diff --git a/doc/v2/howto/capi/compile_paddle_lib_en.md b/doc/v2/howto/capi/compile_paddle_lib_en.md
index 11d69b9b79..6212a30811 100644
--- a/doc/v2/howto/capi/compile_paddle_lib_en.md
+++ b/doc/v2/howto/capi/compile_paddle_lib_en.md
@@ -1,3 +1,175 @@
## Install and Build
-TBD
+### Download & Install
+
+ Download the latest C-API development package from CI system and install. You can find the required version in the table below:
+
+
+### From source
+
+ Users can also compile the C-API library from PaddlePaddle source code by compiling with the following compilation options:
+
+
+
+
+| Options |
+Value |
+
+
+
+
+| WITH_C_API |
+ON |
+
+
+| WITH_PYTHON |
+OFF（recommended） |
+
+
+| WITH_SWIG_PY |
+OFF（recommended） |
+
+
+| WITH_GOLANG |
+OFF（recommended） |
+
+
+| WITH_GPU |
+ON/OFF |
+
+
+| WITH_MKL |
+ON/OFF |
+
+
+It is best to set up with recommended values to avoid linking with unnecessary libraries. Set other compilation options as you need.
+
+Pull the latest following code snippet from github, and configure compilation options(replace PADDLE_ROOT with the installation path of the PaddlePaddle C-API inference library):
+
+```shell
+PADDLE_ROOT=/path/of/capi
+git clone https://github.com/PaddlePaddle/Paddle.git
+cd Paddle
+mkdir build
+cd build
+cmake -DCMAKE_INSTALL_PREFIX=$PADDLE_ROOT \
+ -DCMAKE_BUILD_TYPE=Release \
+ -DWITH_C_API=ON \
+ -DWITH_SWIG_PY=OFF \
+ -DWITH_GOLANG=OFF \
+ -DWITH_PYTHON=OFF \
+ -DWITH_MKL=OFF \
+ -DWITH_GPU=OFF \
+ ..
+```
+
+After running the above code to generate Makefile , run: `make && make install`. After successful compilation, the dependencies required by C-API(includes: (1)PaddlePaddle inference library and header files; (2) Third-party libraries and header files) will be stored in the `PADDLE_ROOT` directory.
+
+If the compilation is successful, see the following directory structure under `PADDLE_ROOT`(includes PaddlePaddle header files and libraries, and third-party libraries and header files(determined by the link methods if necessary)):
+
+```text
+├── include
+│ └── paddle
+│ ├── arguments.h
+│ ├── capi.h
+│ ├── capi_private.h
+│ ├── config.h
+│ ├── error.h
+│ ├── gradient_machine.h
+│ ├── main.h
+│ ├── matrix.h
+│ ├── paddle_capi.map
+│ └── vector.h
+├── lib
+│ ├── libpaddle_capi_engine.a
+│ ├── libpaddle_capi_layers.a
+│ ├── libpaddle_capi_shared.so
+│ └── libpaddle_capi_whole.a
+└── third_party
+ ├── gflags
+ │ ├── include
+ │ │ └── gflags
+ │ │ ├── gflags_completions.h
+ │ │ ├── gflags_declare.h
+ │ │ ...
+ │ └── lib
+ │ └── libgflags.a
+ ├── glog
+ │ ├── include
+ │ │ └── glog
+ │ │ ├── config.h
+ │ │ ...
+ │ └── lib
+ │ └── libglog.a
+ ├── openblas
+ │ ├── include
+ │ │ ├── cblas.h
+ │ │ ...
+ │ └── lib
+ │ ...
+ ├── protobuf
+ │ ├── include
+ │ │ └── google
+ │ │ └── protobuf
+ │ │ ...
+ │ └── lib
+ │ └── libprotobuf-lite.a
+ └── zlib
+ ├── include
+ │ ...
+ └── lib
+ ...
+
+```
+
+### Linking Description:
+
+There are three kinds of linking methods:
+
+1. Linking with dynamic library `libpaddle_capi_shared.so`（This way is much more convenient and easier, **Without special requirements, it is recommended**）, refer to the following：
+ 1. Compiling with CPU version and using `OpenBLAS`; only need to link one library named `libpaddle_capi_shared.so` to develop prediction program through C-API.
+ 1. Compiling with CPU version and using `MKL` lib, you need to link MKL library directly to develop prediction program through PaddlePaddle C-API, due to `MKL` has its own dynamic library.
+ 1. Compiling with GPU version, CUDA library will be loaded dynamically on prediction program run-time, and also set CUDA library to `LD_LIBRARY_PATH` environment variable.
+
+2. Linking with static library `libpaddle_capi_whole.a`，refer to the following：
+ 1. Specify `-Wl,--whole-archive` linking options.
