code optimized

Dockerfile

@@ -24,7 +24,7 @@ COPY ./paddle/scripts/docker/root/ /root/
 
 RUN apt-get update && \
     apt-get install -y --allow-downgrades \
-    git python-pip python-dev openssh-server bison \
+    git python-pip python-dev python-opencv openssh-server bison \
     libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
@@ -76,8 +76,7 @@ RUN easy_install -U pip && \
     pip install sphinx-rtd-theme==0.1.9 recommonmark
 
 RUN pip install pre-commit 'ipython==5.3.0' && \
-    pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
-    pip install opencv-python
+    pip install 'ipykernel==4.6.0' 'jupyter==1.0.0'
 
 #For docstring checker
 RUN pip install pylint pytest astroid isort

benchmark/.gitignore

@@ -7,3 +7,6 @@ paddle/rnn/imdb.pkl
 caffe/image/logs
 tensorflow/image/logs
 tensorflow/rnn/logs
+fluid/models/*.pyc
+fluid/logs
+fluid/nohup.out

benchmark/fluid/Dockerfile

+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 python-opencv
+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 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/

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:

benchmark/fluid/fluid_benchmark.py

@@ -40,10 +40,7 @@ def parse_args():
     parser.add_argument(
         '--batch_size', type=int, default=32, help='The minibatch size.')
     parser.add_argument(
-        '--learning_rate',
-        type=float,
-        default=0.001,
-        help='The minibatch size.')
+        '--learning_rate', type=float, default=0.001, help='The learning rate.')
     # TODO(wuyi): add "--use_fake_data" option back.
     parser.add_argument(
         '--skip_batch_num',
@@ -72,6 +69,11 @@ def parse_args():
         type=int,
         default=1,
         help='If gpus > 1, will use ParallelExecutor to run, else use Executor.')
+    parser.add_argument(
+        '--cpus',
+        type=int,
+        default=1,
+        help='If cpus > 1, will use ParallelDo to run, else use Executor.')
     parser.add_argument(
         '--data_set',
         type=str,
@@ -88,8 +90,8 @@ def parse_args():
         help='If set, use nvprof for CUDA.')
     parser.add_argument(
         '--no_test',
-        action='store_false',
-        help='If set, test the testset during training.')
+        action='store_true',
+        help='If set, do not test the testset during training.')
     parser.add_argument(
         '--memory_optimize',
         action='store_true',
@@ -231,13 +233,10 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
             train_losses.append(loss)
             print("Pass: %d, Iter: %d, Loss: %f\n" %
                   (pass_id, iters, np.mean(train_losses)))
-        train_elapsed = time.time() - start_time
-        examples_per_sec = num_samples / train_elapsed
-        print('\nTotal examples: %d, total time: %.5f, %.5f examples/sec\n' %
-              (num_samples, train_elapsed, examples_per_sec))
-        print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses)))
+        print_train_time(start_time, time.time(), num_samples)
+        print("Pass: %d, Loss: %f" % (pass_id, np.mean(train_losses))),
         # evaluation
-        if not args.no_test and batch_acc != None:
+        if not args.no_test and batch_acc:
             pass_test_acc = test(exe, infer_prog, test_reader, feeder,
                                  batch_acc)
             print(", Test Accuracy: %f" % pass_test_acc)
@@ -315,11 +314,8 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
             if batch_id % 1 == 0:
                 print("Pass %d, batch %d, loss %s" %
                       (pass_id, batch_id, np.array(loss)))
-        train_elapsed = time.time() - start_time
-        examples_per_sec = num_samples / train_elapsed
-        print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
-              (num_samples, train_elapsed, examples_per_sec))
-        if not args.no_test and batch_acc != None:
+        print_train_time(start_time, time.time(), num_samples)
+        if not args.no_test and batch_acc:
             test_acc = test(startup_exe, infer_prog, test_reader, feeder,
                             batch_acc)
             print("Pass: %d, Test Accuracy: %f\n" % (pass_id, test_acc))
@@ -329,12 +325,19 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
 def print_arguments(args):
     vars(args)['use_nvprof'] = (vars(args)['use_nvprof'] and
                                 vars(args)['device'] == 'GPU')
-    print('----------- resnet Configuration Arguments -----------')
+    print('----------- Configuration Arguments -----------')
     for arg, value in sorted(vars(args).iteritems()):
         print('%s: %s' % (arg, value))
     print('------------------------------------------------')
 
 
+def print_train_time(start_time, end_time, num_samples):
+    train_elapsed = end_time - start_time
+    examples_per_sec = num_samples / train_elapsed
+    print('\nTotal examples: %d, total time: %.5f, %.5f examples/sed\n' %
+          (num_samples, train_elapsed, examples_per_sec))
+
+
 def main():
     args = parse_args()
     print_arguments(args)
@@ -342,7 +345,7 @@ def main():
     # the unique trainer id, starting from 0, needed by trainer
     # only
     nccl_id_var, num_trainers, trainer_id = (
-        None, 1, int(os.getenv("PADDLE_TRAINER_ID", "-1")))
+        None, 1, int(os.getenv("PADDLE_TRAINER_ID", "0")))
 
     if args.use_cprof:
         pr = cProfile.Profile()

benchmark/fluid/models/mnist.py

@@ -69,15 +69,30 @@ def get_model(args):
     images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE)
     label = fluid.layers.data(name='label', shape=[1], dtype='int64')
 
-    # Train program
-    predict = cnn_model(images)
-    cost = fluid.layers.cross_entropy(input=predict, label=label)
-    avg_cost = fluid.layers.mean(x=cost)
-
-    # Evaluator
-    batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
-    batch_acc = fluid.layers.accuracy(
-        input=predict, label=label, total=batch_size_tensor)
+    if args.device == 'CPU' and args.cpus > 1:
+        places = fluid.layers.get_places(args.cpus)
+        pd = fluid.layers.ParallelDo(places)
+        with pd.do():
+            predict = cnn_model(pd.read_input(images))
+            label = pd.read_input(label)
+            cost = fluid.layers.cross_entropy(input=predict, label=label)
+            avg_cost = fluid.layers.mean(x=cost)
+            batch_acc = fluid.layers.accuracy(input=predict, label=label)
+
+            pd.write_output(avg_cost)
+            pd.write_output(batch_acc)
+
+        avg_cost, batch_acc = pd()
+        avg_cost = fluid.layers.mean(avg_cost)
+        batch_acc = fluid.layers.mean(batch_acc)
+    else:
+        # Train program
+        predict = cnn_model(images)
+        cost = fluid.layers.cross_entropy(input=predict, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+
+        # Evaluator
+        batch_acc = fluid.layers.accuracy(input=predict, label=label)
 
     # inference program
     inference_program = fluid.default_main_program().clone()

benchmark/fluid/models/resnet.py

@@ -132,18 +132,33 @@ def get_model(args):
 
     input = fluid.layers.data(name='data', shape=dshape, dtype='float32')
     label = fluid.layers.data(name='label', shape=[1], dtype='int64')
-    predict = model(input, class_dim)
-    cost = fluid.layers.cross_entropy(input=predict, label=label)
-    avg_cost = fluid.layers.mean(x=cost)
 
-    batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
-    batch_acc = fluid.layers.accuracy(
-        input=predict, label=label, total=batch_size_tensor)
+    if args.device == 'CPU' and args.cpus > 1:
+        places = fluid.layers.get_places(args.cpus)
+        pd = fluid.layers.ParallelDo(places)
+        with pd.do():
+            predict = model(pd.read_input(input), class_dim)
+            label = pd.read_input(label)
+            cost = fluid.layers.cross_entropy(input=predict, label=label)
+            avg_cost = fluid.layers.mean(x=cost)
+            batch_acc = fluid.layers.accuracy(input=predict, label=label)
+
+            pd.write_output(avg_cost)
+            pd.write_output(batch_acc)
+
+        avg_cost, batch_acc = pd()
+        avg_cost = fluid.layers.mean(avg_cost)
+        batch_acc = fluid.layers.mean(batch_acc)
+    else:
+        predict = model(input, class_dim)
+        cost = fluid.layers.cross_entropy(input=predict, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+        batch_acc = fluid.layers.accuracy(input=predict, label=label)
 
     inference_program = fluid.default_main_program().clone()
     with fluid.program_guard(inference_program):
         inference_program = fluid.io.get_inference_program(
-            target_vars=[batch_acc, batch_size_tensor])
+            target_vars=[batch_acc])
 
     optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
 

benchmark/fluid/models/stacked_dynamic_lstm.py

@@ -101,9 +101,8 @@ def get_model(args):
     loss = fluid.layers.mean(x=loss)
 
     # add acc
-    batch_size_tensor = fluid.layers.create_tensor(dtype='int64')
     batch_acc = fluid.layers.accuracy(input=logit, label=fluid.layers.data(name='label', \
-                shape=[1], dtype='int64'), total=batch_size_tensor)
+                shape=[1], dtype='int64'))
 
     inference_program = fluid.default_main_program().clone()
     with fluid.program_guard(inference_program):

benchmark/fluid/run.sh

@@ -2,6 +2,7 @@
 # This script benchmarking the PaddlePaddle Fluid on
 # single thread single GPU.
 
+mkdir -p logs
 #export FLAGS_fraction_of_gpu_memory_to_use=0.0
 export CUDNN_PATH=/paddle/cudnn_v5
 
@@ -35,71 +36,74 @@ nohup stdbuf -oL nvidia-smi \
       --format=csv \
       --filename=mem.log  \
       -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 \
                --iterations=500 \
-               2>&1 | tee -a mnist_gpu_128.log
+               2>&1 | tee -a logs/mnist_gpu_128.log
 
 # 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 \
                --iterations=30 \
-               2>&1 | tee -a vgg16_gpu_128.log
+               2>&1 | tee -a logs/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 \
                --skip_batch_num=5 \
                --iterations=30 \
-               2>&1 | tee -a vgg16_gpu_flowers_32.log
+               2>&1 | tee -a logs/vgg16_gpu_flowers_32.log
 
 # resnet50
 # resnet50 gpu cifar10 128
-FLAGS_benchmark=true stdbuf -oL python fluid/resnet50.py \
+FLAGS_benchmark=true stdbuf -oL python fluid_benchmark.py \
+               --model=resnet \
                --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
+               2>&1 | tee -a logs/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=resnet \
                --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
+               2>&1 | tee -a logs/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
+               2>&1 | tee -a logs/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 \
                --iterations=30 \
-               2>&1 | tee -a lstm_gpu_128.log
+               2>&1 | tee -a logs/lstm_gpu_128.log

cmake/external/grpc.cmake

@@ -33,10 +33,19 @@ ELSE()
   SET(BUILD_CMD make HAS_SYSTEM_PROTOBUF=false -s -j ${NUM_OF_PROCESSOR} static grpc_cpp_plugin)
 ENDIF()
 
+# FIXME(wuyi): do not build zlib cares protobuf twice, find a way to build grpc with them
 ExternalProject_Add(
     extern_grpc
     DEPENDS protobuf zlib
-    URL "http://paddlepaddledeps.bj.bcebos.com/grpc.tar.xz"
+    # NOTE(wuyi):
+    # this package is generated by following steps:
+    # 1. git clone -b v1.8.x https://github.com/grpc/grpc.git
+    # 2. submodule update --init
+    # 3. keep only zlib, cares, protobuf, boringssl under "third_party",
+    #    checkout and clean other dirs under third_party
+    # 4. remove .git, and package the directory.
+    URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.8.x.tar.gz"
+    URL_MD5  "c9c58ee7d0e8929a63155af6a2ecdbd0"
     PREFIX          ${GRPC_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CONFIGURE_COMMAND ""
@@ -49,7 +58,6 @@ ExternalProject_Add(
     INSTALL_COMMAND make prefix=${GRPC_INSTALL_DIR} install
 )
 
-# FIXME(typhoonzero): hack to get static lib path, try a better way like merge them.
 ADD_LIBRARY(grpc++_unsecure STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET grpc++_unsecure PROPERTY IMPORTED_LOCATION
              "${GRPC_INSTALL_DIR}/lib/libgrpc++_unsecure.a")

doc/fluid/api/io.rst

@@ -59,3 +59,21 @@ get_inference_program
 ..  autofunction:: paddle.fluid.io.get_inference_program
     :noindex:
 
+save_checkpoint
+---------------
+
+..  autofunction:: paddle.fluid.io.save_checkpoint
+    :noindex:
+
+load_checkpoint
+---------------
+
+..  autofunction:: paddle.fluid.io.load_checkpoint
+    :noindex:
+
+clean_checkpoint
+----------------
+
+..  autofunction:: paddle.fluid.io.clean_checkpoint
+    :noindex:
+

doc/fluid/api/layers.rst

@@ -181,6 +181,12 @@ Print
 ..  autofunction:: paddle.fluid.layers.Print
     :noindex:
 
+is_empty
+--------
+
+..  autofunction:: paddle.fluid.layers.is_empty
+    :noindex:
+
 device
 ======
 
@@ -255,6 +261,19 @@ double_buffer
 ..  autofunction:: paddle.fluid.layers.double_buffer
     :noindex:
 
+random_data_generator
+---------------------
+
+..  autofunction:: paddle.fluid.layers.random_data_generator
+    :noindex:
+
+Preprocessor
+------------
+
+..  autoclass:: paddle.fluid.layers.Preprocessor
+    :members:
+    :noindex:
+
 nn
 ==
 
@@ -594,6 +613,29 @@ roi_pool
 ..  autofunction:: paddle.fluid.layers.roi_pool
     :noindex:
 
+dice_loss
+---------
+
+..  autofunction:: paddle.fluid.layers.dice_loss
+    :noindex:
+
+resize_bilinear
+---------------
+
+..  autofunction:: paddle.fluid.layers.resize_bilinear
+    :noindex:
+
+gather
+------
+
+..  autofunction:: paddle.fluid.layers.gather
+    :noindex:
+
+random_crop
+-----------
+
+..  autofunction:: paddle.fluid.layers.random_crop
+    :noindex:
 
 ops
 ===
@@ -742,6 +784,12 @@ sum
 ..  autofunction:: paddle.fluid.layers.sum
     :noindex:
 
+shape
+-----
+
+..  autofunction:: paddle.fluid.layers.shape
+    :noindex:
+
 sigmoid
 -------
 
@@ -991,21 +1039,3 @@ zeros
 ..  autofunction:: paddle.fluid.layers.zeros
     :noindex:
 
-topk
-----
-
-..  autofunction:: paddle.fluid.layers.topk
-    :noindex:
-
-dice_loss
-----
-
-..  autofunction:: paddle.fluid.layers.dice_loss
-    :noindex:
-
-upsampling_bilinear2d
-____
-
-..  autofunction:: paddle.fluid.layers.upsampling_bilinear2d
-    :noindex:
-

doc/fluid/api/optimizer.rst

@@ -47,28 +47,6 @@ DecayedAdagrad
     :members:
     :noindex:
 
-Adadelta
------------------
-
-..  autoclass:: paddle.fluid.optimizer.Adadelta
-    :members:
-    :noindex:
-
-RMSProp
------------------
-
-..  autoclass:: paddle.fluid.optimizer.RMSProp
-    :members:
-    :noindex:
-
-ModelAverage
------------------
-
-..  autoclass:: paddle.fluid.optimizer.ModelAverage
-    :members:
-    :noindex:
-
-
 SGDOptimizer
 ------------
 
@@ -111,25 +89,31 @@ DecayedAdagradOptimizer
     :members:
     :noindex:
 
+RMSPropOptimizer
+----------------
 
-AdadeltaOptimizer
------------------
-
-..  autoclass:: paddle.fluid.optimizer.AdadeltaOptimizer
+..  autoclass:: paddle.fluid.optimizer.RMSPropOptimizer
     :members:
     :noindex:
 
+Adadelta
+--------
 
-RMSPropOptimizer
------------------
+..  autoclass:: paddle.fluid.optimizer.Adadelta
+    :members:
+    :noindex:
 
-..  autoclass:: paddle.fluid.optimizer.RMSPropOptimizer
+ModelAverage
+------------
+
+..  autoclass:: paddle.fluid.optimizer.ModelAverage
     :members:
     :noindex:
-    
+
 Optimizer
 ---------
 
 ..  autoclass:: paddle.fluid.optimizer.Optimizer
     :members:
     :noindex:
+

doc/fluid/api/profiler.rst

@@ -23,3 +23,15 @@ profiler
 ..  autofunction:: paddle.fluid.profiler.profiler
     :noindex:
 
+start_profiler
+--------------
+
+..  autofunction:: paddle.fluid.profiler.start_profiler
+    :noindex:
+
+stop_profiler
+-------------
+
+..  autofunction:: paddle.fluid.profiler.stop_profiler
+    :noindex:
+

doc/fluid/design/concepts/var_desc.md

@@ -35,7 +35,7 @@ The computation `Program` consists of nested `Blocks`. Each `Block` will consist
 
