; int channels = input_t->dims()[1]; int in_h = input_t->dims()[2]; @@ -139,8 +150,8 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel { 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* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto& device_ctx = ctx.template device_context(); @@ -149,6 +160,16 @@ class BilinearInterpGradOpCUDAKernel : public framework::OpKernel { int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + Tensor sizes; + framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes); + auto size_data = sizes.data(); + out_h = size_data[0]; + out_w = size_data[1]; + } + int batch_size = d_input_t->dims()[0]; int channels = d_input_t->dims()[1]; int in_h = d_input_t->dims()[2]; diff --git a/paddle/fluid/operators/bilinear_interp_op.h b/paddle/fluid/operators/bilinear_interp_op.h index f6cd77e4d49b53ecde6a84908cdffc7e1e02ac6a..8b03cd5a0635584a45782fe5a4823c37fe4fa8e8 100644 --- a/paddle/fluid/operators/bilinear_interp_op.h +++ b/paddle/fluid/operators/bilinear_interp_op.h @@ -24,11 +24,18 @@ class BilinearInterpKernel : public framework::OpKernel { 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 out_dims = output_t->dims(); 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"); + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + auto out_size_data = out_size_t->data(); + out_h = out_size_data[0]; + out_w = out_size_data[1]; + } + auto* output = output_t->mutable_data( + {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace()); int batch_size = input_t->dims()[0]; int channels = input_t->dims()[1]; int in_h = input_t->dims()[2]; @@ -83,9 +90,8 @@ class BilinearInterpGradKernel : public framework::OpKernel { 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* d_input = d_input_t->mutable_data(ctx.GetPlace()); auto& device_ctx = ctx.template device_context(); math::SetConstant zero; @@ -93,6 +99,14 @@ class BilinearInterpGradKernel : public framework::OpKernel { int out_h = ctx.Attr("out_h"); int out_w = ctx.Attr("out_w"); + + auto out_size_t = ctx.Input("OutSize"); + if (out_size_t != nullptr) { + auto out_size_data = out_size_t->data(); + out_h = out_size_data[0]; + out_w = out_size_data[1]; + } + int batch_size = d_input_t->dims()[0]; int channels = d_input_t->dims()[1]; int in_h = d_input_t->dims()[2]; diff --git a/paddle/fluid/operators/detail/CMakeLists.txt b/paddle/fluid/operators/detail/CMakeLists.txt index b9a66474c9afc27462f9c47af1a0465e2cec70bc..cf20530513cf6cd420e56b2f6378225f73c2bc8b 100644 --- a/paddle/fluid/operators/detail/CMakeLists.txt +++ b/paddle/fluid/operators/detail/CMakeLists.txt @@ -1,6 +1,7 @@ if(WITH_DISTRIBUTE) grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc - grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor selected_rows) + request_handler_impl.cc rpc_server.cc grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor + selected_rows memory) set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") set_source_files_properties(serde_test.cc grpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) cc_test(serde_test SRCS serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/detail/grpc_client.cc index f7ce7786874285795878b655365974f082c00b44..da9ca1a0c1d55018141f0e4285fe35d7c437fd55 100644 --- a/paddle/fluid/operators/detail/grpc_client.cc +++ b/paddle/fluid/operators/detail/grpc_client.cc @@ -205,6 +205,8 @@ void RPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) { } bool RPCClient::Wait() { + VLOG(3) << "RPCClient begin Wait()" + << " req_count_:" << req_count_; if (req_count_ <= 0) { return true; } diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc index 361cc24b5ba11e2654f1282327730befaeca9f55..e73756d89004bc48339c0aa31dd0857c2ca6722d 100644 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ b/paddle/fluid/operators/detail/grpc_server.cc @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/*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. @@ -12,19 +12,12 @@ 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/detail/grpc_server.h" - #include #include -using ::grpc::ServerAsyncResponseWriter; +#include "paddle/fluid/operators/detail/grpc_server.h" -DEFINE_int32(rpc_server_handle_send_threads, 20, - "Number of threads used to handle send at rpc server."); -DEFINE_int32(rpc_server_handle_get_threads, 20, - "Number of threads used to handle get at rpc server."); -DEFINE_int32(rpc_server_handle_prefetch_threads, 1, - "Number of threads used to handle prefetch at rpc server."); +using ::grpc::ServerAsyncResponseWriter; namespace paddle { namespace operators { @@ -36,49 +29,40 @@ enum CallStatus { PROCESS = 0, FINISH }; class RequestBase { public: explicit RequestBase(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, bool sync_mode, - const platform::DeviceContext* dev_ctx) + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) : service_(service), cq_(cq), - sync_mode_(sync_mode), status_(PROCESS), - dev_ctx_(dev_ctx) { + request_handler_(request_handler), + req_id_(req_id) { PADDLE_ENFORCE(cq_); } virtual ~RequestBase() {} - virtual void Process() { assert(false); } + virtual void Process() = 0; CallStatus Status() { return status_; } void SetStatus(CallStatus status) { status_ = status; } - virtual std::string GetReqName() { - assert(false); - return ""; - } + virtual std::string GetReqName() = 0; protected: ::grpc::ServerContext ctx_; GrpcService::AsyncService* service_; ::grpc::ServerCompletionQueue* cq_; - const bool sync_mode_; CallStatus status_; - const platform::DeviceContext* dev_ctx_; + RequestHandler* request_handler_; + int req_id_; }; class RequestSend final : public RequestBase { public: explicit RequestSend(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, bool sync_mode, - framework::Scope* scope, ReceivedQueue* queue, - const platform::DeviceContext* dev_ctx, int req_id) - : RequestBase(service, cq, sync_mode, dev_ctx), - queue_(queue), - responder_(&ctx_), - req_id_(req_id) { - if (sync_mode_) { - request_.reset(new VariableResponse(scope, dev_ctx_, false)); - } else { - request_.reset(new VariableResponse(scope, dev_ctx_, true)); - } + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { + request_.reset(new VariableResponse(request_handler->scope(), + request_handler->dev_ctx(), + !request_handler->sync_mode())); int method_id = static_cast(detail::GrpcMethod::kSendVariable); service_->RequestAsyncUnary( method_id, &ctx_, request_.get(), &responder_, cq_, cq_, @@ -87,12 +71,17 @@ class RequestSend final : public RequestBase { virtual ~RequestSend() {} - virtual std::string GetReqName() { return request_->Varname(); } + std::string GetReqName() override { return request_->Varname(); } + + void Process() override { + std::string varname = GetReqName(); + VLOG(3) << "RequestSend var_name:" << varname; - virtual void Process() { - std::string var_name = GetReqName(); - VLOG(3) << "RequestSend " << var_name; - queue_->Push(std::make_pair(var_name, request_)); + auto scope = request_->GetMutableLocalScope(); + auto invar = request_->GetVar(); + framework::Variable* outvar = nullptr; + + request_handler_->Handle(varname, scope, invar, &outvar); status_ = FINISH; responder_.Finish(reply_, ::grpc::Status::OK, @@ -102,105 +91,85 @@ class RequestSend final : public RequestBase { protected: sendrecv::VoidMessage reply_; std::shared_ptr request_; - ReceivedQueue* queue_; ServerAsyncResponseWriter responder_; - int req_id_; }; class RequestGet final : public RequestBase { public: explicit RequestGet(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, bool sync_mode, - framework::Scope* scope, - const platform::DeviceContext* dev_ctx, - framework::BlockingQueue* queue, - int req_id) - : RequestBase(service, cq, sync_mode, dev_ctx), - responder_(&ctx_), - scope_(scope), - queue_(queue), - req_id_(req_id) { + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_) { auto method_id = static_cast(detail::GrpcMethod::kGetVariable); service_->RequestAsyncUnary( method_id, &ctx_, &request_, &responder_, cq_, cq_, - reinterpret_cast(static_cast(req_id_))); + reinterpret_cast(static_cast(req_id))); } virtual ~RequestGet() {} - virtual std::string GetReqName() { return request_.varname(); } + std::string GetReqName() override { return request_.varname(); } - virtual void Process() { + void Process() override { // proc request. - std::string var_name = request_.varname(); - VLOG(3) << "RequestGet " << var_name; - auto* var = scope_->FindVar(var_name); + std::string varname = request_.varname(); + VLOG(3) << "RequestGet " << varname; + + auto scope = request_handler_->scope(); + auto invar = scope->FindVar(varname); + framework::Variable* outvar = nullptr; - if (var_name != FETCH_BARRIER_MESSAGE) { - SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply_); + request_handler_->Handle(varname, scope, invar, &outvar); + + if (outvar) { + SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(), + &reply_); } status_ = FINISH; responder_.Finish(reply_, ::grpc::Status::OK, reinterpret_cast(static_cast(req_id_))); - - if (var_name == FETCH_BARRIER_MESSAGE) { - sendrecv::VariableMessage msg; - MessageWithName msg_with_name = std::make_pair(var_name, msg); - queue_->Push(msg_with_name); - } } protected: sendrecv::VariableMessage request_; ::grpc::ByteBuffer reply_; ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; - framework::Scope* scope_; - framework::BlockingQueue* queue_; - int req_id_; }; class RequestPrefetch final : public RequestBase { public: explicit RequestPrefetch(GrpcService::AsyncService* service, - ::grpc::ServerCompletionQueue* cq, bool sync_mode, - framework::Scope* scope, - const platform::DeviceContext* dev_ctx, - framework::Executor* executor, - framework::ProgramDesc* program, - framework::ExecutorPrepareContext* prefetch_ctx, - int req_id) - : RequestBase(service, cq, sync_mode, dev_ctx), + ::grpc::ServerCompletionQueue* cq, + RequestHandler* request_handler, int req_id) + : RequestBase(service, cq, request_handler, req_id), responder_(&ctx_), - scope_(scope), - executor_(executor), - program_(program), - prefetch_ctx_(prefetch_ctx), - req_id_(req_id) { - // prefetch always create a new sub scope - request_.reset(new VariableResponse(scope, dev_ctx_, true)); + local_scope_(nullptr) { + request_.reset(new VariableResponse(request_handler->scope(), + request_handler->dev_ctx(), true)); int method_id = static_cast(detail::GrpcMethod::kPrefetchVariable); service_->RequestAsyncUnary( method_id, &ctx_, request_.get(), &responder_, cq_, cq_, - reinterpret_cast(static_cast(req_id_))); + reinterpret_cast(static_cast(req_id))); } virtual ~RequestPrefetch() {} - virtual std::string GetReqName() { return request_->Varname(); } + std::string GetReqName() override { return request_->Varname(); } - virtual void Process() { + void Process() override { // prefetch process... + std::string varname = request_->OutVarname(); + VLOG(3) << "RequestPrefetch " << varname; + + auto scope = request_->GetMutableLocalScope(); + auto invar = scope->FindVar(varname); + framework::Variable* outvar = nullptr; - std::string var_name = request_->OutVarname(); - VLOG(3) << "RequestPrefetch " << var_name; - auto var_desc = program_->Block(0).FindVar(var_name); - framework::Scope* local_scope = request_->GetMutableLocalScope(); - auto* var = local_scope->FindVar(var_name); - InitializeVariable(var, var_desc->GetType()); - executor_->RunPreparedContext(prefetch_ctx_, local_scope); + request_handler_->Handle(varname, scope, invar, &outvar); - SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply_); + SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(), + &reply_); status_ = FINISH; responder_.Finish(reply_, ::grpc::Status::OK, @@ -211,202 +180,169 @@ class RequestPrefetch final : public RequestBase { std::shared_ptr request_; ::grpc::ByteBuffer reply_; ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; - framework::Scope* scope_; - framework::Executor* executor_; - framework::ProgramDesc* program_; - framework::ExecutorPrepareContext* prefetch_ctx_; - int req_id_; + framework::Scope* local_scope_; }; -void AsyncGRPCServer::WaitClientGet(int count) { - int fetch_barriers = 0; - while (fetch_barriers < count) { - auto msg = var_get_queue_.Pop(); - if (msg.first == FETCH_BARRIER_MESSAGE) { - fetch_barriers++; - } - } -} - void AsyncGRPCServer::WaitServerReady() { + VLOG(3) << "AsyncGRPCServer is wait server ready"; std::unique_lock lock(this->mutex_ready_); condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); + VLOG(3) << "AsyncGRPCServer WaitSeverReady"; } -void AsyncGRPCServer::RunSyncUpdate() { +void AsyncGRPCServer::StartServer() { ::grpc::ServerBuilder