+ 1. Explicitly link third-party libraries such as `gflags`、`glog`、`libz`、`protobuf` .etc, you can find them under `PADDLE_ROOT/third_party` directory.
+ 1. Use OpenBLAS library if compiling C-API，must explicitly link `libopenblas.a`.
+ 1. Use MKL when compiling C-API, must explicitly link MKL dynamic library.
+
+3. Linking with static library `libpaddle_capi_layers.a` and `libpaddle_capi_engine.a`，refer to the following：
+ 1. This linking methods is mainly used for mobile prediction.
+ 1. Split `libpaddle_capi_whole.a` into two static linking library at least to reduce the size of linking libraries.
+ 1. Specify `-Wl,--whole-archive -lpaddle_capi_layers` and `-Wl,--no-whole-archive -lpaddle_capi_engine` for linking.
+ 1. The third-party dependencies need explicitly link same as method 2 above.
diff --git a/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md b/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md
index bc3d50b3ff..dee1b7554f 100644
--- a/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md
+++ b/doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md
@@ -1,3 +1,372 @@
-# Kubernetes Distributed
+# Distributed Training on Kubernetes
-TBD
+We introduced how to create a PaddlePaddle Job with a single node on Kuberentes in the
+previous document.
+In this article, we will introduce how to create a PaddlePaddle job with multiple nodes
+on Kubernetes cluster.
+
+## Overall Architecture
+
+Before creating a training job, the users need to slice the training data and deploy
+the Python scripts along with it into the distributed file system
+(We can use the different type of Kuberentes Volumes to mount different distributed
+file systems). Before training starts, The program will copy the training data into the
+Container and also save the models at the same path during training. The global architecture
+is as follows:
+
+
+
+The above figure describes a distributed training architecture which contains 3 nodes, each
+Pod mounts a folder of the distributed file system to save training data and models
+by Kubernetes Volume. Kubernetes created 3 Pods for this training phase and scheduled these on
+3 nodes, each Pod has a PaddlePaddle container. After the containers car created,
+PaddlePaddle starts up the communication between PServer and Trainer and read training
+data for this training job.
+
+As the description above, we can start up a PaddlePaddle distributed training job on a
+Kubernetes ready cluster with the following steps:
+
+1. [Build PaddlePaddle Docker Image](#Build a Docker Image)
+1. [Split training data and upload to the distributed file system](#Upload Training Data)
+1. [Edit a YAML file and create a Kubernetes Job](#Create a Job)
+1. [Check the output](#Check The Output)
+
+We will introduce these steps as follows:
+
+### Build a Docker Image
+
+Training docker image needs to package the paddle pserver and paddle trainer runtimes, as well as two more processes before we can kick off the training:
+
+- Copying the training data into container.
+- Generating the initialization arguments for `Paddle PServer` and `Paddle Training` processes.
+
+Since the paddlepaddle official docker image already has the runtimes we need, we'll take it as the base image and pack some additional scripts for the processes mentioned above to build our training image. for more detail, please find from the following link:
+- https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/src/k8s_train/Dockerfile
+
+
+```bash
+$ cd doc/howto/usage/k8s/src/k8s_train
+$ docker build -t [YOUR_REPO]/paddle:mypaddle .
+```
+
+And then upload the new Docker Image to a Docker hub:
+
+```bash
+docker push [YOUR_REPO]/paddle:mypaddle
+```
+
+**[NOTE]**, in the above command arguments, `[YOUR_REPO]` represents your Docker repository,
+you need to use your repository instead of it. We will replace it with your respository name to
+represent the Docker Image which built in this step.
+
+### Prepare Training Data
+
+We can download and split the training job by creating a Kubernetes Job, or custom your image
+by editing [k8s_train](./src/k8s_train/).
+
+Before creating a Job, we need to bind a [persistenVolumeClaim](https://kubernetes.io/docs/user-guide/persistent-volumes) by the different type of
+the different file system, the generated dataset would be saved on this volume.
+
+```yaml
+apiVersion: batch/v1
+kind: Job
+metadata:
+ name: paddle-data
+spec:
+ template:
+ metadata:
+ name: pi
+ spec:
+ hostNetwork: true
+ containers:
+ - name: paddle-data
+ image: paddlepaddle/paddle-tutorial:k8s_data
+ imagePullPolicy: Always
+ volumeMounts:
+ - mountPath: "/mnt"
+ name: nfs
+ env:
+ - name: OUT_DIR
+ value: /home/work/mfs/paddle-cluster-job
+ - name: SPLIT_COUNT
+ value: "3"
+ volumes:
+ - name: nfs
+ persistentVolumeClaim:
+ claimName: mfs
+ restartPolicy: Never
+```
+
+Create the Job with the following command:
+
+```bash
+> kubectl create -f xxx.yaml
+```
+
+If created successfully, you can see some information like this:
+
+```base
+[root@paddle-kubernetes-node0 nfsdir]$ tree -d
+.