 ## Definition of VarType
 
-A VarDesc should have a name, type and whether or not it is persistable. The are different kinds of variable types supported in PaddlePaddle, apart from the POD_Types like: `LOD_TENSOR`, `SELECTED_ROWS`, `FEED_MINIBATCH`, `FETCH_LIST`, `STEP_SCOPES`, `LOD_RANK_TABLE`, `LOD_TENSOR_ARRAY`, `PLACE_LIST`, `READER` and `CHANNEL`. These are declared inside `VarType`. A `VarDesc` then looks as the following:
+A VarDesc should have a name, type and whether or not it is persistable. There are different kinds of variable types supported in PaddlePaddle, apart from the POD_Types like: `LOD_TENSOR`, `SELECTED_ROWS`, `FEED_MINIBATCH`, `FETCH_LIST`, `STEP_SCOPES`, `LOD_RANK_TABLE`, `LOD_TENSOR_ARRAY`, `PLACE_LIST`, `READER` and `CHANNEL`. These are declared inside `VarType`. A `VarDesc` then looks as the following:
 
 ```proto
 message VarDesc {

doc/fluid/howto/cluster/fluid_recordio.md

+# How to use RecordIO in Fluid
+
+If you want to use RecordIO as your training data format, you need to convert to your training data
+to RecordIO files and reading them in the process of training, PaddlePaddle Fluid provides some
+interface to deal with the RecordIO files.
+
+## Generate RecordIO File
+
+Before start training with RecordIO files, you need to convert your training data
+to RecordIO format by `fluid.recordio_writer.convert_reader_to_recordio_file`, the sample codes
+as follows:
+
+```python
+    reader = paddle.batch(mnist.train(), batch_size=1)
+    feeder = fluid.DataFeeder(
+        feed_list=[  # order is image and label
+            fluid.layers.data(
+            name='image', shape=[784]),
+            fluid.layers.data(
+            name='label', shape=[1], dtype='int64'),
+        ],
+        place=fluid.CPUPlace())
+    fluid.recordio_writer.convert_reader_to_recordio_file('./mnist.recordio', reader, feeder)
+```
+
+The above code snippet would generate a RecordIO `./mnist.recordio` on your host.
+
+**NOTE**: we recommend users to set `batch_size=1` when generating the recordio files so that users can
+adjust it flexibly while reading it.
+
+## Use the RecordIO file in a Local Training Job
+
+PaddlePaddle Fluid provides an interface `fluid.layers.io.open_recordio_file` to load your RecordIO file
+and then you can use them as a Layer in your network configuration, the sample codes as follows:
+
+```python
+    data_file = fluid.layers.io.open_recordio_file(
+        filename="./mnist.recordio",
+        shapes=[(-1, 784),(-1, 1)],
+        lod_levels=[0, 0],
+        dtypes=["float32", "int32"])
+    data_file = fluid.layers.io.batch(data_file, batch_size=4)
+
+    img, label = fluid.layers.io.read_file(data_file)
+    hidden = fluid.layers.fc(input=img, size=100, act='tanh')
+    prediction = fluid.layers.fc(input=hidden, size=10, act='softmax')
+    loss = fluid.layers.cross_entropy(input=prediction, label=label)
+    avg_loss = fluid.layers.mean(loss)
+
+    fluid.optimizer.Adam(learning_rate=1e-3).minimize(avg_loss)
+
+    place = fluid.CPUPlace()
+
+    exe = fluid.Executor(place)
+    exe.run(fluid.default_startup_program())
+    avg_loss_np = []
+
+    # train a pass
+    batch_id = 0
+    while True:
+        tmp, = exe.run(fetch_list=[avg_loss])
+
+        avg_loss_np.append(tmp)
+        print(batch_id)
+        batch_id += 1
+```
+
+## Use the RecordIO files in Distributed Training
+
+1. generate multiple RecordIO files
+
+For a distributed training job, you may have multiple trainer nodes,
+and one or more RecordIO files for one trainer node, you can use the interface
+`fluid.recordio_writer.convert_reader_to_recordio_files` to convert your training data
+into multiple RecordIO files, the sample codes as follows:
+
+```python
+    reader = paddle.batch(mnist.train(), batch_size=1)
+    feeder = fluid.DataFeeder(
+        feed_list=[  # order is image and label
+            fluid.layers.data(
+            name='image', shape=[784]),
+            fluid.layers.data(
+            name='label', shape=[1], dtype='int64'),
+        ],
+        place=fluid.CPUPlace())
+    fluid.recordio_writer.convert_reader_to_recordio_files(
+          filename_suffix='./mnist.recordio', batch_per_file=100, reader, feeder)
+```
+
+The above codes would generate multiple RecordIO files on your host like:
+
+```bash
+.
+ \_mnist-00000.recordio
+ |-mnist-00001.recordio
+ |-mnist-00002.recordio
+ |-mnist-00003.recordio
+ |-mnist-00004.recordio
+```
+
+2. open multiple RecordIO files by `fluid.layers.io.open_files`
+
+For a distributed training job, the distributed operator system will schedule trainer process on multiple nodes,
+each trainer process reads parts of the whole training data, we usually take the following approach to make the training
+data allocated by each trainer process as uniform as possiable:
+
+```python
+def gen_train_list(file_pattern, trainers, trainer_id):
+   file_list = glob.glob(file_pattern)
+   ret_list = []
+   for idx, f in enumerate(file_list):
+       if (idx + trainers) % trainers == trainer_id:
+           ret_list.append(f)
+   return ret_list
+
+trainers = int(os.getenv("TRAINERS"))
+trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+data_file = fluid.layers.io.open_files(
+    filenames=gen_train_list("./mnist-[0-9]*.recordio", 2, 0),
+    thread_num=1,
+    shapes=[(-1, 784),(-1, 1)],
+    lod_levels=[0, 0],
+    dtypes=["float32", "int32"])
+img, label = fluid.layers.io.read_file(data_files)
+...
+```

doc/fluid/howto/optimization/benchmark/README.md

-../../../../../benchmark/cluster/README.md
\ No newline at end of file

doc/fluid/howto/optimization/benchmark/vgg16/README.md

-../../../../../../benchmark/cluster/vgg16/README.md
\ No newline at end of file

doc/fluid/howto/optimization/host_memory_profiling_cn.md

+## 堆内存分析和优化
+
+计算机程序都可能有内存泄漏的风险。**内存泄漏**一般是由于程序在堆(heap)上分配了内存而没有释放，随着程序的运行占用的内存越来越大，一方面会影响程序的稳定性，可能让运行速度越来越慢，或者造成oom，甚至会影响运行程序的机器的稳定性，造成宕机。
+
+
+目前有很多内存泄漏分析工具，比较经典的有[valgrind](http://valgrind.org/docs/manual/quick-start.html#quick-start.intro), [gperftools](https://gperftools.github.io/gperftools/)。
+
+因为Fluid是用Python驱动C++ core来运行，valgrind直接分析非常困难，需要自己编译debug版本的、带valgrind支持的专用Python版本，而且输出的信息中大部分是Python自己的符号和调用信息，分析起来很困难，另外使用valgrind会让程序运行速度变得非常慢，所以不建议使用。
+
+本教程主要介绍[gperftools](https://gperftools.github.io/gperftools/)的使用。
+
+gperftool主要支持以下四个功能：
+
+- thread-caching malloc
+- heap-checking using tcmalloc
+- heap-profiling using tcmalloc
+- CPU profiler
+
+Paddle也提供了基于gperftool的[CPU性能分析教程](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/optimization/cpu_profiling_cn.md)。
+
+对于堆内存的分析，主要用到thread-caching malloc和heap-profiling using tcmalloc。
+
+## 使用流程
+#### 环境
+本教程基于paddle提供的Docker开发环境paddlepaddle/paddle:latest-dev，基于Ubuntu 16.04.4 LTS环境。
+
+#### 使用流程
+
+- 安装google-perftools
+
+```
+apt-get install libunwind-dev 
+apt-get install google-perftools
+```
+
+- 安装pprof
+
+```
+go get -u github.com/google/pprof
+```
+
+- 设置运行环境
+
+```
+export PPROF_PATH=/root/gopath/bin/pprof
+export PPROF_BINARY_PATH=/root/gopath/bin/pprof
+export LD_PRELOAD=/usr/lib/libtcmalloc.so.4
+```
+
+- 使用heap profile来运行python程序。本质上是周期性的对堆的分配情况做一次快照。
+
+```
+# HEAPPROFILE 设置生成的堆分析文件的目录和文件前缀
+# HEAP_PROFILE_ALLOCATION_INTERVAL 设置每分配多少存储dump一次dump，默认1GB
+env HEAPPROFILE="./perf_log/test.log" HEAP_PROFILE_ALLOCATION_INTERVAL=209715200 python trainer.py
+```
+
+随着程序的运行，会在perf_log这个文件夹下生成很多文件，如下：
+
+```
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0001.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0002.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0003.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0004.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0005.heap
+-rw-r--r-- 1 root root 1.0M Jun  1 15:00 test.log.0006.heap
+```
+
+- 使用pprof对heap文件进行分析。分析有两种模式：
+	- 完整模式。会对当前heap做一个分析，显示目前分配内存一些调用路径。
+
+	```
+	pprof --pdf python test.log.0012.heap
+	```
+	上述命令会生成一个profile00x.pdf的文件，可以直接打开，例如：[memory_cpu_allocator](https://github.com/jacquesqiao/Paddle/blob/bd2ea0e1f84bb6522a66d44a072598153634cade/doc/fluid/howto/optimization/memory_cpu_allocator.pdf)。从下图可以看出，在CPU版本fluid的运行过程中，分配存储最多的模块式CPUAllocator. 而别的模块相对而言分配内存较少，所以被忽略了，这对于分配内存泄漏是很不方便的，因为泄漏是一个缓慢的过程，在这种图中是无法看到的。
+	
+	![result](https://user-images.githubusercontent.com/3048612/40964027-a54033e4-68dc-11e8-836a-144910c4bb8c.png)
+	
+	- Diff模式。可以对两个时刻的heap做diff，把一些内存分配没有发生变化的模块去掉，而把增量部分显示出来。
+	```
+	pprof --pdf --base test.log.0010.heap python test.log.1045.heap
+	```
+	生成的结果为：[`memory_leak_protobuf`](https://github.com/jacquesqiao/Paddle/blob/bd2ea0e1f84bb6522a66d44a072598153634cade/doc/fluid/howto/optimization/memory_leak_protobuf.pdf)
+	
+	从图中可以看出：ProgramDesc这个结构，在两个版本之间增长了200MB+，所以这里有很大的内存泄漏的可能性，最终结果也确实证明是这里造成了泄漏。
+	
+	![result](https://user-images.githubusercontent.com/3048612/40964057-b434d5e4-68dc-11e8-894b-8ab62bcf26c2.png)
+	![result](https://user-images.githubusercontent.com/3048612/40964063-b7dbee44-68dc-11e8-9719-da279f86477f.png)
+	

doc/mobile/cross_compiling_for_android_cn.md

@@ -63,16 +63,16 @@ Android的Docker开发镜像向用户提供两个可配置的参数：
 - 编译`armeabi-v7a`，`Android API 21`的PaddlePaddle库
 
 ```bash
-$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev
+$ docker run -it --rm -v $PWD:/paddle -w /paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev ./paddle/scripts/paddle_build.sh build_android
 ```
 
 - 编译`arm64-v8a`，`Android API 21`的PaddlePaddle库
 
 ```bash
-$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev
+$ docker run -it --rm -v $PWD:/paddle -w /paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev ./paddle/scripts/paddle_build.sh build_android
 ```
 
-执行上述`docker run`命令时，容器默认执行[paddle/scripts/docker/build_android.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文[配置交叉编译参数](#配置交叉编译参数)章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。
+执行上述`docker run`命令时，容器执行[paddle/scripts/paddle_build.sh build_android](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/paddle_build.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文[配置交叉编译参数](#配置交叉编译参数)章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。
 
 ## 基于Linux交叉编译环境的编译方式
 本文档将以Linux x86-64平台为例，介绍交叉编译Android平台上适用的PaddlePaddle库的方法和步骤。

doc/mobile/cross_compiling_for_android_en.md

@@ -36,7 +36,7 @@ $ docker pull docker.paddlepaddlehub.com/paddle:latest-dev-android
 We can run the Docker image we just created to build the inference library of PaddlePaddle for Android using the command below:
 
