Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into cross_channel_norm

benchmark/paddle/image/alexnet.py

@@ -6,8 +6,18 @@ height = 227
 width = 227
 num_class = 1000
 batch_size = get_config_arg('batch_size', int, 128)
+gp = get_config_arg('layer_num', int, 1)
+is_infer = get_config_arg("is_infer", bool, False)
+num_samples = get_config_arg('num_samples', int, 2560)
 
-args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
+args = {
+    'height': height,
+    'width': width,
+    'color': True,
+    'num_class': num_class,
+    'is_infer': is_infer,
+    'num_samples': num_samples
+}
 define_py_data_sources2(
     "train.list", None, module="provider", obj="process", args=args)
 
@@ -31,7 +41,7 @@ net = img_pool_layer(input=net, pool_size=3, stride=2)
 
 # conv2
 net = img_conv_layer(
-    input=net, filter_size=5, num_filters=256, stride=1, padding=2, groups=1)
+    input=net, filter_size=5, num_filters=256, stride=1, padding=2, groups=gp)
 net = img_cmrnorm_layer(input=net, size=5, scale=0.0001, power=0.75)
 net = img_pool_layer(input=net, pool_size=3, stride=2)
 
@@ -40,11 +50,11 @@ net = img_conv_layer(
     input=net, filter_size=3, num_filters=384, stride=1, padding=1)
 # conv4
 net = img_conv_layer(
-    input=net, filter_size=3, num_filters=384, stride=1, padding=1, groups=1)
+    input=net, filter_size=3, num_filters=384, stride=1, padding=1, groups=gp)
 
 # conv5
 net = img_conv_layer(
-    input=net, filter_size=3, num_filters=256, stride=1, padding=1, groups=1)
+    input=net, filter_size=3, num_filters=256, stride=1, padding=1, groups=gp)
 net = img_pool_layer(input=net, pool_size=3, stride=2)
 
 net = fc_layer(
@@ -59,6 +69,9 @@ net = fc_layer(
     layer_attr=ExtraAttr(drop_rate=0.5))
 net = fc_layer(input=net, size=1000, act=SoftmaxActivation())
 
-lab = data_layer('label', num_class)
-loss = cross_entropy(input=net, label=lab)
-outputs(loss)
+if is_infer:
+    outputs(net)
+else:
+    lab = data_layer('label', num_class)
+    loss = cross_entropy(input=net, label=lab)
+    outputs(loss)

benchmark/paddle/image/googlenet.py

@@ -7,13 +7,15 @@ num_class = 1000
 batch_size = get_config_arg('batch_size', int, 128)
 use_gpu = get_config_arg('use_gpu', bool, True)
 is_infer = get_config_arg("is_infer", bool, False)
+num_samples = get_config_arg('num_samples', int, 2560)
 
 args = {
     'height': height,
     'width': width,
     'color': True,
     'num_class': num_class,
-    'is_infer': is_infer
+    'is_infer': is_infer,
+    'num_samples': num_samples
 }
 define_py_data_sources2(
     "train.list" if not is_infer else None,

benchmark/paddle/image/provider.py

@@ -14,6 +14,7 @@ def initHook(settings, height, width, color, num_class, **kwargs):
     else:
         settings.data_size = settings.height * settings.width
     settings.is_infer = kwargs.get('is_infer', False)
+    settings.num_samples = kwargs.get('num_samples', 2560)
     if settings.is_infer:
         settings.slots = [dense_vector(settings.data_size)]
     else:
@@ -23,7 +24,7 @@ def initHook(settings, height, width, color, num_class, **kwargs):
 @provider(
     init_hook=initHook, min_pool_size=-1, cache=CacheType.CACHE_PASS_IN_MEM)
 def process(settings, file_list):
-    for i in xrange(2560 if settings.is_infer else 1024):
+    for i in xrange(settings.num_samples):
         img = np.random.rand(1, settings.data_size).reshape(-1, 1).flatten()
         if settings.is_infer:
             yield img.astype('float32')

benchmark/paddle/image/resnet.py

@@ -7,13 +7,15 @@ num_class = 1000
 batch_size = get_config_arg('batch_size', int, 64)
 layer_num = get_config_arg("layer_num", int, 50)
 is_infer = get_config_arg("is_infer", bool, False)
+num_samples = get_config_arg('num_samples', int, 2560)
 
 args = {
     'height': height,
     'width': width,
     'color': True,
     'num_class': num_class,
-    'is_infer': is_infer
+    'is_infer': is_infer,
+    'num_samples': num_samples
 }
 define_py_data_sources2(
     "train.list" if not is_infer else None,

benchmark/paddle/image/run_mkl_infer.sh

@@ -37,7 +37,7 @@ function infer() {
       --trainer_count=1 \
       --num_passes=1 \
       --save_dir="models/${topology}-${layer_num}" \
-      --config_args="batch_size=128,layer_num=${layer_num}" \
+      --config_args="batch_size=128,layer_num=${layer_num},num_samples=256" \
       > /dev/null 2>&1
     echo "Done"
   fi
@@ -79,8 +79,9 @@ fi
 # inference benchmark
 for use_mkldnn in True False; do
   for batchsize in 1 2 4 8 16; do
-    infer googlenet v1 $batchsize $use_mkldnn
-    infer resnet 50 $batchsize $use_mkldnn
     infer vgg 19 $batchsize $use_mkldnn
+    infer resnet 50 $batchsize $use_mkldnn
+    infer googlenet v1 $batchsize $use_mkldnn
+    infer alexnet 2 $batchsize $use_mkldnn
   done
 done

benchmark/paddle/image/run_mkl_train.sh

@@ -47,5 +47,6 @@ for use_mkldnn in True False; do
     train vgg 19 $batchsize $use_mkldnn
     train resnet 50 $batchsize $use_mkldnn
     train googlenet v1 $batchsize $use_mkldnn
+    train alexnet 2 $batchsize $use_mkldnn
   done
 done

benchmark/paddle/image/run_openblas_infer.sh

@@ -23,24 +23,25 @@ function infer() {
     echo "./run_mkl_infer.sh to save the model first"
     exit 0
   fi
-  log_period=$((256 / bs))
+  log_period=$((32 / bs))
   paddle train --job=test \
     --config="${topology}.py" \
+    --use_mkldnn=False \
     --use_gpu=False \
     --trainer_count=$thread \
     --log_period=$log_period \
-    --config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True" \
+    --config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True,num_samples=256" \
     --init_model_path=$models_in \
     2>&1 | tee ${log}
 
-  # calculate the last 5 logs period time of 1280 samples,
+  # calculate the last 5 logs period time of 160(=32*5) samples,
   # the time before are burning time.
   start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
   end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
   start_sec=`clock_to_seconds $start`
   end_sec=`clock_to_seconds $end`
-  fps=`awk 'BEGIN{printf "%.2f",(1280 / ('$end_sec' - '$start_sec'))}'`
-  echo "Last 1280 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
+  fps=`awk 'BEGIN{printf "%.2f",(160 / ('$end_sec' - '$start_sec'))}'`
+  echo "Last 160 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
   echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
 }
 
@@ -56,7 +57,8 @@ fi
 
 # inference benchmark
 for batchsize in 1 2 4 8 16; do
-  infer googlenet v1 $batchsize
-  infer resnet 50 $batchsize
   infer vgg 19 $batchsize
+  infer resnet 50 $batchsize 
+  infer googlenet v1 $batchsize
+  infer alexnet 2 $batchsize
 done

benchmark/paddle/image/run_openblas_train.sh

@@ -12,10 +12,11 @@ function train() {
   config="${topology}.py"
   paddle train --job=time \
     --config=$config \
+    --use_mkldnn=False \
     --use_gpu=False \
     --trainer_count=$thread \
-    --log_period=10 \
-    --test_period=100 \
+    --log_period=3 \
+    --test_period=30 \
     --config_args=$args \
     2>&1 | tee ${log} 
 
@@ -36,4 +37,5 @@ for batchsize in 64 128 256; do
   train vgg 19 $batchsize
   train resnet 50 $batchsize
   train googlenet v1 $batchsize
+  train alexnet 2 $batchsize
 done

benchmark/paddle/image/vgg.py

@@ -7,13 +7,15 @@ num_class = 1000
 batch_size = get_config_arg('batch_size', int, 64)
 layer_num = get_config_arg('layer_num', int, 19)
 is_infer = get_config_arg("is_infer", bool, False)
+num_samples = get_config_arg('num_samples', int, 2560)
 
 args = {
     'height': height,
     'width': width,
     'color': True,
     'num_class': num_class,
-    'is_infer': is_infer
+    'is_infer': is_infer,
+    'num_samples': num_samples
 }
 define_py_data_sources2(
     "train.list" if not is_infer else None,

cmake/external/protobuf.cmake

@@ -253,9 +253,9 @@ IF(NOT PROTOBUF_FOUND)
     IF(WITH_C_API)
         INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
         IF(ANDROID)
-            INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
+            INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
         ELSE()
-            INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib)
+            INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib)
         ENDIF()
     ENDIF()
 

doc/api/v2/config/layer.rst

@@ -467,7 +467,7 @@ lambda_cost
     :noindex:
 
 square_error_cost
---------
+-----------------
 ..  autoclass:: paddle.v2.layer.square_error_cost
     :noindex:
 
@@ -533,7 +533,7 @@ Miscs
 =====
 
 dropout
---------------
+--------
 ..  autoclass:: paddle.v2.layer.dropout
     :noindex:
 

doc/api/v2/fluid/layers.rst

@@ -19,17 +19,17 @@ dynamic_lstm
     :noindex:
 
 data
----------
+----
 ..  autofunction:: paddle.v2.fluid.layers.data
     :noindex:
 
 mean
----------
+----
 ..  autofunction:: paddle.v2.fluid.layers.mean
     :noindex:
 
 mul
----------
+---
 ..  autofunction:: paddle.v2.fluid.layers.mul
     :noindex:
 
@@ -45,13 +45,13 @@ elementwise_div
 
 
 dropout
----------
+-------
 ..  autofunction:: paddle.v2.fluid.layers.dropout
     :noindex:
 
 
 reshape
----------
+--------
 ..  autofunction:: paddle.v2.fluid.layers.reshape
     :noindex:
 
@@ -81,67 +81,67 @@ transpose
 
 
 sigmoid_cross_entropy_with_logits
----------
+---------------------------------
 ..  autofunction:: paddle.v2.fluid.layers.esigmoid_cross_entropy_with_logits
     :noindex:
 
 
 cast
----------
+----
 ..  autofunction:: paddle.v2.fluid.layers.cast
     :noindex:
 
 
 concat
----------
+-------
 ..  autofunction:: paddle.v2.fluid.layers.concat
     :noindex:
 
 
 sums
----------
+----
 ..  autofunction:: paddle.v2.fluid.layers.sums
     :noindex:
 
 
 linear_chain_crf
----------
+----------------
 ..  autofunction:: paddle.v2.fluid.layers.linear_chain_crf
     :noindex:
 
 
 assign
----------
+-------
 ..  autofunction:: paddle.v2.fluid.layers.embedding
     :noindex:
 
 
 split_lod_tensor
----------
+----------------
 ..  autofunction:: paddle.v2.fluid.layers.split_lod_tensor
     :noindex:
 
 
 merge_lod_tensor
----------
+----------------
 ..  autofunction:: paddle.v2.fluid.layers.merge_lod_tensor
     :noindex:
 
 cos_sim
----------
+--------
 ..  autofunction:: paddle.v2.fluid.layers.cos_sim
     :noindex:
 
 
 cross_entropy
----------
+-------------
 ..  autofunction:: paddle.v2.fluid.layers.cross_entropy
     :noindex:
 
 
 
 square_error_cost
----------
+-----------------
 ..  autofunction:: paddle.v2.fluid.layers.square_error_cost
     :noindex:
 
@@ -153,68 +153,80 @@ accuracy
 
 
 sequence_conv
----------
+-------------
 ..  autofunction:: paddle.v2.fluid.layers.sequence_conv
     :noindex:
 
 
 conv2d
----------
+------
 ..  autofunction:: paddle.v2.fluid.layers.conv2d
     :noindex:
 
 
 sequence_pool
----------
+-------------
 ..  autofunction:: paddle.v2.fluid.layers.sequence_pool
     :noindex:
 
 
+sequence_first_step
+-------------------
+..  autofunction:: paddle.v2.fluid.layers.sequence_first_step
+    :noindex:
+
+
+sequence_last_step
+------------------
+..  autofunction:: paddle.v2.fluid.layers.sequence_last_step
+    :noindex:
+
+
 pool2d
----------
+------
 ..  autofunction:: paddle.v2.fluid.layers.pool2d
     :noindex:
 
 
 batch_norm
----------
+----------
 ..  autofunction:: paddle.v2.fluid.layers.batch_norm
     :noindex:
 
 
 beam_search_decode
----------
+------------------
 ..  autofunction:: paddle.v2.fluid.layers.beam_search_decode
     :noindex:
 
 
 lod_rank_table
----------
+--------------
 ..  autofunction:: paddle.v2.fluid.layers.lod_rank_table
     :noindex:
 
 
 max_sequence_len
----------
+----------------
 ..  autofunction:: paddle.v2.fluid.layers.max_sequence_len
     :noindex:
 
 
 topk
----------
+-----
 ..  autofunction:: paddle.v2.fluid.layers.topk
     :noindex:
 
 
 lod_tensor_to_array
----------
+-------------------
 ..  autofunction:: paddle.v2.fluid.layers.lod_tensor_to_array
     :noindex:
 
 
 
 array_to_lod_tensor
----------
+-------------------
 ..  autofunction:: paddle.v2.fluid.layers.array_to_lod_tensor
     :noindex:
 
@@ -222,26 +234,26 @@ array_to_lod_tensor
 
 
 fill_constant
----------
+-------------
 ..  autofunction:: paddle.v2.fluid.layers.fill_constant
     :noindex:
 
 
 
 fill_constant_batch_size_like
----------
+-----------------------------
 ..  autofunction:: paddle.v2.fluid.layers.fill_constant_batch_size_like
     :noindex:
 
 
 ones
----------
+----
 ..  autofunction:: paddle.v2.fluid.layers.ones
     :noindex:
 
 
 zeros
----------
+-----
 ..  autofunction:: paddle.v2.fluid.layers.zeros
     :noindex:
 
@@ -253,14 +265,14 @@ increment
 
 
 array_write
----------
+-----------
 ..  autofunction:: paddle.v2.fluid.layers.array_write
     :noindex:
 
 
 
 create_array
----------
+------------
 ..  autofunction:: paddle.v2.fluid.layers.create_array
     :noindex:
 
@@ -272,31 +284,31 @@ less_than
 
 
 array_read
----------
+----------
 ..  autofunction:: paddle.v2.fluid.layers.array_read
     :noindex:
 
 
 shrink_memory
----------
+--------------
 ..  autofunction:: paddle.v2.fluid.layers.shrink_memory
     :noindex:
 
 
 array_length
----------
+-------------
 ..  autofunction:: paddle.v2.fluid.layers.array_length
     :noindex:
 
 
 conv2d_transpose
----------
+----------------
 ..  autofunction:: paddle.v2.fluid.layers.conv2d_transpose
     :noindex:
 
 
 sequence_expand
----------
+---------------
 ..  autofunction:: paddle.v2.fluid.layers.sequence_expand
     :noindex:
 
@@ -308,7 +320,19 @@ lstm_unit
 
 
 sequence_softmax
----------
+----------------
 ..  autofunction:: paddle.v2.fluid.layers.sequence_softmax
     :noindex:
 
+
+reduce_sum
+----------
+..  autofunction:: paddle.v2.fluid.layers.reduce_sum
+    :noindex:
+
+
+reduce_mean
+---------
+..  autofunction:: paddle.v2.fluid.layers.reduce_mean
+    :noindex:
+

doc/api/v2/fluid/nets.rst

@@ -3,19 +3,19 @@ Nets
 ===========
 
 simple_img_conv_pool
------------
+--------------------
 ..  autofunction:: paddle.v2.fluid.nets.simple_img_conv_pool
     :noindex:
 
 
 img_conv_group
------------
+---------------
 ..  autofunction:: paddle.v2.fluid.nets.img_conv_group
     :noindex:
 
 
 sequence_conv_pool
------------
+------------------
 ..  autofunction:: paddle.v2.fluid.nets.sequence_conv_pool
     :noindex:
 

doc/api/v2/fluid/optimizer.rst

@@ -18,7 +18,7 @@ SGDOptimizer
 
 
 MomentumOptimizer
------------
+-----------------
 ..  automodule:: paddle.v2.fluid.optimizer
     :members: MomentumOptimizer
     :noindex:
@@ -26,14 +26,14 @@ MomentumOptimizer
 
 
 AdagradOptimizer
------------
+----------------
 ..  automodule:: paddle.v2.fluid.optimizer
     :members: AdagradOptimizer
     :noindex:
 
 
 AdamOptimizer
------------
+-------------
 ..  automodule:: paddle.v2.fluid.optimizer
     :members: AdamOptimizer
     :noindex:
@@ -47,7 +47,7 @@ AdamaxOptimizer
 
 
 DecayedAdagradOptimizer
------------
+-----------------------
 ..  automodule:: paddle.v2.fluid.optimizer
     :members: DecayedAdagradOptimizer
     :noindex:

doc/api/v2/fluid/regularizer.rst

@@ -3,14 +3,14 @@ Regularizer
 ===========
 
 WeightDecayRegularizer
------------
+----------------------
 ..  automodule:: paddle.v2.fluid.regularizer
     :members: WeightDecayRegularizer
     :noindex:
 
 
 L2DecayRegularizer
------------
+------------------
 ..  automodule:: paddle.v2.fluid.regularizer
     :members: L2DecayRegularizer
     :noindex:
@@ -18,7 +18,7 @@ L2DecayRegularizer
 
 
 L1DecayRegularizer
------------
+-------------------
 ..  automodule:: paddle.v2.fluid.regularizer
     :members: L1DecayRegularizer
 

doc/design/kernel_hint_design.md

+## Problem
+In PaddlePaddle's [Design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md), one Operator may have multiple kernels. Users may have some personal preference to choose a certain type of kernel for an operator, such as `force_cpu` to choose a CPU kernel, `use_cudnn` to choose a CUDNN kernel, we need to provide a way for users to do this.
+
+In the current design, we use KernelType to describe one kernel.
+
+```cpp
+struct KernelType {
+  Place place_;
+  DataType data_type_;
+  LayoutType layout_;
+};
+```
+ `place_` `data_type_` and `layout_` can be got from the input tensors of the operator, `GetActualKernelType(inputs)` use inputs to infer the proper kernel key that fit the incoming data, but users can not directly configure it.
+
+The [design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md) also provides a virtual method `GetExpectedKernelType` that user can overload and use to choose the KernelType they want to use.
+
+So we should send the information user defined in proto to `GetExpectedKernelType` for choosing a kernel.
+
+The problem is, how should we define and send the information for `GetExpectedKernelType` to use?
+
+## Solution
+
+### Potential choice
+1. Do nothing, let the user add the information they want to operator‘s attribute and get them inside `GetExpectedKernelType`, this can work properly. But there is a little problem that users may define many kinds of hints for the same purpose, such as `force_cpu`, `use_cpu`, `cpu_kernel` to choose CPU kernel, and `use_cudnn`, `force_cudnn`, `cudnn_kernel` to choose CUDNN kernel.
+
+2. Pre-define all the needed option and use a single attr key such as `kernel_hint` for the user, this is not so flexible if the user wants to define some more kind of hint.
+
+### Final choice
+To provide enough flexibility while avoiding confusion definition, we can define some global constants for these attribute names, such as `force_cpu`, `use_cudnn`, `use_mkldnn` for a user to choose.
+
+In C++
+
+```cpp
+const std::string kForceCPU = "force_cpu";
+const std::string kUseCUDNN = "use_cudnn";
+const std::string kUseMKLDNN = "use_mkldnn";
+
+KernelType GetExpectedKernelType() {
+  if (Attr<bool>(kForceCPU)) {
+    return KernelType(CPUPlace, ...)
+  } else {
+    ...
+  }
+}
+```
+
+In Python code
+
+```python
+FORCE_CPU = core.kForceCPU()
+
+def xx_layer(..., force_cpu=false):
+  layer_helper = LayerHelper(...)
+  layer_helper.append_op(
+    type="xx",
+    attr={FORCE_CPU: force_cpu})
+```

doc/design/operator_kernel_type.md

+# Design Doc: The Keys of Operator Kernel Type
+## Problem
+An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique Kernel. Before an operator runs, an certain kernel must be chosen by a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:
+
+```cpp
+struct OpKernelType {
+  platform::Place place_;
+  proto::DataType data_type_;
+};
+```
+For more details, please refer to [codes](https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374) in github.
+
+It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys are not enough. We need a more complete representation of `OpKernelType`. 
+
+We often implement a kernel of an operator with some computing library in certain device(place). Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices. 
+
+For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.
+
+It's obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. But the data layout of a Tensor will also lead to different implementation. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209). Data Layout should also be taken into consideration.
+
+## Solution
+
+There are four keys to determine a kernel type of an operator: `Place`/`Library`/`DataType`/`Layout`.
+
+```cpp
+struct OpKernelType {
+  platform::Place place_;
+  platform::Library library_;
+  proto::DataType data_type_;
+  framework::Layout layout_;
+};
+```
+
+Following is the details:
+
+### Place
+
+`Place` is defined as follows:
+
+```cpp
+typedef boost::variant<CUDAPlace, ROCmPlace, FPGAPlace, CPUPlace> Place;
+```
+
+`Place` is to represent the device memory where data is locating.
+
+
+### Library
+
+One operator kernel is usually implemented based on one library. `Library` is defined as a enum variable:
+
+```cpp
+enum Library { Plain, MKLDNN, CUDNN };
+```
+
+We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
+A library usually has a corresponding `DeviceContext` which contains some handles needed by computation. Fluid now have two default DeviceContexts in CPU and CUDA, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains a Eigen library handle and `CDUADeviceContext` contains a Eigen library handle and cuBLAS handle.
+
+If we want to support new Library, a new enumerator need to be added to `Library` and a new corresponding `LibraryDeviceContext` will be created.
+
+
+### DataType
+
+
+`DataType` is defined in [framework.proto](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto). Currently, int32/int64/fp32/fp64 are supported.
+
+### Layout
+
+Actually, a Tensor is a view of a block of memory. Besides a pointer to the memory, we also have to get some other descriptions of this block of memory, such as shape(ddim), stride, and layout.
+
+Different layout leads to different implementation of operator kernel. There are mainly 4 principles we have to follow to support layout in our fluid framework.
+
+- We take layout as a data member of Tensor. Layout is actually a enum variable. If fluid is built with MKLDNN, then, the memory format in MKLDNN will be added into this enum variable too.
+
+- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout of generating data. Of course, we can have some default layout, like NCHW.
+
+- The inference of Layout is at run-time, not compile-time.
+
+- Every operator have to implement different kernels for different layouts. Let's take MKLDNN as an example, if we want to implement a MKLDNN convolution operator, we have to realize all the kernels for different layout, list at [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to do registering kernels for MKLDNN operators.
+
+`Layout` is also defined as a enum variable:
+
+```cpp
+enum Layout {
+  kNCHW,
+  kNHWC,
+#ifdef PADDLE_WITH_MKLDNN
+  knChw8c
+  ...
+#endif
+};
+```

doc/design/refactor/multi_cpu.md

+# Design Doc: Execute the Program with Multi CPU
+
+## Abstract
+
+This Design Doc propose an approach to make the user-defined Op graph
+running with multi-CPU, we will use an auto transpiler to convert the user-defined
+Op graph to a multi-CPU Op graph, and run `ParallelDo` Op to run the graph.
+
+## Transpiler
+
+<img src="src/multi-threads/single-thread@3x.png" width="300">
+
+After converted:
+
+<img src="src/multi-threads/multi-threads@3x.png" width="1000">
+
+## Implement
+
+- `Multi-CPU Transpiler` will convert the graph to a multi-CPU graph
+  which would be executed with multi-threads.
+- `BlockingCounter` will `Init/Decrement` an atomic counter, and Blocking `Wait`
+  for the atomic counter become `0`:
+  ```cpp
+  BlockingCounter bc(thread_count);
+  for (int i = 0; i < thread_count; ++i) {
+    thread_pool->Start([&bc] {bc.DecrementCount(); })
+  }
+  bc.Wait();
+  ```
+- `ParallelDo` Operator
+  - Initialize a thread pool which is a Singleton.
+  - Use a block id as the input, and create run the specify Block on independent scope
+    with multi-threads.
+  - Initialize a `BlockingCounter` instance and wait until all threads are done.
+- `Split` Operator will split the Input Tensor into a TensorArray.
+- `Merge` merge all the gradients which calculated in different threads
+  with `mean/sum/max/min...` method, and then run the Optimizer Op to optimize `W`.
+
+## TODO
+
+- Improve the optimizer stage with multi-threads, since we could
+  assign the parameters to the different threads and execute
+  optimizer with multi-threads.