+`-- paddle-cluster-job
+ |-- 0
+ | `-- data
+ |-- 1
+ | `-- data
+ |-- 2
+ | `-- data
+ |-- output
+ |-- quick_start
+```
+
+The `paddle-cluster-job` above is the job name for this training job; we need 3
+PaddlePaddle training nodes and save the split training data in `paddle-cluster-job` path,
+the folder `0`, `1` and `2` represents the `training_id` on each node, `quick_start` folder is used to store training data, `output` folder is used to store the models and logs.
+
+
+### Create a Job
+
+Kubernetes allow users to create objects with YAML files, and we can use a command-line tool
+to create it.
+
+The Job YAML file describes that which Docker Image would be used in this training job, how much nodes would be created, what's the startup arguments of `Paddle PServer/Trainer` process and what's the type of Volumes. You can find the details of the YAML filed in
+[Kubernetes Job API](http://kubernetes.io/docs/api-reference/batch/v1/definitions/#_v1_job).
+The following is an example for this training job:
+
+```yaml
+apiVersion: batch/v1
+kind: Job
+metadata:
+ name: paddle-cluster-job
+spec:
+ parallelism: 3
+ completions: 3
+ template:
+ metadata:
+ name: paddle-cluster-job
+ spec:
+ volumes:
+ - name: jobpath
+ hostPath:
+ path: /home/work/mfs
+ containers:
+ - name: trainer
+ image: [YOUR_REPO]/paddle:mypaddle
+ command: ["bin/bash", "-c", "/root/start.sh"]
+ env:
+ - name: JOB_NAME
+ value: paddle-cluster-job
+ - name: JOB_PATH
+ value: /home/jobpath
+ - name: JOB_NAMESPACE
+ value: default
+ - name: TRAIN_CONFIG_DIR
+ value: recommendation
+ - name: CONF_PADDLE_NIC
+ value: eth0
+ - name: CONF_PADDLE_PORT
+ value: "7164"
+ - name: CONF_PADDLE_PORTS_NUM
+ value: "2"
+ - name: CONF_PADDLE_PORTS_NUM_SPARSE
+ value: "2"
+ - name: CONF_PADDLE_GRADIENT_NUM
+ value: "3"
+ volumeMounts:
+ - name: jobpath
+ mountPath: /home/jobpath
+ restartPolicy: Never
+```
+
+In the above YAML file:
+- `metadata.name`, The job name.
+- `parallelism`, Whether the Kubernetes Job would create `parallelism` Pods at the same time.
+- `completions`, The Job would become the success status only when the number of successful Pod(the exit code is 0)
+ is equal to `completions`.
+- `volumeMounts`, the name field `jobpath` is a key, the `mountPath` field represents
+ the path in the container, and we can define the `jobpath` in `volumes` filed, use `hostPath`
+ to configure the host path we want to mount.
+- `env`, the environment variables in the Container, we pass some startup arguments by
+ this approach, some details are as following:
+ - JOB_PATH：the mount path in the container
+ - JOB_NAME：the job name
+ - TRAIN_CONFIG_DIR：the job path in the container, we can find the training data path by
+ combine with JOB_NAME.
+ - CONF_PADDLE_NIC: the argument `--nics` of `Paddle PServer` process, the network
+ device name.
+ - CONF_PADDLE_PORT: the argument `--port` of `Paddle PServer` process.
+ - CONF_PADDLE_PORTS_NUM: the argument `--ports_num` of `Paddle PServer`, the port number
+ for dense prameter update.
+ - CONF_PADDLE_PORTS_NUM_SPARSE：the argument `--ports_num_for_sparse` of `Paddle PServer`,
+ the port number for sparse parameter update.
+ - CONF_PADDLE_GRADIENT_NUM：the number of training node, the argument
+ `--num_gradient_servers` of `Paddle PServer` and `Paddle Trainer`.
+
+You can find some details information at [here]
+(http://www.paddlepaddle.org/docs/develop/documentation/zh/howto/usage/cmd_parameter/detail_introduction_cn.html)。
+
+We can use the command-line tool of Kubernetes to create a Job when we finish the YAML file:
+
+```bash
+kubectl create -f job.yaml
+```
+
+Upon successful creation, Kubernetes would create 3 Pods as PaddlePaddle training node,
+pull the Docker image and begin to train.
+
+
+### Checkout the Output
+
+At the process of training, we can check the logs and the output models which is stored in
+the `output` folder.
+
+**NOTE**, `node_0`, `node_1` and `node_2` represent the
+`trainer_id` of the PaddlePaddle training job rather than the node id of Kubernetes.
+
+```bash
+[root@paddle-kubernetes-node0 output]# tree -d
+.