builder; - builder.AddListeningPort(address_, ::grpc::InsecureServerCredentials(), + builder.AddListeningPort(bind_address_, ::grpc::InsecureServerCredentials(), &selected_port_); + builder.SetMaxSendMessageSize(std::numeric_limits::max()); builder.SetMaxReceiveMessageSize(std::numeric_limits::max()); builder.RegisterService(&service_); - cq_send_ = builder.AddCompletionQueue(); - cq_get_ = builder.AddCompletionQueue(); - cq_prefetch_ = builder.AddCompletionQueue(); + for (auto t : rpc_call_map_) { + rpc_cq_[t.first].reset(builder.AddCompletionQueue().release()); + } server_ = builder.BuildAndStart(); - LOG(INFO) << "Server listening on " << address_ + LOG(INFO) << "Server listening on " << bind_address_ << " selected port: " << selected_port_; - std::function send_register = std::bind( - &AsyncGRPCServer::TryToRegisterNewSendOne, this, std::placeholders::_1); - std::function get_register = std::bind( - &AsyncGRPCServer::TryToRegisterNewGetOne, this, std::placeholders::_1); - std::function prefetch_register = - std::bind(&AsyncGRPCServer::TryToRegisterNewPrefetchOne, this, - std::placeholders::_1); + std::function f = + std::bind(&AsyncGRPCServer::TryToRegisterNewOne, this, + std::placeholders::_1, std::placeholders::_2); - for (int i = 0; i < kSendReqsBufSize; ++i) { - TryToRegisterNewSendOne(i); - } - for (int i = 0; i < kGetReqsBufSize; ++i) { - TryToRegisterNewGetOne(i); - } - for (int i = 0; i < kPrefetchReqsBufSize; ++i) { - TryToRegisterNewPrefetchOne(i); - } + for (auto& t : rpc_call_map_) { + auto& rpc_name = t.first; + auto& cq = rpc_cq_[rpc_name]; + auto threadnum = rpc_thread_num_[rpc_name]; + auto& reqs = rpc_reqs_[rpc_name]; - for (int i = 0; i < FLAGS_rpc_server_handle_send_threads; ++i) { - t_sends_.emplace_back( - new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, - cq_send_.get(), "cq_send", send_register))); - } - for (int i = 0; i < FLAGS_rpc_server_handle_get_threads; ++i) { - t_gets_.emplace_back( - new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, - cq_get_.get(), "cq_get", get_register))); - } - for (int i = 0; i < FLAGS_rpc_server_handle_prefetch_threads; ++i) { - t_prefetchs_.emplace_back(new std::thread( - std::bind(&AsyncGRPCServer::HandleRequest, this, cq_prefetch_.get(), - "cq_prefetch", prefetch_register))); + reqs.reserve(kRequestBufSize); + + for (int i = 0; i < kRequestBufSize; i++) { + TryToRegisterNewOne(rpc_name, i); + } + + for (int i = 0; i < threadnum; i++) { + rpc_threads_[rpc_name].emplace_back(new std::thread(std::bind( + &AsyncGRPCServer::HandleRequest, this, cq.get(), rpc_name, f))); + VLOG(3) << t.first << " creates threads!"; + } } + { std::lock_guard lock(this->mutex_ready_); ready_ = 1; } condition_ready_.notify_all(); + // wait server server_->Wait(); - for (int i = 0; i < FLAGS_rpc_server_handle_send_threads; ++i) { - t_sends_[i]->join(); - } - for (int i = 0; i < FLAGS_rpc_server_handle_get_threads; ++i) { - t_gets_[i]->join(); - } - for (int i = 0; i < FLAGS_rpc_server_handle_prefetch_threads; ++i) { - t_prefetchs_[i]->join(); + + for (auto& t : rpc_threads_) { + auto& threads = t.second; + for (size_t i = 0; i < threads.size(); ++i) { + threads[i]->join(); + VLOG(3) << t.first << " threads ends!"; + } } } void AsyncGRPCServer::ShutdownQueue() { - std::unique_lock lock(cq_mutex_); - cq_send_->Shutdown(); - cq_get_->Shutdown(); - cq_prefetch_->Shutdown(); + for (auto& t : rpc_cq_) { + t.second->Shutdown(); + VLOG(3) << t.first << " shutdown!"; + } } -// This URL explains why shutdown is complicate: -void AsyncGRPCServer::ShutDown() { +void AsyncGRPCServer::ShutDownImpl() { + std::unique_lock lock(cq_mutex_); is_shut_down_ = true; ShutdownQueue(); + + VLOG(3) << "server_ shutdown!"; server_->Shutdown(); } -void AsyncGRPCServer::TryToRegisterNewSendOne(int i) { +void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, + int req_id) { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { VLOG(3) << "shutdown, do not TryToRegisterNewSendOne"; return; } - RequestSend* send = new RequestSend(&service_, cq_send_.get(), sync_mode_, - scope_, &var_recv_queue_, dev_ctx_, i); - send_reqs_[i] = static_cast(send); - VLOG(4) << "Create RequestSend status:" << send->Status(); -} -void AsyncGRPCServer::TryToRegisterNewGetOne(int req_id) { - std::unique_lock lock(cq_mutex_); - if (is_shut_down_) { - VLOG(3) << "shutdown, do not TryToRegisterNewGetOne"; - return; + VLOG(4) << "register send rpc_name:" << rpc_name + << ", handler:" << rpc_call_map_[kRequestSend]; + + auto& reqs = rpc_reqs_[rpc_name]; + auto& handler = rpc_call_map_[rpc_name]; + auto& cq = rpc_cq_[rpc_name]; + + RequestBase* b = nullptr; + if (rpc_name == kRequestSend) { + b = new RequestSend(&service_, cq.get(), handler, req_id); + } else if (rpc_name == kRequestGet) { + b = new RequestGet(&service_, cq.get(), handler, req_id); + } else if (rpc_name == kRequestPrefetch) { + b = new RequestPrefetch(&service_, cq.get(), handler, req_id); + } else { + PADDLE_ENFORCE(false, "not surpported rpc"); } - RequestGet* get = new RequestGet(&service_, cq_get_.get(), sync_mode_, scope_, - dev_ctx_, &var_get_queue_, req_id); - get_reqs_[req_id] = static_cast(get); - VLOG(4) << "Create RequestGet status:" << get->Status(); -} -void AsyncGRPCServer::TryToRegisterNewPrefetchOne(int req_id) { - std::unique_lock lock(cq_mutex_); - if (is_shut_down_) { - VLOG(3) << "shutdown, do not TryToRegisterNewPrefetchOne"; - return; - } - RequestPrefetch* prefetch = new RequestPrefetch( - &service_, cq_prefetch_.get(), sync_mode_, scope_, dev_ctx_, executor_, - program_, prefetch_ctx_.get(), req_id); - prefetch_reqs_[req_id] = static_cast(prefetch); + reqs[req_id] = b; - VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status(); + VLOG(4) << "Create RequestSend status:" << b->Status(); } -// FIXME(typhoonzero): change cq_name to enum. void AsyncGRPCServer::HandleRequest( - ::grpc::ServerCompletionQueue* cq, const std::string& cq_name, - std::function TryToRegisterNewOne) { + ::grpc::ServerCompletionQueue* cq, const std::string& rpc_name, + std::function TryToRegisterNewOne) { void* tag = NULL; bool ok = false; while (true) { - VLOG(3) << "HandleRequest for " << cq_name << " wait Next"; + VLOG(3) << "HandleRequest " << rpc_name << " wait next"; if (!cq->Next(&tag, &ok)) { - LOG(INFO) << cq_name << " CompletionQueue shutdown!"; + LOG(INFO) << "CompletionQueue " << rpc_name << " shutdown!"; break; } - VLOG(3) << "HandleRequest for " << cq_name << " get Next"; - int req_id = static_cast(reinterpret_cast(tag)); - 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); - VLOG(3) << "HandleRequest for " << cq_name << " after WaitCond"; - } + int req_id = static_cast(reinterpret_cast(tag)); + VLOG(3) << "HandleRequest " << rpc_name << ", req_id:" << req_id + << " get next"; + auto& reqs = rpc_reqs_[rpc_name]; RequestBase* base = nullptr; { - std::lock_guard l(cq_mutex_); - if (cq_name == "cq_get") { - base = get_reqs_[req_id]; - } else if (cq_name == "cq_send") { - base = send_reqs_[req_id]; - } else if (cq_name == "cq_prefetch") { - base = prefetch_reqs_[req_id]; - } + PADDLE_ENFORCE(req_id >= 0 && req_id < kRequestBufSize); + std::unique_lock lock(cq_mutex_); + base = reqs[req_id]; } + // reference: // https://github.com/tensorflow/tensorflow/issues/5596 // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM // https://groups.google.com/forum/#!topic/grpc-io/ywATt88Ef_I if (!ok) { - LOG(WARNING) << cq_name << " recv no regular event:argument name[" + LOG(WARNING) << "completion queue:" << rpc_name + << " recv no regular event:argument name[" << base->GetReqName() << "]"; - TryToRegisterNewOne(req_id); + TryToRegisterNewOne(rpc_name, req_id); delete base; continue; } + VLOG(3) << "queue id:" << rpc_name << ", req_id:" << req_id + << ", status:" << base->Status(); + switch (base->Status()) { case PROCESS: { base->Process(); - VLOG(4) << cq_name << " PROCESS status:" << base->Status(); break; } case FINISH: { - TryToRegisterNewOne(req_id); - VLOG(4) << cq_name << " FINISH status:" << base->Status(); + TryToRegisterNewOne(rpc_name, req_id); delete base; break; } @@ -415,20 +351,6 @@ void AsyncGRPCServer::HandleRequest( } } -void AsyncGRPCServer::WaitCond(int cond) { - std::unique_lock lock(this->barrier_mutex_); - barrier_condition_.wait(lock, - [=] { return this->barrier_cond_step_ == cond; }); -} - -void AsyncGRPCServer::SetCond(int cond) { - { - std::lock_guard lock(this->barrier_mutex_); - barrier_cond_step_ = cond; - } - barrier_condition_.notify_all(); -} - } // namespace detail } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h index bdff9801a928699f8391bfb68c1c7bd2d75aa642..d1fcbc414f123c5c4810d9cecf807a406aa2c405 100644 --- a/paddle/fluid/operators/detail/grpc_server.h +++ b/paddle/fluid/operators/detail/grpc_server.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once +#include +#include #include #include // NOLINT #include @@ -28,6 +30,8 @@ limitations under the License. */ #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/operators/detail/grpc_service.h" +#include "paddle/fluid/operators/detail/request_handler.h" +#include "paddle/fluid/operators/detail/rpc_server.h" #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" @@ -37,106 +41,48 @@ namespace paddle { namespace operators { namespace detail { -typedef std::pair> - ReceivedMessage; -typedef framework::BlockingQueue ReceivedQueue; - -typedef std::pair MessageWithName; class RequestBase; -class AsyncGRPCServer final { +class AsyncGRPCServer final : public RPCServer { public: - explicit AsyncGRPCServer(const std::string &address, bool sync_mode) - : address_(address), sync_mode_(sync_mode), ready_(0) {} - - ~AsyncGRPCServer() {} - void WaitServerReady(); - void RunSyncUpdate(); - - // functions to sync server barrier status. - void WaitCond(int cond); - void SetCond(int cond); - void WaitClientGet(int count); - - void SetScope(framework::Scope *scope) { scope_ = scope; } - - void SetDevCtx(const platform::DeviceContext *dev_ctx) { dev_ctx_ = dev_ctx; } - - void SetProgram(framework::ProgramDesc *program) { program_ = program; } - - void SetExecutor(framework::Executor *executor) { executor_ = executor; } - - void SetPrefetchPreparedCtx( - std::unique_ptr prepared) { - prefetch_ctx_.reset(prepared.release()); - } - - int GetSelectedPort() const { return selected_port_; } - - const ReceivedMessage Get() { return this->var_recv_queue_.Pop(); } + explicit AsyncGRPCServer(const std::string& address, int client_num) + : RPCServer(address, client_num), ready_(0) {} - void Push(const std::string &msg_name) { - this->var_recv_queue_.Push(std::make_pair(msg_name, nullptr)); - } + virtual ~AsyncGRPCServer() {} + void WaitServerReady() override; + void StartServer() override; - void ShutDown(); + private: + void HandleRequest( + ::grpc::ServerCompletionQueue* cq, const std::string& rpc_name, + std::function TryToRegisterNewOne); - protected: - void HandleRequest(::grpc::ServerCompletionQueue *cq, - const std::string &cq_name, - std::function TryToRegisterNewOne); - void TryToRegisterNewSendOne(int req_id); - void TryToRegisterNewGetOne(int req_id); - void TryToRegisterNewPrefetchOne(int req_id); + void TryToRegisterNewOne(const std::string& rpc_name, int req_id); void ShutdownQueue(); + void ShutDownImpl() override; private: - static const int kSendReqsBufSize = 100; - static const int kGetReqsBufSize = 100; - static const int kPrefetchReqsBufSize = 10; + static const int kRequestBufSize = 100; std::mutex cq_mutex_; volatile bool is_shut_down_ = false; - std::unique_ptr<::grpc::ServerCompletionQueue> cq_send_; - std::unique_ptr<::grpc::ServerCompletionQueue> cq_get_; - std::unique_ptr<::grpc::ServerCompletionQueue> cq_prefetch_; - - RequestBase *send_reqs_[kSendReqsBufSize]; - RequestBase *get_reqs_[kGetReqsBufSize]; - RequestBase *prefetch_reqs_[kPrefetchReqsBufSize]; GrpcService::AsyncService service_; 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. - framework::BlockingQueue var_get_queue_; - // client send variable to this queue. - ReceivedQueue var_recv_queue_; - // condition of the sub program std::mutex barrier_mutex_; mutable int barrier_cond_step_; std::condition_variable barrier_condition_; - std::vector> t_sends_; - std::vector> t_gets_; - std::vector> t_prefetchs_; - - std::unique_ptr t_prefetch_; - - std::unique_ptr prefetch_ctx_; - framework::ProgramDesc *program_; - framework::Executor *executor_; - int selected_port_; - std::mutex mutex_ready_; std::condition_variable condition_ready_; + int ready_; + + std::map> rpc_cq_; + std::map>> rpc_threads_; + std::map> rpc_reqs_; }; }; // namespace detail diff --git a/paddle/fluid/operators/detail/grpc_server_test.cc b/paddle/fluid/operators/detail/grpc_server_test.cc index 