doc/design/switch_kernel.md

+## Background
+Every operator has many kernels because there are multiple data types, places, data layout that Fluid supports. We use the `KernelType` to describe kernel types that operators can hold. 
+
+The `KernelType` is as follows.
+
+```
+struct KernelType {
+  Place place_;
+  DataType data_type_;
+  LayoutType layout_;
+};
+```
+
+The `place_` is a descriptor of the device and the computational library, e.g., `MKLDNNPlace`, `CUDAPlace`.
+
+The `data_type_` is the data type that this kernel performs on, e.g., `FP32`, `INT64`. Note that one kernel may have inputs with different data types. However, it will be a major `data_type`. For example, the `cross_entropy` takes `int64` as it label, and `double`/`float` as its input logit and output cost. The major `data_type` of `cross_entropy` is `float`/`double`.
+
+The `layout` is useful for some computational library. One example is that MKLDNN uses many kinds of layout, such as `nChw8c`. Each kind of layout will invoke the different kernel.
+
+## Problem
+
+We register a kernel for every operator and every kernel type ideally. However, it is impracticable for the following situations.
+
+1. Some operators, like CRF, are complicated and inefficient to be implemented on GPU. The CRF operator will only have a CPU kernel.
+2. Some operators will take too many memory. It is better to force them into CPU. However, the rest of operators in this neural network will be performed on GPU, i.e., model parallel problem.
+3. Some layout and place are particular. One example is that MKLDNN uses `nChw8` and there is no other library uses `nChw8c`.
+
+Problems under these situations are similar. We can formalise this problem as follow.
+
+We register kernels with types $KT = \{kt_1, kt_2, kt_3, ...\}$ for one operator. The inputs of this operator should be run on kernel type $kt_{?}$, which the $kt_{?} \notin KT$. How to cast the input of this operator from $kt_{?}$ to any of kernel type in $KT$.
+
+## Solution
+
+It is clearly that transforming inputs of an operator toadapt another kernel type is not related to the particular operator. So we should register these transformation methods as global methods.
+
+We can infer a kernel type from the inputs of an operators. We let this kernel type as `actual kernel type`, which means this kernel type is the actually kernel type that operator should be performed.
+
+We can get a kernel type by 1) The configuration of operator description. (Users may want to force use `MKL` for `conv` operator). 2) The place of the current executor. (Executor is running on GPU). This kernel type is what we expect the operator will be performed on. We let this kernel type as `expect kernel type`.
+
+We transform the input data from `actual` to `expect` if the expect kernel type is not as same as actual kernel type.
+
+The algorithm is described as follow
+
+```cpp
+using DataTransformationFN = std::function<void(const Tensor& in, Tensor* out)>;
+using KernelTypePair = std::pair<KernelType, KernelType>;
+
+map<KernelTypePair, DataTransformationFN> g_data_transformation_;
+
+void OpWithKernel::Run() {
+  vec<Tensor> inputs = ...
+  auto actual_kernel_type = GetActualKernelType(inputs);
+  
+  // The expected kernel type is related to actual kernel type.
+  // For the most operators, the expected kernel type is as same as
+  // actual kernel type.
+  //
+  // So we pass `actual_kernel_type` as a parameter of 
+  // GetExpectedKernelType
+  auto expect_kernel_type = GetExpectedKernelType(actual_kernel_type);
+  
+  auto trans = g_data_transformation_[{actual_kernel_type, expect_kernel_type}];
+  
+  kernel.run(trans(inputs));
+}
+```

doc/getstarted/build_and_install/build_from_source_cn.rst

@@ -70,13 +70,13 @@ PaddlePaddle编译需要使用到下面的依赖（包含但不限于），其
    :header: "依赖", "版本", "说明"
    :widths: 10, 15, 30
 
-   "CMake", ">=3.5", ""
+   "CMake", ">=3.2", ""
    "GCC", "4.8.2", "推荐使用CentOS的devtools2"
-   "Python", "2.7.x", "依赖libpython2.7.so"
-   "pip", ">=9.0", ""
-   "numpy", "", ""
+   "Python", "2.7.x", "依赖libpython2.7.so"
+   "pip", ">=9.0", ""
+   "numpy", "", ""
    "SWIG", ">=2.0", ""
-   "Go", ">=1.8", "可选"
+   "Go", ">=1.8", "可选"
 
 
 .. _build_options:

doc/getstarted/build_and_install/build_from_source_en.rst

@@ -76,13 +76,13 @@ will be downloaded automatically.
    :header: "Dependency", "Version", "Description"
    :widths: 10, 15, 30
 
-   "CMake", ">=3.5", ""
+   "CMake", ">=3.2", ""
    "GCC", "4.8.2", "Recommend devtools2 for CentOS"
-   "Python", "2.7.x", "Need libpython2.7.so"
-   "pip", ">=9.0", ""
-   "numpy", "", ""
+   "Python", "2.7.x", "Need libpython2.7.so"
+   "pip", ">=9.0", ""
+   "numpy", "", ""
    "SWIG", ">=2.0", ""
-   "Go", ">=1.8", "Optional"
+   "Go", ">=1.8", "Optional"
 
 
 .. _build_options:

doc/getstarted/build_and_install/docker_install_cn.rst

@@ -128,7 +128,7 @@ PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Note
 
 AVX是一种CPU指令集，可以加速PaddlePaddle的计算。最新的PaddlePaddle Docker镜像默认
 是开启AVX编译的，所以，如果您的电脑不支持AVX，需要单独
-`编译 <./build_from_source_cn.rst>`_ PaddlePaddle为no-avx版本。
+`编译 <./build_from_source_cn.html>`_ PaddlePaddle为no-avx版本。
 
 以下指令能检查Linux电脑是否支持AVX：
 

doc/getstarted/build_and_install/docker_install_en.rst

@@ -137,7 +137,7 @@ GPU driver installed before move on.
 AVX is a kind of CPU instruction can accelerate PaddlePaddle's calculations.
 The latest PaddlePaddle Docker image turns AVX on by default, so, if your
 computer doesn't support AVX, you'll probably need to
-`build <./build_from_source_en.rst>`_ with :code:`WITH_AVX=OFF`.
+`build <./build_from_source_en.html>`_ with :code:`WITH_AVX=OFF`.
 
 The following command will tell you whether your computer supports AVX.
 

doc/getstarted/build_and_install/pip_install_cn.rst

@@ -37,11 +37,11 @@ PaddlePaddle可以使用常用的Python包管理工具
     :header: "版本说明", "cp27-cp27mu", "cp27-cp27m", "C-API"
     :widths: 1, 3, 3, 3
 
-    "cpu_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
-    "cpu_avx_openblas", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "暂无"
-    "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
-    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
-    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "暂无"
+    "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
 
 .. _pip_dependency:
 

doc/getstarted/build_and_install/pip_install_en.rst

@@ -40,11 +40,11 @@ If the links below shows up the login form, just click "Log in as guest" to star
     :header: "version", "cp27-cp27mu", "cp27-cp27m", "C-API"
     :widths: 1, 3, 3, 3
 
-    "cpu_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
-    "cpu_avx_openblas", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "Not Available"
-    "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
-    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
-    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle-0.10.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <http://guest@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cpu_avx_mkl", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cpu_avx_openblas", "`paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxOpenblas/.lastSuccessful/paddlepaddle-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "Not Available"
+    "cuda7.5_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cuda8.0_cudnn5_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
+    "cuda8.0_cudnn7_avx_mkl", "`paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27mu-linux_x86_64.whl>`_", "`paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddlepaddle_gpu-0.11.0-cp27-cp27m-linux_x86_64.whl>`_", "`paddle.tgz <https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz>`_"
 
 .. _pip_dependency:
 

doc/howto/dev/new_op_cn.md

@@ -53,7 +53,7 @@ Kernel实现       | CPU、CUDA共享Kernel实现在`.h`文件中，否则，CPU
 ```cpp
 class MulOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "(Tensor), 2D tensor of size (M x K)");
     AddInput("Y", "(Tensor), 2D tensor of size (K x N)");
@@ -82,7 +82,7 @@ The equation is: Out = X * Y
 template <typename AttrType>
 class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input tensor of scale operator.").NotInGradient();
     AddOutput("Out", "The output tensor of scale operator.").NotInGradient();

doc/howto/dev/new_op_en.md

@@ -50,7 +50,7 @@ First, define `ProtoMaker` to describe the Operator's input, output, and additio
 ```cpp
 class MulOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "(Tensor), 2D tensor of size (M x K)");
     AddInput("Y", "(Tensor), 2D tensor of size (K x N)");
@@ -79,7 +79,7 @@ An additional example [`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/de
 template <typename AttrType>
 class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  ScaleOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input tensor of scale operator.").NotInGradient();
     AddOutput("Out", "The output tensor of scale operator.").NotInGradient();

doc/howto/usage/cluster/k8s_distributed_cn.md

@@ -2,8 +2,6 @@
 
 前一篇文章介绍了如何在Kubernetes集群上启动一个单机PaddlePaddle训练作业 (Job)。在这篇文章里，我们介绍如何在Kubernetes集群上进行分布式PaddlePaddle训练作业。关于PaddlePaddle的分布式训练，文章 [Cluster Training](http://www.paddlepaddle.org/docs/develop/documentation/zh/howto/usage/cluster/cluster_train_cn.html)介绍了一种通过SSH远程分发任务，进行分布式训练的方法，与此不同的是，本文将介绍在Kubernetes容器管理平台上快速构建PaddlePaddle容器集群，进行分布式训练的方案。
 
-有关Kubernetes相关概念以及如何搭建和配置Kubernetes集群，可以参考[k8s_basis](./k8s_basis_cn.md)。
-
 ## 整体方案
 
 在训练之前，用户将配置与训练数据切分好放在分布式文件系统预先分配好的目录中(不同的分布式文件系统，需要使用其制定的方式挂载后并导入数据)，训练时，程序从此目录拷贝文件到容器内进行训练，将结果保存到此目录里。整体的结构图如下：

paddle/framework/attribute.cc

@@ -19,42 +19,42 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
+Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc) {
   switch (attr_desc.type()) {
-    case framework::AttrType::BOOLEAN: {
+    case proto::AttrType::BOOLEAN: {
       return attr_desc.b();
     }
-    case framework::AttrType::INT: {
+    case proto::AttrType::INT: {
       return attr_desc.i();
     }
-    case framework::AttrType::FLOAT: {
+    case proto::AttrType::FLOAT: {
       return attr_desc.f();
     }
-    case framework::AttrType::STRING: {
+    case proto::AttrType::STRING: {
       return attr_desc.s();
     }
-    case framework::AttrType::BOOLEANS: {
+    case proto::AttrType::BOOLEANS: {
       std::vector<bool> val(attr_desc.bools_size());
       for (int i = 0; i < attr_desc.bools_size(); ++i) {
         val[i] = attr_desc.bools(i);
       }
       return val;
     }
-    case framework::AttrType::INTS: {
+    case proto::AttrType::INTS: {
       std::vector<int> val(attr_desc.ints_size());
       for (int i = 0; i < attr_desc.ints_size(); ++i) {
         val[i] = attr_desc.ints(i);
       }
       return val;
     }
-    case framework::AttrType::FLOATS: {
+    case proto::AttrType::FLOATS: {
       std::vector<float> val(attr_desc.floats_size());
       for (int i = 0; i < attr_desc.floats_size(); ++i) {
         val[i] = attr_desc.floats(i);
       }
       return val;
     }
-    case framework::AttrType::STRINGS: {
+    case proto::AttrType::STRINGS: {
       std::vector<std::string> val(attr_desc.strings_size());
       for (int i = 0; i < attr_desc.strings_size(); ++i) {
         val[i] = attr_desc.strings(i);

paddle/framework/attribute.h

@@ -27,12 +27,12 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 template <typename T>
-inline AttrType AttrTypeID() {
+inline proto::AttrType AttrTypeID() {
   Attribute tmp = T();
-  return static_cast<AttrType>(tmp.which() - 1);
+  return static_cast<proto::AttrType>(tmp.which() - 1);
 }
 
-Attribute GetAttrValue(const OpDesc::Attr& attr_desc);
+Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc);
 
 class AttrReader {
  public:

paddle/framework/backward.cc

@@ -42,7 +42,7 @@ static std::unordered_set<std::string>& CtrlFlowOps() {
 static inline std::unique_ptr<OperatorBase> CreateGradOp(
     const OperatorBase& op, const std::unordered_set<std::string>& no_grad_set,
     std::unordered_map<std::string, std::string>* grad_to_var) {
-  OpDescBind op_desc;
+  OpDesc op_desc;
   op_desc.SetInputMap(op.Inputs());
   op_desc.SetOutputMap(op.Outputs());
   op_desc.SetType(op.Type());
@@ -53,7 +53,7 @@ static inline std::unique_ptr<OperatorBase> CreateGradOp(
   grad_ops.reserve(grad_descs.size());
   std::transform(grad_descs.begin(), grad_descs.end(),
                  std::back_inserter(grad_ops),
-                 [](const std::unique_ptr<OpDescBind>& grad_desc) {
+                 [](const std::unique_ptr<OpDesc>& grad_desc) {
                    return OpRegistry::CreateOp(*grad_desc);
                  });
   PADDLE_ENFORCE(!grad_ops.empty());
@@ -217,7 +217,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
         // If part of input gradient of that operator is not calculated, fill
         // zero variables to that input gradient.
         net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", {{"X", {prefix}}},
-                                           {{"Y", {grad_input}}},
+                                           {{"Out", {grad_input}}},
                                            AttributeMap{}));
       }
       return false;
@@ -296,7 +296,7 @@ static std::string FwdName(const std::string& grad_name) {
 static void CreateGradVarInBlock(
     size_t grad_op_start_index,
     const std::unordered_map<std::string, std::string>& param_name_map,
-    BlockDescBind* block_desc,
+    BlockDesc* block_desc,
     std::unordered_map<std::string, GradVarInfo>* grad_var_record) {
   auto ops = block_desc->AllOps();
   for (size_t op_index = grad_op_start_index; op_index < ops.size();
@@ -341,7 +341,7 @@ static void CreateGradVarInBlock(
       auto* param = block_desc->FindVarRecursive(pname);
       auto* grad = block_desc->FindVar(arg);
       if (param == nullptr) {
-        grad->SetDataType(DataType::FP32);
+        grad->SetDataType(proto::DataType::FP32);
       } else {
         grad->SetDataType(param->GetDataType());
       }
@@ -350,12 +350,11 @@ static void CreateGradVarInBlock(
   }
 }
 
-std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
-    const OpDescBind* op_desc, std::unordered_set<std::string>* no_grad_vars,
+std::vector<std::unique_ptr<OpDesc>> MakeOpGrad(
+    const OpDesc* op_desc, std::unordered_set<std::string>* no_grad_vars,
     std::unordered_map<std::string, std::string>* grad_to_var,
-    const std::vector<BlockDescBind*>& grad_block =
-        std::vector<BlockDescBind*>()) {
-  std::vector<std::unique_ptr<OpDescBind>> grad_op_descs;
+    const std::vector<BlockDesc*>& grad_block = std::vector<BlockDesc*>()) {
+  std::vector<std::unique_ptr<OpDesc>> grad_op_descs;
   // All input gradients of forwarding operator do not need to calculate.
   const std::vector<std::string>& inputs = op_desc->InputArgumentNames();
   if (AllGradInSet(inputs, *no_grad_vars)) {
@@ -386,7 +385,7 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
           .Get(op_desc->Type())
           .GradOpMaker()(*op_desc, *no_grad_vars, grad_to_var, grad_block);
 
-  std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
+  std::list<std::unique_ptr<OpDesc>> pending_fill_zeros_ops;
   for (auto& desc : grad_op_descs) {
     for (const std::string& in_name : desc->InputArgumentNames()) {
       if (no_grad_vars->count(in_name)) {
@@ -394,9 +393,9 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
             0, in_name.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
         std::string new_name = prefix + kZeroVarSuffix;
         desc->Rename(in_name, new_name);
-        std::unique_ptr<OpDescBind> fill_zeros_op(
-            new OpDescBind("fill_zeros_like", {{"X", {prefix}}},
-                           {{"Y", {new_name}}}, AttributeMap{}));
+        std::unique_ptr<OpDesc> fill_zeros_op(
+            new OpDesc("fill_zeros_like", {{"X", {prefix}}},
+                       {{"Out", {new_name}}}, AttributeMap{}));
         pending_fill_zeros_ops.push_back(std::move(fill_zeros_op));
       }
     }
@@ -408,34 +407,33 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
   return grad_op_descs;
 }
 
-static BlockDescBind* CreateStepBlock(
-    ProgramDescBind& program_desc,
-    std::unordered_set<std::string>* no_grad_vars,
+static BlockDesc* CreateStepBlock(
+    ProgramDesc& program_desc, std::unordered_set<std::string>* no_grad_vars,
     std::unordered_map<std::string, std::string>* grad_to_var,
     int step_block_idx);
 
-std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
-    ProgramDescBind& program_desc, int block_idx,
+std::vector<std::unique_ptr<OpDesc>> MakeBlockBackward(
+    ProgramDesc& program_desc, int block_idx,
     std::unordered_set<std::string>* no_grad_vars,
     std::unordered_map<std::string, std::string>* grad_to_var) {
   VLOG(5) << "MakeBlockBackward";
-  BlockDescBind* cur_block = program_desc.MutableBlock(block_idx);
-  std::vector<OpDescBind*> op_descs = cur_block->AllOps();
+  BlockDesc* cur_block = program_desc.MutableBlock(block_idx);
+  std::vector<OpDesc*> op_descs = cur_block->AllOps();
   std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
   size_t grad_desc_idx = 0;
-  std::vector<std::unique_ptr<OpDescBind>> backward_descs;
+  std::vector<std::unique_ptr<OpDesc>> backward_descs;
 
   for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) {
     VLOG(5) << "Making backward " << (*it)->Type() << " op";
-    std::vector<std::unique_ptr<OpDescBind>> op_grads;
+    std::vector<std::unique_ptr<OpDesc>> op_grads;
 
     if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") {
       int step_block_idx = (*it)->GetBlockAttr("sub_block");
-      BlockDescBind* backward_block = CreateStepBlock(
-          program_desc, no_grad_vars, grad_to_var, step_block_idx);
+      BlockDesc* backward_block = CreateStepBlock(program_desc, no_grad_vars,
+                                                  grad_to_var, step_block_idx);
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
     } else if ((*it)->Type() == "conditional_block") {
-      BlockDescBind* backward_block =
+      BlockDesc* backward_block =
           CreateStepBlock(program_desc, no_grad_vars, grad_to_var,
                           (*it)->GetBlockAttr("sub_block"));
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
@@ -463,14 +461,14 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
       }
       ++grad_desc_idx;
     }
-    std::transform(
-        op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs),
-        [](std::unique_ptr<OpDescBind>& ptr) { return std::move(ptr); });
+    std::transform(op_grads.begin(), op_grads.end(),
+                   std::back_inserter(backward_descs),
+                   [](std::unique_ptr<OpDesc>& ptr) { return std::move(ptr); });
   }
 
   VLOG(5) << "Appending Sums";
   // Check whether some variables are written more than once
-  std::list<std::pair<size_t, std::unique_ptr<OpDescBind>>> pending_sum_ops;
+  std::list<std::pair<size_t, std::unique_ptr<OpDesc>>> pending_sum_ops;
   for (const auto& dup : dup_out_ops) {
     const std::string& out_name = dup.first;
     const std::vector<size_t> dup_op = dup.second;
@@ -486,18 +484,17 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
         sum_op_inputs.emplace_back(new_name);
         next_g_name = sum_op_inputs.back();
       }
-      std::unique_ptr<OpDescBind> sum_op(
-          new OpDescBind("sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}},
-                         AttributeMap{}));
+      std::unique_ptr<OpDesc> sum_op(new OpDesc("sum", {{"X", sum_op_inputs}},
+                                                {{"Out", {out_name}}},
+                                                AttributeMap{}));
       pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)});
     }
   }
 
-  pending_sum_ops.sort(
-      [](const std::pair<size_t, std::unique_ptr<OpDescBind>>& a,
-         const std::pair<size_t, std::unique_ptr<OpDescBind>>& b) {
-        return a.first > b.first;
-      });
+  pending_sum_ops.sort([](const std::pair<size_t, std::unique_ptr<OpDesc>>& a,
+                          const std::pair<size_t, std::unique_ptr<OpDesc>>& b) {
+    return a.first > b.first;
+  });
   for (auto& p : pending_sum_ops) {
     backward_descs.insert(backward_descs.begin() + p.first + 1,
                           std::move(p.second));
@@ -508,14 +505,13 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
   return backward_descs;
 }
 