+├── node_0
+│ ├── server.log
+│ └── train.log
+├── node_1
+│ ├── server.log
+│ └── train.log
+├── node_2
+......
+├── pass-00002
+│ ├── done
+│ ├── ___embedding_0__.w0
+│ ├── ___embedding_1__.w0
+......
+```
+
+We can checkout the status of each training Pod by viewing the logs:
+
+```bash
+[root@paddle-kubernetes-node0 node_0]# cat train.log
+I1116 09:10:17.123121 50 Util.cpp:155] commandline:
+ /usr/local/bin/../opt/paddle/bin/paddle_trainer
+ --nics=eth0 --port=7164
+ --ports_num=2 --comment=paddle_process_by_paddle
+ --pservers=192.168.129.66,192.168.223.143,192.168.129.71
+ --ports_num_for_sparse=2 --config=./trainer_config.py
+ --trainer_count=4 --num_passes=10 --use_gpu=0
+ --log_period=50 --dot_period=10 --saving_period=1
+ --local=0 --trainer_id=0
+ --save_dir=/home/jobpath/paddle-cluster-job/output
+I1116 09:10:17.123440 50 Util.cpp:130] Calling runInitFunctions
+I1116 09:10:17.123764 50 Util.cpp:143] Call runInitFunctions done.
+[WARNING 2016-11-16 09:10:17,227 default_decorators.py:40] please use keyword arguments in paddle config.
+[INFO 2016-11-16 09:10:17,239 networks.py:1282] The input order is [movie_id, title, genres, user_id, gender, age, occupation, rating]
+[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__square_error_cost_0__]
+I1116 09:10:17.392917 50 Trainer.cpp:170] trainer mode: Normal
+I1116 09:10:17.613910 50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process
+I1116 09:10:17.680917 50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process
+I1116 09:10:17.681543 50 GradientMachine.cpp:134] Initing parameters..
+I1116 09:10:18.012390 50 GradientMachine.cpp:141] Init parameters done.
+I1116 09:10:18.018641 50 ParameterClient2.cpp:122] pserver 0 192.168.129.66:7164
+I1116 09:10:18.018950 50 ParameterClient2.cpp:122] pserver 1 192.168.129.66:7165
+I1116 09:10:18.019069 50 ParameterClient2.cpp:122] pserver 2 192.168.223.143:7164
+I1116 09:10:18.019492 50 ParameterClient2.cpp:122] pserver 3 192.168.223.143:7165
+I1116 09:10:18.019716 50 ParameterClient2.cpp:122] pserver 4 192.168.129.71:7164
+I1116 09:10:18.019836 50 ParameterClient2.cpp:122] pserver 5 192.168.129.71:7165
+```
+
+## Some Additional Details
+
+### Using Environment Variables
+
+Usually we use the environment varialbes to configurate the PaddlePaddle Job which runs in
+Kubernetes, `start_paddle.py` provides a start up script to convert the environment variable
+to the start up arguments of PaddlePaddle process:
+
+```bash
+API = "/api/v1/namespaces/"
+JOBSELECTOR = "labelSelector=job-name="
+JOB_PATH = os.getenv("JOB_PATH") + "/" + os.getenv("JOB_NAME")
+JOB_PATH_OUTPUT = JOB_PATH + "/output"
+JOBNAME = os.getenv("JOB_NAME")
+NAMESPACE = os.getenv("JOB_NAMESPACE")
+PADDLE_NIC = os.getenv("CONF_PADDLE_NIC")
+PADDLE_PORT = os.getenv("CONF_PADDLE_PORT")
+PADDLE_PORTS_NUM = os.getenv("CONF_PADDLE_PORTS_NUM")
+PADDLE_PORTS_NUM_SPARSE = os.getenv("CONF_PADDLE_PORTS_NUM_SPARSE")
+PADDLE_SERVER_NUM = os.getenv("CONF_PADDLE_GRADIENT_NUM")
+```
+
+### Communication between Pods
+
+At the begin of `start_paddle.py`, it would initializes and parses the arguments.
+
+```python
+parser = argparse.ArgumentParser(prog="start_paddle.py",
+ description='simple tool for k8s')
+ args, train_args_list = parser.parse_known_args()
+ train_args = refine_unknown_args(train_args_list)
+ train_args_dict = dict(zip(train_args[:-1:2], train_args[1::2]))
+ podlist = getPodList()
+```
+
+And then query the status of all the other Pods of this Job by the function `getPodList()`, and fetch `triner_id` by the function `getIdMap(podlist)` if all the Pods status is `RUNNING`.