 ```bash
-$ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" paddle:dev-android
+$ docker run -it --rm -v $PWD:/paddle -w /paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" paddle:dev-android ./paddle/scripts/paddle_build.sh build_android
 ```
 
 The Docker image accepts two arguments `ANDROID_ABI` and `ANDROID_API`:
@@ -70,7 +70,7 @@ The Docker image accepts two arguments `ANDROID_ABI` and `ANDROID_API`:
 
 The ARM-64 architecture (`arm64-v8a`) requires at least level 21 of Android API.
 
-The default entry-point of the Docker image, [`paddle/scripts/docker/build_android.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh) generates the [Android cross-compiling standalone toolchain](https://developer.android.com/ndk/guides/standalone_toolchain.html) based on the argument: `ANDROID_ABI` or `ANDROID_API`.  For information about other configuration arguments, please continue reading.
+The build command, [`paddle/scripts/paddle_build.sh build_android`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/paddle_build.sh) generates the [Android cross-compiling standalone toolchain](https://developer.android.com/ndk/guides/standalone_toolchain.html) based on the argument: `ANDROID_ABI` or `ANDROID_API`.  For information about other configuration arguments, please continue reading.
 
 The above command generates and outputs the inference library in `$PWD/install_android` and puts third-party libraries in `$PWD/install_android/third_party`.
 

doc/v2/build_and_install/build_from_source_cn.rst

@@ -23,7 +23,7 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安
 在 `这里 <https://github.com/PaddlePaddle/Paddle/tree/develop/tools/manylinux1/>`__ 找到 paddle_manylinux_devel
 镜像的编译以及使用方法。或者参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。
 
-如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。
+如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 :ref:`编译依赖 <_compile_deps>` 之后才能开始编译的步骤。
 
 编译PaddlePaddle，需要执行：
 
@@ -106,7 +106,7 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安
 
 - 学习 Docker 有多难？
 
-  理解 Docker 并不难，大概花十分钟看一下[这篇文章](https://zhuanlan.zhihu.com/p/19902938)。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。
+  理解 Docker 并不难，大概花十分钟看一下 `这篇文章 <https://zhuanlan.zhihu.com/p/19902938>`_ 。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。
 
 - 我可以用 IDE 吗？
 
@@ -123,7 +123,7 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安
 
 - 可以并行编译吗？
 
-  是的。我们的 Docker image 运行一个 [Bash 脚本](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。
+  是的。我们的 Docker image 运行一个 `Bash脚本 <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh>`_ 。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。
 
 - Docker 需要 sudo
 
@@ -131,11 +131,11 @@ PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安
 
 - 在 Windows/MacOS 上编译很慢
 
-  Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考[这个issue](https://github.com/PaddlePaddle/Paddle/issues/627)。
+  Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考 `这个issue <https://github.com/PaddlePaddle/Paddle/issues/627>`_ 。
 
 - 磁盘不够
 
-  本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考[这篇文章](https://zaiste.net/posts/removing_docker_containers/)来清理这些内容。
+  本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考 `这篇文章 <https://zaiste.net/posts/removing_docker_containers/>`_ 来清理这些内容。
 
 
 .. _compile_deps:
@@ -211,7 +211,7 @@ PaddlePaddle可以使用cuDNN v5.1之后的任何一个版本来编译运行，
 编译选项的设置
 ++++++++++++++
 
-PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如 
+PaddePaddle通过编译时指定路径来实现引用各种BLAS/CUDA/cuDNN库。cmake编译时，首先在系统路径（ :code:`/usr/lib:/usr/local/lib` ）中搜索这几个库，同时也会读取相关路径变量来进行搜索。 通过使用 ``-D`` 命令可以设置，例如
 
 ..  code-block:: bash
 

doc/v2/build_and_install/build_from_source_en.rst

@@ -11,7 +11,7 @@ To build PaddlePaddle, you need
 1. A computer -- Linux, Windows, MacOS.
 2. Docker.
 
-Nothing else.  Not even Python and GCC, because you can install all build tools into a Docker image. 
+Nothing else.  Not even Python and GCC, because you can install all build tools into a Docker image.
 We run all the tools by running this image.
 
 .. _build_step:
@@ -26,6 +26,8 @@ you can also find how to build and use paddle_manylinux_devel Docker image from
 `here <https://github.com/PaddlePaddle/Paddle/tree/develop/tools/manylinux1/>`__
 Or you can build your own image from source as the optional step below:
 
+If you don't wish to use docker，you need to install several compile dependencies manually as :ref:`Compile Dependencies <_compile_deps>` shows to start compilation.
+
 .. code-block:: bash
 
    # 1. clone the source code
@@ -108,7 +110,7 @@ Frequently Asked Questions
 
 - How difficult is it to learn Docker?
 
-    It takes you ten minutes to read [an introductory article](https://docs.docker.com/get-started) and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools.  Not even to mention the time saved when other people trying to reproduce the issue you have.
+    It takes you ten minutes to read `an introductory article <https://docs.docker.com/get-started>`_ and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools.  Not even to mention the time saved when other people trying to reproduce the issue you have.
 
 - Can I use my favorite IDE?
 
@@ -125,7 +127,7 @@ Frequently Asked Questions
 
 - Does Docker do parallel building?
 
-  Our building Docker image runs a [Bash script](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh), which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores.
+  Our building Docker image runs a  `Bash script <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh>`_ , which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores.
 
 - Docker requires sudo
 
@@ -133,11 +135,11 @@ Frequently Asked Questions
 
 - Docker on Windows/MacOS builds slowly
 
-  On Windows and MacOS, Docker containers run in a Linux VM.  You might want to give this VM some more memory and CPUs so to make the building efficient.  Please refer to [this issue](https://github.com/PaddlePaddle/Paddle/issues/627) for details.
+  On Windows and MacOS, Docker containers run in a Linux VM.  You might want to give this VM some more memory and CPUs so to make the building efficient.  Please refer to `this issue  <https://github.com/PaddlePaddle/Paddle/issues/627>`_ for details.
 
 - Not enough disk space
 
-  Examples in this article use option `--rm` with the `docker run` command.  This option ensures that stopped containers do not exist on hard disks.  We can use `docker ps -a` to list all containers, including stopped.  Sometimes `docker build` generates some intermediate dangling images, which also take disk space.  To clean them, please refer to [this article](https://zaiste.net/posts/removing_docker_containers/).
+  Examples in this article use option `--rm` with the `docker run` command.  This option ensures that stopped containers do not exist on hard disks.  We can use `docker ps -a` to list all containers, including stopped.  Sometimes `docker build` generates some intermediate dangling images, which also take disk space.  To clean them, please refer to `this article <https://zaiste.net/posts/removing_docker_containers/>`_ .
 
 .. _compile_deps:
 

paddle/contrib/inference/CMakeLists.txt

@@ -17,33 +17,77 @@ if(APPLE)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
 endif(APPLE)
 
+set(ANAKIN_INCLUDE "" CACHE STRING "root of Anakin header files")
+set(ANAKIN_LIBRARY "" CACHE STRING "path of Anakin library")
+
+
+set(inference_deps paddle_inference_api paddle_fluid_api)
+
+# if anakin is set enable anakin api implementation
+if(ANAKIN_INCLUDE_DIR AND ANAKIN_LIBRARY)
+    set(ANAKIN_FOUND ON)
+else()
+    set(ANAKIN_FOUND OFF)
+endif()
+
+if (ANAKIN_FOUND)
+    # Anakin's code style doesn't follow google c style.
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=comment
+                                            -Wno-error=reorder
+                                            -Wno-error=format
+                                            -Wno-error=switch
+                                            -Wno-error=return-type
+                                            -Wno-error=non-virtual-dtor
+                                            -Wno-error=cpp")
+
+    message(STATUS "Anakin for inference is enabled")
+    message(STATUS "Anakin is set INCLUDE:${ANAKIN_INCLUDE} LIBRARY:${ANAKIN_LIBRARY}")
+    include_directories("${ANAKIN_INCLUDE}")
+    # Anakin's source path is a mass, need to set sub-directories trivially.
+    include_directories("${ANAKIN_INCLUDE}/saber")
+    link_directories("${ANAKIN_LIBRARY}")
+
+    nv_library(inference_anakin_api SRCS paddle_inference_api_anakin_engine.cc)
+    target_link_libraries(inference_anakin_api anakin)
+    list(APPEND inference_deps inference_anakin_api)
+endif()
+
+
 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 ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
-    cc_test(test_paddle_inference_${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)
-        set_tests_properties(test_paddle_inference_${TARGET_NAME}
-                 PROPERTIES DEPENDS "${inference_test_ARGS}")
-    endif()
-endfunction(inference_api_test)
+    if (WITH_TESTING)
+        set(options "")
+        set(oneValueArgs "")
+        set(multiValueArgs ARGS)
+        cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
 
+        set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
+        cc_test(${TARGET_NAME}
+                SRCS ${TARGET_NAME}.cc
+                DEPS "${inference_deps}"
+                ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
+        if(inference_test_ARGS)
+            set_tests_properties(${TARGET_NAME}
+                    PROPERTIES DEPENDS "${inference_test_ARGS}")
+        endif()
+    endif(WITH_TESTING)
+endfunction(inference_api_test)
 
 cc_library(paddle_inference_api
     SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
     DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
 
-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(api_impl
-                       ARGS test_word2vec test_image_classification)
+inference_api_test(test_paddle_inference_api_impl
+                    ARGS test_word2vec test_image_classification)
+
+if (ANAKIN_FOUND)
+  nv_test(inference_anakin_test SRCS paddle_inference_api_anakin_engine_tester.cc
+    DEPS ${inference_deps} protobuf)
+endif()
+
+if(WITH_TESTING)
+    add_subdirectory(demo)
 endif()

paddle/contrib/inference/demo/CMakeLists.txt

+# 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.
+#
+
+inference_api_test(simple_on_word2vec ARGS test_word2vec)

paddle/contrib/inference/demo/simple_on_word2vec.cc

+/* 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. */
+
+/*
+ * This file contains a simple demo for how to take a model for inference.
+ */
+
+#include <glog/logging.h>
+#include <gtest/gtest.h>
+#include <memory>
+#include "paddle/contrib/inference/paddle_inference_api.h"
+
+namespace paddle {
+namespace demo {
+
+DEFINE_string(dirname, "", "Directory of the inference model.");
+
+void Main(bool use_gpu) {
+  //# 1. Create PaddlePredictor with a config.
+  NativeConfig config;
+  config.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  config.use_gpu = use_gpu;
+  config.fraction_of_gpu_memory = 0.15;
+  config.device = 0;
+  auto predictor =
+      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+
+  for (int batch_id = 0; batch_id < 3; batch_id++) {
+    //# 2. Prepare input.
+    int64_t data[4] = {1, 2, 3, 4};
+
+    PaddleBuf buf{.data = data, .length = sizeof(data)};
+    PaddleTensor tensor{.name = "",
+                        .shape = std::vector<int>({4, 1}),
+                        .data = buf,
+                        .dtype = PaddleDType::INT64};
+
+    // For simplicity, we set all the slots with the same data.
+    std::vector<PaddleTensor> slots(4, tensor);
+
+    //# 3. Run
+    std::vector<PaddleTensor> outputs;
+    CHECK(predictor->Run(slots, &outputs));
+
+    //# 4. Get output.
+    ASSERT_EQ(outputs.size(), 1UL);
+    LOG(INFO) << "output buffer size: " << outputs.front().data.length;
+    const size_t num_elements = outputs.front().data.length / sizeof(float);
+    // The outputs' buffers are in CPU memory.
+    for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
+      LOG(INFO) << static_cast<float*>(outputs.front().data.data)[i];
+    }
+  }
+}
+
+TEST(demo, word2vec_cpu) { Main(false /*use_gpu*/); }
+
+#ifdef PADDLE_WITH_CUDA
+TEST(demo, word2vec_gpu) { Main(true /*use_gpu*/); }
+#endif
+
+}  // namespace demo
+}  // namespace paddle

paddle/contrib/inference/paddle_inference_api.h

 /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
 
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License. */
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
 
 /*
  * This file contains the definition of a simple Inference API for Paddle.
@@ -40,20 +40,30 @@ struct PaddleBuf {
 struct PaddleTensor {
   std::string name;  // variable name.
   std::vector<int> shape;
+  // TODO(Superjomn) for LoD support, add a vector<vector<int>> field if needed.
   PaddleBuf data;  // blob of data.
   PaddleDType dtype;
 };
 
+enum class PaddleEngineKind {
+  kNative = 0,  // Use the native Fluid facility.
+  kAnakin,      // Use Anakin for inference.
+  // TODO(Superjomn) support following engines latter.
+  // 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 can be used by
  * non-sequence scenerios.
- * TODO(Superjomn) Support another API for NLP-related usages.
  */
 class PaddlePredictor {
  public:
   struct Config;
   PaddlePredictor() = default;
   PaddlePredictor(const PaddlePredictor&) = delete;
+  PaddlePredictor& operator=(const PaddlePredictor&) = delete;
 
   // Predict an record.
   // The caller should be responsible for allocating and releasing the memory of
@@ -67,16 +77,7 @@ class PaddlePredictor {
   virtual std::unique_ptr<PaddlePredictor> Clone() = 0;
 
   // Destroy the Predictor.
-  virtual ~PaddlePredictor() {}
-
-  enum class EngineKind {
-    kNative = -1,  // Use the native Fluid facility.
-    // TODO(Superjomn) support latter.
-    // kAnakin,             // Use Anakin for inference.
-    // kTensorRT,           // Use TensorRT for inference.
-    // kAutoMixedAnakin,    // Automatically mix Fluid with Anakin.
-    // kAutoMixedTensorRT,  // Automatically mix Fluid with TensorRT.
-  };
+  virtual ~PaddlePredictor() = default;
 
   // The common configs for all the predictors.
   struct Config {
@@ -86,18 +87,31 @@ class PaddlePredictor {
 };
 
 struct NativeConfig : public PaddlePredictor::Config {
+  // GPU related fields.
   bool use_gpu{false};
-  int device;
-  float fraction_of_gpu_memory;
+  int device{0};
+  float fraction_of_gpu_memory{-1.f};  // Negative to notify initialization.
+
   std::string prog_file;
   std::string param_file;
-  bool share_variables;
 };
 
-// A factory to help create difference predictor.
-template <
-    typename ConfigT,
-    PaddlePredictor::EngineKind engine = PaddlePredictor::EngineKind::kNative>
+// Configurations for Anakin engine.
+struct AnakinConfig : public PaddlePredictor::Config {
+  int device;
+  std::string model_file;
+  int max_batch_size{-1};
+};
+
+// 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 <typename ConfigT, PaddleEngineKind engine = PaddleEngineKind::kNative>
 std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT& config);
 
 }  // namespace paddle

paddle/contrib/inference/paddle_inference_api_anakin_engine.cc

+// 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 <cuda.h>
+
+#include "paddle/contrib/inference/paddle_inference_api_anakin_engine.h"
+
+namespace paddle {
+
+PaddleInferenceAnakinPredictor::PaddleInferenceAnakinPredictor(
+    const AnakinConfig &config) {
+  CHECK(Init(config));
+}
+
+bool PaddleInferenceAnakinPredictor::Init(const AnakinConfig &config) {
+  // TODO(Superjomn) Tell anakin to support return code.
+  engine_.Build(config.model_file, config.max_batch_size);
+  return true;
+}
+
+bool PaddleInferenceAnakinPredictor::Run(
+    const std::vector<PaddleTensor> &inputs,
+    std::vector<PaddleTensor> *output_data) {
+  for (const auto &input : inputs) {
+    if (input.dtype != PaddleDType::FLOAT32) {
+      LOG(ERROR) << "Only support float type inputs. " << input.name
+                 << "'s type is not float";
+      return false;
+    }
+    engine_.SetInputFromCPU(
+        input.name, static_cast<float *>(input.data.data), input.data.length);
+  }
+
+  // TODO(Superjomn) Tell anakin to support return code.
+  engine_.Execute();
+
+  if (output_data->empty()) {
+    LOG(ERROR) << "At least one output should be set with tensors' names.";
+    return false;
+  }
+  for (auto &output : *output_data) {
+    auto *tensor = engine_.GetOutputInGPU(output.name);
+    output.shape = tensor->shape();
+    // Copy data from GPU -> CPU
+    if (cudaMemcpy(output.data.data,
+                   tensor->data(),
+                   tensor->size(),
+                   cudaMemcpyDeviceToHost) != 0) {
+      LOG(ERROR) << "copy data from GPU to CPU error";
+      return false;
+    }
+  }
+  return true;
+}
+
+// TODO(Superjomn) To implement latter.
+std::unique_ptr<PaddlePredictor> PaddleInferenceAnakinPredictor::Clone() {
+  return nullptr;
+}
+
+// A factory to help create difference predictor.
+template <>
+std::unique_ptr<PaddlePredictor>
+CreatePaddlePredictor<AnakinConfig, PaddleEngineKind::kAnakin>(
+    const AnakinConfig &config) {
+  std::unique_ptr<PaddlePredictor> x(
+      new PaddleInferenceAnakinPredictor(config));
+  return x;
+};
+
+}  // namespace paddle

paddle/contrib/inference/paddle_inference_api_anakin_engine.h

+/* 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. */
+
+/*
+ * This file contains the implementation of inference API with Anakin engine
+ * embeded, this API can only support Anakin models.
+ */
+
+#pragma once
+
+// NOTE This header file do not have namespace.
+// TODO(Superjomn) Tell Anakin to provide better APIs.
+#include <test/framework/net/paddle_api.h>
+#include "paddle/contrib/inference/paddle_inference_api.h"