-static BlockDescBind* CreateStepBlock(
-    ProgramDescBind& program_desc,
-    std::unordered_set<std::string>* no_grad_vars,
+static BlockDesc* CreateStepBlock(
+    ProgramDesc& program_desc, std::unordered_set<std::string>* no_grad_vars,
     std::unordered_map<std::string, std::string>* grad_to_var,
     int step_block_idx) {
   auto backward_block_op_descs = MakeBlockBackward(program_desc, step_block_idx,
                                                    no_grad_vars, grad_to_var);
-  BlockDescBind* backward_block =
+  BlockDesc* backward_block =
       program_desc.AppendBlock(*program_desc.MutableBlock(step_block_idx));
   for (auto& ptr : backward_block_op_descs) {
     backward_block->AppendAllocatedOp(move(ptr));
@@ -524,7 +520,7 @@ static BlockDescBind* CreateStepBlock(
 }
 
 ParamGradInfoMap AppendBackward(
-    ProgramDescBind& program_desc, const VarDescBind& target,
+    ProgramDesc& program_desc, const VarDesc& target,
     const std::unordered_set<std::string>& no_grad_vars) {
   std::unordered_set<std::string> no_grad_var_names;
   no_grad_var_names.reserve(no_grad_vars.size() + 1);
@@ -541,11 +537,11 @@ ParamGradInfoMap AppendBackward(
   PADDLE_ENFORCE(is_scalar, "target should be scalar");
   VLOG(3) << "backward from loss=" << target.Name()
           << " data_type=" << target.GetDataType();
-  std::unique_ptr<OpDescBind> fill_one_op(
-      new OpDescBind("fill_constant", {}, {{"Out", {fill_one_op_out}}},
-                     {{"shape", std::vector<int>{1}},
-                      {"value", static_cast<float>(1.0)},
-                      {"dtype", target.GetDataType()}}));
+  std::unique_ptr<OpDesc> fill_one_op(
+      new OpDesc("fill_constant", {}, {{"Out", {fill_one_op_out}}},
+                 {{"shape", std::vector<int>{1}},
+                  {"value", static_cast<float>(1.0)},
+                  {"dtype", target.GetDataType()}}));
   // infer var type of fill_one_op
   fill_one_op->InferVarType(root_block);
 

paddle/framework/backward.h

@@ -49,7 +49,7 @@ using ParamGradInfoMap = std::unordered_map<std::string /*fwd_var_name*/,
                                             GradVarInfo /*grad_var_info*/>;
 
 ParamGradInfoMap AppendBackward(
-    ProgramDescBind& program_desc, const VarDescBind& target,
+    ProgramDesc& program_desc, const VarDesc& target,
     const std::unordered_set<std::string>& no_grad_vars);
 
 }  // namespace framework

paddle/framework/backward_test.cc

@@ -58,13 +58,13 @@ class RowWiseAddGradMaker : public SingleGradOpDescMaker {
   using SingleGradOpDescMaker::SingleGradOpDescMaker;
 
  protected:
-  std::unique_ptr<OpDescBind> Apply() const override {
-    auto grad_op = new OpDescBind();
+  std::unique_ptr<OpDesc> Apply() const override {
+    auto grad_op = new OpDesc();
     grad_op->SetInput(GradVarName("Out"), OutputGrad("Out"));
     grad_op->SetOutput(GradVarName("X"), InputGrad("X"));
     grad_op->SetOutput(GradVarName("b"), InputGrad("b"));
     grad_op->SetType("rowwise_add_grad");
-    return std::unique_ptr<OpDescBind>(grad_op);
+    return std::unique_ptr<OpDesc>(grad_op);
   }
 };
 
@@ -159,14 +159,14 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker {
   FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "x");
-    AddOutput("Y", "out");
+    AddOutput("Out", "out");
     AddComment("");
   }
 };
 
 class SumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SumOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  SumOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "the input tensors of sum operator.").AsDuplicable();
     AddOutput("Out", "the output tensor of sum operator.");
@@ -190,11 +190,11 @@ class MinusGradOpDescMaker : public GradOpDescMakerBase {
  public:
   using GradOpDescMakerBase::GradOpDescMakerBase;
 
-  std::vector<std::unique_ptr<OpDescBind>> operator()() const override {
-    std::vector<std::unique_ptr<OpDescBind>> retv;
+  std::vector<std::unique_ptr<OpDesc>> operator()() const override {
+    std::vector<std::unique_ptr<OpDesc>> retv;
     auto x_g = InputGrad("X");
     if (!x_g.empty()) {
-      auto *op_desc = new OpDescBind();
+      auto *op_desc = new OpDesc();
       op_desc->SetType("scale");
       op_desc->SetInput("X", OutputGrad("Out"));
       op_desc->SetOutput("Out", x_g);
@@ -204,7 +204,7 @@ class MinusGradOpDescMaker : public GradOpDescMakerBase {
 
     auto y_g = InputGrad("Y");
     if (!y_g.empty()) {
-      auto *op_desc = new OpDescBind();
+      auto *op_desc = new OpDesc();
       op_desc->SetType("scale");
       op_desc->SetInput("X", OutputGrad("Out"));
       op_desc->SetOutput("Out", y_g);
@@ -430,8 +430,8 @@ TEST(Backward, op_part_of_output_are_not_need) {
   ASSERT_EQ("fill_zeros_like", fill_zero.Type());
   ASSERT_EQ(1UL, fill_zero.Inputs("X").size());
   ASSERT_EQ("Z", fill_zero.Input("X"));
-  ASSERT_EQ(1UL, fill_zero.Outputs("Y").size());
-  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Y"));
+  ASSERT_EQ(1UL, fill_zero.Outputs("Out").size());
+  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Out"));
 
   auto &d_many_out = *net->ops_[1];
   ASSERT_EQ("many_output_op_grad", d_many_out.Type());
@@ -505,25 +505,25 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
 }
 
 TEST(Backward, simple_single_op) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
 
-  f::OpDescBind *op = block->AppendOp();
+  f::OpDesc *op = block->AppendOp();
   op->SetType("rowwise_add");
   op->SetInput("X", {"x"});
   op->SetInput("b", {"b"});
   op->SetOutput("Out", {"out"});
 
-  auto target = f::VarDescBind("out");
+  auto target = f::VarDesc("out");
   target.SetShape({1});
   auto var_to_grad =
       AppendBackward(program, target, std::unordered_set<std::string>{});
 
   ASSERT_EQ(block->AllOps().size(), 3UL);
-  f::OpDescBind *fill_op = block->AllOps()[1];
+  f::OpDesc *fill_op = block->AllOps()[1];
   EXPECT_EQ(fill_op->Type(), "fill_constant");
 
-  f::OpDescBind *grad_op = block->AllOps()[2];
+  f::OpDesc *grad_op = block->AllOps()[2];
   EXPECT_EQ(grad_op->Type(), "rowwise_add_grad");
   ASSERT_EQ(grad_op->InputNames().size(), 1UL);
   ASSERT_EQ(grad_op->OutputNames().size(), 2UL);
@@ -543,16 +543,16 @@ TEST(Backward, simple_single_op) {
 }
 
 TEST(Backward, default_attribute) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
-  f::OpDescBind *op = block->AppendOp();
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
+  f::OpDesc *op = block->AppendOp();
   op->SetType("mul");
   op->SetInput("X", {"x"});
   op->SetInput("Y", {"y"});
   op->SetOutput("Out", {"out"});
   op->CheckAttrs();
 
-  auto target = f::VarDescBind("out");
+  auto target = f::VarDesc("out");
   target.SetShape({1});
   AppendBackward(program, target, std::unordered_set<std::string>{});
 
@@ -560,47 +560,47 @@ TEST(Backward, default_attribute) {
   EXPECT_EQ(boost::get<int>(op->GetAttr("x_num_col_dims")), 1);
   EXPECT_EQ(boost::get<int>(op->GetAttr("y_num_col_dims")), 1);
 
-  f::OpDescBind *fill_op = block->AllOps()[1];
+  f::OpDesc *fill_op = block->AllOps()[1];
   EXPECT_EQ(fill_op->Type(), "fill_constant");
 
-  f::OpDescBind *grad_op = block->AllOps()[2];
+  f::OpDesc *grad_op = block->AllOps()[2];
   ASSERT_EQ(grad_op->Type(), "mul_grad");
   EXPECT_EQ(boost::get<int>(grad_op->GetAttr("x_num_col_dims")), 1);
   EXPECT_EQ(boost::get<int>(grad_op->GetAttr("y_num_col_dims")), 1);
 }
 
 TEST(Backward, simple_mult_op) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
-  f::OpDescBind *op1 = block->AppendOp();
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
+  f::OpDesc *op1 = block->AppendOp();
   op1->SetType("rowwise_add");
   op1->SetInput("X", {"x1"});
   op1->SetInput("b", {"b1"});
   op1->SetOutput("Out", {"out1"});
 
-  f::OpDescBind *op2 = block->AppendOp();
+  f::OpDesc *op2 = block->AppendOp();
   op2->SetType("mul");
   op2->SetInput("X", {"out1"});
   op2->SetInput("Y", {"y2"});
   op2->SetOutput("Out", {"out2"});
 
-  f::OpDescBind *op3 = block->AppendOp();
+  f::OpDesc *op3 = block->AppendOp();
   op3->SetType("rowwise_add");
   op3->SetInput("X", {"out2"});
   op3->SetInput("b", {"b3"});
   op3->SetOutput("Out", {"out3"});
 
-  auto target = f::VarDescBind("out3");
+  auto target = f::VarDesc("out3");
   target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad =
       AppendBackward(program, target, std::unordered_set<std::string>{});
 
   ASSERT_EQ(block->AllOps().size(), 6UL + 1);
-  f::OpDescBind *fill_op = block->AllOps()[forward_len];
+  f::OpDesc *fill_op = block->AllOps()[forward_len];
   EXPECT_EQ(fill_op->Type(), "fill_constant");
 
-  f::OpDescBind *grad_op1 = block->AllOps()[6];
+  f::OpDesc *grad_op1 = block->AllOps()[6];
   EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad");
   ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
   ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
@@ -611,7 +611,7 @@ TEST(Backward, simple_mult_op) {
   EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
             std::vector<std::string>({f::GradVarName("b1")}));
 
-  f::OpDescBind *grad_op2 = block->AllOps()[5];
+  f::OpDesc *grad_op2 = block->AllOps()[5];
   EXPECT_EQ(grad_op2->Type(), "mul_grad");
   ASSERT_EQ(grad_op2->InputNames().size(), 4UL);
   ASSERT_EQ(grad_op2->OutputNames().size(), 2UL);
@@ -625,7 +625,7 @@ TEST(Backward, simple_mult_op) {
   EXPECT_EQ(grad_op2->Output(f::GradVarName("Y")),
             std::vector<std::string>({f::GradVarName("y2")}));
 
-  f::OpDescBind *grad_op3 = block->AllOps()[4];
+  f::OpDesc *grad_op3 = block->AllOps()[4];
   EXPECT_EQ(grad_op3->Type(), "rowwise_add_grad");
   ASSERT_EQ(grad_op3->InputNames().size(), 1UL);
   ASSERT_EQ(grad_op3->OutputNames().size(), 2UL);
@@ -655,42 +655,42 @@ TEST(Backward, simple_mult_op) {
 }
 
 TEST(Backward, intermedia_var_no_grad) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
-  f::OpDescBind *op1 = block->AppendOp();
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
+  f::OpDesc *op1 = block->AppendOp();
   op1->SetType("rowwise_add");
   op1->SetInput("X", {"x1"});
   op1->SetInput("b", {"b1"});
   op1->SetOutput("Out", {"out1"});
 
-  f::OpDescBind *op2 = block->AppendOp();
+  f::OpDesc *op2 = block->AppendOp();
   op2->SetType("mul");
   op2->SetInput("X", {"x2"});
   op2->SetInput("Y", {"y2"});
   op2->SetOutput("Out", {"out2"});
 
-  f::OpDescBind *op3 = block->AppendOp();
+  f::OpDesc *op3 = block->AppendOp();
   op3->SetType("rowwise_add");
   op3->SetInput("X", {"out2"});
   op3->SetInput("b", {"b3"});
   op3->SetOutput("Out", {"out3"});
 
-  f::OpDescBind *op4 = block->AppendOp();
+  f::OpDesc *op4 = block->AppendOp();
   op4->SetType("mul");
   op4->SetInput("X", {"out1"});
   op4->SetInput("Y", {"out3"});
   op4->SetOutput("Out", {"out4"});
 
-  auto target = f::VarDescBind("out4");
+  auto target = f::VarDesc("out4");
   target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad = AppendBackward(program, target, {"out3"});
 
   ASSERT_EQ(block->AllOps().size(), 7UL);
-  f::OpDescBind *fill_op = block->AllOps()[forward_len];
+  f::OpDesc *fill_op = block->AllOps()[forward_len];
   EXPECT_EQ(fill_op->Type(), "fill_constant");
 
-  f::OpDescBind *grad_op1 = block->AllOps()[6];
+  f::OpDesc *grad_op1 = block->AllOps()[6];
   EXPECT_EQ(grad_op1->Type(), "rowwise_add_grad");
   ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
   ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
@@ -701,7 +701,7 @@ TEST(Backward, intermedia_var_no_grad) {
   EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
             std::vector<std::string>({f::GradVarName("b1")}));
 
-  f::OpDescBind *grad_op4 = block->AllOps()[5];
+  f::OpDesc *grad_op4 = block->AllOps()[5];
   EXPECT_EQ(grad_op4->Type(), "mul_grad");
   ASSERT_EQ(grad_op4->InputNames().size(), 4UL);
   ASSERT_EQ(grad_op4->OutputNames().size(), 2UL);
@@ -726,32 +726,32 @@ TEST(Backward, intermedia_var_no_grad) {
 }
 
 TEST(Backward, var_no_grad) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
-  f::OpDescBind *op1 = block->AppendOp();
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
+  f::OpDesc *op1 = block->AppendOp();
   op1->SetType("mult_in_out");
   op1->SetInput("X", {"x1"});
   op1->SetInput("H", {"h1"});
   op1->SetOutput("Y", {"y1"});
   op1->SetOutput("Z", {"z1"});
 
-  f::OpDescBind *op2 = block->AppendOp();
+  f::OpDesc *op2 = block->AppendOp();
   op2->SetType("mult_in_out");
   op2->SetInput("X", {"y1"});
   op2->SetInput("H", {"z1"});
   op2->SetOutput("Y", {"y2"});
   op2->SetOutput("Z", {"z2"});
 
-  auto target = f::VarDescBind("z2");
+  auto target = f::VarDesc("z2");
   target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad = AppendBackward(program, target, {"z1"});
 
   ASSERT_EQ(block->AllOps().size(), 6UL);
-  f::OpDescBind *fill_op = block->AllOps()[forward_len];
+  f::OpDesc *fill_op = block->AllOps()[forward_len];
   EXPECT_EQ(fill_op->Type(), "fill_constant");
 
-  f::OpDescBind *grad_op2 = block->AllOps()[3];
+  f::OpDesc *grad_op2 = block->AllOps()[3];
   ASSERT_EQ(grad_op2->Type(), "mult_in_out_grad");
   ASSERT_EQ(grad_op2->InputNames().size(), 6UL);
   ASSERT_EQ(grad_op2->OutputNames().size(), 2UL);
@@ -767,15 +767,15 @@ TEST(Backward, var_no_grad) {
             std::vector<std::string>({f::GradVarName("y1")}));
   EXPECT_EQ(grad_op2->Output(f::GradVarName("H")), std::vector<std::string>());
 
-  f::OpDescBind *fill_zero_op = block->AllOps()[4];
+  f::OpDesc *fill_zero_op = block->AllOps()[4];
   ASSERT_EQ(fill_zero_op->Type(), "fill_zeros_like");
   ASSERT_EQ(fill_zero_op->InputNames().size(), 1UL);
   ASSERT_EQ(fill_zero_op->OutputNames().size(), 1UL);
   EXPECT_EQ(fill_zero_op->Input("X"), std::vector<std::string>({"z1"}));
-  EXPECT_EQ(fill_zero_op->Output("Y"),
+  EXPECT_EQ(fill_zero_op->Output("Out"),
             std::vector<std::string>({std::string("z1") + f::kZeroVarSuffix}));
 
-  f::OpDescBind *grad_op1 = block->AllOps()[5];
+  f::OpDesc *grad_op1 = block->AllOps()[5];
   ASSERT_EQ(grad_op1->Type(), "mult_in_out_grad");
   ASSERT_EQ(grad_op1->InputNames().size(), 6UL);
   ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
@@ -803,37 +803,37 @@ TEST(Backward, var_no_grad) {
 }
 
 TEST(Backward, shared_var) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
-  f::OpDescBind *op1 = block->AppendOp();
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
+  f::OpDesc *op1 = block->AppendOp();
   op1->SetType("rowwise_add");
   op1->SetInput("X", {"x1"});
   op1->SetInput("b", {"b1"});
   op1->SetOutput("Out", {"out1"});
 
-  f::OpDescBind *op2 = block->AppendOp();
+  f::OpDesc *op2 = block->AppendOp();
   op2->SetType("mul");
   op2->SetInput("X", {"out1"});
   op2->SetInput("Y", {"y2"});
   op2->SetOutput("Out", {"out2"});
 
-  f::OpDescBind *op3 = block->AppendOp();
+  f::OpDesc *op3 = block->AppendOp();
   op3->SetType("rowwise_add");
   op3->SetInput("X", {"out1"});
   op3->SetInput("b", {"b3"});
   op3->SetOutput("Out", {"out3"});
 
-  auto target = f::VarDescBind("out3");
+  auto target = f::VarDesc("out3");
   target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad =
       AppendBackward(program, target, std::unordered_set<std::string>{});
 
   ASSERT_EQ(block->AllOps().size(), 8UL);
-  f::OpDescBind *fill_op = block->AllOps()[forward_len];
+  f::OpDesc *fill_op = block->AllOps()[forward_len];
   EXPECT_EQ(fill_op->Type(), "fill_constant");
 
-  f::OpDescBind *grad_op3 = block->AllOps()[4];
+  f::OpDesc *grad_op3 = block->AllOps()[4];
   ASSERT_EQ(grad_op3->Type(), "rowwise_add_grad");
   ASSERT_EQ(grad_op3->InputNames().size(), 1UL);
   ASSERT_EQ(grad_op3->OutputNames().size(), 2UL);
@@ -844,7 +844,7 @@ TEST(Backward, shared_var) {
   EXPECT_EQ(grad_op3->Output(f::GradVarName("b")),
             std::vector<std::string>({f::GradVarName("b3")}));
 
-  f::OpDescBind *grad_op4 = block->AllOps()[5];
+  f::OpDesc *grad_op4 = block->AllOps()[5];
   ASSERT_EQ(grad_op4->Type(), "mul_grad");
   ASSERT_EQ(grad_op4->InputNames().size(), 4UL);
   ASSERT_EQ(grad_op4->OutputNames().size(), 2UL);
@@ -858,7 +858,7 @@ TEST(Backward, shared_var) {
   EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")),
             std::vector<std::string>({f::GradVarName("y2")}));
 
-  f::OpDescBind *sum_op = block->AllOps()[6];
+  f::OpDesc *sum_op = block->AllOps()[6];
   ASSERT_EQ(sum_op->Type(), "sum");
   ASSERT_EQ(sum_op->InputNames().size(), 1UL);
   ASSERT_EQ(sum_op->OutputNames().size(), 1UL);
@@ -868,7 +868,7 @@ TEST(Backward, shared_var) {
   EXPECT_EQ(sum_op->Output("Out"),
             std::vector<std::string>({f::GradVarName("out1")}));
 
-  f::OpDescBind *grad_op1 = block->AllOps()[7];
+  f::OpDesc *grad_op1 = block->AllOps()[7];
   ASSERT_EQ(grad_op1->Type(), "rowwise_add_grad");
   ASSERT_EQ(grad_op1->InputNames().size(), 1UL);
   ASSERT_EQ(grad_op1->OutputNames().size(), 2UL);
@@ -895,19 +895,19 @@ TEST(Backward, shared_var) {
 }
 
 TEST(Backward, half_backward) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
   auto *op1 = block->AppendOp();
   op1->SetType("minus");
   op1->SetInput("X", {"a"});
   op1->SetInput("Y", {"b"});
   op1->SetOutput("Out", {"out"});
 
-  auto target = f::VarDescBind("out");
+  auto target = f::VarDesc("out");
   target.SetShape({1});
   size_t forward_len = block->AllOps().size();
   auto var_to_grad = AppendBackward(program, target, {"b"});
-  f::OpDescBind *fill_op = block->AllOps()[forward_len];
+  f::OpDesc *fill_op = block->AllOps()[forward_len];
   EXPECT_EQ(fill_op->Type(), "fill_constant");
   auto ops = block->AllOps();
   ASSERT_EQ(3UL, ops.size());

paddle/framework/block_desc.cc

@@ -19,18 +19,18 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-VarDescBind *BlockDescBind::Var(const std::string &name) {
+VarDesc *BlockDesc::Var(const std::string &name) {
   auto it = vars_.find(name);
   if (it != vars_.end()) {
     return it->second.get();
   }
   need_update_ = true;
-  auto *var = new VarDescBind(name);
+  auto *var = new VarDesc(name);
   vars_[name].reset(var);
   return var;
 }
 
-VarDescBind *BlockDescBind::FindVar(const std::string &name) const {
+VarDesc *BlockDesc::FindVar(const std::string &name) const {
   auto it = vars_.find(name);
   if (it == vars_.end()) {
     return nullptr;
@@ -38,11 +38,11 @@ VarDescBind *BlockDescBind::FindVar(const std::string &name) const {
   return it->second.get();
 }
 
-bool BlockDescBind::HasVar(const std::string &name) const {
+bool BlockDesc::HasVar(const std::string &name) const {
   return vars_.find(name) != vars_.end();
 }
 
-VarDescBind *BlockDescBind::FindVarRecursive(const std::string &name) const {
+VarDesc *BlockDesc::FindVarRecursive(const std::string &name) const {
   if (name == kEmptyVarName) return nullptr;
 
   auto it = vars_.find(name);
@@ -53,53 +53,67 @@ VarDescBind *BlockDescBind::FindVarRecursive(const std::string &name) const {
   return it->second.get();
 }
 
-VarDescBind *BlockDescBind::FindRecursiveOrCreateVar(
-    const std::string &name_bytes) {
-  VarDescBind *res = FindVarRecursive(name_bytes);
+VarDesc *BlockDesc::FindRecursiveOrCreateVar(const std::string &name_bytes) {
+  VarDesc *res = FindVarRecursive(name_bytes);
   if (res == nullptr) {
     res = Var(name_bytes);
   }
   return res;
 }
 
-bool BlockDescBind::HasVarRecursive(const std::string &name) const {
+bool BlockDesc::HasVarRecursive(const std::string &name) const {
   return FindVarRecursive(name) != nullptr;
 }
 