350a7ee1234da5b88d09ea955ce14b7c161d804e..f97f638701cfb263f28dddbdc3bc80fb16468744 100644 --- a/paddle/fluid/operators/detail/grpc_server_test.cc +++ b/paddle/fluid/operators/detail/grpc_server_test.cc @@ -24,13 +24,16 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/operators/detail/request_handler_impl.h" + namespace framework = paddle::framework; namespace platform = paddle::platform; namespace detail = paddle::operators::detail; USE_OP(lookup_table); -std::unique_ptr rpc_service_; +std::unique_ptr g_rpc_service; +std::unique_ptr g_req_handler; framework::BlockDesc* AppendPrefetchBlcok(framework::ProgramDesc* program) { auto root_block = program->MutableBlock(0); @@ -88,8 +91,7 @@ void InitTensorsOnServer(framework::Scope* scope, platform::CPUPlace* place, } } -void StartServer(const std::string& endpoint) { - rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, true)); +void StartServer() { framework::ProgramDesc program; framework::Scope scope; platform::CPUPlace place; @@ -99,42 +101,59 @@ void StartServer(const std::string& endpoint) { auto prepared = exe.Prepare(program, block->ID()); InitTensorsOnServer(&scope, &place, 10); - rpc_service_->SetProgram(&program); - rpc_service_->SetPrefetchPreparedCtx(std::move(prepared)); - rpc_service_->SetDevCtx(&ctx); - rpc_service_->SetScope(&scope); - rpc_service_->SetExecutor(&exe); + g_req_handler->SetProgram(&program); + g_req_handler->SetPrefetchPreparedCtx(std::move(prepared)); + g_req_handler->SetDevCtx(&ctx); + g_req_handler->SetScope(&scope); + g_req_handler->SetExecutor(&exe); + + g_rpc_service->RegisterRPC(detail::kRequestPrefetch, g_req_handler.get()); + g_req_handler->SetRPCServer(g_rpc_service.get()); + + std::thread server_thread( + std::bind(&detail::AsyncGRPCServer::StartServer, g_rpc_service.get())); - rpc_service_->RunSyncUpdate(); + // FIXME(gongwb): don't use hard time. + sleep(10); + LOG(INFO) << "got nccl id and stop server..."; + g_rpc_service->ShutDown(); + server_thread.join(); } -TEST(PREFETCH, DISABLED_CPU) { - // start up a server instance backend - std::thread server_thread(StartServer, "127.0.0.1:8889"); - sleep(2); +TEST(PREFETCH, CPU) { + g_req_handler.reset(new detail::RequestPrefetchHandler(true)); + g_rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", 1)); + + std::thread server_thread(StartServer); + g_rpc_service->WaitServerReady(); + + detail::RPCClient client; + int port = g_rpc_service->GetSelectedPort(); + std::string ep = paddle::string::Sprintf("127.0.0.1:%d", port); + framework::Scope scope; platform::CPUPlace place; platform::CPUDeviceContext ctx(place); - // create var on local scope - int64_t rows_numel = 5; - InitTensorsOnClient(&scope, &place, rows_numel); - std::string in_var_name("ids"); - std::string out_var_name("out"); - - auto client = detail::RPCClient::GetInstance(); - client->AsyncPrefetchVariable("127.0.0.1:8889", ctx, scope, in_var_name, - out_var_name); - client->Wait(); - - auto var = scope.Var(out_var_name); - auto value = var->GetMutable()->value(); - auto ptr = value.mutable_data(place); - - rpc_service_->ShutDown(); - server_thread.join(); - rpc_service_.reset(nullptr); - - for (int64_t i = 0; i < rows_numel; ++i) { - EXPECT_EQ(ptr[0 + i * value.dims()[1]], static_cast(i * 2)); + { + // create var on local scope + int64_t rows_numel = 5; + InitTensorsOnClient(&scope, &place, rows_numel); + std::string in_var_name("ids"); + std::string out_var_name("out"); + + client.AsyncPrefetchVariable(ep, ctx, scope, in_var_name, out_var_name); + client.Wait(); + auto var = scope.Var(out_var_name); + auto value = var->GetMutable()->value(); + auto ptr = value.mutable_data(place); + + for (int64_t i = 0; i < rows_numel; ++i) { + EXPECT_EQ(ptr[0 + i * value.dims()[1]], static_cast(i * 2)); + } } + + server_thread.join(); + LOG(INFO) << "begin reset"; + g_rpc_service.reset(nullptr); + g_req_handler.reset(nullptr); } diff --git a/paddle/fluid/operators/detail/request_handler.h b/paddle/fluid/operators/detail/request_handler.h new file mode 100644 index 0000000000000000000000000000000000000000..4bc5e7f10ee2a8939d230fe96517bd9f56c13933 --- /dev/null +++ b/paddle/fluid/operators/detail/request_handler.h @@ -0,0 +1,127 @@ +// 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/framework/data_type.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/detail/sendrecvop_utils.h" + +namespace paddle { +namespace operators { +namespace detail { + +constexpr char kRequestSend[] = "RequestSend"; +constexpr char kRequestGet[] = "RequestGet"; +constexpr char kRequestPrefetch[] = "RequestPrefetch"; + +class RPCServer; + +class RequestHandler { + public: + explicit RequestHandler(bool sync_mode) + : sync_mode_(sync_mode), + dev_ctx_(nullptr), + executor_(nullptr), + scope_(nullptr), + program_(nullptr), + rpc_server_(nullptr) {} + + virtual ~RequestHandler() {} + + // Set attributes. + void SetScope(framework::Scope* scope) { scope_ = scope; } + void SetDevCtx(const platform::DeviceContext* dev_ctx) { dev_ctx_ = dev_ctx; } + void SetProgram(framework::ProgramDesc* program) { program_ = program; } + void SetExecutor(framework::Executor* executor) { executor_ = executor; } + void SetPrefetchPreparedCtx( + std::unique_ptr prepared) { + prefetch_ctx_.reset(prepared.release()); + } + + // Used for async. + void SetGradToPreparedCtx( + std::unordered_map< + std::string, std::shared_ptr>* g) { + grad_to_prepared_ctx_ = g; + } + + void SetRPCServer(RPCServer* rpc_server) { rpc_server_ = rpc_server; } + + // Get attributes. + bool sync_mode() { return sync_mode_; } + framework::Scope* scope() { return scope_; } + const platform::DeviceContext* dev_ctx() { return dev_ctx_; } + framework::ExecutorPrepareContext* prefetch_ctx() { + return prefetch_ctx_.get(); + } + framework::ProgramDesc* program() { return program_; } + framework::Executor* executor() { return executor_; } + std::vector& sparse_vars() { return sparse_vars_; } + + // This function processes user's rpc request. + // The implemention is in request_handler_impl. + // example: + // std::string varname = request_.varname(); + // + // auto scope = request_handler_->scope(); + // auto invar = scope->FindVar(varname); + // framework::Variable* outvar = nullptr; + // + // request_handler_->Handle(varname, scope, invar, &outvar); + // if (outvar) { + // SerializeToByteBuffer(varname, outvar, + // *request_handler_->dev_ctx(), &reply_); + // } + virtual bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, + framework::Variable** outvar) = 0; + + protected: + const bool sync_mode_; + + const platform::DeviceContext* dev_ctx_; + framework::Executor* executor_; + framework::Scope* scope_; + framework::ProgramDesc* program_; + std::unique_ptr prefetch_ctx_; + + // Used for async. + std::unordered_map>* + grad_to_prepared_ctx_; + + // Record received sparse variables, so that + // we could reset those after execute optimize program + std::vector sparse_vars_; + RPCServer* rpc_server_; + + std::mutex sparse_var_mutex_; +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler_impl.cc b/paddle/fluid/operators/detail/request_handler_impl.cc new file mode 100644 index 0000000000000000000000000000000000000000..f16c06d52f4fb86d51083a8b3b98d05a64c1af74 --- /dev/null +++ b/paddle/fluid/operators/detail/request_handler_impl.cc @@ -0,0 +1,115 @@ +// 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 "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/request_handler_impl.h" +#include "paddle/fluid/operators/detail/rpc_server.h" +#include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "paddle/fluid/operators/detail/variable_response.h" + +namespace paddle { +namespace operators { +namespace detail { + +bool RequestSendHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar) { + VLOG(4) << "RequestSendHandler:" << varname; + + // Async + if (!sync_mode_) { + try { + executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(), + scope); + } catch (std::exception& e) { + LOG(ERROR) << "async: run sub program error " << e.what(); + return false; + } + return true; + } + + // Sync + if (varname == BATCH_BARRIER_MESSAGE) { + VLOG(3) << "sync: recv batch barrier message"; + rpc_server_->IncreaseBatchBarrier(kRequestSend); + } else { + VLOG(3) << "sync: received var_name: " << varname; + if (sync_mode_) { + rpc_server_->WaitCond(kRequestSend); + } + + if (invar == nullptr) { + LOG(ERROR) << "sync: Can not find server side var: " << varname; + PADDLE_THROW("sync: Can not find server side var"); + return false; + } + + if (invar->IsType()) { + std::unique_lock lock(sparse_var_mutex_); + sparse_vars_.push_back(invar); + } + } + + return true; +} + +bool RequestGetHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar) { + VLOG(4) << "RequestGetHandler:" << varname; + + if (varname != FETCH_BARRIER_MESSAGE) { + if (sync_mode_) { + rpc_server_->WaitCond(kRequestGet); + } + *outvar = scope_->FindVar(varname); + return true; + } + + // FETCH_BARRIER_MESSAGE + if (sync_mode_) { + VLOG(3) << "sync: recv fetch barrier message"; + rpc_server_->IncreaseBatchBarrier(kRequestGet); + } + + return true; +} + +bool RequestPrefetchHandler::Handle(const std::string& varname, + framework::Scope* scope, + framework::Variable* invar, + framework::Variable** outvar) { + VLOG(4) << "RequestPrefetchHandler " << varname; + + auto var_desc = program_->Block(0).FindVar(varname); + *outvar = scope->FindVar(varname); + InitializeVariable(*outvar, var_desc->GetType()); + executor_->RunPreparedContext(prefetch_ctx_.get(), scope); + + return true; +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/request_handler_impl.h b/paddle/fluid/operators/detail/request_handler_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..8d0c62232b68ad6c05e751c25103802ee12db57e --- /dev/null +++ b/paddle/fluid/operators/detail/request_handler_impl.h @@ -0,0 +1,64 @@ +// 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/framework/data_type.h" +#include "paddle/fluid/framework/executor.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/detail/request_handler.h" +#include "paddle/fluid/operators/detail/sendrecvop_utils.h" + +namespace paddle { +namespace operators { +namespace detail { + +class RequestSendHandler final : public RequestHandler { + public: + explicit RequestSendHandler(bool sync_mode) : RequestHandler(sync_mode) {} + virtual ~RequestSendHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar) override; +}; + +class RequestGetHandler final : public RequestHandler { + public: + explicit RequestGetHandler(bool sync_mode) : RequestHandler(sync_mode) {} + virtual ~RequestGetHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar) override; +}; + +class RequestPrefetchHandler final : public RequestHandler { + public: + explicit RequestPrefetchHandler(bool sync_mode) : RequestHandler(sync_mode) {} + virtual ~RequestPrefetchHandler() {} + bool Handle(const std::string& varname, framework::Scope* scope, + framework::Variable* var, framework::Variable** outvar) override; +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_server.cc b/paddle/fluid/operators/detail/rpc_server.cc new file mode 100644 index 0000000000000000000000000000000000000000..448763372a8c224cc68319a4a444915896b68234 --- /dev/null +++ b/paddle/fluid/operators/detail/rpc_server.cc @@ -0,0 +1,113 @@ +// 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/operators/detail/rpc_server.h" + +namespace paddle { +namespace operators { +namespace detail { + +void RPCServer::ShutDown() { + LOG(INFO) << "RPCServer ShutDown "; + ShutDownImpl(); + + exit_flag_ = true; + barrier_cond_.notify_all(); + rpc_cond_.notify_all(); +} + +void RPCServer::SavePort() const { + auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); + std::ofstream port_file; + port_file.open(file_path); + port_file << selected_port_; + port_file.close(); + VLOG(4) << "selected port written to " << file_path; +} + +void RPCServer::WaitBarrier(const std::string& rpc_name) { + std::unique_lock lock(this->mutex_); + barrier_cond_.wait(lock, [=] { + return (barrier_counter_[rpc_name] >= client_num_ || exit_flag_.load()); + }); + + VLOG(3) << "batch_barrier_:" << barrier_counter_[rpc_name]; +} + +void RPCServer::IncreaseBatchBarrier(const std::string rpc_name) { + VLOG(3) << "RPCServer begin IncreaseBatchBarrier " << rpc_name; + int b = 0; + { + std::unique_lock lock(mutex_); + b = ++barrier_counter_[rpc_name]; + } + + VLOG(3) << "RPCServer IncreaseBatchBarrier " << rpc_name + << ", barrier_count:" << b << ", fan_in" << client_num_; + + if (b >= client_num_) { + barrier_cond_.notify_all(); + } +} + +void RPCServer::ResetBarrierCounter() { + VLOG(3) << "RPCServer ResetBarrierCounter "; + std::unique_lock lock(mutex_); + for (auto& t : barrier_counter_) { + t.second = 0; + } +} + +void RPCServer::RegisterRPC(const std::string& rpc_name, + RequestHandler* handler, int thread_num) { + rpc_call_map_[rpc_name] = handler; + rpc_thread_num_[rpc_name] = thread_num; + + static int cond = -1; + rpc_cond_map_[rpc_name] = ++cond; + VLOG(4) << "RegisterRPC rpc_name:" << rpc_name << ", handler:" << handler + << ", cond:" << rpc_cond_map_[rpc_name]; +} + +void RPCServer::SetCond(const std::string& rpc_name) { + VLOG(3) << "RPCServer SetCond " << rpc_name; + { + std::unique_lock lock(mutex_); + cur_cond_ = rpc_cond_map_[rpc_name]; + } + + rpc_cond_.notify_all(); +} + +void RPCServer::WaitCond(const std::string& rpc_name) { + VLOG(3) << "RPCServer WaitCond " << rpc_name; + int cond = 0; + { + std::unique_lock lock(mutex_); + cond = rpc_cond_map_[rpc_name]; + } + + std::unique_lock lock(mutex_); + rpc_cond_.wait( + lock, [=] { return (cur_cond_.load() == cond || exit_flag_.load()); }); +} + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/rpc_server.h b/paddle/fluid/operators/detail/rpc_server.h new file mode 100644 index 0000000000000000000000000000000000000000..c2e7ae706c9dc6776e09b25e424b30f110c3855d --- /dev/null +++ b/paddle/fluid/operators/detail/rpc_server.h @@ -0,0 +1,91 @@ +// 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 // NOLINT +#include +#include +#include "paddle/fluid/operators/detail/request_handler.h" + +namespace paddle { +namespace operators { +namespace detail { + +class RPCServer { + public: + explicit RPCServer(const std::string& address, int client_num) + : cur_cond_(0), + bind_address_(address), + exit_flag_(false), + selected_port_(0), + client_num_(client_num) {} + + virtual ~RPCServer() {} + virtual void StartServer() = 0; + virtual void WaitServerReady() = 0; + + void ShutDown(); + + bool IsExit() { return exit_flag_.load(); } + + int GetSelectedPort() const { return selected_port_; } + void SavePort() const; + + // RegisterRPC, register the rpc method name to a handler + // class, and auto generate a condition id for this call + // to be used for the barrier. + void RegisterRPC(const std::string& rpc_name, RequestHandler* handler, + int thread_num = 5); + + // Wait util all the clients have reached the barrier for one + // rpc method. This function should be called in the + // RequestHandler if you want to run the server/client in a + // synchronous mode. + void WaitBarrier(const std::string& rpc_name); + + void SetCond(const std::string& rpc_name); + void WaitCond(const std::string& rpc_name); + void IncreaseBatchBarrier(const std::string rpc_name); + void ResetBarrierCounter(); + + protected: + virtual void ShutDownImpl() = 0; + + private: + std::mutex mutex_; + std::unordered_map barrier_counter_; + std::condition_variable barrier_cond_; + + std::unordered_map rpc_cond_map_; + std::atomic cur_cond_; + std::condition_variable rpc_cond_; + + protected: + std::string bind_address_; + std::atomic exit_flag_; + int selected_port_; + + const int client_num_; + + std::unordered_map rpc_call_map_; + std::unordered_map rpc_thread_num_; + friend class RequestHandler; +}; + +}; // 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 bf624da2a6c26472e47711b3c6409f78afba0a64..69cfd784f8dd4f129f50c6882061e53e8535b949 100644 --- a/paddle/fluid/operators/detail/variable_response.h +++ b/paddle/fluid/operators/detail/variable_response.h @@ -67,8 +67,8 @@ class VariableResponse { framework::Scope* GetMutableLocalScope() const { return local_scope_; } - inline std::string Varname() { return meta_.varname(); } - inline std::string OutVarname() { return meta_.out_varname(); } + inline std::string Varname() const { return meta_.varname(); } + inline std::string OutVarname() const { return meta_.out_varname(); } // should call parse first. framework::Variable* GetVar() { diff --git a/paddle/fluid/operators/gather_op.cc b/paddle/fluid/operators/gather_op.cc index e21b57258928856a10d6e86c3e2c6e81fb241ee3..aa3e05b83b23569a4dd9c83294916e289f993abc 100644 --- a/paddle/fluid/operators/gather_op.cc +++ b/paddle/fluid/operators/gather_op.cc @@ -33,7 +33,6 @@ class GatherOp : public framework::OperatorWithKernel { auto index_dims = ctx->GetInputDim("Index"); PADDLE_ENFORCE(index_dims.size() == 1); int batch_size = ctx->GetInputDim("Index")[0]; - PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0"); framework::DDim output_dims(ctx->GetInputDim("X")); output_dims[0] = batch_size; ctx->SetOutputDim("Out", output_dims); diff --git a/paddle/fluid/operators/gen_nccl_id_op.cc b/paddle/fluid/operators/gen_nccl_id_op.cc index a5678f63466d368b3dd59380c18f9625cabd368b..4bce2d322d825110a446c9bc5eccdacf0ba3c943 100644 --- a/paddle/fluid/operators/gen_nccl_id_op.cc +++ b/paddle/fluid/operators/gen_nccl_id_op.cc @@ -23,6 +23,7 @@ limitations under the License. */ #include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/operators/detail/grpc_client.h" #include "paddle/fluid/operators/detail/grpc_server.h" +#include "paddle/fluid/operators/detail/request_handler_impl.h" #include "paddle/fluid/platform/nccl_helper.h" namespace paddle { @@ -75,19 +76,23 @@ class GenNCCLIdOp : public framework::OperatorBase { // NOTE: Can not use unique_ptr here because the default // deleter will call GRPC Server's base class's dtor and // that will cause a wired crash. - detail::AsyncGRPCServer rpc_service(endpoint, true); + detail::RequestSendHandler rpc_h(true); + detail::AsyncGRPCServer rpc_service(endpoint, 1); + rpc_service.RegisterRPC(detail::kRequestSend, &rpc_h); + rpc_h.SetRPCServer(&rpc_service); + framework::ProgramDesc empty_program; framework::Executor executor(dev_ctx.GetPlace()); - rpc_service.SetScope(scope); - rpc_service.SetDevCtx(&dev_ctx); - rpc_service.SetProgram(&empty_program); - rpc_service.SetExecutor(&executor); + rpc_h.SetScope(scope); + rpc_h.SetDevCtx(&dev_ctx); + rpc_h.SetProgram(&empty_program); + rpc_h.SetExecutor(&executor); std::thread server_thread( - std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, &rpc_service)); - rpc_service.SetCond(0); + std::bind(&detail::AsyncGRPCServer::StartServer, &rpc_service)); + rpc_service.SetCond(detail::kRequestSend); VLOG(3) << "start getting nccl id from trainer 0..."; - auto recv = rpc_service.Get(); + rpc_service.WaitBarrier(detail::kRequestSend); VLOG(3) << "got nccl id and stop server..."; rpc_service.ShutDown(); VLOG(3) << "rpc server stopped"; diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index df5f229acd75ee3df55d46444a63d9f1915f9d22..71e75c25321812c849e205460217b174d80654be 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -19,14 +19,16 @@ limitations under the License. */ #include // NOLINT #include +#include "paddle/fluid/operators/detail/grpc_server.h" +#include "paddle/fluid/operators/detail/request_handler_impl.h" #include "paddle/fluid/operators/listen_and_serv_op.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { -void RunServer(std::shared_ptr service) { - service->RunSyncUpdate(); +void RunServer(std::shared_ptr service) { + service->StartServer(); VLOG(4) << "RunServer thread end"; } static void split(const std::string &str, char sep, @@ -67,8 +69,6 @@ static void ParallelExecuteBlocks( for (size_t i = 0; i < fs.size(); ++i) fs[i].wait(); } -std::atomic_int ListenAndServOp::selected_port_{0}; - ListenAndServOp::ListenAndServOp(const std::string &type, const framework::VariableNameMap &inputs, const framework::VariableNameMap &outputs, @@ -78,7 +78,6 @@ ListenAndServOp::ListenAndServOp(const std::string &type, ListenAndServOp::~ListenAndServOp() { Stop(); } void ListenAndServOp::Stop() { - rpc_service_->Push(LISTEN_TERMINATE_MESSAGE); rpc_service_->ShutDown(); server_thread_->join(); auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); @@ -87,26 +86,13 @@ void ListenAndServOp::Stop() { void ListenAndServOp::SavePort() const { // NOTE: default write file to /tmp/paddle.selected_port - selected_port_ = rpc_service_->GetSelectedPort(); - auto file_path = string::Sprintf("/tmp/paddle.%d.port", ::getpid()); - std::ofstream port_file; - port_file.open(file_path); - port_file << selected_port_.load(); - port_file.close(); - VLOG(4) << "selected port written to " << file_path; -} - -void ListenAndServOp::WaitServerReady() { - while (selected_port_.load() == 0) { - } + rpc_service_->SavePort(); } void ListenAndServOp::RunSyncLoop(framework::Executor *executor, framework::ProgramDesc *program, framework::Scope *recv_scope, framework::BlockDesc *prefetch_block) const { - auto fan_in = Attr("Fanin"); - size_t num_blocks = program->Size(); PADDLE_ENFORCE_GE(num_blocks, 2, "server program should have at least 2 blocks"); @@ -121,49 +107,24 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor, optimize_prepared.begin(), std::shared_ptr(nullptr)); - bool exit_flag = false; + rpc_service_->ResetBarrierCounter(); // Record received sparse variables, so that // we could reset those after execute optimize program std::vector sparse_vars; - while (!exit_flag && !SignalHandler::IsProgramExit()) { + while (true) { // Get from multiple trainers, we don't care about the order in which // the gradients arrives, just add suffix 0~n and merge the gradient. - rpc_service_->SetCond(0); - size_t recv_var_cnt = 0; - int batch_barrier = 0; - while (batch_barrier != fan_in) { - 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 if (recv_var_name == BATCH_BARRIER_MESSAGE) { - VLOG(3) << "recv batch barrier message"; - batch_barrier++; - continue; - } else { - VLOG(3) << "received grad: " << recv_var_name; - recv_var_cnt++; - 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"); - } - if (var->IsType()) { - sparse_vars.push_back(var); - } - } - } - if (exit_flag) { - rpc_service_->SetCond(1); - rpc_service_->ShutDown(); + rpc_service_->SetCond(detail::kRequestSend); + rpc_service_->WaitBarrier(detail::kRequestSend); + + if (rpc_service_->IsExit()) { + LOG(WARNING) << "get exit!rpc_processor break!"; + rpc_service_->SetCond(detail::kRequestGet); break; } // NOTE: if is_gpu_place, CUDA kernels are launched by multiple threads // and this will still work. - // The optimize blocks which have the same parent ID would run parallel // TODO(Yancey1989): need to use ParallelExecutor for future int32_t last_parent_blkid = program->Block(1).Parent(); @@ -194,52 +155,18 @@ void ListenAndServOp::RunSyncLoop(framework::Executor *executor, var->GetMutable()->mutable_rows()->clear(); } - rpc_service_->SetCond(1); - // FIXME(typhoonzero): use another condition to sync wait clients get. - rpc_service_->WaitClientGet(fan_in); - sparse_vars.clear(); + rpc_service_->SetCond(detail::kRequestGet); + rpc_service_->WaitBarrier(detail::kRequestGet); + rpc_service_->ResetBarrierCounter(); } // while(true) } -static void AsyncUpdateThread( - const std::string &var_name, const bool &exit_flag, - const std::shared_ptr &queue, - framework::Executor *executor, - framework::ExecutorPrepareContext *prepared) { - VLOG(3) << "update thread for " << var_name << " started"; - while (!exit_flag && !SignalHandler::IsProgramExit()) { - const detail::ReceivedMessage v = queue->Pop(); - if (SignalHandler::IsProgramExit()) { - VLOG(3) << "update thread for " << var_name << " exit"; - break; - } - auto recv_var_name = v.first; - VLOG(4) << "async update " << 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"); - } - auto fs = framework::Async([var_name, &executor, &v, prepared] { - try { - executor->RunPreparedContext(prepared, - v.second->GetMutableLocalScope()); - } catch (const std::exception &e) { - LOG(ERROR) << "run sub program error " << e.what(); - } - }); - fs.wait(); - } -} - void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, framework::ProgramDesc *program) const { VLOG(3) << "RunAsyncLoop in"; // grad name to block id std::unordered_map grad_to_block_id; std::unordered_map id_to_grad; - std::unordered_map> - grad_to_queue; auto grad_to_block_id_str = Attr>("grad_to_block_id"); @@ -249,13 +176,9 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, 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; - std::shared_ptr queue = - std::make_shared(); - grad_to_queue[pieces[0]] = queue; - // record blocking queue in SignalHandler - SignalHandler::RegisterBlockingQueue(queue); id_to_grad[block_id] = pieces[0]; } size_t num_blocks = program->Size(); @@ -274,39 +197,36 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, grad_to_prepared_ctx[id_to_grad[block_list[i]]] = optimize_prepared[i]; } - bool exit_flag = false; + request_send_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); + request_get_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); + request_prefetch_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); - VLOG(3) << "start async