+
+namespace paddle {
+
+class PaddleInferenceAnakinPredictor : public PaddlePredictor {
+ public:
+  PaddleInferenceAnakinPredictor(const AnakinConfig& config);
+
+  // NOTE Unlike the native engine, the buffers of anakin engine's output_data
+  // should be allocated first.
+  // TODO(Superjomn) should unify all the behaviors of output_data accross all
+  // the engines.
+  bool Run(const std::vector<PaddleTensor>& inputs,
+           std::vector<PaddleTensor>* output_data) override;
+
+  std::unique_ptr<PaddlePredictor> Clone() override;
+
+ private:
+  bool Init(const AnakinConfig& config);
+
+  anakin::AnakinEngine<anakin::NV,
+                       anakin::saber::AK_FLOAT,
+                       anakin::Precision::FP32>
+      engine_;
+};
+
+}  // namespace paddle

paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc

+/* 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/contrib/inference/paddle_inference_api.h"
+#include <gtest/gtest.h>
+
+namespace paddle {
+
+TEST(inference, anakin) {
+  AnakinConfig config;
+
+  auto engine =
+      CreatePaddlePredictor<AnakinConfig, PaddleEngineKind::kAnakin>(config);
+}
+
+}  // namespace paddle

paddle/contrib/inference/paddle_inference_api_impl.cc

 /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
 
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
 
-   http://www.apache.org/licenses/LICENSE-2.0
+http://www.apache.org/licenses/LICENSE-2.0
 
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License. */
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
 
 #include <sys/time.h>
 #include <algorithm>
@@ -54,18 +54,24 @@ std::string num2str(T a) {
 }
 }  // namespace
 
-bool NativePaddlePredictor::Init() {
+bool NativePaddlePredictor::Init(
+    std::shared_ptr<framework::Scope> parent_scope) {
   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();
   }
-  paddle::framework::InitDevices(false);
+  if (parent_scope) {
+    scope_ = parent_scope;
+    sub_scope_ = &(parent_scope->NewScope());
+  } else {
+    paddle::framework::InitDevices(false);
+    scope_.reset(new paddle::framework::Scope());
+  }
+
   executor_.reset(new paddle::framework::Executor(place_));
-  scope_.reset(new paddle::framework::Scope());
 
   // Initialize the inference program
   if (!config_.model_dir.empty()) {
@@ -84,18 +90,22 @@ bool NativePaddlePredictor::Init() {
     return false;
   }
   ctx_ = executor_->Prepare(*inference_program_, 0);
+  executor_->CreateVariables(
+      *inference_program_, sub_scope_ ? sub_scope_ : scope_.get(), 0);
 
-  // Create variables
-  // TODO(panyx0718): Why need to test share_variables here?
-  if (config_.share_variables) {
-    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;
 }
 
+NativePaddlePredictor::~NativePaddlePredictor() {
+  if (sub_scope_) {
+    PADDLE_ENFORCE_NOT_NULL(scope_, "Should have parent scope!");
+    scope_->DeleteScope(sub_scope_);
+  }
+};
+
 bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
                                 std::vector<PaddleTensor> *output_data) {
   VLOG(3) << "Predictor::predict";
@@ -120,11 +130,12 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
   }
   // Run the inference program
   // if share variables, we need not create variables
-  executor_->RunPreparedContext(ctx_.get(),
-                                scope_.get(),
-                                &feed_targets,
-                                &fetch_targets,
-                                !config_.share_variables);
+  executor_->RunPreparedContext(
+      ctx_.get(),
+      sub_scope_ != nullptr ? sub_scope_ : scope_.get(),
+      &feed_targets,
+      &fetch_targets,
+      false /* don't create variable eatch time */);
   if (!GetFetch(fetchs, output_data)) {
     LOG(ERROR) << "fail to get fetchs";
     return false;
@@ -137,7 +148,7 @@ std::unique_ptr<PaddlePredictor> NativePaddlePredictor::Clone() {
   VLOG(3) << "Predictor::clone";
   std::unique_ptr<PaddlePredictor> cls(new NativePaddlePredictor(config_));
 
-  if (!dynamic_cast<NativePaddlePredictor *>(cls.get())->Init()) {
+  if (!dynamic_cast<NativePaddlePredictor *>(cls.get())->Init(scope_)) {
     LOG(ERROR) << "fail to call Init";
     return nullptr;
   }
@@ -242,11 +253,16 @@ bool NativePaddlePredictor::GetFetch(
 
 template <>
 std::unique_ptr<PaddlePredictor>
-CreatePaddlePredictor<NativeConfig, PaddlePredictor::EngineKind::kNative>(
+CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(
     const NativeConfig &config) {
   VLOG(3) << "create NativePaddlePredictor";
   if (config.use_gpu) {
     // 1. GPU memeroy
+    PADDLE_ENFORCE_GT(
+        config.fraction_of_gpu_memory,
+        0.f,
+        "fraction_of_gpu_memory in the config should be set to range (0., 1.]");
+    PADDLE_ENFORCE_GE(config.device, 0, "Invalid device id %d", config.device);
     std::vector<std::string> flags;
     if (config.fraction_of_gpu_memory >= 0.0f ||
         config.fraction_of_gpu_memory <= 0.95f) {
@@ -260,7 +276,7 @@ CreatePaddlePredictor<NativeConfig, PaddlePredictor::EngineKind::kNative>(
   }
 
   std::unique_ptr<PaddlePredictor> predictor(new NativePaddlePredictor(config));
-  if (!dynamic_cast<NativePaddlePredictor *>(predictor.get())->Init()) {
+  if (!dynamic_cast<NativePaddlePredictor *>(predictor.get())->Init(nullptr)) {
     return nullptr;
   }
   return std::move(predictor);

paddle/contrib/inference/paddle_inference_api_impl.h

@@ -34,14 +34,15 @@ class NativePaddlePredictor : public PaddlePredictor {
   explicit NativePaddlePredictor(const NativeConfig &config)
       : config_(config) {}
 
-  bool Init();
+  // will only create sub scope if have global scope
+  bool Init(std::shared_ptr<framework::Scope> parent_scope);
 
   bool Run(const std::vector<PaddleTensor> &inputs,
            std::vector<PaddleTensor> *output_data) override;
 
   std::unique_ptr<PaddlePredictor> Clone() override;
 
-  ~NativePaddlePredictor() override{};
+  ~NativePaddlePredictor() override;
 
  private:
   bool SetFeed(const std::vector<PaddleTensor> &input_datas,
@@ -52,11 +53,13 @@ class NativePaddlePredictor : public PaddlePredictor {
   NativeConfig config_;
   platform::Place place_;
   std::unique_ptr<framework::Executor> executor_;
-  std::unique_ptr<framework::Scope> scope_;
+  std::shared_ptr<framework::Scope> scope_;
   std::unique_ptr<framework::ExecutorPrepareContext> ctx_;
   std::unique_ptr<framework::ProgramDesc> inference_program_;
   std::vector<std::string> feed_target_names_;
   std::vector<std::string> fetch_target_names_;
+  // Do not use unique_ptr, use parent scope to delete
+  framework::Scope *sub_scope_{nullptr};
 };
 
 }  // namespace paddle

paddle/contrib/inference/test_paddle_inference_api_impl.cc

@@ -47,7 +47,6 @@ NativeConfig GetConfig() {
   config.fraction_of_gpu_memory = 0.15;
   config.use_gpu = true;
   config.device = 0;
-  config.share_variables = true;
   return config;
 }
 
@@ -75,7 +74,7 @@ TEST(paddle_inference_api_impl, word2vec) {
   ASSERT_EQ(outputs.size(), 1UL);
   size_t len = outputs[0].data.length;
   float* data = static_cast<float*>(outputs[0].data.data);
-  for (int j = 0; j < len / sizeof(float); ++j) {
+  for (size_t j = 0; j < len / sizeof(float); ++j) {
     ASSERT_LT(data[j], 1.0);
     ASSERT_GT(data[j], -1.0);
   }
@@ -93,7 +92,7 @@ TEST(paddle_inference_api_impl, word2vec) {
   TestInference<platform::CPUPlace>(config.model_dir, cpu_feeds, cpu_fetchs1);
 
   float* lod_data = output1.data<float>();
-  for (size_t i = 0; i < output1.numel(); ++i) {
+  for (int i = 0; i < output1.numel(); ++i) {
     EXPECT_LT(lod_data[i] - data[i], 1e-3);
     EXPECT_GT(lod_data[i] - data[i], -1e-3);
   }

paddle/fluid/framework/CMakeLists.txt

@@ -87,7 +87,7 @@ cc_library(executor SRCS executor.cc DEPS op_registry device_context scope
 framework_proto glog lod_rank_table feed_fetch_method)
 
 
-cc_library(parallel_executor SRCS parallel_executor.cc DEPS multi_devices_graph_builder threaded_ssa_graph_executor)
+cc_library(parallel_executor SRCS parallel_executor.cc DEPS multi_devices_graph_builder threaded_ssa_graph_executor scope_buffered_ssa_graph_executor)
 
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)

paddle/fluid/framework/block_desc.cc

@@ -169,17 +169,13 @@ void BlockDesc::Flush() {
   }
 
   if (need_update_) {
-    auto &op_field = *this->desc_->mutable_ops();
-    this->ClearPBOps();
-    op_field.Reserve(static_cast<int>(ops_.size()));
+    this->desc_->mutable_ops()->Clear();
     for (auto &op_desc : ops_) {
-      op_field.AddAllocated(op_desc->Proto());
+      this->desc_->mutable_ops()->Add()->CopyFrom(*op_desc->Proto());
     }
-    auto &var_field = *this->desc_->mutable_vars();
-    this->ClearPBVars();
-    var_field.Reserve(static_cast<int>(vars_.size()));
+    this->desc_->mutable_vars()->Clear();
     for (auto &var_desc : vars_) {
-      var_field.AddAllocated(var_desc.second->Proto());
+      this->desc_->mutable_vars()->Add()->CopyFrom(*var_desc.second->Proto());
     }
     need_update_ = false;
   }
@@ -217,22 +213,6 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc,
   }
 }
 
-void BlockDesc::ClearPBOps() {
-  auto ops = this->desc_->mutable_ops();
-  while (!ops->empty()) {
-    // we do not own the OpDesc, so release the ownership.
-    ops->ReleaseLast();
-  }
-}
-
-void BlockDesc::ClearPBVars() {
-  auto vars = this->desc_->mutable_vars();
-  while (!vars->empty()) {
-    // we do not own the VarDesc, so release the ownership.
-    vars->ReleaseLast();
-  }
-}
-
 void BlockDesc::SetForwardBlockID(int32_t forward_block_id) {
   PADDLE_ENFORCE(!desc_->has_forward_block_idx(),
                  "Parent block ID has been set to %d. Cannot set to %d",

paddle/fluid/framework/block_desc.h

@@ -41,11 +41,6 @@ class BlockDesc {
 
   BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, ProgramDesc *prog);
 
-  ~BlockDesc() {
-    this->ClearPBVars();
-    this->ClearPBOps();
-  }
-
   int32_t ID() const { return desc_->idx(); }
 
   int32_t Parent() const { return desc_->parent_idx(); }
@@ -113,10 +108,6 @@ class BlockDesc {
 
   ProgramDesc *Program() const { return this->prog_; }
 
- private:
-  void ClearPBOps();
-  void ClearPBVars();
-
  private:
   ProgramDesc *prog_;       // not_own
   proto::BlockDesc *desc_;  // not_own

paddle/fluid/framework/details/CMakeLists.txt

@@ -36,5 +36,6 @@ cc_test(broadcast_op_test SRCS broadcast_op_handle_test.cc DEPS var_handle op_ha
         device_context broadcast_op_handle)
 cc_test(gather_op_test SRCS gather_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
         device_context gather_op_handle)
+cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_executor.cc DEPS ssa_graph_executor)
 #cc_test(reduce_op_handle_test SRCS reduce_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory
 #        device_context reduce_op_handle )

paddle/fluid/framework/details/execution_strategy.h

@@ -22,6 +22,7 @@ struct ExecutionStrategy {
   size_t num_threads_{0};
   bool use_event_{true};
   bool allow_op_delay_{false};
+  size_t num_iteration_per_drop_scope_{100};
 };
 
 }  //  namespace details

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc

+// 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/details/scope_buffered_ssa_graph_executor.h"
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/executor.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+ScopeBufferedSSAGraphExecutor::ScopeBufferedSSAGraphExecutor(
+    ExecutionStrategy strategy, std::vector<Scope *> local_scopes,
+    std::vector<VariableInfo> var_infos, std::vector<platform::Place> places,
+    std::unique_ptr<SSAGraphExecutor> &&underlying_executor)
+    : strategy_(std::move(strategy)),
+      underlying_executor_(std::move(underlying_executor)),
+      local_scopes_(std::move(local_scopes)),
+      var_infos_(std::move(var_infos)),
+      places_(std::move(places)) {}
+
+FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
+    const std::vector<std::string> &fetch_tensors) {
+  if (drop_scope_counter_ == 0) {
+    // Create local scopes.
+    for (auto it = local_scopes_.rbegin(); it != local_scopes_.rend(); ++it) {
+      auto &scope = *it;
+      Scope &local_scope = scope->NewScope();
+      *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>() =
+          &local_scope;
+
+      for (auto &info : var_infos_) {
+        if (scope->FindVar(info.name_) != nullptr) {
+          continue;
+        }
+
+        if (info.persistable_) {  // Persistable
+          InitializeVariable(scope->Var(info.name_), info.type_);
+        } else {
+          InitializeVariable(local_scope.Var(info.name_), info.type_);
+        }
+      }
+    }
+  }
+
+  auto fetch_data = underlying_executor_->Run(fetch_tensors);
+  drop_scope_counter_ += 1;
+  if (!fetch_tensors.empty() ||
+      drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {
+    drop_scope_counter_ = 0;
+    // Wait All computational streams
+    for (auto p : places_) {
+      platform::DeviceContextPool::Instance().Get(p)->Wait();
+    }
+    for (auto &scope : local_scopes_) {
+      auto &local_scope =
+          *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>();
+      scope->DeleteScope(local_scope);
+    }
+  }
+  return fetch_data;
+}
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h

+// 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 <memory>
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/details/execution_strategy.h"
+#include "paddle/fluid/framework/details/ssa_graph_executor.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/platform/place.h"
+namespace paddle {
+namespace framework {
+namespace details {
+
+struct VariableInfo {
+  std::string name_;
+  proto::VarType::Type type_;
+  bool persistable_;
+};
+
+class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor {
+ public:
+  ScopeBufferedSSAGraphExecutor(
+      ExecutionStrategy strategy, std::vector<Scope*> local_scopes,
+      std::vector<VariableInfo> var_infos, std::vector<platform::Place> places,
+      std::unique_ptr<SSAGraphExecutor>&& underlying_executor);
+  FeedFetchList Run(const std::vector<std::string>& fetch_tensors) override;
+
+ private:
+  size_t drop_scope_counter_{0};
+
+  ExecutionStrategy strategy_;
+  std::unique_ptr<SSAGraphExecutor> underlying_executor_;
+  std::vector<Scope*> local_scopes_;
+  std::vector<VariableInfo> var_infos_;
+  std::vector<platform::Place> places_;
+};
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/ssa_graph_executor.cc

@@ -17,10 +17,6 @@
 namespace paddle {
 namespace framework {
 namespace details {
-
-SSAGraphExecutor::SSAGraphExecutor(std::unique_ptr<SSAGraph> &&graph)
-    : graph_(std::move(graph)) {}
-
 SSAGraphExecutor::~SSAGraphExecutor() {}
 
 }  // namespace details

paddle/fluid/framework/details/ssa_graph_executor.h

@@ -28,15 +28,11 @@ class SSAGraphExecutor {
   DISABLE_COPY_AND_ASSIGN(SSAGraphExecutor);
 
  public:
-  // Steal graph inside
-  explicit SSAGraphExecutor(std::unique_ptr<SSAGraph> &&graph);
+  SSAGraphExecutor() {}
 
   virtual ~SSAGraphExecutor();
 
   virtual FeedFetchList Run(const std::vector<std::string> &fetch_tensors) = 0;
-
- protected:
-  std::unique_ptr<SSAGraph> graph_;
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -21,7 +21,7 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
     const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places,
     std::unique_ptr<SSAGraph> &&graph)
-    : SSAGraphExecutor(std::move(graph)),
+    : graph_(std::move(graph)),
       pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_)
                                        : nullptr),
       local_scopes_(local_scopes),
@@ -189,7 +189,9 @@ void ThreadedSSAGraphExecutor::RunOp(
     BlockingQueue<VarHandleBase *> *ready_var_q, details::OpHandleBase *op) {
   auto op_run = [ready_var_q, op, this] {
     try {
-      VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
+      if (VLOG_IS_ON(10)) {
+        VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
+      }
       op->Run(strategy_.use_event_);
       VLOG(10) << op << " " << op->Name() << " Done ";
       running_ops_--;

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -51,6 +51,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
              details::OpHandleBase *op);
 
  private:
+  std::unique_ptr<SSAGraph> graph_;
   std::unique_ptr<::ThreadPool> pool_;
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;

paddle/fluid/framework/executor.cc

@@ -220,8 +220,10 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
       has_fetch_operators(program.Block(0), *fetch_targets, fetch_holder_name);
 
   ProgramDesc* copy_program = const_cast<ProgramDesc*>(&program);
+  std::unique_ptr<ProgramDesc> unique_ptr_of_copy_program;
   if (!has_feed_ops || !has_fetch_ops) {
-    copy_program = std::unique_ptr<ProgramDesc>(new ProgramDesc(program)).get();
+    unique_ptr_of_copy_program.reset(new ProgramDesc(program));
+    copy_program = unique_ptr_of_copy_program.get();
   }
 
   auto* global_block = copy_program->MutableBlock(0);

paddle/fluid/framework/parallel_executor.cc

@@ -23,6 +23,7 @@ limitations under the License. */
 #endif
 
 #include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
+#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
 #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -42,8 +43,6 @@ class ParallelExecutorPrivate {
 #ifdef PADDLE_WITH_CUDA
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
-
-  std::vector<std::tuple<std::string, proto::VarType::Type, bool>> var_types_;
   bool own_local_scope;
 };
 
@@ -92,9 +91,18 @@ ParallelExecutor::ParallelExecutor(
       local_scopes.empty()) {  // Is CUDA
     BCastParamsToGPUs(bcast_vars);
   }
-// Startup Program has been run. All local scopes has correct parameters.
+  // Startup Program has been run. All local scopes has correct parameters.
+
+  // Step 2. Create vars in each scope;
+  std::vector<details::VariableInfo> var_infos;
+  for (auto *var : main_program.Block(0).AllVars()) {
+    var_infos.emplace_back();