-std::vector<VarDescBind *> BlockDescBind::AllVars() const {
-  std::vector<VarDescBind *> res;
+std::vector<VarDesc *> BlockDesc::AllVars() const {
+  std::vector<VarDesc *> res;
   for (const auto &p : vars_) {
     res.push_back(p.second.get());
   }
   return res;
 }
 
-OpDescBind *BlockDescBind::AppendOp() {
+OpDesc *BlockDesc::AppendOp() {
   need_update_ = true;
-  ops_.emplace_back(new OpDescBind());
+  ops_.emplace_back(new OpDesc());
   return ops_.back().get();
 }
 
-void BlockDescBind::AppendAllocatedOp(std::unique_ptr<OpDescBind> &&op_desc) {
+void BlockDesc::AppendAllocatedOp(std::unique_ptr<OpDesc> &&op_desc) {
   need_update_ = true;
   ops_.emplace_back(std::move(op_desc));
 }
 
-OpDescBind *BlockDescBind::PrependOp() {
+OpDesc *BlockDesc::PrependOp() {
   need_update_ = true;
-  ops_.emplace_front(new OpDescBind());
+  ops_.emplace_front(new OpDesc());
   return ops_.front().get();
 }
 
-std::vector<OpDescBind *> BlockDescBind::AllOps() const {
-  std::vector<OpDescBind *> res;
+void BlockDesc::RemoveOp(size_t s, size_t e) {
+  if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) {
+    return;
+  }
+  need_update_ = true;
+  for (auto it = ops_.begin() + s; it != ops_.begin() + e; it++) {
+    auto names = (*it)->InputArgumentNames();
+    for (auto n : names) {
+      // TODO(typhoonzero): delete vars if no other op use it.
+      VLOG(3) << "deleting var " << n;
+    }
+  }
+  ops_.erase(ops_.begin() + s, ops_.begin() + e);
+}
+
+std::vector<OpDesc *> BlockDesc::AllOps() const {
+  std::vector<OpDesc *> res;
   for (const auto &op : ops_) {
     res.push_back(op.get());
   }
   return res;
 }
 
-void BlockDescBind::Flush() {
+void BlockDesc::Flush() {
   for (auto &op_desc : ops_) {
     op_desc->Flush();
   }
@@ -121,43 +135,43 @@ void BlockDescBind::Flush() {
   }
 }
 
-BlockDescBind *BlockDescBind::ParentBlock() const {
+BlockDesc *BlockDesc::ParentBlock() const {
   if (this->desc_->parent_idx() == kNoneBlockIndex) {
     return nullptr;
   }
   return prog_->MutableBlock(static_cast<size_t>(this->desc_->parent_idx()));
 }
 
-BlockDesc *BlockDescBind::Proto() {
+proto::BlockDesc *BlockDesc::Proto() {
   Flush();
   return desc_;
 }
 
-BlockDescBind::BlockDescBind(ProgramDescBind *prog, BlockDesc *desc)
+BlockDesc::BlockDesc(ProgramDesc *prog, proto::BlockDesc *desc)
     : prog_(prog), desc_(desc), need_update_(false) {
-  for (const VarDesc &var_desc : desc_->vars()) {
-    vars_[var_desc.name()].reset(new VarDescBind(var_desc));
+  for (const proto::VarDesc &var_desc : desc_->vars()) {
+    vars_[var_desc.name()].reset(new VarDesc(var_desc));
   }
-  for (const OpDesc &op_desc : desc_->ops()) {
-    ops_.emplace_back(new OpDescBind(op_desc, prog));
+  for (const proto::OpDesc &op_desc : desc_->ops()) {
+    ops_.emplace_back(new OpDesc(op_desc, prog));
   }
 }
 
-BlockDescBind::BlockDescBind(const BlockDescBind &other, BlockDesc *desc,
-                             ProgramDescBind *prog)
+BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc,
+                     ProgramDesc *prog)
     : prog_(prog), desc_(desc) {
   need_update_ = true;
   for (auto &op : other.ops_) {
-    ops_.emplace_back(new OpDescBind(*op));
+    ops_.emplace_back(new OpDesc(*op));
   }
 
   for (auto &it : other.vars_) {
-    auto *var = new VarDescBind(*it.second);
+    auto *var = new VarDesc(*it.second);
     vars_[it.first].reset(var);
   }
 }
 
-void BlockDescBind::ClearPBOps() {
+void BlockDesc::ClearPBOps() {
   auto ops = this->desc_->mutable_ops();
   while (!ops->empty()) {
     // we do not own the OpDesc, so release the ownership.
@@ -165,7 +179,7 @@ void BlockDescBind::ClearPBOps() {
   }
 }
 
-void BlockDescBind::ClearPBVars() {
+void BlockDesc::ClearPBVars() {
   auto vars = this->desc_->mutable_vars();
   while (!vars->empty()) {
     // we do not own the VarDesc, so release the ownership.

paddle/framework/block_desc.h

@@ -28,20 +28,19 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class ProgramDescBind;
+class ProgramDesc;
 
 // Each Protobuf Message, we provide a XXXBind class. In that class, we optimize
 // read/write speed. Only when we want the protobuf message, the local changes
 // will be synchronized (by `Sync` method).
 
-class BlockDescBind {
+class BlockDesc {
  public:
-  BlockDescBind(ProgramDescBind *prog, BlockDesc *desc);
+  BlockDesc(ProgramDesc *prog, proto::BlockDesc *desc);
 
-  BlockDescBind(const BlockDescBind &other, BlockDesc *desc,
-                ProgramDescBind *prog);
+  BlockDesc(const BlockDesc &other, proto::BlockDesc *desc, ProgramDesc *prog);
 
-  ~BlockDescBind() {
+  ~BlockDesc() {
     this->ClearPBVars();
     this->ClearPBOps();
   }
@@ -50,15 +49,15 @@ class BlockDescBind {
 
   int32_t Parent() const { return desc_->parent_idx(); }
 
-  VarDescBind *Var(const std::string &name_bytes);
+  VarDesc *Var(const std::string &name_bytes);
 
-  VarDescBind *FindVar(const std::string &name_bytes) const;
+  VarDesc *FindVar(const std::string &name_bytes) const;
 
   bool HasVar(const std::string &var_name) const;
 
-  VarDescBind *FindVarRecursive(const std::string &name_bytes) const;
+  VarDesc *FindVarRecursive(const std::string &name_bytes) const;
 
-  VarDescBind *FindRecursiveOrCreateVar(const std::string &name_bytes);
+  VarDesc *FindRecursiveOrCreateVar(const std::string &name_bytes);
 
   bool HasVarRecursive(const std::string &var_name) const;
 
@@ -70,41 +69,43 @@ class BlockDescBind {
     return var_names;
   }
 
-  std::vector<VarDescBind *> AllVars() const;
+  std::vector<VarDesc *> AllVars() const;
 
-  BlockDescBind *ParentBlock() const;
+  BlockDesc *ParentBlock() const;
 
-  OpDescBind *AppendOp();
+  OpDesc *AppendOp();
 
-  void AppendAllocatedOp(std::unique_ptr<OpDescBind> &&op_desc);
+  void AppendAllocatedOp(std::unique_ptr<OpDesc> &&op_desc);
 
-  OpDescBind *PrependOp();
+  OpDesc *PrependOp();
 
-  std::vector<OpDescBind *> AllOps() const;
+  void RemoveOp(size_t s, size_t e);
+
+  std::vector<OpDesc *> AllOps() const;
 
   size_t OpSize() const { return ops_.size(); }
 
-  OpDescBind *Op(int idx) { return ops_.at(idx).get(); }
+  OpDesc *Op(int idx) { return ops_.at(idx).get(); }
 
   void Flush();
 
-  BlockDesc *Proto();
+  proto::BlockDesc *Proto();
 
-  ProgramDescBind *Program() { return this->prog_; }
+  ProgramDesc *Program() { return this->prog_; }
 
  private:
   void ClearPBOps();
   void ClearPBVars();
 
  private:
-  ProgramDescBind *prog_;  // not_own
-  BlockDesc *desc_;        // not_own
+  ProgramDesc *prog_;       // not_own
+  proto::BlockDesc *desc_;  // not_own
   bool need_update_;
 
-  std::deque<std::unique_ptr<OpDescBind>> ops_;
-  std::unordered_map<std::string, std::unique_ptr<VarDescBind>> vars_;
+  std::deque<std::unique_ptr<OpDesc>> ops_;
+  std::unordered_map<std::string, std::unique_ptr<VarDesc>> vars_;
 
-  DISABLE_COPY_AND_ASSIGN(BlockDescBind);
+  DISABLE_COPY_AND_ASSIGN(BlockDesc);
 };
 }  // namespace framework
 }  // namespace paddle

paddle/framework/data_layout.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+namespace paddle {
+namespace framework {
+
+enum DataLayout {
+  kNHWC = 0,
+  kNCHW = 1,
+  kAnyLayout = 2,
+};
+
+inline DataLayout StringToDataLayout(const std::string& str) {
+  if (str == "NHWC" || str == "nhwc") {
+    return DataLayout::kNHWC;
+  } else if (str == "NCHW" || str == "nchw") {
+    return DataLayout::kNCHW;
+  } else {
+    PADDLE_THROW("Unknown storage order string: %s", str);
+  }
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/data_type.h

@@ -20,7 +20,8 @@
 namespace paddle {
 namespace framework {
 
-inline DataType ToDataType(std::type_index type) {
+inline proto::DataType ToDataType(std::type_index type) {
+  using namespace paddle::framework::proto;
   if (typeid(float).hash_code() == type.hash_code()) {
     return DataType::FP32;
   } else if (typeid(double).hash_code() == type.hash_code()) {
@@ -36,7 +37,8 @@ inline DataType ToDataType(std::type_index type) {
   }
 }
 
-inline std::type_index ToTypeIndex(DataType type) {
+inline std::type_index ToTypeIndex(proto::DataType type) {
+  using namespace paddle::framework::proto;
   switch (type) {
     case DataType::FP32:
       return typeid(float);
@@ -54,7 +56,8 @@ inline std::type_index ToTypeIndex(DataType type) {
 }
 
 template <typename Visitor>
-inline void VisitDataType(DataType type, Visitor visitor) {
+inline void VisitDataType(proto::DataType type, Visitor visitor) {
+  using namespace paddle::framework::proto;
   switch (type) {
     case DataType::FP32:
       visitor.template operator()<float>();

paddle/framework/details/op_registry.h

@@ -90,7 +90,7 @@ struct OpInfoFiller<T, kOperator> {
 template <typename T>
 struct OpInfoFiller<T, kOpProtoAndCheckerMaker> {
   void operator()(const char* op_type, OpInfo* info) const {
-    info->proto_ = new OpProto;
+    info->proto_ = new proto::OpProto;
     info->checker_ = new OpAttrChecker();
     auto maker = T(info->proto_, info->checker_);
     maker.Validate();
@@ -106,10 +106,10 @@ template <typename T>
 struct OpInfoFiller<T, kGradOpDescMaker> {
   void operator()(const char* op_type, OpInfo* info) const {
     info->grad_op_maker_ = [](
-        const OpDescBind& fwd_op,
+        const OpDesc& fwd_op,
         const std::unordered_set<std::string>& no_grad_set,
         std::unordered_map<std::string, std::string>* grad_to_var,
-        const std::vector<BlockDescBind*>& grad_block) {
+        const std::vector<BlockDesc*>& grad_block) {
       T maker(fwd_op, no_grad_set, grad_to_var, grad_block);
       return maker();
     };
@@ -119,7 +119,7 @@ struct OpInfoFiller<T, kGradOpDescMaker> {
 template <typename T>
 struct OpInfoFiller<T, kVarTypeInference> {
   void operator()(const char* op_type, OpInfo* info) const {
-    info->infer_var_type_ = [](const OpDescBind& fwd_op, BlockDescBind* block) {
+    info->infer_var_type_ = [](const OpDesc& fwd_op, BlockDesc* block) {
       T inference;
       inference(fwd_op, block);
     };

paddle/framework/executor.cc

@@ -41,20 +41,20 @@ Executor::Executor(const std::vector<platform::Place>& places) {
   device_contexts_.swap(borrowed_contexts);
 }
 
-static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
-  if (var_type == VarDesc::LOD_TENSOR) {
+static void CreateTensor(Variable* var, proto::VarDesc::VarType var_type) {
+  if (var_type == proto::VarDesc::LOD_TENSOR) {
     var->GetMutable<LoDTensor>();
-  } else if (var_type == VarDesc::SELECTED_ROWS) {
+  } else if (var_type == proto::VarDesc::SELECTED_ROWS) {
     var->GetMutable<SelectedRows>();
-  } else if (var_type == VarDesc::FEED_MINIBATCH) {
+  } else if (var_type == proto::VarDesc::FEED_MINIBATCH) {
     var->GetMutable<FeedFetchList>();
-  } else if (var_type == VarDesc::FETCH_LIST) {
+  } else if (var_type == proto::VarDesc::FETCH_LIST) {
     var->GetMutable<FeedFetchList>();
-  } else if (var_type == VarDesc::STEP_SCOPES) {
+  } else if (var_type == proto::VarDesc::STEP_SCOPES) {
     var->GetMutable<std::vector<framework::Scope>>();
-  } else if (var_type == VarDesc::LOD_RANK_TABLE) {
+  } else if (var_type == proto::VarDesc::LOD_RANK_TABLE) {
     var->GetMutable<LoDRankTable>();
-  } else if (var_type == VarDesc::LOD_TENSOR_ARRAY) {
+  } else if (var_type == proto::VarDesc::LOD_TENSOR_ARRAY) {
     var->GetMutable<LoDTensorArray>();
   } else {
     PADDLE_THROW(
@@ -64,8 +64,8 @@ static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
   }
 }
 
-void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
-                   bool create_local_scope) {
+void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
+                   bool create_local_scope, bool create_vars) {
   // TODO(tonyyang-svail):
   //    - only runs on the first device (i.e. no interdevice communication)
   //    - will change to use multiple blocks for RNN op and Cond Op
@@ -74,33 +74,35 @@ void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
   auto& device = device_contexts_[0];
 
   Scope* local_scope = scope;
-  if (create_local_scope) {
-    local_scope = &scope->NewScope();
-    for (auto& var : block.AllVars()) {
-      if (var->Name() == framework::kEmptyVarName) {
-        continue;
+  if (create_vars) {
+    if (create_local_scope) {
+      local_scope = &scope->NewScope();
+      for (auto& var : block.AllVars()) {
+        if (var->Name() == framework::kEmptyVarName) {
+          continue;
+        }
+
+        if (var->Persistable()) {
+          auto* ptr = scope->Var(var->Name());
+          CreateTensor(ptr, var->GetType());
+          VLOG(3) << "Create Variable " << var->Name()
+                  << " global, which pointer is " << ptr;
+        } else {
+          auto* ptr = local_scope->Var(var->Name());
+          CreateTensor(ptr, var->GetType());
+          VLOG(3) << "Create Variable " << var->Name()
+                  << " locally, which pointer is " << ptr;
+        }
       }
-
-      if (var->Persistable()) {
-        auto* ptr = scope->Var(var->Name());
-        CreateTensor(ptr, var->GetType());
-        VLOG(3) << "Create Variable " << var->Name()
-                << " global, which pointer is " << ptr;
-      } else {
+    } else {
+      for (auto& var : block.AllVars()) {
         auto* ptr = local_scope->Var(var->Name());
         CreateTensor(ptr, var->GetType());
-        VLOG(3) << "Create Variable " << var->Name()
-                << " locally, which pointer is " << ptr;
+        VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
+                << ptr;
       }
-    }
-  } else {
-    for (auto& var : block.AllVars()) {
-      auto* ptr = local_scope->Var(var->Name());
-      CreateTensor(ptr, var->GetType());
-      VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
-              << ptr;
-    }
-  }
+    }  // if (create_local_scope)
+  }    // if (create_vars)
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);

paddle/framework/executor.h

@@ -40,6 +40,16 @@ class DeviceContextPool {
     return *pool;
   }
 
+  const platform::DeviceContext* Borrow(const platform::Place& place) {
+    auto range = device_contexts_.equal_range(place);
+    if (range.first == range.second) {
+      PADDLE_THROW(
+          "'Place' is not supported, Please re-compile with WITH_GPU "
+          "option");
+    }
+    return range.first->second;
+  }
+
   std::vector<const platform::DeviceContext*> Borrow(
       const std::vector<platform::Place>& places) {
     PADDLE_ENFORCE_GT(places.size(), 0);
@@ -114,7 +124,8 @@ class Executor {
    *  ProgramDesc
    *  Scope
    */
-  void Run(const ProgramDescBind&, Scope*, int, bool create_local_scope = true);
+  void Run(const ProgramDesc&, Scope*, int, bool create_local_scope = true,
+           bool create_vars = true);
 
  private:
   std::vector<const platform::DeviceContext*> device_contexts_;

paddle/framework/framework.proto

@@ -14,7 +14,7 @@ limitations under the License. */
 
 syntax = "proto2";
 option optimize_for = LITE_RUNTIME;
-package paddle.framework;
+package paddle.framework.proto;
 
 enum AttrType {
   INT = 0;

paddle/framework/grad_op_desc_maker.h

@@ -22,21 +22,27 @@
 namespace paddle {
 namespace framework {
 
+/*
+  This functor class is responsible for creating the gradient ops for the given
+  operator fwd_op. After it is called (through operator()), the pairs of
+  (gradient variable, corresponding input variable of fwd_op) will be added to
+  grad_to_var. If an input variable of fwd_op is contained in no_grad_set, its
+  gradient varialbe will be ignored or kEmptyVarName depending on the template
+  argument DropEmptyIG in the derived classes.
+ */
 class GradOpDescMakerBase {
  public:
   explicit GradOpDescMakerBase(
-      const OpDescBind& fwd_op,
-      const std::unordered_set<std::string>& no_grad_set,
+      const OpDesc& fwd_op, const std::unordered_set<std::string>& no_grad_set,
       std::unordered_map<std::string, std::string>* grad_to_var,
-      const std::vector<BlockDescBind*>& grad_block =
-          std::vector<BlockDescBind*>())
+      const std::vector<BlockDesc*>& grad_block = std::vector<BlockDesc*>())
       : fwd_op_(fwd_op),
         no_grad_set_(no_grad_set),
         grad_to_var_(grad_to_var),
         grad_block_(grad_block) {}
 
   virtual ~GradOpDescMakerBase() = default;
-  virtual std::vector<std::unique_ptr<OpDescBind>> operator()() const = 0;
+  virtual std::vector<std::unique_ptr<OpDesc>> operator()() const = 0;
 
  protected:
   std::vector<std::string> InputGrad(const std::string& name,
@@ -58,6 +64,16 @@ class GradOpDescMakerBase {
     if (!drop_empty_grad) {
       return ret_val;
     }
+    PADDLE_ENFORCE_LE(var_names.size(), 1UL,
+                      "BUG from operator developer:"
+                      " for input argument with a list of variables, "
+                      " drop_empty_grad is not allowed because it makes"
+                      " the correspondence bewteen a variable and its gradient"
+                      " ambiguous. Use REGISTER_OP_EX to register the op"
+                      " or call InputGrad(?,false) in GradOpDescMaker."
+                      " Op type %s",
+                      fwd_op_.Type());
+
     std::vector<std::string> dropped_ret_val;
     dropped_ret_val.reserve(ret_val.size());
     std::copy_if(ret_val.begin(), ret_val.end(),
@@ -105,26 +121,26 @@ class GradOpDescMakerBase {
   std::string ForwardOpType() const { return this->fwd_op_.Type(); }
 
  private:
-  const OpDescBind& fwd_op_;
+  const OpDesc& fwd_op_;
   const std::unordered_set<std::string>& no_grad_set_;
   std::unordered_map<std::string, std::string>* grad_to_var_;
 
  protected:
-  std::vector<BlockDescBind*> grad_block_;
+  std::vector<BlockDesc*> grad_block_;
 };
 
 class SingleGradOpDescMaker : public GradOpDescMakerBase {
  public:
   using GradOpDescMakerBase::GradOpDescMakerBase;
 
-  std::vector<std::unique_ptr<OpDescBind>> operator()() const {
-    std::vector<std::unique_ptr<OpDescBind>> retv;
+  std::vector<std::unique_ptr<OpDesc>> operator()() const {
+    std::vector<std::unique_ptr<OpDesc>> retv;
     retv.emplace_back(this->Apply());
     return retv;
   }
 
  protected:
-  virtual std::unique_ptr<OpDescBind> Apply() const = 0;
+  virtual std::unique_ptr<OpDesc> Apply() const = 0;
 };
 
 template <bool DropEmptyIG = true>
@@ -133,8 +149,8 @@ class DefaultGradOpDescMaker : public SingleGradOpDescMaker {
   using SingleGradOpDescMaker::SingleGradOpDescMaker;
 
  protected:
-  virtual std::unique_ptr<OpDescBind> Apply() const {
-    auto* grad = new OpDescBind();
+  virtual std::unique_ptr<OpDesc> Apply() const {
+    auto* grad = new OpDesc();
     grad->SetType(this->GradOpType());
 
     for (auto& input_param : this->InputNames()) {
@@ -150,7 +166,7 @@ class DefaultGradOpDescMaker : public SingleGradOpDescMaker {
 
     grad->SetAttrMap(this->Attrs());
 
-    return std::unique_ptr<OpDescBind>(grad);
+    return std::unique_ptr<OpDesc>(grad);
   }
 
   virtual std::string GradOpType() const {
@@ -161,7 +177,7 @@ class DefaultGradOpDescMaker : public SingleGradOpDescMaker {
 class EmptyGradOpMaker : public GradOpDescMakerBase {
  public:
   using GradOpDescMakerBase::GradOpDescMakerBase;
-  std::vector<std::unique_ptr<OpDescBind>> operator()() const override {
+  std::vector<std::unique_ptr<OpDesc>> operator()() const override {
     return {};
   }
 };

paddle/framework/library_type.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+namespace paddle {
+namespace framework {
+
+// For more details about the design of LibraryType, Please refer to
+// https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md#library
+
+enum LibraryType { kPlain = 0; kMKLDNN = 1; kCUDNN = 2; }
+
+}  // namespace
+}  // framework

paddle/framework/lod_rank_table.cc

@@ -46,4 +46,13 @@ void LoDRankTable::Reset(const LoD& lod, size_t level) {
 }
 
 }  // namespace framework
+
+std::ostream& operator<<(std::ostream& out,
+                         const framework::LoDRankTable& table) {
+  out << "NumOfSequence " << table.items().size() << "\n";
+  for (auto& each_item : table.items()) {
+    out << "\tSeq #" << each_item.index << ", Len=" << each_item.length << "\n";
+  }
+  return out;
+}
 }  // namespace paddle

paddle/framework/lod_rank_table.h

@@ -13,6 +13,7 @@
    limitations under the License. */
 
 #pragma once
+#include <iosfwd>
 #include "paddle/framework/lod_tensor.h"
 
 namespace paddle {
@@ -52,4 +53,8 @@ class LoDRankTable {
 };
 