optimize threads"; - std::vector> fs; - for (auto iter = grad_to_queue.begin(); iter != grad_to_queue.end(); iter++) { - std::string grad_name = iter->first; - VLOG(3) << "create async update thread for " << grad_name; - fs.push_back(framework::AsyncIO([grad_name, &exit_flag, &executor, - &grad_to_queue, &grad_to_prepared_ctx]() { - AsyncUpdateThread(grad_name, exit_flag, grad_to_queue[grad_name], - executor, grad_to_prepared_ctx[grad_name].get()); - })); - } VLOG(3) << "RunAsyncLoop into while"; - while (!exit_flag && !SignalHandler::IsProgramExit()) { - 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; + while (true) { + if (rpc_service_->IsExit()) { + LOG(INFO) << "get exit!rpc_processor break!"; break; - } else { - VLOG(3) << "received grad: " << recv_var_name; - grad_to_queue[recv_var_name]->Push(v); } - if (exit_flag) { - rpc_service_->ShutDown(); - break; - } + sleep(1); } // while(true) } +static void FillRequestCtx(detail::RequestHandler *h, framework::Scope *scope, + platform::DeviceContext *dev_ctx, + framework::Executor *executor, + framework::ProgramDesc *program, + framework::ExecutorPrepareContext *prefetch_ctx, + detail::RPCServer *rpc_server) { + h->SetScope(scope); + h->SetDevCtx(dev_ctx); + h->SetExecutor(executor); + h->SetProgram(program); + h->SetPrefetchPreparedCtx(std::move( + std::unique_ptr(prefetch_ctx))); + h->SetRPCServer(rpc_server); +} + void ListenAndServOp::RunImpl(const framework::Scope &scope, const platform::Place &dev_place) const { // Mark this as PS that it should decide profiling by listening from trainer. @@ -316,27 +236,42 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, framework::Scope &recv_scope = scope.NewScope(); bool sync_mode = Attr("sync_mode"); + auto fan_in = Attr("Fanin"); PADDLE_ENFORCE(!rpc_service_); std::string endpoint = Attr("endpoint"); - rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, sync_mode)); + LOG(INFO) << "sync_mode:" << sync_mode << ", fan_in:" << fan_in + << ", end_point:" << endpoint; + + // request_handler_.reset(new detail::GRPCRequestSendHandler(sync_mode)); + rpc_service_.reset(new detail::AsyncGRPCServer(endpoint, fan_in)); + request_send_handler_.reset(new detail::RequestSendHandler(sync_mode)); + request_get_handler_.reset(new detail::RequestGetHandler(sync_mode)); + request_prefetch_handler_.reset( + new detail::RequestPrefetchHandler(sync_mode)); + + rpc_service_->RegisterRPC(detail::kRequestSend, request_send_handler_.get()); + rpc_service_->RegisterRPC(detail::kRequestGet, request_get_handler_.get()); + rpc_service_->RegisterRPC(detail::kRequestPrefetch, + request_prefetch_handler_.get()); auto *optimize_block = Attr(kOptimizeBlock); auto *prefetch_block = Attr(kPrefetchBlock); auto *program = optimize_block->Program(); framework::Executor executor(dev_place); - // prepare rpc_service - rpc_service_->SetScope(&recv_scope); - rpc_service_->SetDevCtx(&dev_ctx); - rpc_service_->SetProgram(program); - rpc_service_->SetExecutor(&executor); - // prepare for prefetch VLOG(3) << "prefetch block id is " << prefetch_block->ID(); auto prefetch_prepared = executor.Prepare(*program, prefetch_block->ID()); - rpc_service_->SetPrefetchPreparedCtx(std::move(prefetch_prepared)); + + auto f = std::bind(FillRequestCtx, std::placeholders::_1, &recv_scope, + &dev_ctx, &executor, program, prefetch_prepared.release(), + rpc_service_.get()); + + f(request_send_handler_.get()); + f(request_get_handler_.get()); + f(request_prefetch_handler_.get()); // start the server listening after all member initialized. server_thread_.reset(new std::thread(RunServer, rpc_service_)); @@ -348,8 +283,6 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, signal(SIGTERM, SignalHandler::StopAndExit); // Write to a file of server selected port for python use. - std::string file_path = string::Sprintf("/tmp/paddle.%d.selected_port", - static_cast(::getpid())); SavePort(); if (sync_mode) { RunSyncLoop(&executor, program, &recv_scope, prefetch_block); @@ -385,27 +318,9 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { } }; -bool SignalHandler::program_exit_flag_ = false; - -SignalHandler::BlockingQueueSet SignalHandler::blocking_queue_set_{}; - void SignalHandler::StopAndExit(int signal_num) { VLOG(3) << "Catch interrupt signal: " << signal_num << ", program will exit"; - - program_exit_flag_ = true; - - // awake all blocking queues - for (BlockingQueueSet::iterator iter = blocking_queue_set_.begin(); - iter != blocking_queue_set_.end(); iter++) { - iter->get()->Push( - std::make_pair(std::string(LISTEN_TERMINATE_MESSAGE), nullptr)); - } - - exit(EXIT_SUCCESS); -} - -void SignalHandler::RegisterBlockingQueue(BlockingQueue &queue) { - blocking_queue_set_.insert(queue); + exit(0); } } // namespace operators diff --git a/paddle/fluid/operators/listen_and_serv_op.h b/paddle/fluid/operators/listen_and_serv_op.h index 6f868369dcf2067fd71f4107d20c79ead0cf9f56..87952cb0e683596b2b0395890b6e25b15f74d7e2 100644 --- a/paddle/fluid/operators/listen_and_serv_op.h +++ b/paddle/fluid/operators/listen_and_serv_op.h @@ -23,7 +23,8 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/threadpool.h" -#include "paddle/fluid/operators/detail/grpc_server.h" +#include "paddle/fluid/operators/detail/request_handler.h" +#include "paddle/fluid/operators/detail/rpc_server.h" namespace paddle { namespace operators { @@ -31,7 +32,7 @@ namespace operators { constexpr char kOptimizeBlock[] = "OptimizeBlock"; constexpr char kPrefetchBlock[] = "PrefetchBlock"; -void RunServer(std::shared_ptr service); +void RunServer(std::shared_ptr service); class ListenAndServOp : public framework::OperatorBase { public: @@ -52,41 +53,27 @@ class ListenAndServOp : public framework::OperatorBase { void SavePort() const; - void WaitServerReady(); - - int GetSelectedPort() { return selected_port_; } + int GetSelectedPort() { return rpc_service_->GetSelectedPort(); } void Stop() override; void RunImpl(const framework::Scope& scope, const platform::Place& dev_place) const override; - static void ResetPort() { selected_port_ = 0; } - protected: - mutable std::shared_ptr rpc_service_; + mutable std::shared_ptr rpc_service_; + mutable std::shared_ptr request_send_handler_; + mutable std::shared_ptr request_get_handler_; + mutable std::shared_ptr request_prefetch_handler_; + mutable std::shared_ptr server_thread_; - // FIXME(wuyi): it's static so that the operator can be cloned. - static std::atomic_int selected_port_; }; class SignalHandler { - public: - typedef std::shared_ptr BlockingQueue; - typedef std::unordered_set BlockingQueueSet; - public: static void StopAndExit(int signal_num); - static void RegisterBlockingQueue(BlockingQueue&); - - static inline bool IsProgramExit() { return program_exit_flag_; } - private: - static bool program_exit_flag_; - - static BlockingQueueSet blocking_queue_set_; - DISABLE_COPY_AND_ASSIGN(SignalHandler); }; diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index 2c77ee2e2792d6fdd76bacd68b6c3b4a296b2e3a..bcd8e81609a37cc544f5a5cc4188400c1632a668 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -46,6 +46,8 @@ class SendBarrierOp : public framework::OperatorBase { auto rpc_client = detail::RPCClient::GetInstance(); + VLOG(3) << "SendBarrierOp sync_mode:" << sync_mode; + // need to wait before sending send_barrier message PADDLE_ENFORCE(rpc_client->Wait()); if (sync_mode) { diff --git a/paddle/fluid/operators/shape_op.cc b/paddle/fluid/operators/shape_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c75fce7959d1af51afd52af23fe657d10a2f3988 --- /dev/null +++ b/paddle/fluid/operators/shape_op.cc @@ -0,0 +1,54 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/shape_op.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class ShapeOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input (Input) of get_shape op should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output (Out) of get_shape op should not be null."); + auto in_dim = ctx->GetInputDim("Input"); + ctx->SetOutputDim("Out", {in_dim.size()}); + } +}; + +class ShapeOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Input", "(Tensor), The input tensor."); + AddOutput("Out", "(Tensor), The shape of input tensor."); + AddComment(R"DOC( +Shape Operator. +Get the shape of input tensor. +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(shape, ops::ShapeOp, ops::ShapeOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(shape, ops::ShapeKernel, ops::ShapeKernel, + ops::ShapeKernel, ops::ShapeKernel); diff --git a/paddle/fluid/operators/shape_op.cu b/paddle/fluid/operators/shape_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..7736a2a1e13cfa5d445411b3efac7669a7bf23a2 --- /dev/null +++ b/paddle/fluid/operators/shape_op.cu @@ -0,0 +1,20 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/shape_op.h" + +REGISTER_OP_CUDA_KERNEL(shape, paddle::operators::ShapeKernel, + paddle::operators::ShapeKernel, + paddle::operators::ShapeKernel, + paddle::operators::ShapeKernel); diff --git a/paddle/fluid/operators/shape_op.h b/paddle/fluid/operators/shape_op.h new file mode 100644 index 0000000000000000000000000000000000000000..3be86b66a538e7b38a5d59095fee7e7636364bce --- /dev/null +++ b/paddle/fluid/operators/shape_op.h @@ -0,0 +1,38 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class ShapeKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in_t = ctx.Input("Input"); + auto* out_t = ctx.Output("Out"); + auto out_data = out_t->mutable_data(platform::CPUPlace()); + auto in_dims = in_t->dims(); + for (int i = 0; i < in_dims.size(); ++i) { + out_data[i] = in_dims[i]; + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/tensorrt_engine_op.cc b/paddle/fluid/operators/tensorrt_engine_op.cc index 83e768b4dc9c607b0f73d7183462d772ae7ab994..855157e7c4c5c4a43091d28d3a5414e6e386b727 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.cc +++ b/paddle/fluid/operators/tensorrt_engine_op.cc @@ -31,8 +31,9 @@ void paddle::operators::TensorRTEngineKernel::Prepare( auto max_workspace = context.Attr("max_workspace"); engine_.reset(new inference::tensorrt::TensorRTEngine( max_batch_, max_workspace, nullptr)); + // TODO(Superjomn) parameters should be passed after analysised from outside. inference::Singleton::Global().ConvertBlock( - block, engine_.get()); + block, {}, context.scope(), engine_.get()); engine_->FreezeNetwork(); } diff --git a/paddle/fluid/operators/test_send_nccl_id.cc b/paddle/fluid/operators/test_send_nccl_id.cc index 719f039a0f5fcd7445bf1589a683f122e6d62ba0..a845ba2eb038fa6a8e70dfbac06c31c19dbb9e3e 100644 --- a/paddle/fluid/operators/test_send_nccl_id.cc +++ b/paddle/fluid/operators/test_send_nccl_id.cc @@ -21,6 +21,8 @@ limitations under the License. */ #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/grpc_server.h" +#include "paddle/fluid/operators/detail/request_handler_impl.h" #include "paddle/fluid/operators/listen_and_serv_op.