+    var_infos.back().name_ = var->Name();
+    var_infos.back().type_ = var->GetType();
+    var_infos.back().persistable_ = var->Persistable();
+  }
 
-// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
+// Step 3. Convert main_program to SSA form and dependency graph. Also, insert
 // ncclOp
 #ifdef PADDLE_WITH_CUDA
   details::MultiDevSSAGraphBuilder builder(
@@ -105,16 +113,15 @@ ParallelExecutor::ParallelExecutor(
                                            params, member_->local_scopes_,
                                            build_strategy);
 #endif
+
   auto graph = builder.Build(main_program);
 
   member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
       exec_strategy, member_->local_scopes_, places, std::move(graph)));
 
-  // Step 3. Create vars in each scope;
-  for (auto *var : main_program.Block(0).AllVars()) {
-    member_->var_types_.emplace_back(var->Name(), var->GetType(),
-                                     var->Persistable());
-  }
+  member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
+      exec_strategy, member_->local_scopes_, std::move(var_infos),
+      member_->places_, std::move(member_->executor_)));
 }
 
 void ParallelExecutor::BCastParamsToGPUs(
@@ -169,42 +176,9 @@ void ParallelExecutor::BCastParamsToGPUs(
 void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                            const std::string &fetched_var_name) {
   platform::RecordBlock b(0);
-  // Create local scopes.
-  for (auto it = member_->local_scopes_.rbegin();
-       it != member_->local_scopes_.rend(); ++it) {
-    auto &scope = *it;
-    Scope &local_scope = scope->NewScope();
-    *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>() =
-        &local_scope;
-
-    for (auto &name_type_pair : member_->var_types_) {
-      if (scope->FindVar(std::get<0>(name_type_pair)) != nullptr) {
-        continue;
-      }
-
-      if (std::get<2>(name_type_pair)) {  // Persistable
-        InitializeVariable(scope->Var(std::get<0>(name_type_pair)),
-                           std::get<1>(name_type_pair));
-      } else {
-        InitializeVariable(local_scope.Var(std::get<0>(name_type_pair)),
-                           std::get<1>(name_type_pair));
-      }
-    }
-  }
-
   auto fetch_data = member_->executor_->Run(fetch_tensors);
   *member_->global_scope_->Var(fetched_var_name)->GetMutable<FeedFetchList>() =
       fetch_data;
-
-  // Wait All computational streams
-  for (auto p : member_->places_) {
-    platform::DeviceContextPool::Instance().Get(p)->Wait();
-  }
-  for (auto &scope : member_->local_scopes_) {
-    auto &local_scope =
-        *scope->Var(details::kLocalExecScopeName)->GetMutable<Scope *>();
-    scope->DeleteScope(local_scope);
-  }
 }
 
 void ParallelExecutor::FeedTensorsIntoLocalScopes(

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -15,3 +15,9 @@ cc_test(test_subgraph_splitter
         DEPS analysis paddle_fluid tensor
         ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model)
 set_tests_properties(test_subgraph_splitter PROPERTIES DEPENDS test_word2vec)
+
+cc_test(test_dfg_graphviz_draw_pass
+        SRCS dfg_graphviz_draw_pass_tester.cc
+        DEPS analysis
+        ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model)
+set_tests_properties(test_dfg_graphviz_draw_pass PROPERTIES DEPENDS test_word2vec)

paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h

+/* 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. */
+
+/*
+ * This file create an DFG_GraphvizDrawPass which helps to draw a data flow
+ * graph's structure using graphviz.
+ */
+
+#pragma once
+
+#include <fstream>
+#include <string>
+#include "paddle/fluid/inference/analysis/pass.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+/*
+ * Output a dot file and write to some place.
+ */
+class DFG_GraphvizDrawPass : public DataFlowGraphPass {
+ public:
+  DFG_GraphvizDrawPass(const std::string& dir, const std::string& id)
+      : dir_(dir), id_(id) {}
+
+  bool Initialize() override { return Pass::Initialize(); }
+  void Run(DataFlowGraph* graph) override {
+    auto content = Draw(graph);
+    std::ofstream file(GenDotPath());
+    file.write(content.c_str(), content.size());
+    file.close();
+    LOG(INFO) << "draw dot to " << GenDotPath();
+  }
+
+  bool Finalize() override { return Pass::Finalize(); }
+
+  Pass* CreatePrinterPass(std::ostream& os,
+                          const std::string& banner) const override {
+    return nullptr;
+  }
+
+ private:
+  // Path of the dot file to output.
+  std::string GenDotPath() const {
+    return dir_ + "/" + "graph_" + id_ + ".dot";
+  }
+
+  std::string Draw(DataFlowGraph* graph) { return graph->DotString(); }
+
+  std::string dir_;
+  std::string id_;
+};
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h"
+
+#include <gtest/gtest.h>
+#include <fstream>
+#include <string>
+#include "paddle/fluid/inference/analysis/ut_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+TEST_F(DFG_Tester, dfg_graphviz_draw_pass_tester) {
+  auto dfg = ProgramDescToDFG(desc);
+  DFG_GraphvizDrawPass pass("./", "test");
+  pass.Initialize();
+  pass.Run(&dfg);
+
+  // test content
+  std::ifstream file("./graph_test.dot");
+  ASSERT_TRUE(file.is_open());
+
+  std::string line;
+  int no{0};
+  while (std::getline(file, line)) {
+    no++;
+  }
+  ASSERT_EQ(no, 82);
+}
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

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)

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";
   }

paddle/fluid/inference/tensorrt/convert/fc_op.cc

+/* 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 <typename T>
+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<float const*>(iweights.get().values), istrides,
+           static_cast<float*>(const_cast<void*>(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);
+    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<framework::LoDTensor>();
+    // 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<float>(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<float>(platform::CPUPlace()), Y_t->data<float>(),
+           Y_t->dims()[0] * Y_t->dims()[1]);
+
+    TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT,
+                                  static_cast<void*>(weight_data),
+                                  Y_t->memory_size() / sizeof(float)};
+    TensorRTEngine::Weight tmp_weight(nvinfer1::DataType::kFLOAT,
+                                      static_cast<void*>(tmp.data<float>()),
+                                      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<nvinfer1::ITensor*>(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);

paddle/fluid/inference/tensorrt/convert/mul_op.cc

@@ -24,8 +24,9 @@ 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);
     // Declare inputs

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<OpConverter>::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<std::string>& parameters,
+                 const framework::Scope& scope, TensorRTEngine* engine) {
+    framework::OpDesc op_desc(op, 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<OpConverter>::Lookup("fc");
+      }
+    }
+    if (!it) {
+      it = Registry<OpConverter>::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<std::string>& 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);
     }
   }
 

paddle/fluid/inference/tensorrt/convert/test_fc_op.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "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<std::string> 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

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<std::string> 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));

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 <gtest/gtest.h>
 #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

paddle/fluid/inference/tensorrt/convert/ut_helper.h

@@ -61,7 +61,10 @@ class TRTConvertValidation {
  public:
   TRTConvertValidation() = delete;
 
-  explicit TRTConvertValidation(int batch_size, int workspace_size = 1024) {
+  TRTConvertValidation(int batch_size,
+                       const std::unordered_set<std::string>& 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();
@@ -76,19 +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<int> dim_vec(dims.nbDims);
-    for (int i = 0; i < dims.nbDims; i++) {
-      dim_vec[i] = dims.d[i];
-    }
+    std::vector<int> dim_vec(dims.d, dims.d + dims.nbDims);
     auto* x = scope_.Var(name);
     auto* x_tensor = x->GetMutable<framework::LoDTensor>();
     x_tensor->Resize(framework::make_ddim(dim_vec));
@@ -99,7 +105,7 @@ 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();
 
@@ -108,11 +114,13 @@ class TRTConvertValidation {
 
     // 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<framework::LoDTensor>();
+
       engine_->SetInputFromCPU(
-          input, static_cast<void*>(tensor->data<float>()),
+          input, static_cast<void*>(tensor->data<void>()),
           sizeof(float) *
               analysis::AccuDims(tensor->dims(), tensor->dims().size()));
     }
@@ -120,18 +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<float> fluid_out;
-      std::vector<float> trt_out(200);
-      engine_->GetOutputInCPU(output, &trt_out[0], 200 * sizeof(float));
+      std::vector<float> 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<framework::LoDTensor>();
@@ -139,7 +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);
       }
     }
   }
@@ -149,9 +160,10 @@ class TRTConvertValidation {
  private:
   std::unique_ptr<TensorRTEngine> engine_;
   cudaStream_t stream_;
-  framework::Scope scope_;
   std::unique_ptr<framework::OperatorBase> op_;
   std::unique_ptr<framework::OpDesc> op_desc_;
+  const std::unordered_set<std::string>& parameters_;
+  framework::Scope& scope_;
 };
 
 }  // namespace tensorrt

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);

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<int64_t> dims;
+
    private:
     nvinfer1::Weights w_;
   };

paddle/fluid/inference/tests/book/CMakeLists.txt

@@ -38,3 +38,11 @@ inference_test(recommender_system)
 #inference_test(rnn_encoder_decoder)
 #inference_test(understand_sentiment ARGS conv)
 inference_test(word2vec)
+
+# This is an unly work around to make this test run
+# TODO(TJ): clean me up
+cc_test(test_inference_nlp
+  SRCS test_inference_nlp.cc
+  DEPS paddle_fluid
+  ARGS
+  --model_path=${PADDLE_BINARY_DIR}/python/paddle/fluid/tests/book/recognize_digits_mlp.inference.model)

paddle/fluid/inference/tests/book/test_inference_nlp.cc

+/* 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 <sys/time.h>
+#include <time.h>
+#include <fstream>
+#include <thread>  // NOLINT
+#include "gflags/gflags.h"
+#include "gtest/gtest.h"
+#include "paddle/fluid/inference/tests/test_helper.h"
+#ifdef PADDLE_WITH_MKLML
+#include <mkl_service.h>
+#include <omp.h>
+#endif
+
+DEFINE_string(model_path, "", "Directory of the inference model.");
+DEFINE_string(data_file, "", "File of input index data.");
+DEFINE_int32(repeat, 100, "Running the inference program repeat times");
+DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference");
+DEFINE_bool(prepare_vars, true, "Prepare variables before executor");
+DEFINE_int32(num_threads, 1, "Number of threads should be used");
+
+inline double GetCurrentMs() {
+  struct timeval time;
+  gettimeofday(&time, NULL);
+  return 1e+3 * time.tv_sec + 1e-3 * time.tv_usec;
+}
+
+// This function just give dummy data for recognize_digits model.
+size_t DummyData(std::vector<paddle::framework::LoDTensor>* out) {
+  paddle::framework::LoDTensor input;
+  SetupTensor<float>(&input, {1, 1, 28, 28}, -1.f, 1.f);
+  out->emplace_back(input);
+  return 1;
+}
+
+// Load the input word index data from file and save into LodTensor.
+// Return the size of words.
+size_t LoadData(std::vector<paddle::framework::LoDTensor>* out,
+                const std::string& filename) {
+  if (filename.empty()) {
+    return DummyData(out);
+  }
+
+  size_t sz = 0;
+  std::fstream fin(filename);
+  std::string line;
+  out->clear();
+  while (getline(fin, line)) {
+    std::istringstream iss(line);
+    std::vector<int64_t> ids;
+    std::string field;
+    while (getline(iss, field, ' ')) {
+      ids.push_back(stoi(field));
+    }
+    if (ids.size() >= 1024) {
+      // Synced with NLP guys, they will ignore input larger then 1024
+      continue;
+    }
+
+    paddle::framework::LoDTensor words;
+    paddle::framework::LoD lod{{0, ids.size()}};
+    words.set_lod(lod);
+    int64_t* pdata = words.mutable_data<int64_t>(
+        {static_cast<int64_t>(ids.size()), 1}, paddle::platform::CPUPlace());
+    memcpy(pdata, ids.data(), words.numel() * sizeof(int64_t));
+    out->emplace_back(words);
+    sz += ids.size();
+  }
+  return sz;
+}
+
+// Split input data samples into small pieces jobs as balanced as possible,
+// according to the number of threads.
+void SplitData(
+    const std::vector<paddle::framework::LoDTensor>& datasets,
+    std::vector<std::vector<const paddle::framework::LoDTensor*>>* jobs,
+    const int num_threads) {
+  size_t s = 0;
+  jobs->resize(num_threads);
+  while (s < datasets.size()) {
+    for (auto it = jobs->begin(); it != jobs->end(); it++) {
+      it->emplace_back(&datasets[s]);
+      s++;
+      if (s >= datasets.size()) {
+        break;
+      }
+    }
+  }
+}
+
+void ThreadRunInfer(
+    const int tid, paddle::framework::Executor* executor,
+    paddle::framework::Scope* scope,
+    const std::unique_ptr<paddle::framework::ProgramDesc>& inference_program,
+    const std::vector<std::vector<const paddle::framework::LoDTensor*>>& jobs) {
+  auto copy_program = std::unique_ptr<paddle::framework::ProgramDesc>(
+      new paddle::framework::ProgramDesc(*inference_program));
+  auto& sub_scope = scope->NewScope();
+
+  std::string feed_holder_name = "feed_" + paddle::string::to_string(tid);
+  std::string fetch_holder_name = "fetch_" + paddle::string::to_string(tid);
+  copy_program->SetFeedHolderName(feed_holder_name);
+  copy_program->SetFetchHolderName(fetch_holder_name);
+
+  const std::vector<std::string>& feed_target_names =
+      copy_program->GetFeedTargetNames();
+  const std::vector<std::string>& fetch_target_names =
+      copy_program->GetFetchTargetNames();
+
+  PADDLE_ENFORCE_EQ(fetch_target_names.size(), 1UL);
+  std::map<std::string, paddle::framework::LoDTensor*> fetch_targets;
+  paddle::framework::LoDTensor outtensor;
+  fetch_targets[fetch_target_names[0]] = &outtensor;
+
+  std::map<std::string, const paddle::framework::LoDTensor*> feed_targets;
+  PADDLE_ENFORCE_EQ(feed_target_names.size(), 1UL);
+
+  auto& inputs = jobs[tid];
+  auto start_ms = GetCurrentMs();
+  for (size_t i = 0; i < inputs.size(); ++i) {
+    feed_targets[feed_target_names[0]] = inputs[i];
+    executor->Run(*copy_program, &sub_scope, &feed_targets, &fetch_targets,
+                  true /*create_local_scope*/, true /*create_vars*/,
+                  feed_holder_name, fetch_holder_name);
+  }
+  auto stop_ms = GetCurrentMs();
+  scope->DeleteScope(&sub_scope);
+  LOG(INFO) << "Tid: " << tid << ", process " << inputs.size()
+            << " samples, avg time per sample: "
+            << (stop_ms - start_ms) / inputs.size() << " ms";
+}
+
+TEST(inference, nlp) {
+  if (FLAGS_model_path.empty()) {
+    LOG(FATAL) << "Usage: ./example --model_path=path/to/your/model";
+  }
+  if (FLAGS_data_file.empty()) {
+    LOG(WARNING) << "No data file provided, will use dummy data!"
+                 << "Note: if you use nlp model, please provide data file.";
+  }
+  LOG(INFO) << "Model Path: " << FLAGS_model_path;
+  LOG(INFO) << "Data File: " << FLAGS_data_file;
+
+  std::vector<paddle::framework::LoDTensor> datasets;
+  size_t num_total_words = LoadData(&datasets, FLAGS_data_file);
+  LOG(INFO) << "Number of samples (seq_len<1024): " << datasets.size();
+  LOG(INFO) << "Total number of words: " << num_total_words;
+
+  const bool model_combined = false;
+  // 0. Call `paddle::framework::InitDevices()` initialize all the devices
+  // 1. Define place, executor, scope
+  auto place = paddle::platform::CPUPlace();
+  auto executor = paddle::framework::Executor(place);
+  std::unique_ptr<paddle::framework::Scope> scope(
+      new paddle::framework::Scope());
+
+  // 2. Initialize the inference_program and load parameters
+  std::unique_ptr<paddle::framework::ProgramDesc> inference_program;
+  inference_program =
+      InitProgram(&executor, scope.get(), FLAGS_model_path, model_combined);
+  if (FLAGS_use_mkldnn) {
+    EnableMKLDNN(inference_program);
+  }
+
+#ifdef PADDLE_WITH_MKLML
+  // only use 1 thread number per std::thread
+  omp_set_dynamic(0);
+  omp_set_num_threads(1);
+  mkl_set_num_threads(1);
+#endif
+
+  double start_ms = 0, stop_ms = 0;
+  if (FLAGS_num_threads > 1) {
+    std::vector<std::vector<const paddle::framework::LoDTensor*>> jobs;
+    SplitData(datasets, &jobs, FLAGS_num_threads);
+    std::vector<std::unique_ptr<std::thread>> threads;
+    start_ms = GetCurrentMs();
+    for (int i = 0; i < FLAGS_num_threads; ++i) {
+      threads.emplace_back(
+          new std::thread(ThreadRunInfer, i, &executor, scope.get(),
+                          std::ref(inference_program), std::ref(jobs)));
+    }
+    for (int i = 0; i < FLAGS_num_threads; ++i) {
+      threads[i]->join();
+    }
+    stop_ms = GetCurrentMs();
+  } else {
+    if (FLAGS_prepare_vars) {
+      executor.CreateVariables(*inference_program, scope.get(), 0);
+    }
+    // always prepare context
+    std::unique_ptr<paddle::framework::ExecutorPrepareContext> ctx;
+    ctx = executor.Prepare(*inference_program, 0);
+
+    // preapre fetch
+    const std::vector<std::string>& fetch_target_names =
+        inference_program->GetFetchTargetNames();
+    PADDLE_ENFORCE_EQ(fetch_target_names.size(), 1UL);
+    std::map<std::string, paddle::framework::LoDTensor*> fetch_targets;
+    paddle::framework::LoDTensor outtensor;
+    fetch_targets[fetch_target_names[0]] = &outtensor;