 }  // namespace framework
+
+std::ostream& operator<<(std::ostream& out,
+                         const framework::LoDRankTable& table);
+
 }  // namespace paddle

paddle/framework/lod_tensor.cc

@@ -197,7 +197,7 @@ void SerializeToStream(std::ostream &os, const LoDTensor &tensor,
   {  // the 2nd field, tensor description
      // int32_t  size
      // void*    protobuf message
-    framework::TensorDesc desc;
+    proto::TensorDesc desc;
     desc.set_data_type(framework::ToDataType(tensor.type()));
     auto dims = framework::vectorize(tensor.dims());
     auto *pb_dims = desc.mutable_dims();
@@ -262,7 +262,7 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor) {
   uint32_t version;
   is.read(reinterpret_cast<char *>(&version), sizeof(version));
   PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
-  framework::TensorDesc desc;
+  proto::TensorDesc desc;
   {  // int32_t size
      // proto buffer
     int32_t size;
@@ -281,16 +281,16 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor) {
     void *buf;
     platform::Place cpu = platform::CPUPlace();
     switch (desc.data_type()) {
-      case framework::FP32:
+      case proto::FP32:
         buf = tensor->mutable_data<float>(cpu);
         break;
-      case framework::FP64:
+      case proto::FP64:
         buf = tensor->mutable_data<double>(cpu);
         break;
-      case framework::INT32:
+      case proto::INT32:
         buf = tensor->mutable_data<int>(cpu);
         break;
-      case framework::INT64:
+      case proto::INT64:
         buf = tensor->mutable_data<int64_t>(cpu);
         break;
       default:

paddle/framework/lod_tensor.h

@@ -184,6 +184,18 @@ LoDTensor LodExpand(const LoDTensor& source, const LoD& lod, size_t level,
   return tensor;
 }
 
+// Get the absolute offset of a lod[start_level][start_idx:end_idx] and
+// relative length of details for every levels(i.e., [start_level: ]).
+//
+// For example,
+//   lod = [[0, 3, 4, 8], [0, 9, 10, 11, 13, 17, 19, 22, 24]]
+//   start_level = 0
+//   start_idx = 1
+//   end_idx = 3
+//
+// Returns:
+//  LoD = [[1, 4], [2, 4, 2, 3, 2]]
+//  pair<size_t, size_t> = {11, 24}
 std::pair<LoD, std::pair<size_t, size_t>> GetSubLoDAndAbsoluteOffset(
     const LoD& lod, size_t start_idx, size_t end_idx, size_t start_level);
 

paddle/framework/op_desc.cc

@@ -25,12 +25,11 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class OpDescBind;
-class BlockDescBind;
+class OpDesc;
+class BlockDesc;
 class CompileTimeInferShapeContext : public InferShapeContext {
  public:
-  CompileTimeInferShapeContext(const OpDescBind &op,
-                               const BlockDescBind &block);
+  CompileTimeInferShapeContext(const OpDesc &op, const BlockDesc &block);
 
   bool HasInput(const std::string &name) const override;
 
@@ -58,11 +57,11 @@ class CompileTimeInferShapeContext : public InferShapeContext {
     PADDLE_ENFORCE_LT(j, Outputs(out).size());
     auto *in_var = block_.FindVarRecursive(Inputs(in)[i]);
     auto *out_var = block_.FindVarRecursive(Outputs(out)[j]);
-    if (in_var->GetType() != VarDesc::LOD_TENSOR) {
+    if (in_var->GetType() != proto::VarDesc::LOD_TENSOR) {
       VLOG(3) << "input " << in << " is not LodTensor";
       return;
     }
-    PADDLE_ENFORCE_EQ(in_var->GetType(), VarDesc::LOD_TENSOR,
+    PADDLE_ENFORCE_EQ(in_var->GetType(), proto::VarDesc::LOD_TENSOR,
                       "The %d-th output of Output(%s) must be LoDTensor.", j,
                       out);
     out_var->SetLoDLevel(in_var->GetLodLevel());
@@ -70,19 +69,18 @@ class CompileTimeInferShapeContext : public InferShapeContext {
   bool IsRuntime() const override;
 
  protected:
-  VarDesc::VarType GetVarType(const std::string &name) const override;
+  proto::VarDesc::VarType GetVarType(const std::string &name) const override;
 
   DDim GetDim(const std::string &name) const override;
 
   void SetDim(const std::string &name, const DDim &dim) override;
 
-  const OpDescBind &op_;
-  const BlockDescBind &block_;
+  const OpDesc &op_;
+  const BlockDesc &block_;
 };
 
-OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs,
-                       const VariableNameMap &outputs,
-                       const AttributeMap &attrs) {
+OpDesc::OpDesc(const std::string &type, const VariableNameMap &inputs,
+               const VariableNameMap &outputs, const AttributeMap &attrs) {
   desc_.set_type(type);
   inputs_ = inputs;
   outputs_ = outputs;
@@ -90,12 +88,12 @@ OpDescBind::OpDescBind(const std::string &type, const VariableNameMap &inputs,
   need_update_ = true;
 }
 
-OpDescBind::OpDescBind(const OpDesc &desc, ProgramDescBind *prog)
+OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog)
     : desc_(desc), need_update_(false) {
   // restore inputs_
   int input_size = desc_.inputs_size();
   for (int i = 0; i < input_size; ++i) {
-    const OpDesc::Var &var = desc_.inputs(i);
+    const proto::OpDesc::Var &var = desc_.inputs(i);
     std::vector<std::string> &args = inputs_[var.parameter()];
     int argu_size = var.arguments_size();
     args.reserve(argu_size);
@@ -106,7 +104,7 @@ OpDescBind::OpDescBind(const OpDesc &desc, ProgramDescBind *prog)
   // restore outputs_
   int output_size = desc_.outputs_size();
   for (int i = 0; i < output_size; ++i) {
-    const OpDesc::Var &var = desc_.outputs(i);
+    const proto::OpDesc::Var &var = desc_.outputs(i);
     std::vector<std::string> &args = outputs_[var.parameter()];
     int argu_size = var.arguments_size();
     args.reserve(argu_size);
@@ -115,9 +113,9 @@ OpDescBind::OpDescBind(const OpDesc &desc, ProgramDescBind *prog)
     }
   }
   // restore attrs_
-  for (const OpDesc::Attr &attr : desc_.attrs()) {
+  for (const proto::OpDesc::Attr &attr : desc_.attrs()) {
     std::string attr_name = attr.name();
-    if (attr.type() != AttrType::BLOCK) {
+    if (attr.type() != proto::AttrType::BLOCK) {
       attrs_[attr_name] = GetAttrValue(attr);
     } else {
       auto bid = attr.block_idx();
@@ -126,20 +124,19 @@ OpDescBind::OpDescBind(const OpDesc &desc, ProgramDescBind *prog)
   }
 }
 
-OpDesc *OpDescBind::Proto() {
+proto::OpDesc *OpDesc::Proto() {
   Flush();
   return &desc_;
 }
 
-const std::vector<std::string> &OpDescBind::Input(
-    const std::string &name) const {
+const std::vector<std::string> &OpDesc::Input(const std::string &name) const {
   auto it = inputs_.find(name);
   PADDLE_ENFORCE(it != inputs_.end(), "Input %s cannot be found in Op %s", name,
                  Type());
   return it->second;
 }
 
-std::vector<std::string> OpDescBind::InputArgumentNames() const {
+std::vector<std::string> OpDesc::InputArgumentNames() const {
   std::vector<std::string> retv;
   for (auto &ipt : this->inputs_) {
     retv.insert(retv.end(), ipt.second.begin(), ipt.second.end());
@@ -147,21 +144,20 @@ std::vector<std::string> OpDescBind::InputArgumentNames() const {
   return retv;
 }
 
-void OpDescBind::SetInput(const std::string &param_name,
-                          const std::vector<std::string> &args) {
+void OpDesc::SetInput(const std::string &param_name,
+                      const std::vector<std::string> &args) {
   need_update_ = true;
   inputs_[param_name] = args;
 }
 
-const std::vector<std::string> &OpDescBind::Output(
-    const std::string &name) const {
+const std::vector<std::string> &OpDesc::Output(const std::string &name) const {
   auto it = outputs_.find(name);
   PADDLE_ENFORCE(it != outputs_.end(), "Output %s cannot be found in Op %s",
                  name, Type());
   return it->second;
 }
 
-std::vector<std::string> OpDescBind::OutputArgumentNames() const {
+std::vector<std::string> OpDesc::OutputArgumentNames() const {
   std::vector<std::string> retv;
   for (auto &ipt : this->outputs_) {
     retv.insert(retv.end(), ipt.second.begin(), ipt.second.end());
@@ -169,19 +165,19 @@ std::vector<std::string> OpDescBind::OutputArgumentNames() const {
   return retv;
 }
 
-void OpDescBind::SetOutput(const std::string &param_name,
-                           const std::vector<std::string> &args) {
+void OpDesc::SetOutput(const std::string &param_name,
+                       const std::vector<std::string> &args) {
   need_update_ = true;
   this->outputs_[param_name] = args;
 }
 
-AttrType OpDescBind::GetAttrType(const std::string &name) const {
+proto::AttrType OpDesc::GetAttrType(const std::string &name) const {
   auto it = attrs_.find(name);
   PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);
-  return static_cast<AttrType>(it->second.which() - 1);
+  return static_cast<proto::AttrType>(it->second.which() - 1);
 }
 
-std::vector<std::string> OpDescBind::AttrNames() const {
+std::vector<std::string> OpDesc::AttrNames() const {
   std::vector<std::string> retv;
   retv.reserve(attrs_.size());
   for (auto &attr : attrs_) {
@@ -190,41 +186,39 @@ std::vector<std::string> OpDescBind::AttrNames() const {
   return retv;
 }
 
-void OpDescBind::SetAttr(const std::string &name, const Attribute &v) {
+void OpDesc::SetAttr(const std::string &name, const Attribute &v) {
   this->attrs_[name] = v;
   need_update_ = true;
 }
 
-void OpDescBind::SetBlockAttr(const std::string &name, BlockDescBind &block) {
+void OpDesc::SetBlockAttr(const std::string &name, BlockDesc &block) {
   this->attrs_[name] = &block;
   need_update_ = true;
 }
 
-void OpDescBind::SetAttrMap(
+void OpDesc::SetAttrMap(
     const std::unordered_map<std::string, Attribute> &attr_map) {
   attrs_ = attr_map;
   need_update_ = true;
 }
 
-Attribute OpDescBind::GetAttr(const std::string &name) const {
+Attribute OpDesc::GetAttr(const std::string &name) const {
   auto it = attrs_.find(name);
   PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);
   return it->second;
 }
 
-int OpDescBind::GetBlockAttr(const std::string &name) const {
+int OpDesc::GetBlockAttr(const std::string &name) const {
   auto it = attrs_.find(name);
   PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);
-  return boost::get<BlockDescBind *>(it->second)->ID();
+  return boost::get<BlockDesc *>(it->second)->ID();
 }
 
-const std::unordered_map<std::string, Attribute> &OpDescBind::GetAttrMap()
-    const {
+const std::unordered_map<std::string, Attribute> &OpDesc::GetAttrMap() const {
   return attrs_;
 }
 
-void OpDescBind::Rename(const std::string &old_name,
-                        const std::string &new_name) {
+void OpDesc::Rename(const std::string &old_name, const std::string &new_name) {
   for (auto &input : inputs_) {
     std::replace(input.second.begin(), input.second.end(), old_name, new_name);
   }
@@ -235,8 +229,8 @@ void OpDescBind::Rename(const std::string &old_name,
   need_update_ = true;
 }
 
-void OpDescBind::RenameOutput(const std::string &old_name,
-                              const std::string &new_name) {
+void OpDesc::RenameOutput(const std::string &old_name,
+                          const std::string &new_name) {
   for (auto &output : outputs_) {
     std::replace(output.second.begin(), output.second.end(), old_name,
                  new_name);
@@ -244,8 +238,8 @@ void OpDescBind::RenameOutput(const std::string &old_name,
   need_update_ = true;
 }
 
-void OpDescBind::RenameInput(const std::string &old_name,
-                             const std::string &new_name) {
+void OpDesc::RenameInput(const std::string &old_name,
+                         const std::string &new_name) {
   for (auto &input : inputs_) {
     std::replace(input.second.begin(), input.second.end(), old_name, new_name);
   }
@@ -253,8 +247,8 @@ void OpDescBind::RenameInput(const std::string &old_name,
 }
 
 struct SetAttrDescVisitor : public boost::static_visitor<void> {
-  explicit SetAttrDescVisitor(OpDesc::Attr *attr) : attr_(attr) {}
-  mutable OpDesc::Attr *attr_;
+  explicit SetAttrDescVisitor(proto::OpDesc::Attr *attr) : attr_(attr) {}
+  mutable proto::OpDesc::Attr *attr_;
   void operator()(int v) const { attr_->set_i(v); }
   void operator()(float v) const { attr_->set_f(v); }
   void operator()(const std::string &v) const { attr_->set_s(v); }
@@ -272,11 +266,13 @@ struct SetAttrDescVisitor : public boost::static_visitor<void> {
   void operator()(const std::vector<bool> &v) const {
     VectorToRepeated(v, attr_->mutable_bools());
   }
-  void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->idx()); }
+  void operator()(proto::BlockDesc *desc) const {
+    attr_->set_block_idx(desc->idx());
+  }
   void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }
 };
 
-void OpDescBind::Flush() {
+void OpDesc::Flush() {
   if (need_update_) {
     this->desc_.mutable_inputs()->Clear();
     for (auto &ipt : inputs_) {
@@ -297,7 +293,7 @@ void OpDescBind::Flush() {
       auto *attr_desc = desc_.add_attrs();
       attr_desc->set_name(attr.first);
       attr_desc->set_type(
-          static_cast<framework::AttrType>(attr.second.which() - 1));
+          static_cast<proto::AttrType>(attr.second.which() - 1));
       SetAttrDescVisitor visitor(attr_desc);
       boost::apply_visitor(visitor, attr.second);
     }
@@ -328,7 +324,7 @@ static void InitInferShapeFuncs() {
   });
 }
 
-void OpDescBind::CheckAttrs() {
+void OpDesc::CheckAttrs() {
   PADDLE_ENFORCE(!Type().empty(),
                  "CheckAttr() can not be called before type is setted.");
   auto *checker = OpInfoMap::Instance().Get(Type()).Checker();
@@ -340,7 +336,7 @@ void OpDescBind::CheckAttrs() {
   checker->Check(attrs_);
 }
 
-void OpDescBind::InferShape(const BlockDescBind &block) const {
+void OpDesc::InferShape(const BlockDesc &block) const {
   VLOG(3) << "CompileTime infer shape on " << Type();
   InitInferShapeFuncs();
   auto &infer_shape = OpInfoMap::Instance().Get(this->Type()).infer_shape_;
@@ -363,7 +359,7 @@ void OpDescBind::InferShape(const BlockDescBind &block) const {
   infer_shape(&ctx);
 }
 
-void OpDescBind::InferVarType(BlockDescBind *block) const {
+void OpDesc::InferVarType(BlockDesc *block) const {
   auto &info = OpInfoMap::Instance().Get(this->Type());
   if (info.infer_var_type_) {
     info.infer_var_type_(*this, block);
@@ -375,14 +371,14 @@ void OpDescBind::InferVarType(BlockDescBind *block) const {
     for (auto &out_pair : this->outputs_) {
       for (auto &out_var_name : out_pair.second) {
         block->FindRecursiveOrCreateVar(out_var_name)
-            ->SetType(VarDesc::LOD_TENSOR);
+            ->SetType(proto::VarDesc::LOD_TENSOR);
       }
     }
   }
 }
 
 CompileTimeInferShapeContext::CompileTimeInferShapeContext(
-    const OpDescBind &op, const BlockDescBind &block)
+    const OpDesc &op, const BlockDesc &block)
     : op_(op), block_(block) {}
 
 bool CompileTimeInferShapeContext::HasInput(const std::string &name) const {
@@ -484,7 +480,7 @@ void CompileTimeInferShapeContext::SetDim(const std::string &name,
 }
 bool CompileTimeInferShapeContext::IsRuntime() const { return false; }
 
-VarDesc::VarType CompileTimeInferShapeContext::GetVarType(
+proto::VarDesc::VarType CompileTimeInferShapeContext::GetVarType(
     const std::string &name) const {
   return block_.FindVarRecursive(name)->GetType();
 }

paddle/framework/op_desc.h

@@ -23,19 +23,19 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class BlockDescBind;
-class ProgramDescBind;
+class BlockDesc;
+class ProgramDesc;
 
-class OpDescBind {
+class OpDesc {
  public:
-  OpDescBind() {}
+  OpDesc() {}
 
-  OpDescBind(const std::string &type, const VariableNameMap &inputs,
-             const VariableNameMap &outputs, const AttributeMap &attrs);
+  OpDesc(const std::string &type, const VariableNameMap &inputs,
+         const VariableNameMap &outputs, const AttributeMap &attrs);
 
-  OpDescBind(const OpDesc &desc, ProgramDescBind *prog);
+  OpDesc(const proto::OpDesc &desc, ProgramDesc *prog);
 
-  OpDesc *Proto();
+  proto::OpDesc *Proto();
 
   std::string Type() const { return desc_.type(); }
 
@@ -59,13 +59,13 @@ class OpDescBind {
     return attrs_.find(name) != attrs_.end();
   }
 
-  AttrType GetAttrType(const std::string &name) const;
+  proto::AttrType GetAttrType(const std::string &name) const;
 
   std::vector<std::string> AttrNames() const;
 
   void SetAttr(const std::string &name, const Attribute &v);
 
-  void SetBlockAttr(const std::string &name, BlockDescBind &block);
+  void SetBlockAttr(const std::string &name, BlockDesc &block);
 
   Attribute GetAttr(const std::string &name) const;
 
@@ -107,9 +107,9 @@ class OpDescBind {
 
   void CheckAttrs();
 
-  void InferShape(const BlockDescBind &block) const;
+  void InferShape(const BlockDesc &block) const;
 
-  void InferVarType(BlockDescBind *block) const;
+  void InferVarType(BlockDesc *block) const;
 
   void MarkAsTarget() { desc_.set_is_target(true); }
 
@@ -126,8 +126,10 @@ class OpDescBind {
     return ret_val;
   }
 
-  OpDesc desc_;
+  proto::OpDesc desc_;
+  // input arg name => output variable names
   VariableNameMap inputs_;
+  // output arg name => output variable names
   VariableNameMap outputs_;
   AttributeMap attrs_;
 

paddle/framework/op_info.h

@@ -34,7 +34,7 @@ class InferShapeBase {
 struct OpInfo {
   OpCreator creator_;
   GradOpMakerFN grad_op_maker_;
-  OpProto* proto_{nullptr};
+  proto::OpProto* proto_{nullptr};
   OpAttrChecker* checker_{nullptr};
   InferVarTypeFN infer_var_type_;
   InferShapeFN infer_shape_;
@@ -43,7 +43,7 @@ struct OpInfo {
     return proto_ != nullptr && checker_ != nullptr;
   }
 
-  const OpProto& Proto() const {
+  const proto::OpProto& Proto() const {
     PADDLE_ENFORCE_NOT_NULL(proto_, "Operator Proto has not been registered");
     PADDLE_ENFORCE(proto_->IsInitialized(),
                    "Operator Proto must be initialized in op info");

paddle/framework/op_proto_maker.h

@@ -22,6 +22,8 @@ namespace framework {
 // this class not only make proto but also init attribute checkers.
 class OpProtoAndCheckerMaker {
  public:
+  using OpProto = proto::OpProto;
+  using OpAttrChecker = framework::OpAttrChecker;
   OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
       : proto_(proto), op_checker_(op_checker) {}
 
@@ -80,7 +82,7 @@ class OpProtoAndCheckerMaker {
 
 class NOPMaker : public OpProtoAndCheckerMaker {
  public:
-  NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+  NOPMaker(OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {}
 };
 

paddle/framework/op_proto_maker_test.cc

@@ -18,7 +18,7 @@ limitations under the License. */
 
 class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
  public:
-  TestAttrProtoMaker(paddle::framework::OpProto* proto,
+  TestAttrProtoMaker(paddle::framework::proto::OpProto* proto,
                      paddle::framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddAttr<float>("scale", "scale of test op");
@@ -27,7 +27,7 @@ class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
 };
 
 TEST(ProtoMaker, DuplicatedAttr) {
-  paddle::framework::OpProto op_proto;
+  paddle::framework::proto::OpProto op_proto;
   paddle::framework::OpAttrChecker op_checker;
   auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker);
   ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
@@ -35,7 +35,7 @@ TEST(ProtoMaker, DuplicatedAttr) {
 
 class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
  public:
-  TestInOutProtoMaker(paddle::framework::OpProto* proto,
+  TestInOutProtoMaker(paddle::framework::proto::OpProto* proto,
                       paddle::framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("input", "input of test op");
@@ -44,7 +44,7 @@ class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
 };
 
 TEST(ProtoMaker, DuplicatedInOut) {
-  paddle::framework::OpProto op_proto;
+  paddle::framework::proto::OpProto op_proto;
   paddle::framework::OpAttrChecker op_checker;
   auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker);
   ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);

paddle/framework/op_registry.cc

@@ -31,7 +31,8 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
 }
 
 static VariableNameMap ConvertOpDescVarsToVarNameMap(
-    const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
+    const google::protobuf::RepeatedPtrField<proto::OpDesc::Var>&
+        op_desc_vars) {
   VariableNameMap ret_val;
   for (auto& var : op_desc_vars) {
     auto& var_names = ret_val[var.parameter()];
@@ -43,9 +44,10 @@ static VariableNameMap ConvertOpDescVarsToVarNameMap(
   return ret_val;
 }
 
-std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
+    const proto::OpDesc& op_desc) {
   VLOG(1) << "CreateOp directly from OpDesc is deprecated. It should only be"
-             "used in unit tests. Use CreateOp(const OpDescBind& op_desc) "
+             "used in unit tests. Use CreateOp(const OpDesc& op_desc) "
              "instead.";
   VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
   VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
@@ -57,7 +59,7 @@ std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
   return CreateOp(op_desc.type(), inputs, outputs, attrs);
 }
 
-std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDescBind& op_desc) {
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
   return CreateOp(op_desc.Type(), op_desc.Inputs(), op_desc.Outputs(),
                   op_desc.GetAttrMap());
 }

paddle/framework/op_registry.h

@@ -77,9 +77,9 @@ class OpRegistry {
                                                 const VariableNameMap& outputs,
                                                 AttributeMap attrs);
 
-  static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
+  static std::unique_ptr<OperatorBase> CreateOp(const proto::OpDesc& op_desc);
 