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" @@ -35,42 +37,44 @@ namespace m = paddle::operators::math; namespace detail = paddle::operators::detail; namespace string = paddle::string; -std::unique_ptr rpc_service; +std::unique_ptr g_rpc_service; +std::unique_ptr g_req_handler; -void StartServer(std::atomic* initialized) { +void StartServer() { f::Scope scope; p::CPUPlace place; scope.Var(NCCL_ID_VARNAME); p::DeviceContextPool& pool = p::DeviceContextPool::Instance(); auto& dev_ctx = *pool.Get(p::CPUPlace()); - rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", true)); - f::ProgramDesc empty_program; f::Executor executor(dev_ctx.GetPlace()); - rpc_service->SetScope(&scope); - rpc_service->SetDevCtx(&dev_ctx); - rpc_service->SetProgram(&empty_program); - rpc_service->SetExecutor(&executor); + g_req_handler->SetScope(&scope); + g_req_handler->SetDevCtx(&dev_ctx); + g_req_handler->SetProgram(&empty_program); + g_req_handler->SetExecutor(&executor); + + g_rpc_service->RegisterRPC(detail::kRequestSend, g_req_handler.get()); + g_req_handler->SetRPCServer(g_rpc_service.get()); std::thread server_thread( - std::bind(&detail::AsyncGRPCServer::RunSyncUpdate, rpc_service.get())); - *initialized = true; - rpc_service->SetCond(0); - auto recv = rpc_service->Get(); + std::bind(&detail::AsyncGRPCServer::StartServer, g_rpc_service.get())); + + g_rpc_service->SetCond(detail::kRequestSend); + std::cout << "before WaitFanInOfSend" << std::endl; + g_rpc_service->WaitBarrier(detail::kRequestSend); + LOG(INFO) << "got nccl id and stop server..."; - rpc_service->ShutDown(); + g_rpc_service->ShutDown(); server_thread.join(); } -TEST(SendNcclId, DISABLED_Normal) { - std::atomic initialized{false}; - std::thread server_thread(StartServer, &initialized); - while (!initialized) { - } - // wait server to start - // sleep(2); - rpc_service->WaitServerReady(); +TEST(SendNcclId, GrpcServer) { + g_req_handler.reset(new detail::RequestSendHandler(true)); + g_rpc_service.reset(new detail::AsyncGRPCServer("127.0.0.1:0", 1)); + + std::thread server_thread(StartServer); + g_rpc_service->WaitServerReady(); f::Scope scope; p::CPUPlace place; @@ -78,17 +82,20 @@ TEST(SendNcclId, DISABLED_Normal) { auto& dev_ctx = *pool.Get(p::CPUPlace()); auto var = scope.Var(NCCL_ID_VARNAME); - // var->SetType(f::proto::VarType_Type_RAW); auto id = var->GetMutable(); p::dynload::ncclGetUniqueId(id); - int port = rpc_service->GetSelectedPort(); + int port = g_rpc_service->GetSelectedPort(); + std::string ep = string::Sprintf("127.0.0.1:%d", port); detail::RPCClient client; - + LOG(INFO) << "connect to server" << ep; client.AsyncSendVariable(ep, dev_ctx, scope, NCCL_ID_VARNAME); client.Wait(); + client.AsyncSendBatchBarrier(ep); + client.Wait(); + server_thread.join(); - auto* ptr = rpc_service.release(); - delete ptr; + g_rpc_service.reset(nullptr); + g_req_handler.reset(nullptr); } diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h index 09367889a9517956ad01ad2847c31e2633cc643d..6f8e3f22db54d166cf97cfdd3d009058207a7ca5 100644 --- a/paddle/fluid/platform/nccl_helper.h +++ b/paddle/fluid/platform/nccl_helper.h @@ -15,6 +15,7 @@ #pragma once #include +#include #include // NOLINT #include #include diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index fd3834ee21d8858016c3039cfea152904ac573e2..8eeea1805d8610f6f27f422337f3526688b73de3 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -183,7 +183,7 @@ function build() { ============================================ EOF make clean - make -j `nproc` + make install -j `nproc` } function build_android() { diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 63ec83151477770ea64070cae4f5e4fcc497f7af..56f5c6b4bedb6ae864c5b6f54afc758b8be8c415 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -82,6 +82,7 @@ __all__ = [ 'roi_pool', 'dice_loss', 'upsampling_bilinear2d', + 'gather', 'random_crop', ] @@ -3889,7 +3890,6 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): def dice_loss(input, label, epsilon=0.00001): """ - **Dice loss Layer** Dice loss for comparing the similarity of two batch of data, usually is used for binary image segmentation i.e. labels are binary. The dice loss can be defined as below equation: @@ -3944,7 +3944,7 @@ def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None): input (Variable): The input tensor of bilinear interpolation, This is a 4-D tensor of the shape (num_batches, channels, in_h, in_w). - out_shape(list|tuple|None): Output shape of bilinear interpolation + out_shape(list|tuple|Variable|None): Output shape of bilinear interpolation layer, the shape is (out_h, out_w). Default: None scale(int|None): The multiplier for the input height or width. @@ -3971,13 +3971,20 @@ def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None): def _is_list_or_turple_(data): return (isinstance(data, list) or isinstance(data, tuple)) + out_h = 0 + out_w = 0 + inputs = {"X": input} if out_shape is not None: - if not (_is_list_or_turple_(out_shape) and len(out_shape) == 2): + if not (_is_list_or_turple_(out_shape) and len(out_shape) == 2) and ( + out_shape is not Variable): raise ValueError('out_shape should be a list or tuple ', 'with length 2, (out_h, out_w).') - out_shape = list(map(int, out_shape)) - out_h = out_shape[0] - out_w = out_shape[1] + if _is_list_or_turple_(out_shape): + out_shape = list(map(int, out_shape)) + out_h = out_shape[0] + out_w = out_shape[1] + else: + inputs['OutSize'] = out_shape else: out_h = int(input.shape[2] * scale) out_w = int(input.shape[3] * scale) @@ -3985,13 +3992,62 @@ def upsampling_bilinear2d(input, out_shape=None, scale=None, name=None): out = helper.create_tmp_variable(dtype) helper.append_op( type="bilinear_interp", - inputs={"X": input}, + inputs=inputs, outputs={"Out": out}, attrs={"out_h": out_h, "out_w": out_w}) return out +def gather(input, index): + """ + Output is obtained by gathering entries of the outer-most dimension + of X indexed by `index` and concatenate them together. + + .. math:: + + Out = X[Index] + + + .. code-block:: text + + + Given: + + X = [[1, 2], + [3, 4], + [5, 6]] + + Index = [1, 2] + + Then: + + Out = [[3, 4], + [5, 6]] + + Args: + input (Variable): The source input with rank>=1. + index (Variable): The index input with rank=1. + + Returns: + output (Variable): The output is a tensor with the same rank as input. + + Examples: + .. code-block:: python + + output = fluid.layers.gather(x, index) + """ + helper = LayerHelper('gather', **locals()) + dtype = helper.input_dtype() + out = helper.create_tmp_variable(dtype) + helper.append_op( + type="gather", + inputs={"X": input, + "Index": index}, + outputs={"Out": out}) + return out + + def random_crop(input, shape, seed=1): helper = LayerHelper("random_crop", **locals()) dtype = helper.input_dtype() diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index a9fe25744cc0b385479c9366af1b731ec221dd5a..60f8cbbfa714e8500606fdf68b7a23e1ffb9d37a 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -71,6 +71,7 @@ __all__ = [ 'cumsum', 'scatter', 'sum', + 'shape', ] + __activations__ for _OP in set(__all__): diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py index bffb4f3b666a7ddcc133b7c30fab132b49aa1d0e..87c11e7880e73b911f21dda77c1cc2b4850b3591 100644 --- a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py +++ b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py @@ -17,7 +17,10 @@ import numpy as np from op_test import OpTest -def bilinear_interp_np(input, out_h, out_w): +def bilinear_interp_np(input, out_h, out_w, out_size): + if out_size is not None: + out_h = out_size[0] + out_w = out_size[1] batch_size, channel, in_h, in_w = input.shape if out_h > 1: ratio_h = (in_h - 1.0) / (out_h - 1.0) @@ -49,12 +52,15 @@ def bilinear_interp_np(input, out_h, out_w): class TestBilinearInterpOp(OpTest): def setUp(self): + self.out_size = None 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) - + output_np = bilinear_interp_np(input_np, self.out_h, self.out_w, + self.out_size) self.inputs = {'X': input_np} + if self.out_size is not None: + self.inputs['OutSize'] = self.out_size self.attrs = {'out_h': self.out_h, 'out_w': self.out_w} self.outputs = {'Out': output_np} @@ -68,6 +74,7 @@ class TestBilinearInterpOp(OpTest): self.input_shape = [2, 3, 4, 4] self.out_h = 2 self.out_w = 2 + self.out_size = np.array([3, 3]).astype("int32") class TestCase1(TestBilinearInterpOp): @@ -91,5 +98,29 @@ class TestCase3(TestBilinearInterpOp): self.out_w = 128 +class TestCase4(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [4, 1, 7, 8] + self.out_h = 1 + self.out_w = 1 + self.out_size = np.array([2, 2]).astype("int32") + + +class TestCase5(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [3, 3, 9, 6] + self.out_h = 12 + self.out_w = 12 + self.out_size = np.array([11, 11]).astype("int32") + + +class TestCase6(TestBilinearInterpOp): + def init_test_case(self): + self.input_shape = [1, 1, 128, 64] + self.out_h = 64 + self.out_w = 128 + self.out_size = np.array([65, 129]).astype("int32") + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_gather_op.py b/python/paddle/fluid/tests/unittests/test_gather_op.py index 6fd043c27e27db53c95be3630b6c08216e8e35f4..4ae90864806204197c52bbbdc5516f141afd4613 100644 --- a/python/paddle/fluid/tests/unittests/test_gather_op.py +++ b/python/paddle/fluid/tests/unittests/test_gather_op.py @@ -20,8 +20,9 @@ from op_test import OpTest class TestGatherOp(OpTest): def setUp(self): self.op_type = "gather" - xnp = np.random.random((10, 20)).astype("float32") - self.inputs = {'X': xnp, 'Index': np.array([1, 3, 5]).astype("int32")} + self.config() + xnp = np.random.random(self.x_shape).astype("float32") + self.inputs = {'X': xnp, 'Index': np.array(self.index).astype("int32")} self.outputs = {'Out': self.inputs["X"][self.inputs["Index"]]} def test_check_output(self): @@ -30,6 +31,16 @@ class TestGatherOp(OpTest): def test_check_grad(self): self.check_grad(['X'], 'Out') + def config(self): + self.x_shape = (10, 20) + self.index = [1, 3, 5] + + +class TestCase1(TestGatherOp): + def config(self): + self.x_shape = (10) + self.index = [1, 3, 5] + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_shape_op.py b/python/paddle/fluid/tests/unittests/test_shape_op.py new file mode 100644 index 0000000000000000000000000000000000000000..a62ee050075cb8c9f8817c142825a89c24bdfedf --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_shape_op.py @@ -0,0 +1,47 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + + +class TestShapeOp(OpTest): + def setUp(self): + self.op_type = "shape" + self.config() + self.shape = [2, 3] + input = np.zeros(self.shape) + self.inputs = {'Input': input} + self.outputs = {'Out': np.array(self.shape)} + + def config(self): + self.shape = [2, 3] + + def test_check_output(self): + self.check_output() + + +class case1(TestShapeOp): + def config(self): + self.shape = [2] + + +class case2(TestShapeOp): + def config(self): + self.shape = [1, 2, 3] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_split_var.py b/python/paddle/fluid/tests/unittests/test_split_var.py index 0c5e8901b903375c7d4de32943e657b205d8fae9..157def9b56e44092a86023035d1ab444c938aa07 100644 --- a/python/paddle/fluid/tests/unittests/test_split_var.py +++ b/python/paddle/fluid/tests/unittests/test_split_var.py @@ -14,7 +14,7 @@ import math import unittest -from paddle.fluid.transpiler.distribute_transpiler import split_dense_variable +from paddle.fluid.transpiler.distribute_transpiler import split_variable import paddle.fluid as fluid import paddle.fluid.core as core import random @@ -31,7 +31,7 @@ class TestSplitVar(unittest.TestCase): # dtype=core.VarDesc.VarType.LOD_TENSOR, shape=shape) var_list.append(var) - blocks = split_dense_variable(var_list, 10, min_size) + blocks = split_variable(var_list, 10, min_size) all_sizes = [] for s in expected_sizes: for s2 in s: diff --git a/python/paddle/fluid/transpiler/details/__init__.py b/python/paddle/fluid/transpiler/details/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..dc597c33849dc06cc975b245099672f64c3539d3 --- /dev/null +++ b/python/paddle/fluid/transpiler/details/__init__.py @@ -0,0 +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 +# +# 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. + +from program_utils import * +from ufind import * diff --git a/python/paddle/fluid/transpiler/details/program_utils.py b/python/paddle/fluid/transpiler/details/program_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f10b496306a002ee131d01798a0698b807d379ca --- /dev/null +++ b/python/paddle/fluid/transpiler/details/program_utils.py @@ -0,0 +1,37 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +def delete_ops(block, ops): + try: + start = list(block.ops).index(ops[0]) + end = list(block.ops).index(ops[-1]) + [block.remove_op(start) for _ in xrange(end - start + 1)] + except Exception, e: + raise e + block.program.sync_with_cpp() + + +def find_op_by_input_arg(block, arg_name): + for index, op in enumerate(block.ops): + if arg_name in op.input_arg_names: + return index + return -1 + + +def find_op_by_output_arg(block, arg_name): + for index, op in enumerate(block.ops): + if arg_name in op.output_arg_names: + return index + return -1 diff --git a/python/paddle/fluid/transpiler/details/ufind.py b/python/paddle/fluid/transpiler/details/ufind.py new file mode 100644 index 0000000000000000000000000000000000000000..0e30d0e3f9c5712c494daf17b2b4bcec86f69c23 --- /dev/null +++ b/python/paddle/fluid/transpiler/details/ufind.py @@ -0,0 +1,64 @@ +# 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. + + +class UnionFind(object): + """ Union-find data structure. + + Union-find is a data structure that keeps track of a set of elements partitioned + into a number of disjoint (non-overlapping) subsets. + + Reference: + https://en.wikipedia.org/wiki/Disjoint-set_data_structure + + Args: + elements(list): The initialize element list. + """ + + def __init__(self, elementes=None): + self._parents = [] # index -> parent index + self._index = {} # element -> index + self._curr_idx = 0 + if not elementes: + elementes = [] + for ele in elementes: + self._parents.append(self._curr_idx) + self._index.update({ele: self._curr_idx}) + self._curr_idx += 1 + + def find(self, x): + # Find the root index of given element x, + # execute the path compress while findind the root index + if not x in self._index: + return -1 + idx = self._index[x] + while idx != self._parents[idx]: + t = self._parents[idx] + self._parents[idx] = self._parents[t] + idx = t + return idx + + def