+
+```python
+ podlist = getPodList()
+ # need to wait until all pods are running
+ while not isPodAllRunning(podlist):
+ time.sleep(10)
+ podlist = getPodList()
+ idMap = getIdMap(podlist)
+```
+
+**NOTE**: `getPodList()` would prefetch all the Pods in the current namespace, if some
+Pods are alreay running, it may cause some error. We will use [statfulesets](https://kubernetes.io/docs/concepts/abstractions/controllers/statefulsets) instead of
+Kubernetes Pod or Replicaset in the future.
+
+The function `getIdMap(podlist)` fetches IPs addresses of `podlist` and then sort them
+to generate `trainer_id`.
+
+```python
+def getIdMap(podlist):
+ '''
+ generate tainer_id by ip
+ '''
+ ips = []
+ for pod in podlist["items"]:
+ ips.append(pod["status"]["podIP"])
+ ips.sort()
+ idMap = {}
+ for i in range(len(ips)):
+ idMap[ips[i]] = i
+ return idMap
+```
+
+After getting the `idMap`, we can generate the arguments of `Paddle PServer` and `Paddle Trainer`
+so that we can start up them by `startPaddle(idMap, train_args_dict)`.
+
+### Create Job
+
+The main goal of `startPaddle` is generating the arguments of `Paddle PServer` and
+`Paddle Trainer` processes. Take `Paddle Trainer` as an example, we parse the
+environment variable and then get `PADDLE_NIC`, `PADDLE_PORT`, `PADDLE_PORTS_NUM` and etc...,
+finally find `trainerId` from `idMap` according to its IP address.
+
+```python
+ program = 'paddle train'
+ args = " --nics=" + PADDLE_NIC
+ args += " --port=" + str(PADDLE_PORT)
+ args += " --ports_num=" + str(PADDLE_PORTS_NUM)
+ args += " --comment=" + "paddle_process_by_paddle"
+ ip_string = ""
+ for ip in idMap.keys():
+ ip_string += (ip + ",")
+ ip_string = ip_string.rstrip(",")
+ args += " --pservers=" + ip_string
+ args_ext = ""
+ for key, value in train_args_dict.items():
+ args_ext += (' --' + key + '=' + value)
+ localIP = socket.gethostbyname(socket.gethostname())
+ trainerId = idMap[localIP]
+ args += " " + args_ext + " --trainer_id=" + \
+ str(trainerId) + " --save_dir=" + JOB_PATH_OUTPUT
+```
diff --git a/paddle/fluid/framework/block_desc.cc b/paddle/fluid/framework/block_desc.cc
index fbe08349c3..b8847e4b90 100644
--- a/paddle/fluid/framework/block_desc.cc
+++ b/paddle/fluid/framework/block_desc.cc
@@ -13,11 +13,10 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/block_desc.h"
+#include
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
-#include
-
namespace paddle {
namespace framework {
@@ -147,52 +146,7 @@ void BlockDesc::RemoveOp(size_t s, size_t e) {
if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) {
return;
}
- auto get_vars = [](std::deque>::iterator &op,
- std::vector &v) {
- auto in_names = (*op)->InputArgumentNames();
- v.insert(v.end(), in_names.begin(), in_names.end());
- auto out_names = (*op)->OutputArgumentNames();
- v.insert(v.end(), out_names.begin(), out_names.end());
- std::sort(v.begin(), v.end());
- auto last = std::unique(v.begin(), v.end());
- v.erase(last, v.end());
- };
- need_update_ = true;
-
- for (size_t i = s; i < e; i++) {
- // since remove op one by one, every time remove the first op.
- auto op = ops_.begin() + s;
-
- // collect input and output variables from current delete op
- std::vector cur_vars;
- get_vars(op, cur_vars);
-
- // remove current op
- ops_.erase(ops_.begin() + s);
-
- // collect input and output variables from other ops
- std::vector other_vars;
- for (auto it = ops_.begin(); it != ops_.end(); it++) {
- get_vars(it, other_vars);
- }
-
- // variables should be deleted
- std::vector delete_vars;
- // delete_vars = cur_vars - cur_vars ^ other_input_vars
- std::set_difference(cur_vars.begin(), cur_vars.end(), other_vars.begin(),
- other_vars.end(),
- std::inserter(delete_vars, delete_vars.end()));
- // remove variables
- for (size_t i = 0; i < delete_vars.size(); i++) {
- auto name = delete_vars[i];
- auto it = vars_.find(name);
- PADDLE_ENFORCE(it != vars_.end(),
- "%s is not in variable list, it should not be deleted",
- name);
- vars_.erase(it);
- VLOG(3) << "deleting variable " << name;
- }
- }
+ ops_.erase(ops_.begin() + s, ops_.begin() + e);
}
std::vector BlockDesc::AllOps() const {
diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
index 99b3065d8d..f393105fe8 100644
--- a/paddle/fluid/framework/parallel_executor.cc
+++ b/paddle/fluid/framework/parallel_executor.cc
@@ -181,10 +181,10 @@ void ParallelExecutor::SplitTensorToPlaces(
member_->places_.size(), lod_tensors.size());
for (size_t j = 0; j < member_->places_.size(); ++j) {
// TODO(panxy0718): Do I need to delete this var?