+
+    // prepare feed
+    const std::vector<std::string>& feed_target_names =
+        inference_program->GetFeedTargetNames();
+    PADDLE_ENFORCE_EQ(feed_target_names.size(), 1UL);
+    std::map<std::string, const paddle::framework::LoDTensor*> feed_targets;
+
+    // feed data and run
+    start_ms = GetCurrentMs();
+    for (size_t i = 0; i < datasets.size(); ++i) {
+      feed_targets[feed_target_names[0]] = &(datasets[i]);
+      executor.RunPreparedContext(ctx.get(), scope.get(), &feed_targets,
+                                  &fetch_targets, !FLAGS_prepare_vars);
+    }
+    stop_ms = GetCurrentMs();
+    LOG(INFO) << "Tid: 0, process " << datasets.size()
+              << " samples, avg time per sample: "
+              << (stop_ms - start_ms) / datasets.size() << " ms";
+  }
+  LOG(INFO) << "Total inference time with " << FLAGS_num_threads
+            << " threads : " << (stop_ms - start_ms) / 1000.0
+            << " sec, QPS: " << datasets.size() / ((stop_ms - start_ms) / 1000);
+}

paddle/fluid/operators/activation_mkldnn_op.cc

@@ -222,35 +222,35 @@ struct MKLDNNActivationGradFunc : public BaseActivationFunctor<T> {
 };
 
 template <typename T>
-using ReluMkldnnFunctor =
+using ReluMKLDNNFunctor =
     MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_relu>;
 
 template <typename T>
-using TanhMkldnnFunctor =
+using TanhMKLDNNFunctor =
     MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_tanh>;
 
 template <typename T>
-using SqrtMkldnnFunctor =
+using SqrtMKLDNNFunctor =
     MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_sqrt>;
 
 template <typename T>
-using AbsMkldnnFunctor =
+using AbsMKLDNNFunctor =
     MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_abs>;
 
 template <typename T>
-using ReluMkldnnGradFunctor =
+using ReluMKLDNNGradFunctor =
     MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_relu>;
 
 template <typename T>
-using TanhMkldnnGradFunctor =
+using TanhMKLDNNGradFunctor =
     MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_tanh>;
 
 template <typename T>
-using SqrtMkldnnGradFunctor =
+using SqrtMKLDNNGradFunctor =
     MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_sqrt>;
 
 template <typename T>
-using AbsMkldnnGradFunctor =
+using AbsMKLDNNGradFunctor =
     MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_abs>;
 }  // namespace operators
 }  // namespace paddle
@@ -265,9 +265,9 @@ namespace ops = paddle::operators;
       ops::MKLDNNActivationGradKernel<ops::grad_functor<float>>);
 
 #define FOR_EACH_MKLDNN_KERNEL_FUNCTOR(__macro)            \
-  __macro(relu, ReluMkldnnFunctor, ReluMkldnnGradFunctor); \
-  __macro(tanh, TanhMkldnnFunctor, TanhMkldnnGradFunctor); \
-  __macro(sqrt, SqrtMkldnnFunctor, SqrtMkldnnGradFunctor); \
-  __macro(abs, AbsMkldnnFunctor, AbsMkldnnGradFunctor);
+  __macro(relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor); \
+  __macro(tanh, TanhMKLDNNFunctor, TanhMKLDNNGradFunctor); \
+  __macro(sqrt, SqrtMKLDNNFunctor, SqrtMKLDNNGradFunctor); \
+  __macro(abs, AbsMKLDNNFunctor, AbsMKLDNNGradFunctor);
 
 FOR_EACH_MKLDNN_KERNEL_FUNCTOR(REGISTER_ACTIVATION_MKLDNN_KERNEL);

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/cudnn_helper.h"
 #include "paddle/fluid/platform/float16.h"
 
-DEFINE_bool(cudnn_algo_use_autotune, true,
+DEFINE_bool(cudnn_deterministic, true,
             "Whether allow using an autotuning algorithm for convolution "
             "operator. The autotuning algorithm may be non-deterministic. If "
             "false, the algorithm is deterministic.");
@@ -272,7 +272,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
     auto handle = dev_ctx.cudnn_handle();
     if (input_grad) {
-      if (FLAGS_cudnn_algo_use_autotune) {
+      if (FLAGS_cudnn_deterministic) {
         PADDLE_ENFORCE(
             platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
                 handle, cudnn_filter_desc,
@@ -297,7 +297,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     }
 
     if (filter_grad) {
-      if (FLAGS_cudnn_algo_use_autotune) {
+      if (FLAGS_cudnn_deterministic) {
         PADDLE_ENFORCE(
             platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
                 handle, cudnn_input_desc, cudnn_output_grad_desc,

paddle/fluid/operators/detail/grpc_client.cc

@@ -38,6 +38,25 @@ void RPCClient::Init() {
   if (rpc_client_.get() == nullptr) {
     rpc_client_.reset(new RPCClient());
   }
+  rpc_client_->InitEventLoop();
+}
+
+void RPCClient::InitEventLoop() {
+  // start the client process thread
+  // TODO(wuyi): can make this in a threadpool
+  client_thread_.reset(new std::thread(std::bind(&RPCClient::Proceed, this)));
+}
+
+RPCClient::~RPCClient() {
+  Wait();
+  cq_.Shutdown();
+  {
+    std::lock_guard<std::mutex> guard(chan_mutex_);
+    for (auto& it : channels_) {
+      it.second.reset();
+    }
+  }
+  client_thread_->join();
 }
 
 bool RPCClient::AsyncSendVariable(const std::string& ep,
@@ -204,70 +223,37 @@ void RPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) {
   req_count_++;
 }
 
-bool RPCClient::Wait() {
-  VLOG(3) << "RPCClient begin Wait()"
-          << " req_count_:" << req_count_;
-  if (req_count_ <= 0) {
-    return true;
-  }
-  const size_t kReqCnt = req_count_;
-  bool a[kReqCnt];
-  std::vector<std::future<void>> waits(req_count_);
-  std::mutex mu;
-
-  for (int i = 0; i < req_count_; i++) {
-    waits[i] = framework::AsyncIO([i, &a, &mu, this] {
-      bool ret = Proceed();
-      std::lock_guard<std::mutex> l(mu);
-      a[i] = ret;
-    });
-  }
-
-  for (int i = 0; i < req_count_; i++) {
-    waits[i].wait();
-  }
-
-  int last_req_count = req_count_;
-  req_count_ = 0;
-
-  for (int i = 0; i < last_req_count; i++) {
-    if (!a[i]) {
-      return false;
-    }
-  }
-
-  return true;
+void RPCClient::Wait() {
+  std::unique_lock<std::mutex> lk(sync_mutex_);
+  sync_cond_.wait(lk, [this] { return req_count_ == 0; });
 }
 
-bool RPCClient::Proceed() {
-  void* tag = NULL;
+void RPCClient::Proceed() {
+  void* tag = nullptr;
   bool ok = false;
 
-  // request counts.
-  if (!cq_.Next(&tag, &ok)) {
-    LOG(ERROR) << "Get meets CompletionQueue error";
-    return false;
-  }
-
-  GPR_ASSERT(ok);
-  PADDLE_ENFORCE(tag);
-
-  // TODO(gongwb): add more retries.
-  BaseProcessor* c = static_cast<BaseProcessor*>(tag);
-  if (!c->status_.ok()) {
-    LOG(ERROR) << "proc param error:" << c->var_h_.String()
-               << " grpc error:" << c->status_.error_message();
+  while (cq_.Next(&tag, &ok)) {
+    BaseProcessor* c = static_cast<BaseProcessor*>(tag);
+    GPR_ASSERT(ok);
+    PADDLE_ENFORCE(c);
+    if (c->status_.ok()) {
+      c->Process();
+    } else {
+      LOG(ERROR) << "var: " << c->var_h_.String()
+                 << " grpc error:" << c->status_.error_message();
+    }
     delete c;
-    return false;
+    {
+      std::lock_guard<std::mutex> lk(sync_mutex_);
+      req_count_--;
+    }
+    sync_cond_.notify_all();
   }
-
-  c->Process();
-  delete c;
-  return true;
 }
+
 std::shared_ptr<grpc::Channel> RPCClient::GetChannel(const std::string& ep) {
   // TODO(Yancey1989): make grpc client completely thread-safe
-  std::unique_lock<std::mutex> lock(mutex_);
+  std::lock_guard<std::mutex> guard(chan_mutex_);
   auto it = channels_.find(ep);
   if (it != channels_.end()) {
     return it->second;

paddle/fluid/operators/detail/grpc_client.h

@@ -16,15 +16,18 @@ limitations under the License. */
 
 #include <time.h>
 
-#include <chrono>  // NOLINT
+#include <chrono>              // NOLINT
+#include <condition_variable>  // NOLINT
 #include <ctime>
 #include <functional>
 #include <iostream>
 #include <map>
 #include <mutex>  // NOLINT
 #include <string>
+#include <thread>  // NOLINT
 #include <vector>
 
+#include "grpc++/channel.h"
 #include "grpc++/generic/generic_stub.h"
 #include "grpc++/grpc++.h"
 #include "grpc++/support/byte_buffer.h"
@@ -164,6 +167,7 @@ class FetchBarrierProcessor : public BaseProcessor {
 class RPCClient {
  public:
   RPCClient() {}
+  ~RPCClient();
 
   static RPCClient* GetInstance();
 
@@ -192,19 +196,28 @@ class RPCClient {
   void AsyncSendFetchBarrier(const std::string& ep,
                              int64_t time_out = 600 * 1000);
 
-  bool Wait();
+  void Wait();
+  // InitEventLoop should only be called by Init()
+  void InitEventLoop();
 
  private:
-  bool Proceed();
+  void Proceed();
   std::shared_ptr<grpc::Channel> GetChannel(const std::string& ep);
   // Init is called by GetInstance.
   static void Init();
 
  private:
   grpc::CompletionQueue cq_;
-  std::map<std::string, std::shared_ptr<grpc::Channel>> channels_;
+  std::unordered_map<std::string, std::shared_ptr<grpc::Channel>> channels_;
+  std::unique_ptr<std::thread> client_thread_;
+
+  // mutex for Wait client sync
+  std::mutex sync_mutex_;
+  std::condition_variable sync_cond_;
   std::atomic<int64_t> req_count_{0};
-  std::mutex mutex_;
+
+  // mutex for GetChannel thread safety
+  std::mutex chan_mutex_;
   static std::unique_ptr<RPCClient> rpc_client_;
   static std::once_flag init_flag_;
   DISABLE_COPY_AND_ASSIGN(RPCClient);

paddle/fluid/operators/detail/grpc_server.cc

@@ -68,9 +68,7 @@ class RequestSend final : public RequestBase {
         method_id, &ctx_, request_.get(), &responder_, cq_, cq_,
         reinterpret_cast<void*>(static_cast<intptr_t>(req_id)));
   }
-
   virtual ~RequestSend() {}
-
   std::string GetReqName() override { return request_->Varname(); }
 
   void Process() override {
@@ -82,7 +80,6 @@ class RequestSend final : public RequestBase {
     framework::Variable* outvar = nullptr;
 
     request_handler_->Handle(varname, scope, invar, &outvar);
-
     status_ = FINISH;
     responder_.Finish(reply_, ::grpc::Status::OK,
                       reinterpret_cast<void*>(static_cast<intptr_t>(req_id_)));
@@ -125,7 +122,6 @@ class RequestGet final : public RequestBase {
       SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(),
                             &reply_);
     }
-
     status_ = FINISH;
     responder_.Finish(reply_, ::grpc::Status::OK,
                       reinterpret_cast<void*>(static_cast<intptr_t>(req_id_)));
@@ -170,10 +166,9 @@ class RequestPrefetch final : public RequestBase {
 
     SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(),
                           &reply_);
-
-    status_ = FINISH;
     responder_.Finish(reply_, ::grpc::Status::OK,
                       reinterpret_cast<void*>(static_cast<intptr_t>(req_id_)));
+    status_ = FINISH;
   }
 
  protected:

paddle/fluid/operators/detail/grpc_server.h

@@ -71,8 +71,6 @@ class AsyncGRPCServer final : public RPCServer {
   std::unique_ptr<::grpc::Server> server_;
 
   // condition of the sub program
-  std::mutex barrier_mutex_;
-  mutable int barrier_cond_step_;
   std::condition_variable barrier_condition_;
 
   std::mutex mutex_ready_;

paddle/fluid/operators/detail/grpc_server_test.cc

@@ -113,10 +113,6 @@ void StartServer() {
   std::thread server_thread(
       std::bind(&detail::AsyncGRPCServer::StartServer, g_rpc_service.get()));
 
-  // FIXME(gongwb): don't use hard time.
-  sleep(10);
-  LOG(INFO) << "got nccl id and stop server...";
-  g_rpc_service->ShutDown();
   server_thread.join();
 }
 
@@ -127,7 +123,7 @@ TEST(PREFETCH, CPU) {
   std::thread server_thread(StartServer);
   g_rpc_service->WaitServerReady();
 
-  detail::RPCClient client;
+  detail::RPCClient* client = detail::RPCClient::GetInstance();
   int port = g_rpc_service->GetSelectedPort();
   std::string ep = paddle::string::Sprintf("127.0.0.1:%d", port);
 
@@ -141,8 +137,8 @@ TEST(PREFETCH, CPU) {
     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();
+    client->AsyncPrefetchVariable(ep, ctx, scope, in_var_name, out_var_name);
+    client->Wait();
     auto var = scope.Var(out_var_name);
     auto value = var->GetMutable<framework::SelectedRows>()->value();
     auto ptr = value.mutable_data<float>(place);
@@ -152,6 +148,7 @@ TEST(PREFETCH, CPU) {
     }
   }
 
+  g_rpc_service->ShutDown();
   server_thread.join();
   LOG(INFO) << "begin reset";
   g_rpc_service.reset(nullptr);

paddle/fluid/operators/detection/box_coder_op.cc

@@ -22,21 +22,21 @@ class BoxCoderOp : public framework::OperatorWithKernel {
   void InferShape(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasInput("PriorBox"),
                    "Input(PriorBox) of BoxCoderOp should not be null.");
-    PADDLE_ENFORCE(ctx->HasInput("PriorBoxVar"),
-                   "Input(PriorBoxVar) of BoxCoderOp should not be null.");
     PADDLE_ENFORCE(ctx->HasInput("TargetBox"),
                    "Input(TargetBox) of BoxCoderOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("OutputBox"),
                    "Output(OutputBox) of BoxCoderOp should not be null.");
 
     auto prior_box_dims = ctx->GetInputDim("PriorBox");
-    auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar");
     auto target_box_dims = ctx->GetInputDim("TargetBox");
 
     PADDLE_ENFORCE_EQ(prior_box_dims.size(), 2,
                       "The rank of Input of PriorBoxVar must be 2");
     PADDLE_ENFORCE_EQ(prior_box_dims[1], 4, "The shape of PriorBox is [N, 4]");
-    PADDLE_ENFORCE_EQ(prior_box_dims, prior_box_var_dims);
+    if (ctx->HasInput("PriorBoxVar")) {
+      auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar");
+      PADDLE_ENFORCE_EQ(prior_box_dims, prior_box_var_dims);
+    }
 
     auto code_type = GetBoxCodeType(ctx->Attrs().Get<std::string>("code_type"));
     if (code_type == BoxCodeType::kEncodeCenterSize) {
@@ -71,9 +71,11 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker {
         "of the coordinate system. [xmax, ymax] is the right bottom "
         "coordinate of the anchor box.");
     AddInput("PriorBoxVar",
-             "(Tensor, default Tensor<float>) "
+             "(Tensor, default Tensor<float>, optional) "
              "PriorBoxVar is a 2-D Tensor with shape [M, 4] holds M group "
-             "of variance.");
+             "of variance. PriorBoxVar will set all elements to 1 by "
+             "default.")
+        .AsDispensable();
     AddInput(
         "TargetBox",
         "(LoDTensor or Tensor) This input can be a 2-D LoDTensor with shape "
@@ -91,6 +93,10 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker {
                          "the code type used with the target box")
         .SetDefault("encode_center_size")
         .InEnum({"encode_center_size", "decode_center_size"});
+    AddAttr<bool>("box_normalized",
+                  "(bool, default true) "
+                  "whether treat the priorbox as a noramlized box")
+        .SetDefault(true);
     AddOutput("OutputBox",
               "(LoDTensor or Tensor) "
               "When code_type is 'encode_center_size', the output tensor of "
@@ -127,5 +133,6 @@ width and height.
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker,
                   paddle::framework::EmptyGradOpMaker);
-REGISTER_OP_CPU_KERNEL(box_coder, ops::BoxCoderKernel<float>,
-                       ops::BoxCoderKernel<double>);
+REGISTER_OP_CPU_KERNEL(
+    box_coder, ops::BoxCoderKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::BoxCoderKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/detection/box_coder_op.cu

@@ -20,15 +20,16 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data,
                                        const T* prior_box_var_data,
                                        const T* target_box_data, const int row,
                                        const int col, const int len,
-                                       T* output) {
+                                       const bool normalized, T* output) {
   const int idx = threadIdx.x + blockIdx.x * blockDim.x;
   if (idx < row * col) {
     const int row_idx = idx / col;
     const int col_idx = idx % col;
-    T prior_box_width =
-        prior_box_data[col_idx * len + 2] - prior_box_data[col_idx * len];
-    T prior_box_height =
-        prior_box_data[col_idx * len + 3] - prior_box_data[col_idx * len + 1];
+    T prior_box_width = prior_box_data[col_idx * len + 2] -
+                        prior_box_data[col_idx * len] + (normalized == false);
+    T prior_box_height = prior_box_data[col_idx * len + 3] -
+                         prior_box_data[col_idx * len + 1] +
+                         (normalized == false);
     T prior_box_center_x =
         (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2;
     T prior_box_center_y = (prior_box_data[col_idx * len + 3] +
@@ -41,20 +42,24 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data,
     T target_box_center_y = (target_box_data[row_idx * len + 3] +
                              target_box_data[row_idx * len + 1]) /
                             2;
-    T target_box_width =
-        target_box_data[row_idx * len + 2] - target_box_data[row_idx * len];
-    T target_box_height =
-        target_box_data[row_idx * len + 3] - target_box_data[row_idx * len + 1];
+    T target_box_width = target_box_data[row_idx * len + 2] -
+                         target_box_data[row_idx * len] + (normalized == false);
+    T target_box_height = target_box_data[row_idx * len + 3] -
+                          target_box_data[row_idx * len + 1] +
+                          (normalized == false);
 
-    output[idx * len] = (target_box_center_x - prior_box_center_x) /
-                        prior_box_width / prior_box_var_data[col_idx * len];
-    output[idx * len + 1] = (target_box_center_y - prior_box_center_y) /
-                            prior_box_height /
-                            prior_box_var_data[col_idx * len + 1];
-    output[idx * len + 2] = log(fabs(target_box_width / prior_box_width)) /
-                            prior_box_var_data[col_idx * len + 2];
-    output[idx * len + 3] = log(fabs(target_box_height / prior_box_height)) /
-                            prior_box_var_data[col_idx * len + 3];
+    output[idx * len] =
+        (target_box_center_x - prior_box_center_x) / prior_box_width;