union(self, x, y): + # Union two given element + x_root = self.find(x) + y_root = self.find(y) + + if x_root == y_root: + return + self._parents[x_root] = y_root + + def is_connected(self, x, y): + # If two given elements have the same root index, + # then they are connected. + return self.find(x) == self.find(y) diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index e9b7d9e9d2dea54a33068d5c3fe3fbf22620d1ea..06b0a1375ce6568cca864cd8a2dd69ee46b223a7 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -11,6 +11,30 @@ # 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. +""" +Transpile the program to distributed data-parallelism programs. +The main_program will be transformed to use a remote parameter server +to do parameter optimization. And the optimization graph will be put +into a parameter server program. + +Use different methods to split trainable variables to different +parameter servers. + +Steps to transpile trainer: +1. split variable to multiple blocks, aligned by product(dim[1:]) (width). +2. rename splited grad variables to add trainer_id suffix ".trainer_%d". +3. modify trainer program add split_op to each grad variable. +4. append send_op to send splited variables to server and fetch + params(splited blocks or origin param) from server. +5. append concat_op to merge splited blocks to update local weights. + +Steps to transpile pserver: +1. create new program for parameter server. +2. create params and grad variables that assigned to current server instance. +3. create a sub-block in the server side program +4. append ops that should run on current server instance. +5. add listen_and_serv op +""" from __future__ import print_function @@ -21,9 +45,11 @@ from .. import core, framework from ..framework import Program, default_main_program, \ default_startup_program, \ Variable, Parameter, grad_var_name +from details import * LOOKUP_TABLE_TYPE = "lookup_table" LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad" +OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName() RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName( ) RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC @@ -40,62 +66,11 @@ class VarBlock: return "%s:%d:%d" % (self.varname, self.offset, self.size) -class UnionFind(object): - """ Union-find data structure. - - Union-find is a data structure that keeps track of a set of elements partitioned - into a number of disjoint (non-overlapping) subsets. - - Reference: - https://en.wikipedia.org/wiki/Disjoint-set_data_structure - - Args: - elements(list): The initialize element list. - """ - - def __init__(self, elementes=None): - self._parents = [] # index -> parent index - self._index = {} # element -> index - self._curr_idx = 0 - if not elementes: - elementes = [] - for ele in elementes: - self._parents.append(self._curr_idx) - self._index.update({ele: self._curr_idx}) - self._curr_idx += 1 - - def find(self, x): - # Find the root index of given element x, - # execute the path compress while findind the root index - if not x in self._index: - return -1 - idx = self._index[x] - while idx != self._parents[idx]: - t = self._parents[idx] - self._parents[idx] = self._parents[t] - idx = t - return idx - - def union(self, x, y): - # Union two given element - x_root = self.find(x) - y_root = self.find(y) - - if x_root == y_root: - return - self._parents[x_root] = y_root - - def is_connected(self, x, y): - # If two given elements have the same root index, - # then they are connected. - return self.find(x) == self.find(y) - - def same_or_split_var(p_name, var_name): return p_name == var_name or p_name.startswith(var_name + ".block") -def split_dense_variable(var_list, service_count, min_block_size=8192): +def split_variable(var_list, service_count, min_block_size=8192): """ We may need to split dense tensor to one or more blocks and put them equally onto parameter server. One block is a sub-tensor @@ -141,99 +116,15 @@ def split_dense_variable(var_list, service_count, min_block_size=8192): return blocks -def delete_ops(block, ops): - try: - start = list(block.ops).index(ops[0]) - end = list(block.ops).index(ops[-1]) - [block.remove_op(start) for _ in xrange(end - start + 1)] - except Exception, e: - raise e - block.program.sync_with_cpp() - - -def find_op_by_input_arg(block, arg_name): - for index, op in enumerate(block.ops): - if arg_name in op.input_arg_names: - return index - return -1 - - -def find_op_by_output_arg(block, arg_name): - for index, op in enumerate(block.ops): - if arg_name in op.output_arg_names: - return index - return -1 - - class DistributeTranspiler: - def transpile(self, - trainer_id, - program=None, - pservers="127.0.0.1:6174", - trainers=1, - split_method=RoundRobin, - sync_mode=True): - """ - Transpile the program to distributed data-parallelism programs. - The main_program will be transformed to use a remote parameter server - to do parameter optimization. And the optimization graph will be put - into a parameter server program. - - Use different methods to split trainable variables to different - parameter servers. - - Steps to transpile trainer: - 1. split variable to multiple blocks, aligned by product(dim[1:]) (width). - 2. rename splited grad variables to add trainer_id suffix ".trainer_%d". - 3. modify trainer program add split_op to each grad variable. - 4. append send_op to send splited variables to server and fetch - params(splited blocks or origin param) from server. - 5. append concat_op to merge splited blocks to update local weights. - - Steps to transpile pserver: - 1. create new program for parameter server. - 2. create params and grad variables that assigned to current server instance. - 3. create a sub-block in the server side program - 4. append ops that should run on current server instance. - 5. add listen_and_serv op - - :param trainer_id: one unique id for each trainer in a job. - :type trainer_id: int - :param program: program to transpile, default is default_main_program - :type program: Program - :param pservers: parameter server endpoints like "m1:6174,m2:6174" - :type pservers: string - :param trainers: total number of workers/trainers in the job - :type trainers: int - :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 (split_method.__bases__[0] == PSDispatcher) - if program is None: - program = default_main_program() - self.origin_program = program - self.trainer_num = trainers - 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. - self.trainer_id = trainer_id - pserver_endpoints = pservers.split(",") - self.pserver_endpoints = pserver_endpoints - self.optimize_ops, params_grads = self._get_optimize_pass() - ps_dispatcher = split_method(pserver_endpoints) - + def _has_distributed_lookup_table(self): # process lookup_table_op # 1. check all lookup_table_op is distributed # 2. check all lookup_table_op share the same table. distributed_lookup_table_ops = [] # support only one distributed_lookup_table now self.table_name = None - for op in program.global_block().ops: + for op in self.origin_program.global_block().ops: if op.type == LOOKUP_TABLE_TYPE: if op.attrs['is_distributed'] is True: if self.table_name is None: @@ -246,20 +137,13 @@ class DistributeTranspiler: if self.table_name is not None: assert op.input("W")[0] != self.table_name - self.has_distributed_lookup_table = len( - distributed_lookup_table_ops) > 0 - - # step1: For large parameters and gradients, split them into smaller - # blocks. - param_list = [] - grad_list = [] - for p, g in params_grads: - # skip parameter marked not trainable - if type(p) == Parameter and p.trainable == False: - continue - param_list.append(p) - grad_list.append(g) + return len(distributed_lookup_table_ops) > 0 + def _update_dist_lookup_table_vars(self, param_list, grad_list, + params_grads): + # TODO(wuyi): put find a way to put dist lookup table stuff all together. + # update self.table_param_grad and self.trainer_side_table_grad_list + program = self.origin_program if self.has_distributed_lookup_table: param_list = [ param for param in param_list if param.name != self.table_name @@ -277,7 +161,7 @@ class DistributeTranspiler: self.trainer_side_table_grad_list = [ program.global_block().create_var( name="%s.trainer_%d.pserver_%d" % - (table_grad_var.name, trainer_id, index), + (table_grad_var.name, self.trainer_id, index), type=table_grad_var.type, shape=table_grad_var.shape, dtype=table_grad_var.dtype) @@ -293,23 +177,41 @@ class DistributeTranspiler: for index in range(len(self.pserver_endpoints)) ] - grad_blocks = split_dense_variable(grad_list, len(pserver_endpoints)) - param_blocks = split_dense_variable(param_list, len(pserver_endpoints)) + def _init_splited_vars(self, split_method): + # update these mappings for further transpile: + # 1. param_var_mapping: param var name -> [splited params vars] + # 2. grad_var_mapping: grad var name -> [splited grads vars] + # 3. grad_param_mapping: grad.blockx -> param.blockx + # 4. param_grad_ep_mapping: ep -> {"params": [], "grads": []} + + param_list = [] + grad_list = [] + for p, g in self.params_grads: + # skip parameter marked not trainable + if type(p) == Parameter and p.trainable == False: + continue + param_list.append(p) + grad_list.append(g) + + self._update_dist_lookup_table_vars(param_list, grad_list, + self.params_grads) + + grad_blocks = split_variable(grad_list, len(self.pserver_endpoints)) + param_blocks = split_variable(param_list, len(self.pserver_endpoints)) assert (len(grad_blocks) == len(param_blocks)) - # step2: Create new vars for the parameters and gradients blocks and - # add ops to do the split. - param_var_mapping = self._create_vars_from_blocklist(program, - param_blocks) - grad_var_mapping = self._create_vars_from_blocklist( - program, grad_blocks, add_trainer_suffix=self.trainer_num > 1) - grad_param_mapping = dict() + # origin_varname -> [splited_var] + self.param_var_mapping = self._create_vars_from_blocklist( + self.origin_program, param_blocks) + self.grad_var_mapping = self._create_vars_from_blocklist( + self.origin_program, + grad_blocks, + add_trainer_suffix=self.trainer_num > 1) + self.grad_param_mapping = dict() for g, p in zip(grad_blocks, param_blocks): g_name, g_bid, _ = g.split(":") p_name, p_bid, _ = p.split(":") - grad_param_mapping[grad_var_mapping[g_name][int(g_bid)]] = \ - param_var_mapping[p_name][int(p_bid)] - - # step 3: transpile trainer side program, insert recv op and send op. + self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \ + self.param_var_mapping[p_name][int(p_bid)] # create mapping of endpoint -> split var to create pserver side program self.param_grad_ep_mapping = dict() @@ -322,10 +224,50 @@ class DistributeTranspiler: }) for ep in self.pserver_endpoints ] + def transpile(self, + trainer_id, + program=None, + pservers="127.0.0.1:6174", + trainers=1, + split_method=RoundRobin, + sync_mode=True): + """ + :param trainer_id: one unique id for each trainer in a job. + :type trainer_id: int + :param program: program to transpile, default is default_main_program + :type program: Program + :param pservers: parameter server endpoints like "m1:6174,m2:6174" + :type pservers: string + :param trainers: total number of workers/trainers in the job + :type trainers: int + :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 (split_method.