- member_->local_scopes_[j]
- ->Var(it.first)
- ->GetMutable()
- ->ShareDataWith(lod_tensors[j]);
+ auto t =
+ member_->local_scopes_[j]->Var(it.first)->GetMutable();
+ t->ShareDataWith(lod_tensors[j]);
+ t->set_lod(lod_tensors[j].lod());
}
}
}
diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc
index 36049ee6a4..c9939e8602 100644
--- a/paddle/fluid/operators/batch_norm_op.cc
+++ b/paddle/fluid/operators/batch_norm_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/batch_norm_op.h"
+#include
#include "paddle/fluid/framework/data_layout.h"
namespace paddle {
diff --git a/paddle/fluid/operators/batch_norm_op.cu.cc b/paddle/fluid/operators/batch_norm_op.cu.cc
index 6ceacc3992..eecb58e11e 100644
--- a/paddle/fluid/operators/batch_norm_op.cu.cc
+++ b/paddle/fluid/operators/batch_norm_op.cu.cc
@@ -13,9 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/batch_norm_op.h"
-#include "paddle/fluid/framework/data_layout.h"
-
#include
+#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/cudnn_helper.h"
#include "paddle/fluid/platform/float16.h"
diff --git a/paddle/fluid/operators/batch_size_like.h b/paddle/fluid/operators/batch_size_like.h
index 0bdf27e620..dd51a11fbe 100644
--- a/paddle/fluid/operators/batch_size_like.h
+++ b/paddle/fluid/operators/batch_size_like.h
@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
-
+#include
+#include
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
diff --git a/paddle/fluid/operators/box_coder_op.h b/paddle/fluid/operators/box_coder_op.h
index 3c7cac1cd1..77fc6c2b62 100644
--- a/paddle/fluid/operators/box_coder_op.h
+++ b/paddle/fluid/operators/box_coder_op.h
@@ -10,6 +10,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
+#include
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
diff --git a/paddle/fluid/operators/compare_op.cc b/paddle/fluid/operators/compare_op.cc
index 9a139ab27e..3a6a357e81 100644
--- a/paddle/fluid/operators/compare_op.cc
+++ b/paddle/fluid/operators/compare_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/compare_op.h"
+#include
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
diff --git a/paddle/fluid/operators/concat_op.cc b/paddle/fluid/operators/concat_op.cc
index 0eedd8ee51..d65a7b3467 100644
--- a/paddle/fluid/operators/concat_op.cc
+++ b/paddle/fluid/operators/concat_op.cc
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/concat_op.h"
+#include
#include
namespace paddle {
diff --git a/paddle/fluid/operators/cond_op.h b/paddle/fluid/operators/cond_op.h
index a04fae2182..d3888923db 100644
--- a/paddle/fluid/operators/cond_op.h
+++ b/paddle/fluid/operators/cond_op.h
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
+#include
#include
#include "glog/logging.h"
#include "paddle/fluid/framework/ddim.h"
diff --git a/paddle/fluid/operators/conv_transpose_op.cc b/paddle/fluid/operators/conv_transpose_op.cc
index b2a3cfc89f..08f5939d42 100644
--- a/paddle/fluid/operators/conv_transpose_op.cc
+++ b/paddle/fluid/operators/conv_transpose_op.cc
@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/conv_transpose_op.h"
+#include
+#include
namespace paddle {
namespace operators {
diff --git a/paddle/fluid/operators/conv_transpose_op.h b/paddle/fluid/operators/conv_transpose_op.h
index d4e4b641ec..bfc0177c2a 100644
--- a/paddle/fluid/operators/conv_transpose_op.h
+++ b/paddle/fluid/operators/conv_transpose_op.h
@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
-
+#include
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/im2col.h"
diff --git a/paddle/fluid/operators/crf_decoding_op.h b/paddle/fluid/operators/crf_decoding_op.h
index 2b2a733fb9..3f5fab3b38 100644
--- a/paddle/fluid/operators/crf_decoding_op.h
+++ b/paddle/fluid/operators/crf_decoding_op.h
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
+#include
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
diff --git a/paddle/fluid/operators/crop_op.h b/paddle/fluid/operators/crop_op.h
index c5ac684978..f05c2e2328 100644
--- a/paddle/fluid/operators/crop_op.h
+++ b/paddle/fluid/operators/crop_op.h
@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