+    output[idx * len + 1] =
+        (target_box_center_y - prior_box_center_y) / prior_box_height;
+    output[idx * len + 2] = log(fabs(target_box_width / prior_box_width));
+    output[idx * len + 3] = log(fabs(target_box_height / prior_box_height));
+    if (prior_box_var_data) {
+      output[idx * len] /= prior_box_var_data[col_idx * len];
+      output[idx * len + 1] /= prior_box_var_data[col_idx * len + 1];
+      output[idx * len + 2] /= prior_box_var_data[col_idx * len + 2];
+      output[idx * len + 3] /= prior_box_var_data[col_idx * len + 3];
+    }
   }
 }
 
@@ -63,42 +68,56 @@ __global__ void DecodeCenterSizeKernel(const T* prior_box_data,
                                        const T* prior_box_var_data,
                                        const T* target_box_data, const int row,
                                        const int col, const int len,
-                                       T* output) {
+                                       const bool normalized, T* output) {
   const int idx = threadIdx.x + blockIdx.x * blockDim.x;
   if (idx < row * col) {
     const int col_idx = idx % col;
-    T prior_box_width =
-        prior_box_data[col_idx * len + 2] - prior_box_data[col_idx * len];
-    T prior_box_height =
-        prior_box_data[col_idx * len + 3] - prior_box_data[col_idx * len + 1];
+    T prior_box_width = prior_box_data[col_idx * len + 2] -
+                        prior_box_data[col_idx * len] + (normalized == false);
+    T prior_box_height = prior_box_data[col_idx * len + 3] -
+                         prior_box_data[col_idx * len + 1] +
+                         (normalized == false);
     T prior_box_center_x =
         (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2;
     T prior_box_center_y = (prior_box_data[col_idx * len + 3] +
                             prior_box_data[col_idx * len + 1]) /
                            2;
-
-    T target_box_width = exp(prior_box_var_data[col_idx * len + 2] *
+    T target_box_width, target_box_height;
+    T target_box_center_x, target_box_center_y;
+    if (prior_box_var_data) {
+      target_box_width = exp(prior_box_var_data[col_idx * len + 2] *
                              target_box_data[idx * len + 2]) *
                          prior_box_width;
-    T target_box_height = exp(prior_box_var_data[col_idx * len + 3] *
+      target_box_height = exp(prior_box_var_data[col_idx * len + 3] *
                               target_box_data[idx * len + 3]) *
                           prior_box_height;
-    T target_box_center_x = prior_box_var_data[col_idx * len] *
+      target_box_center_x = prior_box_var_data[col_idx * len] *
                                 target_box_data[idx * len] * prior_box_width +
                             prior_box_center_x;
-    T target_box_center_y = prior_box_var_data[col_idx * len + 1] *
+      target_box_center_y = prior_box_var_data[col_idx * len + 1] *
                                 target_box_data[idx * len + 1] *
                                 prior_box_height +
                             prior_box_center_y;
+    } else {
+      target_box_width = exp(target_box_data[idx * len + 2]) * prior_box_width;
+      target_box_height =
+          exp(target_box_data[idx * len + 3]) * prior_box_height;
+      target_box_center_x =
+          target_box_data[idx * len] * prior_box_width + prior_box_center_x;
+      target_box_center_y = target_box_data[idx * len + 1] * prior_box_height +
+                            prior_box_center_y;
+    }
 
     output[idx * len] = target_box_center_x - target_box_width / 2;
     output[idx * len + 1] = target_box_center_y - target_box_height / 2;
-    output[idx * len + 2] = target_box_center_x + target_box_width / 2;
-    output[idx * len + 3] = target_box_center_y + target_box_height / 2;
+    output[idx * len + 2] =
+        target_box_center_x + target_box_width / 2 - (normalized == false);
+    output[idx * len + 3] =
+        target_box_center_y + target_box_height / 2 - (normalized == false);
   }
 }
 
-template <typename T>
+template <typename DeviceContext, typename T>
 class BoxCoderCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -109,6 +128,11 @@ class BoxCoderCUDAKernel : public framework::OpKernel<T> {
     auto* target_box = context.Input<framework::LoDTensor>("TargetBox");
     auto* output_box = context.Output<framework::Tensor>("OutputBox");
 
+    const T* prior_box_data = prior_box->data<T>();
+    const T* target_box_data = target_box->data<T>();
+    const T* prior_box_var_data = nullptr;
+    if (prior_box_var) prior_box_var_data = prior_box_var->data<T>();
+
     if (target_box->lod().size()) {
       PADDLE_ENFORCE_EQ(target_box->lod().size(), 1,
                         "Only support 1 level of LoD.");
@@ -120,22 +144,19 @@ class BoxCoderCUDAKernel : public framework::OpKernel<T> {
     int grid = (row * col + block - 1) / block;
     auto& device_ctx = context.cuda_device_context();
 
-    const T* prior_box_data = prior_box->data<T>();
-    const T* prior_box_var_data = prior_box_var->data<T>();
-    const T* target_box_data = target_box->data<T>();
-
     output_box->mutable_data<T>({row, col, len}, context.GetPlace());
     T* output = output_box->data<T>();
 
     auto code_type = GetBoxCodeType(context.Attr<std::string>("code_type"));
+    bool normalized = context.Attr<bool>("box_normalized");
     if (code_type == BoxCodeType::kEncodeCenterSize) {
       EncodeCenterSizeKernel<T><<<grid, block, 0, device_ctx.stream()>>>(
           prior_box_data, prior_box_var_data, target_box_data, row, col, len,
-          output);
+          normalized, output);
     } else if (code_type == BoxCodeType::kDecodeCenterSize) {
       DecodeCenterSizeKernel<T><<<grid, block, 0, device_ctx.stream()>>>(
           prior_box_data, prior_box_var_data, target_box_data, row, col, len,
-          output);
+          normalized, output);
     }
   }
 };
@@ -144,5 +165,7 @@ class BoxCoderCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP_CUDA_KERNEL(box_coder, ops::BoxCoderCUDAKernel<float>,
-                        ops::BoxCoderCUDAKernel<double>);
+REGISTER_OP_CUDA_KERNEL(
+    box_coder,
+    ops::BoxCoderCUDAKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::BoxCoderCUDAKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/operators/detection/box_coder_op.h

@@ -28,26 +28,28 @@ inline BoxCodeType GetBoxCodeType(const std::string& type) {
   PADDLE_THROW("Not support type %s.", type);
 }
 
-template <typename T>
+template <typename DeviceContext, typename T>
 class BoxCoderKernel : public framework::OpKernel<T> {
  public:
-  void EncodeCenterSize(const framework::Tensor& target_box,
-                        const framework::Tensor& prior_box,
-                        const framework::Tensor& prior_box_var,
-                        T* output) const {
-    int64_t row = target_box.dims()[0];
-    int64_t col = prior_box.dims()[0];
-    int64_t len = prior_box.dims()[1];
-    auto* target_box_data = target_box.data<T>();
-    auto* prior_box_data = prior_box.data<T>();
-    auto* prior_box_var_data = prior_box_var.data<T>();
+  void EncodeCenterSize(const framework::Tensor* target_box,
+                        const framework::Tensor* prior_box,
+                        const framework::Tensor* prior_box_var,
+                        const bool normalized, T* output) const {
+    int64_t row = target_box->dims()[0];
+    int64_t col = prior_box->dims()[0];
+    int64_t len = prior_box->dims()[1];
+    auto* target_box_data = target_box->data<T>();
+    auto* prior_box_data = prior_box->data<T>();
+    const T* prior_box_var_data = nullptr;
+    if (prior_box_var) prior_box_var_data = prior_box_var->data<T>();
 
     for (int64_t i = 0; i < row; ++i) {
       for (int64_t j = 0; j < col; ++j) {
-        T prior_box_width =
-            prior_box_data[j * len + 2] - prior_box_data[j * len];
-        T prior_box_height =
-            prior_box_data[j * len + 3] - prior_box_data[j * len + 1];
+        T prior_box_width = prior_box_data[j * len + 2] -
+                            prior_box_data[j * len] + (normalized == false);
+        T prior_box_height = prior_box_data[j * len + 3] -
+                             prior_box_data[j * len + 1] +
+                             (normalized == false);
         T prior_box_center_x =
             (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
         T prior_box_center_y =
@@ -57,67 +59,89 @@ class BoxCoderKernel : public framework::OpKernel<T> {
             (target_box_data[i * len + 2] + target_box_data[i * len]) / 2;
         T target_box_center_y =
             (target_box_data[i * len + 3] + target_box_data[i * len + 1]) / 2;
-        T target_box_width =
-            target_box_data[i * len + 2] - target_box_data[i * len];
-        T target_box_height =
-            target_box_data[i * len + 3] - target_box_data[i * len + 1];
+        T target_box_width = target_box_data[i * len + 2] -
+                             target_box_data[i * len] + (normalized == false);
+        T target_box_height = target_box_data[i * len + 3] -
+                              target_box_data[i * len + 1] +
+                              (normalized == false);
 
         size_t offset = i * col * len + j * len;
-        output[offset] = (target_box_center_x - prior_box_center_x) /
-                         prior_box_width / prior_box_var_data[j * len];
-        output[offset + 1] = (target_box_center_y - prior_box_center_y) /
-                             prior_box_height / prior_box_var_data[j * len + 1];
+        output[offset] =
+            (target_box_center_x - prior_box_center_x) / prior_box_width;
+        output[offset + 1] =
+            (target_box_center_y - prior_box_center_y) / prior_box_height;
         output[offset + 2] =
-            std::log(std::fabs(target_box_width / prior_box_width)) /
-            prior_box_var_data[j * len + 2];
+            std::log(std::fabs(target_box_width / prior_box_width));
         output[offset + 3] =
-            std::log(std::fabs(target_box_height / prior_box_height)) /
-            prior_box_var_data[j * len + 3];
+            std::log(std::fabs(target_box_height / prior_box_height));
+        if (prior_box_var) {
+          output[offset] /= prior_box_var_data[j * len];
+          output[offset + 1] /= prior_box_var_data[j * len + 1];
+          output[offset + 2] /= prior_box_var_data[j * len + 2];
+          output[offset + 3] /= prior_box_var_data[j * len + 3];
+        }
       }
     }
   }
-  void DecodeCenterSize(const framework::Tensor& target_box,
-                        const framework::Tensor& prior_box,
-                        const framework::Tensor& prior_box_var,
-                        T* output) const {
-    int64_t row = target_box.dims()[0];
-    int64_t col = prior_box.dims()[0];
-    int64_t len = prior_box.dims()[1];
-
-    auto* target_box_data = target_box.data<T>();
-    auto* prior_box_data = prior_box.data<T>();
-    auto* prior_box_var_data = prior_box_var.data<T>();
+  void DecodeCenterSize(const framework::Tensor* target_box,
+                        const framework::Tensor* prior_box,
+                        const framework::Tensor* prior_box_var,
+                        const bool normalized, T* output) const {
+    int64_t row = target_box->dims()[0];
+    int64_t col = prior_box->dims()[0];
+    int64_t len = prior_box->dims()[1];
+
+    auto* target_box_data = target_box->data<T>();
+    auto* prior_box_data = prior_box->data<T>();
+    const T* prior_box_var_data = nullptr;
+    if (prior_box_var) prior_box_var_data = prior_box_var->data<T>();
 
     for (int64_t i = 0; i < row; ++i) {
       for (int64_t j = 0; j < col; ++j) {
         size_t offset = i * col * len + j * len;
-        T prior_box_width =
-            prior_box_data[j * len + 2] - prior_box_data[j * len];
-        T prior_box_height =
-            prior_box_data[j * len + 3] - prior_box_data[j * len + 1];
+        T prior_box_width = prior_box_data[j * len + 2] -
+                            prior_box_data[j * len] + (normalized == false);
+        T prior_box_height = prior_box_data[j * len + 3] -
+                             prior_box_data[j * len + 1] +
+                             (normalized == false);
         T prior_box_center_x =
             (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
         T prior_box_center_y =
             (prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2;
 
-        T target_box_center_x = prior_box_var_data[j * len] *
+        T target_box_center_x = 0, target_box_center_y = 0;
+        T target_box_width = 0, target_box_height = 0;
+        if (prior_box_var) {
+          target_box_center_x = prior_box_var_data[j * len] *
                                     target_box_data[offset] * prior_box_width +
                                 prior_box_center_x;
-        T target_box_center_y = prior_box_var_data[j * len + 1] *
+          target_box_center_y = prior_box_var_data[j * len + 1] *
                                     target_box_data[offset + 1] *
                                     prior_box_height +
                                 prior_box_center_y;
-        T target_box_width = std::exp(prior_box_var_data[j * len + 2] *
+          target_box_width = std::exp(prior_box_var_data[j * len + 2] *
                                       target_box_data[offset + 2]) *
                              prior_box_width;
-        T target_box_height = std::exp(prior_box_var_data[j * len + 3] *
+          target_box_height = std::exp(prior_box_var_data[j * len + 3] *
                                        target_box_data[offset + 3]) *
                               prior_box_height;
+        } else {
+          target_box_center_x =
+              target_box_data[offset] * prior_box_width + prior_box_center_x;
+          target_box_center_y = target_box_data[offset + 1] * prior_box_height +
+                                prior_box_center_y;
+          target_box_width =
+              std::exp(target_box_data[offset + 2]) * prior_box_width;
+          target_box_height =
+              std::exp(target_box_data[offset + 3]) * prior_box_height;
+        }
 
         output[offset] = target_box_center_x - target_box_width / 2;
         output[offset + 1] = target_box_center_y - target_box_height / 2;
-        output[offset + 2] = target_box_center_x + target_box_width / 2;
-        output[offset + 3] = target_box_center_y + target_box_height / 2;
+        output[offset + 2] =
+            target_box_center_x + target_box_width / 2 - (normalized == false);
+        output[offset + 3] =
+            target_box_center_y + target_box_height / 2 - (normalized == false);
       }
     }
   }
@@ -139,11 +163,14 @@ class BoxCoderKernel : public framework::OpKernel<T> {
     output_box->mutable_data<T>({row, col, len}, context.GetPlace());
 
     auto code_type = GetBoxCodeType(context.Attr<std::string>("code_type"));
+    bool normalized = context.Attr<bool>("box_normalized");
     T* output = output_box->data<T>();
     if (code_type == BoxCodeType::kEncodeCenterSize) {
-      EncodeCenterSize(*target_box, *prior_box, *prior_box_var, output);
+      EncodeCenterSize(target_box, prior_box, prior_box_var, normalized,
+                       output);
     } else if (code_type == BoxCodeType::kDecodeCenterSize) {
-      DecodeCenterSize(*target_box, *prior_box, *prior_box_var, output);
+      DecodeCenterSize(target_box, prior_box, prior_box_var, normalized,
+                       output);
     }
   }
 };

paddle/fluid/operators/fetch_barrier_op.cc

@@ -45,13 +45,13 @@ class FetchBarrierOp : public framework::OperatorBase {
 
     auto rpc_client = detail::RPCClient::GetInstance();
 
-    PADDLE_ENFORCE(rpc_client->Wait());
+    rpc_client->Wait();
 
     for (auto& ep : eps) {
       VLOG(3) << "fetch barrier, ep: " << ep;
       rpc_client->AsyncSendFetchBarrier(ep);
     }
-    PADDLE_ENFORCE(rpc_client->Wait());
+    rpc_client->Wait();
   }
 };
 

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);

paddle/fluid/operators/listen_and_serv_op.cc

@@ -222,8 +222,8 @@ static void FillRequestCtx(detail::RequestHandler *h, framework::Scope *scope,
   h->SetDevCtx(dev_ctx);
   h->SetExecutor(executor);
   h->SetProgram(program);
-  h->SetPrefetchPreparedCtx(std::move(
-      std::unique_ptr<framework::ExecutorPrepareContext>(prefetch_ctx)));
+  h->SetPrefetchPreparedCtx(
+      std::unique_ptr<framework::ExecutorPrepareContext>(prefetch_ctx));
   h->SetRPCServer(rpc_server);
 }
 

paddle/fluid/operators/pool_cudnn_op.cu.cc

@@ -135,7 +135,11 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel<T> {
 
     PoolingMode pooling_mode;
     if (pooling_type == "max") {
-      pooling_mode = PoolingMode::kMaximum;
+      if (FLAGS_cudnn_deterministic) {
+        pooling_mode = PoolingMode::kMaximumDeterministic;
+      } else {
+        pooling_mode = PoolingMode::kMaximum;
+      }
     } else {
       pooling_mode = PoolingMode::kAverage;
     }

paddle/fluid/operators/prefetch_op.cc

@@ -53,7 +53,7 @@ class PrefetchOp : public framework::OperatorBase {
         VLOG(3) << "don't send no-initialied variable: " << ins[i];
       }
     }
-    PADDLE_ENFORCE(rpc_client->Wait());
+    rpc_client->Wait();
   }
 };
 

paddle/fluid/operators/reader/open_files_op.cc

@@ -26,7 +26,11 @@ class MultiFileReader : public framework::ReaderBase {
   MultiFileReader(const std::vector<std::string>& file_names,
                   const std::vector<framework::DDim>& dims, size_t thread_num,
                   size_t buffer_size)
-      : file_names_(file_names), dims_(dims), buffer_size_(buffer_size) {
+      : buffer_size_(buffer_size) {
+    readers_.reserve(file_names.size());
+    for (const std::string& f_name : file_names) {
+      readers_.emplace_back(CreateReaderByFileName(f_name, dims));
+    }
     prefetchers_.resize(thread_num);
     StartNewScheduler();