-  static std::unique_ptr<OperatorBase> CreateOp(const OpDescBind& op_desc);
+  static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 };
 
 template <typename PlaceType, bool at_end, size_t I, typename... KernelType>
@@ -126,6 +126,14 @@ class OpKernelRegistrar : public Registrar {
                              __test_global_namespace_##uniq_name##__>::value, \
                 msg)
 
+/*
+  The variadic arguments should be class types derived from one of the
+  following classes:
+    OpProtoAndCheckerMaker
+    GradOpDescMakerBase
+    VarTypeInference
+    InferShapeBase
+*/
 #define REGISTER_OPERATOR(op_type, op_class, ...)                      \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                      \
       __reg_op__##op_type,                                             \
@@ -144,20 +152,29 @@ class OpKernelRegistrar : public Registrar {
   }
 
 /**
- * Macro to register Operator.
+ * Macro to register Operator. When the input is duplicable, you should
+ * use REGISTER_OP_EX with deop_empty_grad=false instead.
  */
-#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type,       \
-                    grad_op_class)                                         \
-  REGISTER_OPERATOR(grad_op_type, grad_op_class);                          \
-  class _GradOpDescMaker_##grad_op_type##_                                 \
-      : public ::paddle::framework::DefaultGradOpDescMaker<true> {         \
-    using ::paddle::framework::DefaultGradOpDescMaker<                     \
-        true>::DefaultGradOpDescMaker;                                     \
-                                                                           \
-   protected:                                                              \
-    virtual std::string GradOpType() const { return #grad_op_type; }       \
-  };                                                                       \
-  REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_, \
+#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, \
+                    grad_op_class)                                   \
+  REGISTER_OP_EX(op_type, op_class, op_maker_class, grad_op_type,    \
+                 grad_op_class, true)
+
+// When an argument is duplicable, we need to use this version.
+// Perhaps we can omit DropEmptyIG template parameter and
+// only have one version of REGISTER_OP.
+#define REGISTER_OP_EX(op_type, op_class, op_maker_class, grad_op_type,       \
+                       grad_op_class, drop_empty_grad)                        \
+  REGISTER_OPERATOR(grad_op_type, grad_op_class);                             \
+  class _GradOpDescMaker_##grad_op_type##_                                    \
+      : public ::paddle::framework::DefaultGradOpDescMaker<drop_empty_grad> { \
+    using ::paddle::framework::DefaultGradOpDescMaker<                        \
+        drop_empty_grad>::DefaultGradOpDescMaker;                             \
+                                                                              \
+   protected:                                                                 \
+    virtual std::string GradOpType() const { return #grad_op_type; }          \
+  };                                                                          \
+  REGISTER_OPERATOR(op_type, op_class, _GradOpDescMaker_##grad_op_type##_,    \
                     op_maker_class);
 
 #define REGISTER_OP_WITH_KERNEL(op_type, ...)                         \

paddle/framework/op_registry_test.cc

@@ -51,7 +51,7 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 
 static void BuildVar(const std::string& param_name,
                      std::initializer_list<const char*> arguments,
-                     paddle::framework::OpDesc::Var* var) {
+                     paddle::framework::proto::OpDesc::Var* var) {
   var->set_parameter(param_name);
   for (auto& arg_name : arguments) {
     var->add_arguments(arg_name);
@@ -63,7 +63,7 @@ REGISTER_OP_WITHOUT_GRADIENT(my_test_op, paddle::framework::MyTestOp,
                              paddle::framework::MyTestOpProtoAndCheckerMaker);
 
 TEST(OpRegistry, CreateOp) {
-  paddle::framework::OpDesc op_desc;
+  paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("cos_sim");
   BuildVar("input", {"aa"}, op_desc.add_inputs());
   BuildVar("output", {"bb"}, op_desc.add_outputs());
@@ -71,7 +71,7 @@ TEST(OpRegistry, CreateOp) {
   float scale = 3.3;
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
-  attr->set_type(paddle::framework::AttrType::FLOAT);
+  attr->set_type(paddle::framework::proto::AttrType::FLOAT);
   attr->set_f(scale);
 
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
@@ -83,14 +83,14 @@ TEST(OpRegistry, CreateOp) {
 }
 
 TEST(OpRegistry, IllegalAttr) {
-  paddle::framework::OpDesc op_desc;
+  paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("cos_sim");
   BuildVar("input", {"aa"}, op_desc.add_inputs());
   BuildVar("output", {"bb"}, op_desc.add_outputs());
 
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
-  attr->set_type(paddle::framework::AttrType::FLOAT);
+  attr->set_type(paddle::framework::proto::AttrType::FLOAT);
   attr->set_f(-2.0);
 
   bool caught = false;
@@ -108,7 +108,7 @@ TEST(OpRegistry, IllegalAttr) {
 }
 
 TEST(OpRegistry, DefaultValue) {
-  paddle::framework::OpDesc op_desc;
+  paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("cos_sim");
   BuildVar("input", {"aa"}, op_desc.add_inputs());
   BuildVar("output", {"bb"}, op_desc.add_outputs());
@@ -123,7 +123,7 @@ TEST(OpRegistry, DefaultValue) {
 }
 
 TEST(OpRegistry, CustomChecker) {
-  paddle::framework::OpDesc op_desc;
+  paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("my_test_op");
   BuildVar("input", {"ii"}, op_desc.add_inputs());
   BuildVar("output", {"oo"}, op_desc.add_outputs());
@@ -145,7 +145,7 @@ TEST(OpRegistry, CustomChecker) {
   // set 'test_attr' set to an illegal value
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("test_attr");
-  attr->set_type(paddle::framework::AttrType::INT);
+  attr->set_type(paddle::framework::proto::AttrType::INT);
   attr->set_i(3);
   caught = false;
   try {
@@ -164,7 +164,7 @@ TEST(OpRegistry, CustomChecker) {
   op_desc.mutable_attrs()->Clear();
   attr = op_desc.mutable_attrs()->Add();
   attr->set_name("test_attr");
-  attr->set_type(paddle::framework::AttrType::INT);
+  attr->set_type(paddle::framework::proto::AttrType::INT);
   attr->set_i(4);
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::platform::CPUDeviceContext dev_ctx;

paddle/framework/operator.cc

@@ -377,7 +377,7 @@ class RuntimeInferShapeContext : public InferShapeContext {
     }
   }
 
-  VarDesc::VarType GetVarType(const std::string& name) const override {
+  proto::VarDesc::VarType GetVarType(const std::string& name) const override {
     auto* var = scope_.FindVar(name);
     return ToVarType(var->Type());
   }
@@ -417,7 +417,7 @@ OpKernelType OperatorWithKernel::GetKernelType(
     const ExecutionContext& ctx) const {
   return OpKernelType(IndicateDataType(ctx), ctx.GetPlace());
 }
-DataType OperatorWithKernel::IndicateDataType(
+proto::DataType OperatorWithKernel::IndicateDataType(
     const ExecutionContext& ctx) const {
   auto& scope = ctx.scope();
   int data_type = -1;
@@ -443,7 +443,7 @@ DataType OperatorWithKernel::IndicateDataType(
     }
   }
   PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input");
-  return static_cast<DataType>(data_type);
+  return static_cast<proto::DataType>(data_type);
 }
 
 }  // namespace framework

paddle/framework/operator.h

@@ -358,12 +358,13 @@ struct OpKernelType {
   };
 
   platform::Place place_;
-  DataType data_type_;
+  proto::DataType data_type_;
 
-  OpKernelType(DataType data_type, platform::Place place)
+  OpKernelType(proto::DataType data_type, platform::Place place)
       : place_(place), data_type_(data_type) {}
 
-  OpKernelType(DataType data_type, const platform::DeviceContext& dev_ctx)
+  OpKernelType(proto::DataType data_type,
+               const platform::DeviceContext& dev_ctx)
       : place_(dev_ctx.GetPlace()), data_type_(data_type) {}
 
   bool operator==(const OpKernelType& o) const {
@@ -409,7 +410,7 @@ class OperatorWithKernel : public OperatorBase {
  private:
   // indicate kernel DataType by input data. Defaultly all input data must be
   // same.
-  DataType IndicateDataType(const ExecutionContext& ctx) const;
+  proto::DataType IndicateDataType(const ExecutionContext& ctx) const;
 };
 
 std::ostream& operator<<(std::ostream& os, const OpKernelType& kernel_key);

paddle/framework/operator_test.cc

@@ -58,7 +58,7 @@ class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 
 static void BuildVar(const std::string& param_name,
                      std::initializer_list<const char*> arguments,
-                     paddle::framework::OpDesc::Var* var) {
+                     paddle::framework::proto::OpDesc::Var* var) {
   var->set_parameter(param_name);
   for (auto& arg_name : arguments) {
     *var->mutable_arguments()->Add() = arg_name;
@@ -70,14 +70,14 @@ REGISTER_OP_WITHOUT_GRADIENT(
     paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
 
 TEST(OperatorBase, all) {
-  paddle::framework::OpDesc op_desc;
+  paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("test_operator");
   BuildVar("input", {"IN1"}, op_desc.add_inputs());
   BuildVar("output", {"OUT1"}, op_desc.add_outputs());
 
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
-  attr->set_type(paddle::framework::AttrType::FLOAT);
+  attr->set_type(paddle::framework::proto::AttrType::FLOAT);
   attr->set_f(3.14);
 
   paddle::platform::CPUDeviceContext device_context;
@@ -115,7 +115,7 @@ class OpWithKernelTest : public OperatorWithKernel {
  protected:
   void InferShape(framework::InferShapeContext* ctx) const override {}
   OpKernelType GetKernelType(const ExecutionContext& ctx) const override {
-    return OpKernelType(DataType::FP32, ctx.GetPlace());
+    return OpKernelType(proto::DataType::FP32, ctx.GetPlace());
   }
 };
 
@@ -195,14 +195,14 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel,
 
 // test with single input
 TEST(OpKernel, all) {
-  paddle::framework::OpDesc op_desc;
+  paddle::framework::proto::OpDesc op_desc;
   op_desc.set_type("op_with_kernel");
   BuildVar("x", {"IN1"}, op_desc.add_inputs());
   BuildVar("y", {"OUT1"}, op_desc.add_outputs());
 
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
-  attr->set_type(paddle::framework::AttrType::FLOAT);
+  attr->set_type(paddle::framework::proto::AttrType::FLOAT);
   attr->set_f(3.14);
 
   paddle::platform::CPUDeviceContext cpu_device_context;
@@ -224,7 +224,7 @@ REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
 TEST(OpKernel, multi_inputs) {
   using namespace paddle::framework;
 
-  OpDesc op_desc;
+  proto::OpDesc op_desc;
   op_desc.set_type("op_multi_inputs_with_kernel");
   BuildVar("xs", {"x0", "x1", "x2"}, op_desc.add_inputs());
   BuildVar("k", {"k0"}, op_desc.add_inputs());
@@ -232,7 +232,7 @@ TEST(OpKernel, multi_inputs) {
 
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
-  attr->set_type(paddle::framework::AttrType::FLOAT);
+  attr->set_type(paddle::framework::proto::AttrType::FLOAT);
   attr->set_f(3.14);
 
   paddle::platform::CPUDeviceContext cpu_device_context;

paddle/framework/program_desc.cc

@@ -18,49 +18,49 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-BlockDescBind *ProgramDescBind::AppendBlock(const BlockDescBind &parent) {
+BlockDesc *ProgramDesc::AppendBlock(const BlockDesc &parent) {
   auto *b = desc_.add_blocks();
   b->set_parent_idx(parent.ID());
   b->set_idx(desc_.blocks_size() - 1);
-  blocks_.emplace_back(new BlockDescBind(this, b));
+  blocks_.emplace_back(new BlockDesc(this, b));
   return blocks_.back().get();
 }
 
-ProgramDesc *ProgramDescBind::Proto() {
+proto::ProgramDesc *ProgramDesc::Proto() {
   for (auto &block : blocks_) {
     block->Flush();
   }
   return &desc_;
 }
 
-ProgramDescBind::ProgramDescBind() {
+ProgramDesc::ProgramDesc() {
   auto *block = desc_.mutable_blocks()->Add();
   block->set_idx(kRootBlockIndex);
   block->set_parent_idx(kNoneBlockIndex);
-  blocks_.emplace_back(new BlockDescBind(this, block));
+  blocks_.emplace_back(new BlockDesc(this, block));
 }
 
-ProgramDescBind::ProgramDescBind(const ProgramDescBind &o) {
+ProgramDesc::ProgramDesc(const ProgramDesc &o) {
   desc_ = o.desc_;
 
   for (int i = 0; i < desc_.blocks_size(); ++i) {
     auto *block = desc_.mutable_blocks(i);
-    blocks_.emplace_back(new BlockDescBind(*o.blocks_[i], block, this));
+    blocks_.emplace_back(new BlockDesc(*o.blocks_[i], block, this));
   }
 }
 
-ProgramDescBind::ProgramDescBind(const ProgramDesc &desc) {
+ProgramDesc::ProgramDesc(const proto::ProgramDesc &desc) {
   desc_ = desc;
   for (auto &block_desc : *desc_.mutable_blocks()) {
-    blocks_.emplace_back(new BlockDescBind(this, &block_desc));
+    blocks_.emplace_back(new BlockDesc(this, &block_desc));
   }
 }
 
-ProgramDescBind::ProgramDescBind(const std::string &binary_str) {
+ProgramDesc::ProgramDesc(const std::string &binary_str) {
   PADDLE_ENFORCE(desc_.ParseFromString(binary_str),
                  "Fail to parse program_desc from binary string.");
   for (auto &block_desc : *desc_.mutable_blocks()) {
-    blocks_.emplace_back(new BlockDescBind(this, &block_desc));
+    blocks_.emplace_back(new BlockDesc(this, &block_desc));
   }
 }
 

paddle/framework/program_desc.h

@@ -23,32 +23,32 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class BlockDescBind;
+class BlockDesc;
 
-class ProgramDescBind {
+class ProgramDesc {
  public:
-  ProgramDescBind();
+  ProgramDesc();
 
-  explicit ProgramDescBind(const ProgramDesc &desc);
+  explicit ProgramDesc(const proto::ProgramDesc &desc);
 
-  ProgramDescBind(const ProgramDescBind &o);
+  ProgramDesc(const ProgramDesc &o);
 
-  explicit ProgramDescBind(const std::string &binary_str);
+  explicit ProgramDesc(const std::string &binary_str);
 
-  BlockDescBind *AppendBlock(const BlockDescBind &parent);
+  BlockDesc *AppendBlock(const BlockDesc &parent);
 
-  BlockDescBind *MutableBlock(size_t idx) { return blocks_[idx].get(); }
+  BlockDesc *MutableBlock(size_t idx) { return blocks_[idx].get(); }
 
-  const BlockDescBind &Block(size_t idx) const { return *blocks_[idx]; }
+  const BlockDesc &Block(size_t idx) const { return *blocks_[idx]; }
 
   size_t Size() const { return blocks_.size(); }
 
-  ProgramDesc *Proto();
+  proto::ProgramDesc *Proto();
 
  private:
-  ProgramDesc desc_;
+  proto::ProgramDesc desc_;
 
-  std::vector<std::unique_ptr<BlockDescBind>> blocks_;
+  std::vector<std::unique_ptr<BlockDesc>> blocks_;
 };
 }  // namespace framework
 }  // namespace paddle

paddle/framework/program_desc_test.cc

@@ -19,18 +19,18 @@
 namespace paddle {
 namespace framework {
 TEST(ProgramDesc, copy_ctor) {
-  ProgramDescBind program;
+  ProgramDesc program;
   auto* global_block = program.MutableBlock(0);
   auto* x = global_block->Var("X");
-  x->SetType(VarDesc_VarType_LOD_TENSOR);
+  x->SetType(proto::VarDesc_VarType_LOD_TENSOR);
   x->SetLoDLevel(0);
-  x->SetDataType(FP32);
+  x->SetDataType(proto::FP32);
   x->SetShape({1000, 784});
 
   auto* y = global_block->Var("Y");
-  y->SetType(VarDesc_VarType_LOD_TENSOR);
+  y->SetType(proto::VarDesc_VarType_LOD_TENSOR);
   y->SetLoDLevel(0);
-  y->SetDataType(FP32);
+  y->SetDataType(proto::FP32);
   y->SetShape({784, 100});
 
   auto* op = global_block->AppendOp();
@@ -39,15 +39,15 @@ TEST(ProgramDesc, copy_ctor) {
   op->SetInput("Y", {y->Name()});
 
   auto* out = global_block->Var("Out");
-  out->SetType(VarDesc_VarType_LOD_TENSOR);
+  out->SetType(proto::VarDesc_VarType_LOD_TENSOR);
   op->SetOutput("Y", {out->Name()});
 
-  ProgramDescBind program_copy(program);
+  ProgramDesc program_copy(program);
 
   auto* global_block_copy = program_copy.MutableBlock(0);
   ASSERT_NE(global_block, global_block_copy);
 
-  auto assert_same_var = [&](const std::string& name, VarDescBind* var_before) {
+  auto assert_same_var = [&](const std::string& name, VarDesc* var_before) {
     ASSERT_TRUE(global_block_copy->HasVar(name));
     auto* copy = global_block_copy->Var(name);
     ASSERT_NE(copy, var_before);
@@ -81,18 +81,18 @@ TEST(ProgramDesc, copy_ctor) {
 }
 
 TEST(ProgramDescBind, serialize_and_deserialize) {
-  ProgramDescBind program_origin;
+  ProgramDesc program_origin;
   auto* global_block = program_origin.MutableBlock(0);
   auto* x = global_block->Var("X");
-  x->SetType(VarDesc_VarType_LOD_TENSOR);
+  x->SetType(proto::VarDesc_VarType_LOD_TENSOR);
   x->SetLoDLevel(0);
-  x->SetDataType(FP32);
+  x->SetDataType(proto::FP32);
   x->SetShape({1000, 784});
 
   auto* y = global_block->Var("Y");
-  y->SetType(VarDesc_VarType_LOD_TENSOR);
+  y->SetType(proto::VarDesc_VarType_LOD_TENSOR);
   y->SetLoDLevel(0);
-  y->SetDataType(FP32);
+  y->SetDataType(proto::FP32);
   y->SetShape({784, 100});
 
   auto* op = global_block->AppendOp();
@@ -101,17 +101,17 @@ TEST(ProgramDescBind, serialize_and_deserialize) {
   op->SetInput("Y", {y->Name()});
 
   auto* out = global_block->Var("Out");
-  out->SetType(VarDesc_VarType_LOD_TENSOR);
+  out->SetType(proto::VarDesc_VarType_LOD_TENSOR);
   op->SetOutput("Y", {out->Name()});
 
   std::string binary_str;
   program_origin.Proto()->SerializeToString(&binary_str);
 
-  ProgramDescBind program_restored(binary_str);
+  ProgramDesc program_restored(binary_str);
   auto* global_block_restored = program_restored.MutableBlock(0);
   ASSERT_NE(global_block, global_block_restored);
 
-  auto assert_same_var = [&](const std::string& name, VarDescBind* var_before) {
+  auto assert_same_var = [&](const std::string& name, VarDesc* var_before) {
     ASSERT_TRUE(global_block_restored->HasVar(name));
     auto* restored = global_block_restored->Var(name);
     ASSERT_NE(restored, var_before);

paddle/framework/prune.cc

@@ -29,7 +29,7 @@ const std::string kFetchOpType = "fetch";
 const std::string kDropOutOpType = "dropout";
 const std::string kBatchNormOpType = "batch_norm";
 
-bool HasDependentVar(const OpDesc& op_desc,
+bool HasDependentVar(const proto::OpDesc& op_desc,
                      const std::set<std::string>& dependent_vars) {
   for (auto& var : op_desc.outputs()) {
     for (auto& argu : var.arguments()) {
@@ -41,14 +41,15 @@ bool HasDependentVar(const OpDesc& op_desc,
   return false;
 }
 
-bool IsTarget(const OpDesc& op_desc) {
+bool IsTarget(const proto::OpDesc& op_desc) {
   if (op_desc.has_is_target()) {
     return op_desc.is_target();
   }
   return false;
 }
 
-void prune_impl(const ProgramDesc& input, ProgramDesc* output, int block_id) {
+void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
+                int block_id) {
   // TODO(tonyyang-svail):
   //    - will change to use multiple blocks for RNN op and Cond Op
 
@@ -104,12 +105,12 @@ void prune_impl(const ProgramDesc& input, ProgramDesc* output, int block_id) {
 }
 
 // TODO(fengjiayi): Prune() could be inplaced to avoid unnecessary copies
-void Prune(const ProgramDesc& input, ProgramDesc* output) {
+void Prune(const proto::ProgramDesc& input, proto::ProgramDesc* output) {
   prune_impl(input, output, 0);
 }
 
-void inference_optimize_impl(const ProgramDesc& input, ProgramDesc* output,
-                             int block_id) {
+void inference_optimize_impl(const proto::ProgramDesc& input,
+                             proto::ProgramDesc* output, int block_id) {
   *output = input;
   auto* op_field = output->mutable_blocks(block_id)->mutable_ops();
   for (auto& op_desc : *op_field) {
@@ -125,7 +126,8 @@ void inference_optimize_impl(const ProgramDesc& input, ProgramDesc* output,
   }
 }
 
-void InferenceOptimize(const ProgramDesc& input, ProgramDesc* output) {
+void InferenceOptimize(const proto::ProgramDesc& input,
+                       proto::ProgramDesc* output) {
   inference_optimize_impl(input, output, 0);
 }
 

paddle/framework/prune.h

@@ -20,9 +20,10 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-void Prune(const ProgramDesc& input, ProgramDesc* output);
+void Prune(const proto::ProgramDesc& input, proto::ProgramDesc* output);
 
-void InferenceOptimize(const ProgramDesc& input, ProgramDesc* output);
+void InferenceOptimize(const proto::ProgramDesc& input,
+                       proto::ProgramDesc* output);
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/prune_test.cc

@@ -29,12 +29,12 @@ namespace ops = paddle::operators;
 
 void AddOp(const std::string &type, const f::VariableNameMap &inputs,
            const f::VariableNameMap &outputs, f::AttributeMap attrs,
-           paddle::framework::BlockDescBind *block) {
+           paddle::framework::BlockDesc *block) {
   // insert output
   for (auto kv : outputs) {
     for (auto v : kv.second) {
       auto var = block->Var(v);
-      var->SetDataType(paddle::framework::DataType::FP32);
+      var->SetDataType(paddle::framework::proto::DataType::FP32);
     }
   }
 
@@ -51,26 +51,26 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs,
 }
 
 TEST(Prune, one_operator) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
 
   AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, f::AttributeMap{},
         block);
 
-  f::ProgramDesc *pdesc = program.Proto();
-  f::ProgramDesc pruned;
+  f::proto::ProgramDesc *pdesc = program.Proto();
+  f::proto::ProgramDesc pruned;
 