__bases__[0] == PSDispatcher) + if program is None: + program = default_main_program() + self.origin_program = program + self.trainer_num = trainers + self.sync_mode = sync_mode + self.trainer_id = trainer_id + pserver_endpoints = pservers.split(",") + self.pserver_endpoints = pserver_endpoints + self.optimize_ops, self.params_grads = self._get_optimize_pass() + + ps_dispatcher = split_method(self.pserver_endpoints) + self.has_distributed_lookup_table = self._has_distributed_lookup_table() + + # split and create vars, then put splited vars in dicts for later use. + self._init_splited_vars(split_method) + # step 3.1: insert send op to send gradient vars to parameter servers ps_dispatcher.reset() send_vars = [] - for orig_varname, splited_vars in grad_var_mapping.items(): + for orig_varname, splited_vars in self.grad_var_mapping.items(): eplist = ps_dispatcher.dispatch(splited_vars) if len(splited_vars) == 1: orig_varname = splited_vars[0].name @@ -367,7 +309,7 @@ class DistributeTranspiler: # step 3.2: insert recv op to receive parameters from parameter server recv_vars = [] for _, var in enumerate(send_vars): - recv_vars.append(grad_param_mapping[var]) + recv_vars.append(self.grad_param_mapping[var]) ps_dispatcher.reset() eplist = ps_dispatcher.dispatch(recv_vars) @@ -375,7 +317,7 @@ class DistributeTranspiler: self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i]) self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i]) # step4: Concat the parameters splits together after recv. - for varname, splited_var in param_var_mapping.iteritems(): + for varname, splited_var in self.param_var_mapping.iteritems(): eps = [] for var in splited_var: index = [v.name for v in recv_vars].index(var.name) @@ -399,7 +341,7 @@ class DistributeTranspiler: RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) - for varname, splited_var in param_var_mapping.iteritems(): + for varname, splited_var in self.param_var_mapping.iteritems(): if len(splited_var) <= 1: continue orig_param = program.global_block().vars[varname] @@ -440,7 +382,6 @@ class DistributeTranspiler: # we don't need to create them when grad arrives. # change client side var name to origin name by # removing ".trainer_%d" suffix - suff_idx = v.name.find(".trainer_") if suff_idx >= 0: orig_var_name = v.name[:suff_idx] @@ -477,24 +418,14 @@ class DistributeTranspiler: # located on current pserver opt_op_on_pserver = [] for _, op in enumerate(self.optimize_ops): - if self._is_opt_op(op) and self._is_opt_op_on_pserver(endpoint, op): + if self._is_optimizer_op(op) and self._is_opt_op_on_pserver( + endpoint, op): opt_op_on_pserver.append(op) # step 3.3 # Iterate through the ops, and if an op and the optimize ops # which located on current pserver are in one set, then # append it into the sub program. - # We try to put optimization program run parallelly, assume - # optimization program always looks like: - # - # prevop -> prevop -> opt op -> following op -> following op; -> - # prevop -> prevop -> opt op -> following op -> following op; -> - # global op -> global op - # - # we put operators that can run parallelly to many program blocks. - # in above example, we seperate ops by the ";". Global ops must run - # after all the optimize ops finished. - global_ops = [] # HACK: optimization global ops only used to scale beta1 and beta2 # replace it with dependency engine. @@ -502,12 +433,18 @@ class DistributeTranspiler: if self._is_adam_connected_op(op): global_ops.append(op) - def __append_optimize_op__(op, block, grad_to_block_id): - if self._is_opt_op(op): + def __append_optimize_op__(op, block, grad_to_block_id, merged_var): + if self._is_optimizer_op(op): self._append_pserver_ops(block, op, endpoint, grad_to_block_id, - self.origin_program) + self.origin_program, merged_var) else: - self._append_pserver_non_opt_ops(block, op) + self._append_pserver_non_opt_ops(block, op, endpoint) + + def __op_have_grad_input__(op): + for varname in op.input_arg_names: + if varname.find("@GRAD") >= 0: + return varname + return "" # append lr decay ops to the child block if exists lr_ops = self._get_lr_ops() @@ -515,17 +452,26 @@ class DistributeTranspiler: lr_decay_block = pserver_program.create_block( pserver_program.num_blocks - 1) for _, op in enumerate(lr_ops): - self._append_pserver_non_opt_ops(lr_decay_block, op) + self._append_pserver_non_opt_ops(lr_decay_block, op, endpoint) # 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) + # append grad merging ops before clip and weight decay + for _, op in enumerate(self.optimize_ops): + # find the origin @GRAD var before clipping + grad_varname_for_block = __op_have_grad_input__(op) + if ufind.is_connected(op, opt_op) and grad_varname_for_block: + merged_var = self._append_pserver_grad_merge_ops( + per_opt_block, grad_varname_for_block, endpoint, + grad_to_block_id, self.origin_program) 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, grad_to_block_id) + __append_optimize_op__(op, per_opt_block, grad_to_block_id, + merged_var) # append global ops if global_ops: @@ -533,15 +479,7 @@ class DistributeTranspiler: pserver_program.num_blocks - 1) for glb_op in global_ops: __append_optimize_op__(glb_op, opt_state_block, - grad_to_block_id) - - # NOT USED: single block version: - # - # for _, op in enumerate(self.optimize_ops): - # for _, opt_op in enumerate(opt_op_on_pserver): - # if ufind.is_connected(op, opt_op): - # __append_optimize_op__(glb_op, optimize_block) - # break + grad_to_block_id, None) # process distributed lookup_table prefetch_block = None @@ -631,6 +569,8 @@ class DistributeTranspiler: attrs=op.attrs) return s_prog + # ====================== private transpiler functions ===================== + # transpiler function for dis lookup_table def _replace_lookup_table_op_with_prefetch(self, program, pserver_endpoints): @@ -836,7 +776,6 @@ class DistributeTranspiler: return table_opt_block - # ====================== private transpiler functions ===================== def _create_vars_from_blocklist(self, program, block_list, @@ -979,17 +918,74 @@ class DistributeTranspiler: pass return orig_shape - def _orig_varname(self, varname): - suff_idx = varname.find(".trainer_") + def _get_varname_parts(self, varname): + # returns origin, blockid, trainerid orig_var_name = "" - if suff_idx >= 0: - orig_var_name = varname[:suff_idx] + trainer_part = "" + block_part = "" + trainer_idx = varname.find(".trainer_") + if trainer_idx >= 0: + trainer_part = varname[trainer_idx + 1:] + else: + trainer_idx = len(varname) + block_index = varname.find(".block") + if block_index >= 0: + block_part = varname[block_index + 1:trainer_idx] else: - orig_var_name = varname - return orig_var_name + block_index = len(varname) + orig_var_name = varname[0:min(block_index, trainer_idx)] + return orig_var_name, block_part, trainer_part + + def _orig_varname(self, varname): + orig, _, _ = self._get_varname_parts(varname) + return orig + + def _append_pserver_grad_merge_ops(self, optimize_block, + grad_varname_for_block, endpoint, + grad_to_block_id, origin_program): + program = optimize_block.program + pserver_block = program.global_block() + grad_block = None + for g in self.param_grad_ep_mapping[endpoint]["grads"]: + if self._orig_varname(g.name) == \ + self._orig_varname(grad_varname_for_block): + grad_block = g + break + if not grad_block: + # do not append this op if current endpoint + # is not dealing with this grad block + return + orig_varname, block_name, trainer_name = self._get_varname_parts( + grad_block.name) + if block_name: + merged_var_name = '.'.join([orig_varname, block_name]) + else: + merged_var_name = orig_varname + merged_var = \ + pserver_block.vars[merged_var_name] + 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" % \ + (merged_var_name, i) + vars2merge.append(pserver_block.vars[per_trainer_name]) + + optimize_block.append_op( + type="sum", + inputs={"X": vars2merge}, + outputs={"Out": merged_var}) + # TODO(panyx0718): What if it's SELECTED_ROWS. + if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: + optimize_block.append_op( + type="scale", + inputs={"X": merged_var}, + outputs={"Out": merged_var}, + attrs={"scale": 1.0 / float(self.trainer_num)}) + return merged_var def _append_pserver_ops(self, optimize_block, opt_op, endpoint, - grad_to_block_id, origin_program): + grad_to_block_id, origin_program, merged_var): program = optimize_block.program pserver_block = program.global_block() new_inputs = dict() @@ -997,40 +993,6 @@ class DistributeTranspiler: # moment can use the updated shape for key in opt_op.input_names: if key == "Grad": - grad_block = None - for g in self.param_grad_ep_mapping[endpoint]["grads"]: - if same_or_split_var( - self._orig_varname(g.name), - self._orig_varname(opt_op.input(key)[0])): - grad_block = g - break - if not grad_block: - # do not append this op if current endpoint - # is not dealing with this grad block - return - merged_var = \ - pserver_block.vars[self._orig_varname(grad_block.name)] - 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" % \ - (self._orig_varname(grad_block.name), i) - vars2merge.append(pserver_block.vars[per_trainer_name]) - - optimize_block.append_op( - type="sum", - inputs={"X": vars2merge}, - outputs={"Out": merged_var}) - # TODO(panyx0718): What if it's SELECTED_ROWS. - if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS: - optimize_block.append_op( - type="scale", - 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 @@ -1089,17 +1051,31 @@ class DistributeTranspiler: outputs=outputs, attrs=opt_op.attrs) - def _append_pserver_non_opt_ops(self, optimize_block, opt_op): + def _is_splited_grad_var(self, var, var_dict): + grad_block = None + for _, g in var_dict.iteritems(): + if self._orig_varname(g.name) == self._orig_varname(var.name): + if g.name.find(".trainer_") == -1: + grad_block = g + break + return grad_block + + def _append_pserver_non_opt_ops(self, optimize_block, opt_op, endpoint): program = optimize_block.program # Append the ops for parameters that do not need to be optimized/updated inputs = self._get_input_map_from_op( self.origin_program.global_block().vars, opt_op) - for varlist in inputs.itervalues(): + for key, varlist in inputs.iteritems(): if not isinstance(varlist, list): varlist = [varlist] - for var in varlist: - if not program.global_block().vars.has_key(var.name): + # for ops like clipping and weight decay, get the splited var + # for inputs/outputs + grad_block = self._is_splited_grad_var( + var, program.global_block().vars) + if grad_block: + inputs[key] = grad_block + elif not program.global_block().vars.has_key(var.name): program.global_block().create_var( name=var.name, persistable=var.persistable, @@ -1108,13 +1084,16 @@ class DistributeTranspiler: outputs = self._get_output_map_from_op( self.origin_program.global_block().vars, opt_op) - - for varlist in outputs.itervalues(): + for key, varlist in outputs.iteritems(): if not isinstance(varlist, list): varlist = [varlist] - for var in varlist: - program.global_block().clone_variable(var) + grad_block = self._is_splited_grad_var( + var, program.global_block().vars) + if grad_block: + outputs[key] = grad_block + elif not program.global_block().vars.has_key(var.name): + program.global_block().clone_variable(var) optimize_block.append_op( type=opt_op.type, @@ -1160,9 +1139,17 @@ class DistributeTranspiler: ufind.union(op1, op2) return ufind - def _is_opt_op(self, op): - # NOTE: It's a HACK implement. - # optimize op: SGDOptimize, MomentumOptimizer, AdamOptimizer and etc... + def _is_opt_role_op(self, op): + # NOTE: depend on oprole to find out whether this op is for + # optimize + op_maker = core.op_proto_and_checker_maker + optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize + if op_maker.kOpRoleAttrName() in op.attrs and \ + int(op.attrs[op_maker.kOpRoleAttrName()]) == int(optimize_role): + return True + return False + + def _is_optimizer_op(self, op): if "Param" in op.input_names and \ "LearningRate" in op.input_names: return True @@ -1212,7 +1199,7 @@ class DistributeTranspiler: # find learning rate variables by optimize op lr_vars = set() for op in self.optimize_ops: - if self._is_opt_op(op): + if self._is_optimizer_op(op): lr_vars.add(op.input("LearningRate")[0]) find_ops = [] @@ -1229,7 +1216,7 @@ class DistributeTranspiler: # NOTE: we need to skip all optimize ops, since it is connected # with forward/backward ops and lr ops, we only need the lr ops. if op1 != op2 and self._is_op_connected(op1, op2) and \ - not self._is_opt_op(op1) and not self._is_opt_op(op2): + not self._is_optimizer_op(op1) and not self._is_optimizer_op(op2): ufind.union(op1, op2) # find all ops which is related with lr var for op1 in block.ops: @@ -1250,13 +1237,21 @@ class DistributeTranspiler: block = self.origin_program.global_block() opt_ops = [] params_grads = [] + origin_var_dict = self.origin_program.global_block().vars for op in block.ops: - if self._is_opt_op(op): + if self._is_opt_role_op(op): opt_ops.append(op) - params_grads.append((self.origin_program.global_block().var( - op.input("Param")[0]), - self.origin_program.global_block().var( - op.input("Grad")[0]))) + # HACK(wuyi): if we find grad vars from input of optimize + # ops, we may get the output of clip op. Use syntax "@GRAD" + # and op_role_var to get the pair. + for input_name in op.input_arg_names: + if input_name.find("@GRAD") != -1 and \ + op.attrs[RPC_OP_ROLE_ATTR_NAME]: + param_name = op.attrs[OP_ROLE_VAR_ATTR_NAME][0] + params_grads.append([ + origin_var_dict[param_name], + origin_var_dict[input_name] + ]) elif self._is_adam_connected_op(op): opt_ops.append(op) else: diff --git a/python/setup.py.in b/python/setup.py.in index c42601d335f01491156dc3591341c1a3213aecfe..8257f1d5e212a84188a4c51bc2d0f4d4c7af91fb 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -69,7 +69,8 @@ packages=['paddle', 'paddle.fluid.proto', 'paddle.fluid.proto.profiler', 'paddle.fluid.layers', - 'paddle.fluid.transpiler'] + 'paddle.fluid.transpiler', + 'paddle.fluid.transpiler.details'] if '${WITH_FLUID_ONLY}'== 'OFF': packages+=['paddle.proto',