-
+#include
+#include
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/strided_memcpy.h"
diff --git a/paddle/fluid/operators/math/math_function.cu b/paddle/fluid/operators/math/math_function.cu
index 82e1294314..e53183603f 100644
--- a/paddle/fluid/operators/math/math_function.cu
+++ b/paddle/fluid/operators/math/math_function.cu
@@ -39,13 +39,14 @@ void gemm(
cublasOperation_t cuTransB =
(transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
- float h_alpha = static_cast(alpha);
- float h_beta = static_cast(beta);
-
// TODO(kexinzhao): add processing code for compute capability < 53 case
PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53,
"cublas fp16 gemm requires GPU compute capability >= 53");
+#if CUDA_VERSION >= 8000
+ float h_alpha = static_cast(alpha);
+ float h_beta = static_cast(beta);
+
cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT;
#if CUDA_VERSION >= 9000
if (context.GetComputeCapability() >= 70) {
@@ -56,7 +57,7 @@ void gemm(
PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(),
CUBLAS_DEFAULT_MATH));
}
-#endif
+#endif // CUDA_VERSION >= 9000
// cublasHgemm does true FP16 computation which is slow for non-Volta
// GPUs. So use cublasGemmEx instead which does pesudo FP16 computation:
@@ -66,6 +67,18 @@ void gemm(
context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, B,
CUDA_R_16F, ldb, A, CUDA_R_16F, lda, &h_beta, C, CUDA_R_16F, N,
CUDA_R_32F, algo));
+#else
+ // CUDA 7.5 does not support cublasGemmEx, hence we fall back to use hgemm
+ const half h_alpha = static_cast(alpha);
+ const half h_beta = static_cast(beta);
+ const half* h_A = reinterpret_cast(A);
+ const half* h_B = reinterpret_cast(B);
+ half* h_C = reinterpret_cast(C);
+
+ PADDLE_ENFORCE(platform::dynload::cublasHgemm(
+ context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb,
+ h_A, lda, &h_beta, h_C, N));
+#endif // CUDA_VERSION >= 8000
}
template <>
diff --git a/paddle/fluid/platform/dynload/cublas.cc b/paddle/fluid/platform/dynload/cublas.cc
index eb541579a1..361d3439b8 100644
--- a/paddle/fluid/platform/dynload/cublas.cc
+++ b/paddle/fluid/platform/dynload/cublas.cc
@@ -28,6 +28,10 @@ CUBLAS_BLAS_ROUTINE_EACH(DEFINE_WRAP);
CUBLAS_BLAS_ROUTINE_EACH_R2(DEFINE_WRAP);
#endif
+#ifdef CUBLAS_BLAS_ROUTINE_EACH_R3
+CUBLAS_BLAS_ROUTINE_EACH_R3(DEFINE_WRAP);
+#endif
+
} // namespace dynload
} // namespace platform
} // namespace paddle
diff --git a/paddle/fluid/platform/dynload/cublas.h b/paddle/fluid/platform/dynload/cublas.h
index a41018d350..1ab55d6b9b 100644
--- a/paddle/fluid/platform/dynload/cublas.h
+++ b/paddle/fluid/platform/dynload/cublas.h
@@ -71,7 +71,6 @@ extern void *cublas_dso_handle;
__macro(cublasDgemm_v2); \
__macro(cublasHgemm); \
__macro(cublasSgemmEx); \
- __macro(cublasGemmEx); \
__macro(cublasSgeam_v2); \
__macro(cublasDgeam_v2); \
__macro(cublasCreate_v2); \
@@ -83,11 +82,6 @@ extern void *cublas_dso_handle;
__macro(cublasDgemmBatched); \
__macro(cublasCgemmBatched); \
__macro(cublasZgemmBatched); \
- __macro(cublasSgemmStridedBatched); \
- __macro(cublasDgemmStridedBatched); \
- __macro(cublasCgemmStridedBatched); \
- __macro(cublasZgemmStridedBatched); \
- __macro(cublasHgemmStridedBatched); \
__macro(cublasSgetrfBatched); \
__macro(cublasSgetriBatched); \
__macro(cublasDgetrfBatched); \
@@ -95,10 +89,24 @@ extern void *cublas_dso_handle;
CUBLAS_BLAS_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
+// APIs available after CUDA 8.0
+#if CUDA_VERSION >= 8000
+#define CUBLAS_BLAS_ROUTINE_EACH_R2(__macro) \
+ __macro(cublasGemmEx); \
+ __macro(cublasSgemmStridedBatched); \
+ __macro(cublasDgemmStridedBatched); \
+ __macro(cublasCgemmStridedBatched); \
+ __macro(cublasZgemmStridedBatched); \
+ __macro(cublasHgemmStridedBatched);
+
+CUBLAS_BLAS_ROUTINE_EACH_R2(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
+#endif
+
// APIs available after CUDA 9.0
#if CUDA_VERSION >= 9000
-#define CUBLAS_BLAS_ROUTINE_EACH_R2(__macro) __macro(cublasSetMathMode);
-CUBLAS_BLAS_ROUTINE_EACH_R2(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