-  Prune(*pdesc, &pruned);
+  f::Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 0);
 
   pdesc->mutable_blocks(0)->mutable_ops(0)->set_is_target(true);
-  Prune(*pdesc, &pruned);
+  f::Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 1);
 }
 
 TEST(Prune, forward) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
 
   AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, f::AttributeMap{},
         block);
@@ -81,19 +81,19 @@ TEST(Prune, forward) {
   AddOp("one_one", {{"input", {"d"}}}, {{"output", {"e"}}}, f::AttributeMap{},
         block);
 
-  f::ProgramDesc *pdesc = program.Proto();
+  f::proto::ProgramDesc *pdesc = program.Proto();
 
   for (int i = 0; i < pdesc->blocks(0).ops_size(); ++i) {
-    f::ProgramDesc pruned;
+    f::proto::ProgramDesc pruned;
     pdesc->mutable_blocks(0)->mutable_ops(i)->set_is_target(true);
-    Prune(*pdesc, &pruned);
+    f::Prune(*pdesc, &pruned);
     PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), i + 1);
   }
 }
 
 TEST(Prune, multi_input_op) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
 
   AddOp("one_one", {{"input", {"a0"}}}, {{"output", {"b0"}}}, f::AttributeMap{},
         block);
@@ -104,17 +104,17 @@ TEST(Prune, multi_input_op) {
   AddOp("three_one", {{"input", {"b0", "b1", "b2"}}}, {{"output", {"c"}}},
         f::AttributeMap{}, block);
 
-  f::ProgramDesc *pdesc = program.Proto();
+  f::proto::ProgramDesc *pdesc = program.Proto();
   pdesc->mutable_blocks(0)->mutable_ops(3)->set_is_target(true);
 
-  f::ProgramDesc pruned;
-  Prune(*pdesc, &pruned);
+  f::proto::ProgramDesc pruned;
+  f::Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 4);
 }
 
 TEST(Prune, multi_output_op) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
 
   AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}},
         f::AttributeMap{}, block);
@@ -123,17 +123,17 @@ TEST(Prune, multi_output_op) {
   AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, f::AttributeMap{},
         block);
 
-  f::ProgramDesc *pdesc = program.Proto();
+  f::proto::ProgramDesc *pdesc = program.Proto();
   pdesc->mutable_blocks(0)->mutable_ops(2)->set_is_target(true);
 
-  f::ProgramDesc pruned;
-  Prune(*pdesc, &pruned);
+  f::proto::ProgramDesc pruned;
+  f::Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 2);
 }
 
 TEST(Prune, multi_target) {
-  f::ProgramDescBind program;
-  f::BlockDescBind *block = program.MutableBlock(0);
+  f::ProgramDesc program;
+  f::BlockDesc *block = program.MutableBlock(0);
 
   AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}},
         f::AttributeMap{}, block);
@@ -142,11 +142,11 @@ TEST(Prune, multi_target) {
   AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, f::AttributeMap{},
         block);
 
-  f::ProgramDesc *pdesc = program.Proto();
+  f::proto::ProgramDesc *pdesc = program.Proto();
   pdesc->mutable_blocks(0)->mutable_ops(1)->set_is_target(true);
   pdesc->mutable_blocks(0)->mutable_ops(2)->set_is_target(true);
 
-  f::ProgramDesc pruned;
-  Prune(*pdesc, &pruned);
+  f::proto::ProgramDesc pruned;
+  f::Prune(*pdesc, &pruned);
   PADDLE_ENFORCE_EQ(pruned.blocks(0).ops_size(), 3);
 }

paddle/framework/shape_inference.cc

@@ -57,17 +57,17 @@ void InferShapeContext::SetDims(const std::vector<std::string> &names,
     SetDim(names[i], dims[i]);
   }
 }
-std::vector<VarDesc::VarType> InferShapeContext::GetInputsVarType(
+std::vector<proto::VarDesc::VarType> InferShapeContext::GetInputsVarType(
     const std::string &name) const {
   return GetVarTypes(Inputs(name));
 }
-std::vector<VarDesc::VarType> InferShapeContext::GetOutputsVarType(
+std::vector<proto::VarDesc::VarType> InferShapeContext::GetOutputsVarType(
     const std::string &name) const {
   return GetVarTypes(Outputs(name));
 }
-std::vector<VarDesc::VarType> InferShapeContext::GetVarTypes(
+std::vector<proto::VarDesc::VarType> InferShapeContext::GetVarTypes(
     const std::vector<std::string> &names) const {
-  std::vector<VarDesc::VarType> retv;
+  std::vector<proto::VarDesc::VarType> retv;
   retv.resize(names.size());
   std::transform(names.begin(), names.end(), retv.begin(),
                  std::bind(std::mem_fn(&InferShapeContext::GetVarType), this,

paddle/framework/shape_inference.h

@@ -27,8 +27,9 @@ class InferShapeContext {
   virtual bool HasInput(const std::string &name) const = 0;
   virtual bool HasOutput(const std::string &name) const = 0;
 
-  std::vector<VarDesc::VarType> GetInputsVarType(const std::string &name) const;
-  std::vector<VarDesc::VarType> GetOutputsVarType(
+  std::vector<proto::VarDesc::VarType> GetInputsVarType(
+      const std::string &name) const;
+  std::vector<proto::VarDesc::VarType> GetOutputsVarType(
       const std::string &name) const;
 
   virtual bool HasInputs(const std::string &name) const = 0;
@@ -65,10 +66,10 @@ class InferShapeContext {
   std::vector<framework::DDim> GetDims(
       const std::vector<std::string> &names) const;
 
-  std::vector<VarDesc::VarType> GetVarTypes(
+  std::vector<proto::VarDesc::VarType> GetVarTypes(
       const std::vector<std::string> &names) const;
 
-  virtual VarDesc::VarType GetVarType(const std::string &name) const = 0;
+  virtual proto::VarDesc::VarType GetVarType(const std::string &name) const = 0;
 };
 
 }  // namespace framework

paddle/framework/type_defs.h

@@ -25,11 +25,9 @@
 namespace paddle {
 namespace framework {
 class OperatorBase;
-class OpDescBind;
-class BlockDescBind;
-class BlockDesc;
+class OpDesc;
 class InferShapeContext;
-class BlockDescBind;
+class BlockDesc;
 
 using VariableNameMap = std::map<std::string, std::vector<std::string>>;
 
@@ -37,7 +35,7 @@ using VariableNameMap = std::map<std::string, std::vector<std::string>>;
 using Attribute =
     boost::variant<boost::blank, int, float, std::string, std::vector<int>,
                    std::vector<float>, std::vector<std::string>, bool,
-                   std::vector<bool>, BlockDescBind*>;
+                   std::vector<bool>, BlockDesc*>;
 
 using AttributeMap = std::unordered_map<std::string, Attribute>;
 
@@ -45,13 +43,13 @@ using OpCreator = std::function<OperatorBase*(
     const std::string& /*type*/, const VariableNameMap& /*inputs*/,
     const VariableNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
 
-using GradOpMakerFN = std::function<std::vector<std::unique_ptr<OpDescBind>>(
-    const OpDescBind&, const std::unordered_set<std::string>& /*no_grad_set*/,
+using GradOpMakerFN = std::function<std::vector<std::unique_ptr<OpDesc>>(
+    const OpDesc&, const std::unordered_set<std::string>& /*no_grad_set*/,
     std::unordered_map<std::string, std::string>* /*grad_to_var*/,
-    const std::vector<BlockDescBind*>& grad_block)>;
+    const std::vector<BlockDesc*>& grad_block)>;
 
-using InferVarTypeFN = std::function<void(const OpDescBind& /*op_desc*/,
-                                          BlockDescBind* /*block*/)>;
+using InferVarTypeFN =
+    std::function<void(const OpDesc& /*op_desc*/, BlockDesc* /*block*/)>;
 
 using InferShapeFN = std::function<void(InferShapeContext*)>;
 

paddle/framework/var_desc.cc

@@ -18,30 +18,32 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-VarDesc::VarType VarDescBind::GetType() const { return desc_.type(); }
+proto::VarDesc::VarType VarDesc::GetType() const { return desc_.type(); }
 
-void VarDescBind::SetType(VarDesc::VarType type) { desc_.set_type(type); }
+void VarDesc::SetType(proto::VarDesc::VarType type) { desc_.set_type(type); }
 
-void VarDescBind::SetShape(const std::vector<int64_t> &dims) {
+void VarDesc::SetShape(const std::vector<int64_t> &dims) {
   VectorToRepeated(dims, mutable_tensor_desc()->mutable_dims());
 }
 
-void VarDescBind::SetDataType(DataType data_type) {
+void VarDesc::SetDataType(proto::DataType data_type) {
   mutable_tensor_desc()->set_data_type(data_type);
 }
 
-std::vector<int64_t> VarDescBind::Shape() const {
+std::vector<int64_t> VarDesc::Shape() const {
   return RepeatedToVector(tensor_desc().dims());
 }
 
-DataType VarDescBind::GetDataType() const { return tensor_desc().data_type(); }
+proto::DataType VarDesc::GetDataType() const {
+  return tensor_desc().data_type();
+}
 
-void VarDescBind::SetLoDLevel(int32_t lod_level) {
+void VarDesc::SetLoDLevel(int32_t lod_level) {
   switch (desc_.type()) {
-    case VarDesc::LOD_TENSOR:
+    case proto::VarDesc::LOD_TENSOR:
       desc_.mutable_lod_tensor()->set_lod_level(lod_level);
       break;
-    case VarDesc::LOD_TENSOR_ARRAY:
+    case proto::VarDesc::LOD_TENSOR_ARRAY:
       desc_.mutable_tensor_array()->set_lod_level(lod_level);
       break;
     default:
@@ -50,11 +52,11 @@ void VarDescBind::SetLoDLevel(int32_t lod_level) {
   }
 }
 
-int32_t VarDescBind::GetLodLevel() const {
+int32_t VarDesc::GetLodLevel() const {
   switch (desc_.type()) {
-    case VarDesc::LOD_TENSOR:
+    case proto::VarDesc::LOD_TENSOR:
       return desc_.lod_tensor().lod_level();
-    case VarDesc::LOD_TENSOR_ARRAY:
+    case proto::VarDesc::LOD_TENSOR_ARRAY:
       return desc_.tensor_array().lod_level();
     default:
       PADDLE_THROW("Tensor type=%d does not support LoDLevel",
@@ -62,29 +64,29 @@ int32_t VarDescBind::GetLodLevel() const {
   }
 }
 
-const TensorDesc &VarDescBind::tensor_desc() const {
+const proto::TensorDesc &VarDesc::tensor_desc() const {
   PADDLE_ENFORCE(desc_.has_type(), "invoke TensorDesc must after set type");
   switch (desc_.type()) {
-    case VarDesc::SELECTED_ROWS:
+    case proto::VarDesc::SELECTED_ROWS:
       return desc_.selected_rows();
-    case VarDesc::LOD_TENSOR:
+    case proto::VarDesc::LOD_TENSOR:
       return desc_.lod_tensor().tensor();
-    case VarDesc::LOD_TENSOR_ARRAY:
+    case proto::VarDesc::LOD_TENSOR_ARRAY:
       return desc_.tensor_array().tensor();
     default:
       PADDLE_THROW("Unexpected branch.");
   }
 }
 
-TensorDesc *VarDescBind::mutable_tensor_desc() {
+proto::TensorDesc *VarDesc::mutable_tensor_desc() {
   PADDLE_ENFORCE(desc_.has_type(),
                  "invoke MutableTensorDesc must after set type");
   switch (desc_.type()) {
-    case VarDesc::SELECTED_ROWS:
+    case proto::VarDesc::SELECTED_ROWS:
       return desc_.mutable_selected_rows();
-    case VarDesc::LOD_TENSOR:
+    case proto::VarDesc::LOD_TENSOR:
       return desc_.mutable_lod_tensor()->mutable_tensor();
-    case VarDesc::LOD_TENSOR_ARRAY:
+    case proto::VarDesc::LOD_TENSOR_ARRAY:
       return desc_.mutable_tensor_array()->mutable_tensor();
     default:
       PADDLE_THROW("Unexpected branch.");

paddle/framework/var_desc.h

@@ -53,44 +53,44 @@ inline void VectorToRepeated(const std::vector<bool> &vec,
   }
 }
 
-class VarDescBind {
+class VarDesc {
  public:
-  explicit VarDescBind(const std::string &name) {
+  explicit VarDesc(const std::string &name) {
     desc_.set_name(name);
-    desc_.set_type(VarDesc::LOD_TENSOR);
+    desc_.set_type(proto::VarDesc::LOD_TENSOR);
   }
 
-  explicit VarDescBind(const VarDesc &desc) : desc_(desc) {}
+  explicit VarDesc(const proto::VarDesc &desc) : desc_(desc) {}
 
-  VarDesc *Proto() { return &desc_; }
+  proto::VarDesc *Proto() { return &desc_; }
 
   std::string Name() const { return desc_.name(); }
 
   void SetShape(const std::vector<int64_t> &dims);
 
-  void SetDataType(DataType data_type);
+  void SetDataType(proto::DataType data_type);
 
   std::vector<int64_t> Shape() const;
 
-  DataType GetDataType() const;
+  proto::DataType GetDataType() const;
 
   void SetLoDLevel(int32_t lod_level);
 
   int32_t GetLodLevel() const;
 
-  VarDesc::VarType GetType() const;
+  proto::VarDesc::VarType GetType() const;
 
-  void SetType(VarDesc::VarType type);
+  void SetType(proto::VarDesc::VarType type);
 
   bool Persistable() const { return desc_.persistable(); }
 
   void SetPersistable(bool persistable) { desc_.set_persistable(persistable); }
 
  private:
-  const TensorDesc &tensor_desc() const;
-  TensorDesc *mutable_tensor_desc();
+  const proto::TensorDesc &tensor_desc() const;
+  proto::TensorDesc *mutable_tensor_desc();
 
-  VarDesc desc_;
+  proto::VarDesc desc_;
 };
 }  // namespace framework
 }  // namespace paddle

paddle/framework/var_type.h

@@ -20,15 +20,15 @@
 
 namespace paddle {
 namespace framework {
-inline VarDesc::VarType ToVarType(std::type_index type) {
+inline proto::VarDesc::VarType ToVarType(std::type_index type) {
   if (type.hash_code() == typeid(LoDTensor).hash_code()) {
-    return VarDesc_VarType_LOD_TENSOR;
+    return proto::VarDesc_VarType_LOD_TENSOR;
   } else if (type.hash_code() == typeid(LoDRankTable).hash_code()) {
-    return VarDesc_VarType_LOD_RANK_TABLE;
+    return proto::VarDesc_VarType_LOD_RANK_TABLE;
   } else if (type.hash_code() == typeid(LoDTensorArray).hash_code()) {
-    return VarDesc_VarType_LOD_TENSOR_ARRAY;
+    return proto::VarDesc_VarType_LOD_TENSOR_ARRAY;
   } else if (type.hash_code() == typeid(SelectedRows).hash_code()) {
-    return VarDesc_VarType_SELECTED_ROWS;
+    return proto::VarDesc_VarType_SELECTED_ROWS;
   } else {
     PADDLE_THROW("ToVarType:Unsupported type %s", type.name());
   }
@@ -37,16 +37,16 @@ inline VarDesc::VarType ToVarType(std::type_index type) {
 template <typename Visitor>
 inline void VisitVarType(const Variable& var, Visitor visitor) {
   switch (ToVarType(var.Type())) {
-    case VarDesc_VarType_LOD_TENSOR:
+    case proto::VarDesc_VarType_LOD_TENSOR:
       visitor(var.Get<framework::LoDTensor>());
       return;
-    case VarDesc_VarType_LOD_RANK_TABLE:
+    case proto::VarDesc_VarType_LOD_RANK_TABLE:
       visitor(var.Get<LoDRankTable>());
       return;
-    case VarDesc_VarType_LOD_TENSOR_ARRAY:
+    case proto::VarDesc_VarType_LOD_TENSOR_ARRAY:
       visitor(var.Get<LoDTensorArray>());
       return;
-    case VarDesc_VarType_SELECTED_ROWS:
+    case proto::VarDesc_VarType_SELECTED_ROWS:
       visitor(var.Get<SelectedRows>());
       return;
     default:

paddle/framework/var_type_inference.h

@@ -21,8 +21,7 @@ namespace framework {
 class VarTypeInference {
  public:
   virtual ~VarTypeInference() {}
-  virtual void operator()(const OpDescBind& op_desc,
-                          BlockDescBind* block) const = 0;
+  virtual void operator()(const OpDesc& op_desc, BlockDesc* block) const = 0;
 };
 
 }  // namespace framework

paddle/framework/var_type_inference_test.cc

@@ -33,17 +33,16 @@ class SumOpMaker : public OpProtoAndCheckerMaker {
 
 class SumOpVarTypeInference : public VarTypeInference {
  public:
-  void operator()(const OpDescBind &op_desc,
-                  BlockDescBind *block) const override {
+  void operator()(const OpDesc &op_desc, BlockDesc *block) const override {
     auto &inputs = op_desc.Input("X");
-    auto default_var_type = VarDesc::SELECTED_ROWS;
+    auto default_var_type = proto::VarDesc::SELECTED_ROWS;
 
     bool any_input_is_lod_tensor = std::any_of(
         inputs.begin(), inputs.end(), [block](const std::string &name) {
-          return block->Var(name)->GetType() == VarDesc::LOD_TENSOR;
+          return block->Var(name)->GetType() == proto::VarDesc::LOD_TENSOR;
         });
     if (any_input_is_lod_tensor) {
-      default_var_type = VarDesc::LOD_TENSOR;
+      default_var_type = proto::VarDesc::LOD_TENSOR;
     }
 
     auto out_var_name = op_desc.Output("Out").front();
@@ -62,43 +61,43 @@ namespace paddle {
 namespace framework {
 
 TEST(InferVarType, sum_op) {
-  ProgramDescBind prog;
+  ProgramDesc prog;
   auto *op = prog.MutableBlock(0)->AppendOp();
   op->SetType("sum");
   op->SetInput("X", {"test_a", "test_b", "test_c"});
   op->SetOutput("Out", {"test_out"});
 
-  prog.MutableBlock(0)->Var("test_a")->SetType(VarDesc::SELECTED_ROWS);
-  prog.MutableBlock(0)->Var("test_b")->SetType(VarDesc::SELECTED_ROWS);
-  prog.MutableBlock(0)->Var("test_c")->SetType(VarDesc::SELECTED_ROWS);
+  prog.MutableBlock(0)->Var("test_a")->SetType(proto::VarDesc::SELECTED_ROWS);
+  prog.MutableBlock(0)->Var("test_b")->SetType(proto::VarDesc::SELECTED_ROWS);
+  prog.MutableBlock(0)->Var("test_c")->SetType(proto::VarDesc::SELECTED_ROWS);
   prog.MutableBlock(0)->Var("test_out");
 
   op->InferVarType(prog.MutableBlock(0));
 
-  ASSERT_EQ(VarDesc::SELECTED_ROWS,
+  ASSERT_EQ(proto::VarDesc::SELECTED_ROWS,
             prog.MutableBlock(0)->Var("test_out")->GetType());
 
-  prog.MutableBlock(0)->Var("test_b")->SetType(VarDesc::LOD_TENSOR);
+  prog.MutableBlock(0)->Var("test_b")->SetType(proto::VarDesc::LOD_TENSOR);
   op->InferVarType(prog.MutableBlock(0));
-  ASSERT_EQ(VarDesc::LOD_TENSOR,
+  ASSERT_EQ(proto::VarDesc::LOD_TENSOR,
             prog.MutableBlock(0)->Var("test_out")->GetType());
 }
 
 TEST(InferVarType, sum_op_without_infer_var_type) {
-  ProgramDescBind prog;
+  ProgramDesc prog;
   auto *op = prog.MutableBlock(0)->AppendOp();
   op->SetType("sum_without_infer_var_type");
   op->SetInput("X", {"test2_a", "test2_b", "test2_c"});
   op->SetOutput("Out", {"test2_out"});
 
-  prog.MutableBlock(0)->Var("test2_a")->SetType(VarDesc::SELECTED_ROWS);
-  prog.MutableBlock(0)->Var("test2_b")->SetType(VarDesc::SELECTED_ROWS);
-  prog.MutableBlock(0)->Var("test2_c")->SetType(VarDesc::SELECTED_ROWS);
+  prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarDesc::SELECTED_ROWS);
+  prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarDesc::SELECTED_ROWS);
+  prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarDesc::SELECTED_ROWS);
   prog.MutableBlock(0)->Var("test2_out");
 
   op->InferVarType(prog.MutableBlock(0));
 
-  ASSERT_EQ(VarDesc_VarType_LOD_TENSOR,
+  ASSERT_EQ(proto::VarDesc_VarType_LOD_TENSOR,
             prog.MutableBlock(0)->Var("test2_out")->GetType());
 }
 

paddle/memory/memcpy.cc

@@ -62,33 +62,6 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
   }
 }
 
-template <>
-void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::GPUPlace src_place,
-                                                  const void* src, size_t num) {
-  platform::SetDeviceId(src_place.device);
-  platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost);
-}
-
-template <>
-void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::CPUPlace src_place,
-                                                  const void* src, size_t num) {
-  platform::SetDeviceId(dst_place.device);
-  platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice);
-}
-
-template <>
-void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
-                                                  void* dst,
-                                                  platform::GPUPlace src_place,
-                                                  const void* src, size_t num) {
-  platform::SetDeviceId(dst_place.device);
-  platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToDevice);
-}
-
 #endif
 
 }  // namespace memory

paddle/operators/accuracy_op.cc

@@ -63,8 +63,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
 
 class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AccuracyOpMaker(framework::OpProto *proto,
-                  framework::OpAttrChecker *op_checker)
+  AccuracyOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     // TODO(typhoonzero): support both inference value and indices.
     AddInput("Out", "The network output of topk (inferences)");

paddle/operators/accuracy_op.cu

@@ -26,7 +26,7 @@ template <int BlockSize>
 __global__ void AccuracyCudaKernel(const int N, const int D,
                                    const int64_t* Xdata,
                                    const int64_t* labeldata, int* correct_data,
-                                   float* accuracy) {
+                                   float* accuracy, int* total_data) {
   int count = 0;
   __shared__ int total[BlockSize];
 
@@ -47,6 +47,7 @@ __global__ void AccuracyCudaKernel(const int N, const int D,
   if (threadIdx.x == 0) {
     *correct_data = result;
     *accuracy = static_cast<float>(result) / static_cast<float>(N);
+    *total_data = N;
   }
 }
 
@@ -80,22 +81,11 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
     if (num_samples == 0) {
       return;
     }
-    platform::GpuMemcpyAsync(total_data, &num_samples, sizeof(int),
-                             cudaMemcpyHostToDevice, stream);
 