   }
@@ -40,14 +44,13 @@ class MultiFileReader : public framework::ReaderBase {
   void StartNewScheduler();
   void EndScheduler();
   void ScheduleThreadFunc();
-  void PrefetchThreadFunc(std::string file_name, size_t thread_idx);
+  void PrefetchThreadFunc(size_t reader_idx, size_t thread_idx);
 
-  std::vector<std::string> file_names_;
-  std::vector<framework::DDim> dims_;
+  std::vector<std::unique_ptr<framework::ReaderBase>> readers_;
   std::thread scheduler_;
   std::vector<std::thread> prefetchers_;
   size_t buffer_size_;
-  reader::BlockingQueue<size_t>* waiting_file_idx_;
+  reader::BlockingQueue<size_t>* waiting_reader_idx_;
   reader::BlockingQueue<size_t>* available_thread_idx_;
   reader::BlockingQueue<std::vector<framework::LoDTensor>>* buffer_;
 };
@@ -65,15 +68,15 @@ void MultiFileReader::ReInit() {
 
 void MultiFileReader::StartNewScheduler() {
   size_t thread_num = prefetchers_.size();
-  waiting_file_idx_ = new reader::BlockingQueue<size_t>(file_names_.size());
+  waiting_reader_idx_ = new reader::BlockingQueue<size_t>(readers_.size());
   available_thread_idx_ = new reader::BlockingQueue<size_t>(thread_num);
   buffer_ = new reader::BlockingQueue<std::vector<framework::LoDTensor>>(
       buffer_size_);
 
-  for (size_t i = 0; i < file_names_.size(); ++i) {
-    waiting_file_idx_->Send(i);
+  for (size_t i = 0; i < readers_.size(); ++i) {
+    waiting_reader_idx_->Send(i);
   }
-  waiting_file_idx_->Close();
+  waiting_reader_idx_->Close();
   for (size_t i = 0; i < thread_num; ++i) {
     available_thread_idx_->Send(i);
   }
@@ -84,13 +87,13 @@ void MultiFileReader::StartNewScheduler() {
 void MultiFileReader::EndScheduler() {
   available_thread_idx_->Close();
   buffer_->Close();
-  waiting_file_idx_->Close();
+  waiting_reader_idx_->Close();
   if (scheduler_.joinable()) {
     scheduler_.join();
   }
   delete buffer_;
   delete available_thread_idx_;
-  delete waiting_file_idx_;
+  delete waiting_reader_idx_;
 }
 
 void MultiFileReader::ScheduleThreadFunc() {
@@ -102,12 +105,11 @@ void MultiFileReader::ScheduleThreadFunc() {
     if (prefetcher.joinable()) {
       prefetcher.join();
     }
-    size_t file_idx;
-    if (waiting_file_idx_->Receive(&file_idx)) {
+    size_t reader_idx;
+    if (waiting_reader_idx_->Receive(&reader_idx)) {
       // Still have files to read. Start a new prefetch thread.
-      std::string file_name = file_names_[file_idx];
-      prefetcher = std::thread([this, file_name, thread_idx] {
-        PrefetchThreadFunc(file_name, thread_idx);
+      prefetcher = std::thread([this, reader_idx, thread_idx] {
+        PrefetchThreadFunc(reader_idx, thread_idx);
       });
     } else {
       // No more file to read.
@@ -129,23 +131,22 @@ void MultiFileReader::ScheduleThreadFunc() {
   VLOG(5) << "MultiFileReader schedule thread terminates.";
 }
 
-void MultiFileReader::PrefetchThreadFunc(std::string file_name,
-                                         size_t thread_idx) {
-  VLOG(5) << "The prefetch thread of file '" << file_name << "' starts.";
-  std::unique_ptr<framework::ReaderBase> reader =
-      CreateReaderByFileName(file_name, dims_);
+void MultiFileReader::PrefetchThreadFunc(size_t reader_idx, size_t thread_idx) {
+  VLOG(5) << "The prefetch thread of file idx '" << reader_idx << "' starts.";
+  std::unique_ptr<framework::ReaderBase>& reader = readers_[reader_idx];
   while (true) {
     std::vector<framework::LoDTensor> ins;
     reader->ReadNext(&ins);
     if (ins.empty()) {
+      reader->ReInit();
       break;
     }
     try {
       buffer_->Send(std::move(ins));
     } catch (paddle::platform::EnforceNotMet e) {
       VLOG(5) << "WARNING: The buffer channel has been closed. The prefetch "
-                 "thread of file '"
-              << file_name << "' will terminate.";
+                 "thread of file idx '"
+              << reader_idx << "' will terminate.";
       break;
     }
   }
@@ -154,7 +155,8 @@ void MultiFileReader::PrefetchThreadFunc(std::string file_name,
     VLOG(5) << "WARNING: The available_thread_idx_ channel has been closed. "
                "Fail to send thread_idx.";
   }
-  VLOG(5) << "The prefetch thread of file '" << file_name << "' terminates.";
+  VLOG(5) << "The prefetch thread of file idx '" << reader_idx
+          << "' terminates.";
 }
 
 class OpenFilesOp : public framework::OperatorBase {

paddle/fluid/operators/recv_op.cc

@@ -51,7 +51,7 @@ class RecvOp : public framework::OperatorBase {
       rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]);
     }
     if (sync_mode) {
-      PADDLE_ENFORCE(rpc_client->Wait());
+      rpc_client->Wait();
     }
   }
 };

paddle/fluid/operators/reduce_op.h

@@ -135,15 +135,16 @@ class ReduceKernel : public framework::OpKernel<T> {
     } else {
       int ndim = context.Input<Tensor>("X")->dims().size();
       int rdim = context.Attr<std::vector<int>>("dim").size();
-      HANDLE_DIM(6, 5);
-      HANDLE_DIM(6, 4);
-      HANDLE_DIM(6, 3);
-      HANDLE_DIM(6, 2);
-      HANDLE_DIM(6, 1);
-      HANDLE_DIM(5, 4);
-      HANDLE_DIM(5, 3);
-      HANDLE_DIM(5, 2);
-      HANDLE_DIM(5, 1);
+      // comments for accelerating compiling temporarily.
+      //      HANDLE_DIM(6, 5);
+      //      HANDLE_DIM(6, 4);
+      //      HANDLE_DIM(6, 3);
+      //      HANDLE_DIM(6, 2);
+      //      HANDLE_DIM(6, 1);
+      //      HANDLE_DIM(5, 4);
+      //      HANDLE_DIM(5, 3);
+      //      HANDLE_DIM(5, 2);
+      //      HANDLE_DIM(5, 1);
       HANDLE_DIM(4, 3);
       HANDLE_DIM(4, 2);
       HANDLE_DIM(4, 1);

paddle/fluid/operators/send_barrier_op.cc

@@ -49,13 +49,13 @@ class SendBarrierOp : public framework::OperatorBase {
     VLOG(3) << "SendBarrierOp sync_mode:" << sync_mode;
 
     // need to wait before sending send_barrier message
-    PADDLE_ENFORCE(rpc_client->Wait());
+    rpc_client->Wait();
     if (sync_mode) {
       for (auto& ep : eps) {
         VLOG(3) << "send barrier, ep: " << ep;
         rpc_client->AsyncSendBatchBarrier(ep);
       }
-      PADDLE_ENFORCE(rpc_client->Wait());
+      rpc_client->Wait();
     }
   }
 };

paddle/fluid/operators/send_op.cc

@@ -59,14 +59,14 @@ class SendOp : public framework::OperatorBase {
         VLOG(3) << "don't send no-initialied variable: " << ins[i];
       }
     }
-    PADDLE_ENFORCE(rpc_client->Wait());
+    rpc_client->Wait();
 
     if (sync_mode) {
       for (auto& ep : endpoints) {
         VLOG(3) << "batch barrier, ep: " << ep;
         rpc_client->AsyncSendBatchBarrier(ep);
       }
-      PADDLE_ENFORCE(rpc_client->Wait());
+      rpc_client->Wait();
     }
 
     if (outs.size() > 0) {
@@ -74,13 +74,13 @@ class SendOp : public framework::OperatorBase {
         VLOG(2) << "getting " << outs[i] << " from " << epmap[i];
         rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]);
       }
-      PADDLE_ENFORCE(rpc_client->Wait());
+      rpc_client->Wait();
       // tell pservers that current trainer have called fetch
       for (auto& ep : endpoints) {
         VLOG(2) << "send fetch barrier, ep: " << ep;
         rpc_client->AsyncSendFetchBarrier(ep);
       }
-      PADDLE_ENFORCE(rpc_client->Wait());
+      rpc_client->Wait();
     }
   }
 };

paddle/fluid/operators/sgd_op.h

@@ -114,7 +114,7 @@ class SGDOpKernel : public framework::OpKernel<T> {
         int64_t id_index = param.Index(grad.rows()[i]);
         PADDLE_ENFORCE_GE(id_index, static_cast<int64_t>(0),
                           "id should be in the table");
-        for (size_t j = 0; j < grad_row_width; j++) {
+        for (int64_t j = 0; j < grad_row_width; j++) {
           out_data[id_index * grad_row_width + j] -=
               lr[0] * grad_data[i * grad_row_width + j];
         }

paddle/fluid/operators/shape_op.cc

+/* 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<int>, ops::ShapeKernel<int64_t>,
+                       ops::ShapeKernel<float>, ops::ShapeKernel<double>);

paddle/fluid/operators/shape_op.cu

+/* 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<int>,
+                        paddle::operators::ShapeKernel<int64_t>,
+                        paddle::operators::ShapeKernel<float>,
+                        paddle::operators::ShapeKernel<double>);

paddle/fluid/operators/shape_op.h

+/* 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 <algorithm>
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+class ShapeKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in_t = ctx.Input<Tensor>("Input");
+    auto* out_t = ctx.Output<Tensor>("Out");
+    auto out_data = out_t->mutable_data<int64_t>(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

paddle/fluid/operators/tensorrt_engine_op.cc

@@ -31,8 +31,9 @@ void paddle::operators::TensorRTEngineKernel<DeviceContext, T>::Prepare(
   auto max_workspace = context.Attr<int>("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<inference::tensorrt::OpConverter>::Global().ConvertBlock(
-      block, engine_.get());
+      block, {}, context.scope(), engine_.get());
   engine_->FreezeNetwork();
 }
 

paddle/fluid/operators/test_send_nccl_id.cc

@@ -61,7 +61,6 @@ void StartServer() {
       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...";
@@ -88,12 +87,12 @@ TEST(SendNcclId, GrpcServer) {
   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();
+  detail::RPCClient* client = detail::RPCClient::GetInstance();
+  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();
   g_rpc_service.reset(nullptr);

paddle/fluid/platform/cudnn_helper.h

@@ -22,6 +22,8 @@ limitations under the License. */
 #include "paddle/fluid/platform/float16.h"
 #include "paddle/fluid/platform/macros.h"
 
+DECLARE_bool(cudnn_deterministic);
+
 namespace paddle {
 namespace platform {
 
@@ -76,8 +78,22 @@ enum class DataLayout {  // Not use
 enum class PoolingMode {
   kMaximum,
   kAverage,
+  kMaximumDeterministic,
 };
 
+inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) {
+  switch (mode) {
+    case PoolingMode::kMaximumDeterministic:
+      return CUDNN_POOLING_MAX_DETERMINISTIC;
+    case PoolingMode::kAverage:
+      return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING;
+    case PoolingMode::kMaximum:
+      return CUDNN_POOLING_MAX;
+    default:
+      PADDLE_THROW("Unexpected pooling mode.");
+  }
+}
+
 template <typename T>
 class CudnnDataType;
 
@@ -293,9 +309,7 @@ class ScopedPoolingDescriptor {
     PADDLE_ENFORCE_EQ(kernel.size(), pads.size());
     PADDLE_ENFORCE_EQ(kernel.size(), strides.size());
     PADDLE_ENFORCE(dynload::cudnnSetPoolingNdDescriptor(
-        desc_, (mode == PoolingMode::kMaximum
-                    ? CUDNN_POOLING_MAX
-                    : CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING),
+        desc_, (GetPoolingMode(mode)),
         CUDNN_PROPAGATE_NAN,  // Always propagate nans.
         kernel.size(), kernel.data(), pads.data(), strides.data()));
     return desc_;

paddle/fluid/platform/device_context.cc

@@ -175,7 +175,6 @@ CUDADeviceContext::~CUDADeviceContext() {
 Place CUDADeviceContext::GetPlace() const { return place_; }
 
 void CUDADeviceContext::Wait() const {
-  std::lock_guard<std::recursive_mutex> guard(mutex_);
   PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
   PADDLE_ENFORCE(cudaGetLastError());
 }

paddle/fluid/platform/device_context.h

@@ -100,7 +100,6 @@ class CUDADeviceContext : public DeviceContext {
 
   template <typename Callback>
   void RecordEvent(cudaEvent_t ev, Callback callback) {
-    std::lock_guard<std::recursive_mutex> guard(mutex_);
     callback();
     PADDLE_ENFORCE(cudaEventRecord(ev, stream_));
   }
@@ -110,8 +109,6 @@ class CUDADeviceContext : public DeviceContext {
 
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
   std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
-
-  mutable std::recursive_mutex mutex_;
   cudaStream_t stream_;
   cudnnHandle_t cudnn_handle_;
   cublasHandle_t cublas_handle_;

paddle/fluid/platform/dynload/cublas.h

@@ -45,7 +45,7 @@ extern void *cublas_dso_handle;
       std::call_once(cublas_dso_flag, []() {                                 \
         cublas_dso_handle = paddle::platform::dynload::GetCublasDsoHandle(); \
       });                                                                    \
-      void *p_##__name = dlsym(cublas_dso_handle, #__name);                  \
+      static void *p_##__name = dlsym(cublas_dso_handle, #__name);           \
       return reinterpret_cast<FUNC_TYPE>(p_##__name)(args...);               \
     }                                                                        \
   };                                                                         \

paddle/fluid/platform/dynload/cudnn.h

@@ -39,7 +39,7 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
         cudnn_dso_handle = paddle::platform::dynload::GetCUDNNDsoHandle(); \
       });                                                                  \
       EnforceCUDNNLoaded(#__name);                                         \
-      void* p_##__name = dlsym(cudnn_dso_handle, #__name);                 \
+      static void* p_##__name = dlsym(cudnn_dso_handle, #__name);          \
       return reinterpret_cast<cudnn_func>(p_##__name)(args...);            \
     }                                                                      \
   };                                                                       \

paddle/fluid/platform/dynload/cupti.h

@@ -45,7 +45,7 @@ extern void *cupti_dso_handle;
       std::call_once(cupti_dso_flag, []() {                                \
         cupti_dso_handle = paddle::platform::dynload::GetCUPTIDsoHandle(); \
       });                                                                  \
-      void *p_##__name = dlsym(cupti_dso_handle, #__name);                 \
+      static void *p_##__name = dlsym(cupti_dso_handle, #__name);          \
       return reinterpret_cast<cuptiFunc>(p_##__name)(args...);             \
     }                                                                      \
   };                                                                       \

paddle/fluid/platform/dynload/curand.h

@@ -34,7 +34,7 @@ extern void *curand_dso_handle;
       std::call_once(curand_dso_flag, []() {                                 \
         curand_dso_handle = paddle::platform::dynload::GetCurandDsoHandle(); \
       });                                                                    \
-      void *p_##__name = dlsym(curand_dso_handle, #__name);                  \
+      static void *p_##__name = dlsym(curand_dso_handle, #__name);           \
       return reinterpret_cast<curandFunc>(p_##__name)(args...);              \
     }                                                                        \
   };                                                                         \

paddle/fluid/platform/dynload/nccl.h

@@ -37,7 +37,7 @@ extern void* nccl_dso_handle;
       std::call_once(nccl_dso_flag, []() {                               \
         nccl_dso_handle = paddle::platform::dynload::GetNCCLDsoHandle(); \
       });                                                                \
-      void* p_##__name = dlsym(nccl_dso_handle, #__name);                \
+      static void* p_##__name = dlsym(nccl_dso_handle, #__name);         \
       return reinterpret_cast<nccl_func>(p_##__name)(args...);           \
     }                                                                    \
   };                                                                     \

paddle/fluid/platform/dynload/tensorrt.h

@@ -40,7 +40,7 @@ extern void* tensorrt_dso_handle;
             paddle::platform::dynload::GetTensorRtDsoHandle();          \
         PADDLE_ENFORCE(tensorrt_dso_handle, "load tensorrt so failed"); \
       });                                                               \
-      void* p_##__name = dlsym(tensorrt_dso_handle, #__name);           \
+      static void* p_##__name = dlsym(tensorrt_dso_handle, #__name);    \
       PADDLE_ENFORCE(p_##__name, "load %s failed", #__name);            \
       return reinterpret_cast<tensorrt_func>(p_##__name)(args...);      \
     }                                                                   \

paddle/fluid/platform/dynload/warpctc.h

@@ -40,7 +40,7 @@ extern void* warpctc_dso_handle;
       std::call_once(warpctc_dso_flag, []() {                                  \
         warpctc_dso_handle = paddle::platform::dynload::GetWarpCTCDsoHandle(); \
       });                                                                      \
-      void* p_##_name = dlsym(warpctc_dso_handle, #__name);                    \
+      static void* p_##_name = dlsym(warpctc_dso_handle, #__name);             \
       return reinterpret_cast<warpctcFunc>(p_##_name)(args...);                \
     }                                                                          \
   };                                                                           \

paddle/fluid/platform/profiler.cc

@@ -127,6 +127,7 @@ double Event::CpuElapsedMs(const Event& e) const {
 
 double Event::CudaElapsedMs(const Event& e) const {
 #ifdef PADDLE_WITH_CUDA
+  if (!has_cuda_) return 0.0;
   PADDLE_ENFORCE(e.has_cuda() && has_cuda());
   PADDLE_ENFORCE(e.device() == device());
   PADDLE_ENFORCE(cudaEventSynchronize(event_));

paddle/fluid/pybind/pybind.cc

@@ -519,6 +519,14 @@ All parameter, weight, gradient are variables in Paddle.
           [](const ExecutionStrategy &self) { return self.allow_op_delay_; },
           [](ExecutionStrategy &self, bool allow_op_delay) {
             self.allow_op_delay_ = allow_op_delay;
+          })
+      .def_property(
+          "num_iteration_per_drop_scope",
+          [](const ExecutionStrategy &self) {
+            return self.num_iteration_per_drop_scope_;
+          },
+          [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) {
+            self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope;
           });
   py::class_<BuildStrategy> build_strategy(pe, "BuildStrategy");