+#define CUBLAS_BLAS_ROUTINE_EACH_R3(__macro) __macro(cublasSetMathMode);
+
+CUBLAS_BLAS_ROUTINE_EACH_R3(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
#endif
#undef DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP
diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h
index 2999004320..3a2a423486 100644
--- a/paddle/fluid/platform/nccl_helper.h
+++ b/paddle/fluid/platform/nccl_helper.h
@@ -14,8 +14,9 @@
#pragma once
-#include
+#include // NOLINT
#include
+#include
#include "paddle/fluid/platform/dynload/nccl.h"
#include "paddle/fluid/platform/enforce.h"
@@ -29,6 +30,8 @@ inline ncclDataType_t ToNCCLDataType(std::type_index type) {
return ncclDouble;
} else if (type == typeid(int)) { // NOLINT
return ncclInt;
+ } else if (type == typeid(int64_t)) { // NOLINT
+ return ncclInt64;
} else {
PADDLE_THROW("Not supported");
}
@@ -66,23 +69,23 @@ struct NCCLContext {
return boost::get(ctx_->GetPlace()).device;
}
- static void InitNCCLContext(std::unordered_map &contexts,
+ static void InitNCCLContext(std::unordered_map *contexts,
const std::vector &places) {
std::vector comms;
std::vector devs;
- comms.resize(contexts.size());
- devs.reserve(contexts.size());
+ comms.resize(contexts->size());
+ devs.reserve(contexts->size());
for (auto &p : places) {
devs.push_back(boost::get(p).device);
}
PADDLE_ENFORCE(platform::dynload::ncclCommInitAll(
- &comms[0], static_cast(contexts.size()), &devs[0]));
+ &comms[0], static_cast(contexts->size()), &devs[0]));
int i = 0;
for (auto &dev_id : devs) {
- contexts.at(dev_id).comm_ = comms[i++];
+ contexts->at(dev_id).comm_ = comms[i++];
}
}
};
@@ -91,7 +94,7 @@ struct NCCLContextMap {
std::unordered_map contexts_;
std::vector order_;
- NCCLContextMap(const std::vector &places) {
+ explicit NCCLContextMap(const std::vector &places) {
order_.reserve(places.size());
for (auto &p : places) {
int dev_id = boost::get(p).device;
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index 33cf691817..793421a22f 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -818,6 +818,11 @@ class Block(object):
del self.vars[name]
self.sync_with_cpp()
+ def remove_var(self, name):
+ self.sync_with_cpp()
+ self.desc.remove_var(name)
+ del self.vars[name]
+
def create_parameter(self, *args, **kwargs):
global_block = self.program.global_block()
param = Parameter(global_block, *args, **kwargs)
@@ -838,6 +843,11 @@ class Block(object):
self.ops.insert(index, op)
return op
+ def remove_op(self, index):
+ self.sync_with_cpp()
+ self.desc.remove_op(index, index + 1)
+ del self.ops[index]
+
def delete_ops(self, ops):
# remove from cpp
# FIXME(typhoonzero): remove only the first occurrence.
@@ -846,6 +856,7 @@ class Block(object):
end = list(self.ops).index(ops[-1])
except Exception, e:
raise e
+
self.desc.remove_op(start, end + 1)
def slice_ops(self, start, end):
diff --git a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py
index f98a8bbc68..3f9059fb5b 100644
--- a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py
+++ b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py
@@ -201,24 +201,6 @@ class TestBlockDesc(unittest.TestCase):
op1.set_type("test")
op2.set_type("test")
- var0 = block.var("var0")
- var1 = block.var("var1")
- var2 = block.var("var2")
- var3 = block.var("var3")
- var4 = block.var("var4")
- var5 = block.var("var5")
-
- op0.set_input("X", ["var0"])
- op0.set_output("Y", ["var0"])
- op1.set_input("X", ["var1", "var2"])
- op1.set_output("Y", ["var3", "var4"])
- op2.set_input("X", ["var1"])
- op2.set_output("Y", ["var4", "var5"])
-
- program.sync_with_cpp()
-
- # remove op1, its input var2 and output var3 will be removed at the same time,
- # but its input var1 and output var4 will not be removed since they are used for op2.
block.remove_op(1, 2)
program.sync_with_cpp()
@@ -226,8 +208,6 @@ class TestBlockDesc(unittest.TestCase):
for idx in xrange(0, block.op_size()):
all_ops.append(block.op(idx))
self.assertEqual(all_ops, [op0, op2])
- all_vars = block.all_vars()
- self.assertEqual(set(all_vars), {var0, var1, var4, var5})
if __name__ == '__main__':