     AccuracyCudaKernel<
         PADDLE_CUDA_NUM_THREADS><<<1, PADDLE_CUDA_NUM_THREADS, 0, stream>>>(
         num_samples, infer_width, indices_data, label_data, correct_data,
-        accuracy_data);
-
-    int d_num_samples, d_num_correct;
-    float d_accuracy;
-    platform::GpuMemcpyAsync(&d_num_correct, correct_data, sizeof(int),
-                             cudaMemcpyDeviceToHost, stream);
-    platform::GpuMemcpyAsync(&d_num_samples, total_data, sizeof(int),
-                             cudaMemcpyDeviceToHost, stream);
-    platform::GpuMemcpyAsync(&d_accuracy, accuracy_data, sizeof(float),
-                             cudaMemcpyDeviceToHost, stream);
+        accuracy_data, total_data);
   }
 };
 

paddle/operators/activation_op.cc

@@ -38,9 +38,8 @@ class ActivationOpGrad : public framework::OperatorWithKernel {
 
 class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SigmoidOpMaker(framework::OpProto *proto,
-                 framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Sigmoid operator");
     AddOutput("Y", "Output of Sigmoid operator");
     AddComment(R"DOC(
@@ -54,9 +53,8 @@ $$y = \frac{1}{1 + e^{-x}}$$
 
 class LogSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LogSigmoidOpMaker(framework::OpProto *proto,
-                    framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  LogSigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of LogSigmoid operator");
     AddOutput("Y", "Output of LogSigmoid operator");
     AddComment(R"DOC(
@@ -70,8 +68,8 @@ $$y = \log \frac{1}{1 + e^{-x}}$$
 
 class ExpOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ExpOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  ExpOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Exp operator");
     AddOutput("Y", "Output of Exp operator");
     AddComment(R"DOC(
@@ -85,8 +83,8 @@ $y = e^x$
 
 class ReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ReluOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  ReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Relu operator");
     AddOutput("Y", "Output of Relu operator");
     AddComment(R"DOC(
@@ -100,9 +98,8 @@ $y = \max(x, 0)$
 
 class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LeakyReluOpMaker(framework::OpProto *proto,
-                   framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  LeakyReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of LeakyRelu operator");
     AddOutput("Y", "Output of LeakyRelu operator");
     AddAttr<float>("alpha", "The small negative slope").SetDefault(0.02f);
@@ -117,9 +114,8 @@ $y = \max(x, \alpha * x)$
 
 class SoftShrinkOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SoftShrinkOpMaker(framework::OpProto *proto,
-                    framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  SoftShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Softshrink operator");
     AddOutput("Y", "Output of Softshrink operator");
     AddAttr<float>("lambda", "non-negative offset").SetDefault(0.5f);
@@ -140,8 +136,8 @@ $$
 
 class TanhOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  TanhOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  TanhOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Tanh operator");
     AddOutput("Y", "Output of Tanh operator");
     AddComment(R"DOC(
@@ -155,9 +151,8 @@ $$y = \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
 
 class TanhShrinkOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  TanhShrinkOpMaker(framework::OpProto *proto,
-                    framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  TanhShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of TanhShrink operator");
     AddOutput("Y", "Output of TanhShrink operator");
     AddComment(R"DOC(
@@ -171,9 +166,8 @@ $$y = x - \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
 
 class HardShrinkOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  HardShrinkOpMaker(framework::OpProto *proto,
-                    framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  HardShrinkOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of HardShrink operator");
     AddOutput("Y", "Output of HardShrink operator");
     AddAttr<float>("threshold", "The value of threshold for HardShrink")
@@ -195,8 +189,8 @@ $$
 
 class SqrtOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SqrtOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  SqrtOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Sqrt operator");
     AddOutput("Y", "Output of Sqrt operator");
     AddComment(R"DOC(
@@ -210,8 +204,8 @@ $y = \sqrt{x}$
 
 class AbsOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AbsOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  AbsOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Abs operator");
     AddOutput("Y", "Output of Abs operator");
     AddComment(R"DOC(
@@ -225,8 +219,8 @@ $y = |x|$
 
 class CeilOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CeilOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  CeilOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Ceil operator");
     AddOutput("Y", "Output of Ceil operator");
     AddComment(R"DOC(
@@ -240,8 +234,8 @@ $y = ceil(x)$
 
 class FloorOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  FloorOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  FloorOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Floor operator");
     AddOutput("Y", "Output of Floor operator");
     AddComment(R"DOC(
@@ -255,8 +249,8 @@ $y = floor(x)$
 
 class RoundOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  RoundOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  RoundOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Round operator");
     AddOutput("Y", "Output of Round operator");
     AddComment(R"DOC(
@@ -270,9 +264,8 @@ $y = [x]$
 
 class ReciprocalOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ReciprocalOpMaker(framework::OpProto *proto,
-                    framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  ReciprocalOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Reciprocal operator");
     AddOutput("Y", "Output of Reciprocal operator");
     AddComment(R"DOC(
@@ -286,8 +279,8 @@ $$y = \frac{1}{x}$$
 
 class LogOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  LogOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  LogOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Log operator");
     AddOutput("Y", "Output of Log operator");
     AddComment(R"DOC(
@@ -303,8 +296,8 @@ Natural logarithm of x.
 
 class SquareOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SquareOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  SquareOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Square operator");
     AddOutput("Y", "Output of Square operator");
     AddComment(R"DOC(
@@ -318,9 +311,8 @@ $y = x^2$
 
 class SoftplusOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SoftplusOpMaker(framework::OpProto *proto,
-                  framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  SoftplusOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Softplus operator");
     AddOutput("Y", "Output of Softplus operator");
     AddComment(R"DOC(
@@ -334,9 +326,8 @@ $y = \ln(1 + e^{x})$
 
 class SoftsignOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SoftsignOpMaker(framework::OpProto *proto,
-                  framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  SoftsignOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Softsign operator");
     AddOutput("Y", "Output of Softsign operator");
     AddComment(R"DOC(
@@ -350,8 +341,8 @@ $$y = \frac{x}{1 + |x|}$$
 
 class BReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BReluOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  BReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of BRelu operator");
     AddOutput("Y", "Output of BRelu operator");
     AddAttr<float>("t_min", "The min marginal value of BRelu")
@@ -369,9 +360,8 @@ $y = \max(\min(x, t_{min}), t_{max})$
 
 class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SoftReluOpMaker(framework::OpProto *proto,
-                  framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  SoftReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of SoftRelu operator");
     AddOutput("Y", "Output of SoftRelu operator");
     AddAttr<float>("threshold", "The threshold value of SoftRelu")
@@ -387,8 +377,8 @@ $y = \ln(1 + \exp(\max(\min(x, threshold), threshold))$
 
 class ELUOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ELUOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  ELUOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of ELU operator");
     AddOutput("Y", "Output of ELU operator");
     AddAttr<float>("alpha", "The alpha value of ELU").SetDefault(1.0f);
@@ -406,8 +396,8 @@ $y = \max(0, x) + \min(0, \alpha * (e^x - 1))$
 
 class Relu6OpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Relu6OpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  Relu6OpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Relu6 operator");
     AddOutput("Y", "Output of Relu6 operator");
     AddAttr<float>("threshold", "The threshold value of Relu6")
@@ -423,8 +413,8 @@ $y = \min(\max(0, x), 6)$
 
 class PowOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  PowOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  PowOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Pow operator");
     AddOutput("Y", "Output of Pow operator");
     AddAttr<float>("factor", "The exponential factor of Pow").SetDefault(1.0f);
@@ -439,8 +429,8 @@ $y = x^{factor}$
 
 class STanhOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  STanhOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  STanhOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of STanh operator");
     AddOutput("Y", "Output of STanh operator");
     AddAttr<float>("scale_a", "The scale parameter of a for the input")
@@ -458,9 +448,8 @@ $$y = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$
 
 class ThresholdedReluOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ThresholdedReluOpMaker(framework::OpProto *proto,
-                         framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  ThresholdedReluOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of ThresholdedRelu operator");
     AddOutput("Y", "Output of ThresholdedRelu operator");
     AddAttr<float>("threshold", "The threshold location of activation")
@@ -481,9 +470,8 @@ $$
 
 class HardSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  HardSigmoidOpMaker(framework::OpProto *proto,
-                     framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  HardSigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of HardSigmoid operator");
     AddOutput("Y", "Output of HardSigmoid operator");
     AddAttr<float>("slope", "Slope for linear approximation of sigmoid")
@@ -508,8 +496,8 @@ It is recommended to use the defaults for this activation.
 
 class SwishOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  SwishOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  SwishOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input of Swish operator");
     AddOutput("Y", "Output of Swish operator");
     AddAttr<float>("beta", "Constant beta of swish operator").SetDefault(1.0f);

paddle/operators/adadelta_op.cc

@@ -59,8 +59,7 @@ class AdadeltaOp : public framework::OperatorWithKernel {
 
 class AdadeltaOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AdadeltaOpMaker(framework::OpProto *proto,
-                  framework::OpAttrChecker *op_checker)
+  AdadeltaOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");

paddle/operators/adagrad_op.cc

@@ -59,8 +59,7 @@ class AdagradOp : public framework::OperatorWithKernel {
 
 class AdagradOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AdagradOpMaker(framework::OpProto* proto,
-                 framework::OpAttrChecker* op_checker)
+  AdagradOpMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");

paddle/operators/adam_op.cc

@@ -73,7 +73,7 @@ class AdamOp : public framework::OperatorWithKernel {
 
 class AdamOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AdamOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  AdamOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");

paddle/operators/adamax_op.cc

@@ -67,7 +67,7 @@ class AdamaxOp : public framework::OperatorWithKernel {
 
 class AdamaxOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AdamaxOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  AdamaxOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("Param", "(Tensor) Input parameter");
     AddInput("Grad", "(Tensor) Input gradient");

paddle/operators/array_to_lod_tensor_op.cc

@@ -114,8 +114,7 @@ class ArrayToLoDTensorOp : public framework::OperatorBase {
 
 class ArrayToLoDTensorOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ArrayToLoDTensorOpProtoMaker(framework::OpProto *proto,
-                               framework::OpAttrChecker *op_checker)
+  ArrayToLoDTensorOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X",
              "(std::vector<LodTensor>) A vector of tensors that is going to "
@@ -150,14 +149,14 @@ class ArrayToLoDTensorGradMaker : public framework::SingleGradOpDescMaker {
   using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
 
  protected:
-  std::unique_ptr<framework::OpDescBind> Apply() const override {
-    auto *grad_op = new framework::OpDescBind();
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto *grad_op = new framework::OpDesc();
     grad_op->SetType("lod_tensor_to_array");
     grad_op->SetInput("X", OutputGrad("Out"));
     grad_op->SetInput("RankTable", Input("RankTable"));
     grad_op->SetOutput("Out", InputGrad("X"));
     grad_op->SetAttrMap(Attrs());
-    return std::unique_ptr<framework::OpDescBind>(grad_op);
+    return std::unique_ptr<framework::OpDesc>(grad_op);
   }
 };
 

paddle/operators/assign_op.cc

@@ -86,8 +86,7 @@ class AssignOp : public framework::OperatorBase {
 
 class AssignOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AssignOpProtoMaker(framework::OpProto *proto,
-                     framework::OpAttrChecker *op_checker)
+  AssignOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X",
              "(LoDTensor, SelectedRows or LoDTensorArray) The input variable "
@@ -109,8 +108,8 @@ class AssignInferShape : public framework::InferShapeBase {
   void operator()(framework::InferShapeContext *context) const override {
     if (context->HasInput("X")) {
       auto type = context->GetInputsVarType("X")[0];
-      if (type == framework::VarDesc_VarType_SELECTED_ROWS ||
-          type == framework::VarDesc_VarType_LOD_TENSOR) {
+      if (type == framework::proto::VarDesc_VarType_SELECTED_ROWS ||
+          type == framework::proto::VarDesc_VarType_LOD_TENSOR) {
         context->SetOutputDim("Out", context->GetInputDim("X"));
       }
     }
@@ -122,12 +121,12 @@ class AssignGradMaker : public framework::SingleGradOpDescMaker {
   using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
 
  protected:
-  std::unique_ptr<framework::OpDescBind> Apply() const override {
-    auto *op = new framework::OpDescBind();
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto *op = new framework::OpDesc();
     op->SetType("assign");
     op->SetInput("X", OutputGrad("Out"));
     op->SetOutput("Out", InputGrad("X"));
-    return std::unique_ptr<framework::OpDescBind>(op);
+    return std::unique_ptr<framework::OpDesc>(op);
   }
 };
 

paddle/operators/auc_op.cc

@@ -49,7 +49,7 @@ class AucOp : public framework::OperatorWithKernel {
 
 class AucOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AucOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  AucOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("Out",
              "A floating point 2D tensor, values are in the range [0, 1]."

paddle/operators/batch_norm_op.cc

@@ -13,12 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/batch_norm_op.h"
+#include "paddle/framework/data_layout.h"
 
 namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
+using DataLayout = framework::DataLayout;
 
 template <typename T>
 using EigenArrayMap =
@@ -60,15 +62,15 @@ class BatchNormOp : public framework::OperatorWithKernel {
                       "Variance and VarianceOut should share the same memory");
 
     const auto x_dims = ctx->GetInputDim("X");
-    const TensorFormat tensor_format =
-        StringToTensorFormat(ctx->Attrs().Get<std::string>("tensor_format"));
+    const DataLayout data_layout = framework::StringToDataLayout(
+        ctx->Attrs().Get<std::string>("data_layout"));
 
     PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
                    "Input X must have 2 to 5 dimensions.");
 
     const int C =
-        (tensor_format == TensorFormat::NCHW ? x_dims[1]
-                                             : x_dims[x_dims.size() - 1]);
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
+                                          : x_dims[x_dims.size() - 1]);
 
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL);
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C);
@@ -85,13 +87,12 @@ class BatchNormOp : public framework::OperatorWithKernel {
 
 class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BatchNormOpMaker(framework::OpProto *proto,
-                   framework::OpAttrChecker *op_checker)
+  BatchNormOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddAttr<bool>("is_test", "").SetDefault(false);
     AddAttr<float>("momentum", "").SetDefault(0.9);
     AddAttr<float>("epsilon", "").SetDefault(1e-5);
-    AddAttr<std::string>("tensor_format", "").SetDefault("NCHW");
+    AddAttr<std::string>("data_layout", "").SetDefault("NCHW");
     AddInput("X", "The input tensor");
     AddInput("Scale",
              "Scale is a 1-dimensional tensor of size C "
@@ -142,9 +143,9 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
     const float epsilon = ctx.Attr<float>("epsilon");
     const float momentum = ctx.Attr<float>("momentum");
     const bool is_test = ctx.Attr<bool>("is_test");
-    const std::string tensor_format_str =
-        ctx.Attr<std::string>("tensor_format");
-    const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str);
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
 
     const auto *x = ctx.Input<Tensor>("X");
     const auto &x_dims = x->dims();
@@ -152,8 +153,8 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
                    "The Input dim size should be between 2 and 5");
     const int N = x_dims[0];
     const int C =
-        (tensor_format == TensorFormat::NCHW ? x_dims[1]
-                                             : x_dims[x_dims.size() - 1]);
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
+                                          : x_dims[x_dims.size() - 1]);
     const int sample_size = x->numel() / N / C;
 
     auto *y = ctx.Output<Tensor>("Y");
@@ -178,8 +179,8 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
       saved_mean_e.setZero();
       saved_variance_e.setZero();
 
-      switch (tensor_format) {
-        case TensorFormat::NCHW: {
+      switch (data_layout) {
+        case DataLayout::kNCHW: {
           ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, N * C);
           for (int nc = 0; nc < N * C; ++nc) {
             saved_mean_e(nc % C) += x_arr.col(nc).sum();
@@ -192,7 +193,7 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
           saved_variance_e /= N * sample_size;
           break;
         }
-        case TensorFormat::NHWC: {
+        case DataLayout::kNHWC: {
           ConstEigenArrayMap<T> x_arr(x->data<T>(), C, N * sample_size);
           for (int i = 0; i < N * sample_size; ++i) {
             saved_mean_e += x_arr.col(i);
@@ -206,7 +207,7 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
           break;
         }
         default:
-          PADDLE_THROW("Unknown storage order: %s", tensor_format_str);
+          PADDLE_THROW("Unknown storage order: %s", data_layout_str);
       }
 
       EigenVectorArrayMap<T> running_mean_arr(
@@ -248,8 +249,8 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
     Eigen::Array<T, Eigen::Dynamic, 1> new_bias =
         bias_arr - mean_arr * inv_std * scale_arr;
 
-    switch (tensor_format) {
-      case TensorFormat::NCHW: {
+    switch (data_layout) {
+      case DataLayout::kNCHW: {
         EigenArrayMap<T> y_arr(y->mutable_data<T>(ctx.GetPlace()), sample_size,
                                N * C);
         ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, N * C);
@@ -258,7 +259,7 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
         }
         break;
       }
-      case TensorFormat::NHWC: {
+      case DataLayout::kNHWC: {
         EigenArrayMap<T>(y->mutable_data<T>(ctx.GetPlace()), C,
                          N * sample_size) =
             (ConstEigenArrayMap<T>(x->data<T>(), C, N * sample_size).colwise() *
@@ -268,7 +269,7 @@ class BatchNormKernel<platform::CPUDeviceContext, T>
         break;
       }
       default:
-        PADDLE_THROW("Unknown storage order: %d", tensor_format);
+        PADDLE_THROW("Unknown storage order: %d", data_layout);
     }
   }
 };
@@ -291,11 +292,11 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), "");
 
     const auto x_dims = ctx->GetInputDim("X");
-    const TensorFormat tensor_format =
-        StringToTensorFormat(ctx->Attrs().Get<std::string>("tensor_format"));
+    const DataLayout data_layout = framework::StringToDataLayout(
+        ctx->Attrs().Get<std::string>("data_layout"));
     const int C =
-        (tensor_format == TensorFormat::NCHW ? x_dims[1]
-                                             : x_dims[x_dims.size() - 1]);
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
+                                          : x_dims[x_dims.size() - 1]);
 
     ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
     ctx->SetOutputDim(framework::GradVarName("Scale"), {C});
@@ -334,9 +335,9 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
     const auto *saved_mean = ctx.Input<Tensor>("SavedMean");
     // SavedVariance have been reverted in forward operator
     const auto *saved_inv_variance = ctx.Input<Tensor>("SavedVariance");
-    const std::string tensor_format_str =
-        ctx.Attr<std::string>("tensor_format");
-    const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str);
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
 
     // Get the size for each dimension.
     // NCHW [batch_size, in_channels, in_height, in_width]
@@ -345,8 +346,8 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
                    "The Input dim size should be between 2 and 5");
     const int N = x_dims[0];
     const int C =
-        (tensor_format == TensorFormat::NCHW ? x_dims[1]
-                                             : x_dims[x_dims.size() - 1]);
+        (data_layout == DataLayout::kNCHW ? x_dims[1]
+                                          : x_dims[x_dims.size() - 1]);
     const int sample_size = x->numel() / N / C;
 
     ConstEigenVectorArrayMap<T> scale_arr(scale->data<T>(), C);
@@ -377,8 +378,8 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
 
     const auto scale_inv_var_nhw = scale_arr * inv_var_arr / (N * sample_size);
 
-    switch (tensor_format) {
-      case TensorFormat::NCHW: {
+    switch (data_layout) {
+      case DataLayout::kNCHW: {
         ConstEigenArrayMap<T> x_arr(x->data<T>(), sample_size, N * C);
         ConstEigenArrayMap<T> d_y_arr(d_y->data<T>(), sample_size, N * C);
         EigenArrayMap<T> d_x_arr(d_x->mutable_data<T>(ctx.GetPlace()),
@@ -401,7 +402,7 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
         }
         break;
       }
-      case TensorFormat::NHWC: {
+      case DataLayout::kNHWC: {
         ConstEigenArrayMap<T> x_arr(x->data<T>(), C, N * sample_size);
         ConstEigenArrayMap<T> d_y_arr(d_y->data<T>(), C, N * sample_size);
         EigenArrayMap<T> d_x_arr(d_x->mutable_data<T>(ctx.GetPlace()), C,
@@ -426,7 +427,7 @@ class BatchNormGradKernel<platform::CPUDeviceContext, T>
         break;
       }
       default:
-        PADDLE_THROW("Unknown storage order: %s", tensor_format_str);
+        PADDLE_THROW("Unknown storage order: %s", data_layout_str);
     }
   }
 };

paddle/operators/batch_norm_op.cu.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/batch_norm_op.h"
+#include "paddle/framework/data_layout.h"
 
 #include <cfloat>
 #include "paddle/operators/math/math_function.h"
@@ -22,12 +23,12 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+using DataLayout = framework::DataLayout;
 template <typename T>
 using CudnnDataType = platform::CudnnDataType<T>;
 
-void ExtractNCWHD(const framework::DDim &dims,
-                  const TensorFormat &tensor_format, int *N, int *C, int *H,
-                  int *W, int *D) {
+void ExtractNCWHD(const framework::DDim &dims, const DataLayout &data_layout,
+                  int *N, int *C, int *H, int *W, int *D) {
   *N = dims[0];
   if (dims.size() == 2) {
     *C = dims[1];
@@ -35,13 +36,13 @@ void ExtractNCWHD(const framework::DDim &dims,
     *W = 1;
     *D = 1;
   } else {
-    *C = tensor_format == TensorFormat::NCHW ? dims[1] : dims[dims.size() - 1];
-    *H = tensor_format == TensorFormat::NCHW ? dims[2] : dims[1];
+    *C = data_layout == DataLayout::kNCHW ? dims[1] : dims[dims.size() - 1];
+    *H = data_layout == DataLayout::kNCHW ? dims[2] : dims[1];
     *W = dims.size() > 3
-             ? (tensor_format == TensorFormat::NCHW ? dims[3] : dims[2])
+             ? (data_layout == DataLayout::kNCHW ? dims[3] : dims[2])
              : 1;
     *D = dims.size() > 4
-             ? (tensor_format == TensorFormat::NCHW ? dims[4] : dims[3])
+             ? (data_layout == DataLayout::kNCHW ? dims[4] : dims[3])
              : 1;
   }
 }
@@ -56,9 +57,9 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
     double epsilon = static_cast<double>(ctx.Attr<float>("epsilon"));
     const float momentum = ctx.Attr<float>("momentum");
     const bool is_test = ctx.Attr<bool>("is_test");
-    const std::string tensor_format_str =
-        ctx.Attr<std::string>("tensor_format");
-    const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str);
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
 
     // Get the size for each dimension.
     // NCHW [batch_size, in_channels, in_height, in_width]
@@ -67,7 +68,7 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
     PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
                    "The Input dim size should be between 2 and 5");
     int N, C, H, W, D;
-    ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D);
+    ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D);
 
     // ------------------- cudnn descriptors ---------------------
     cudnnTensorDescriptor_t data_desc_;
@@ -93,7 +94,7 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
     VLOG(1) << "Setting descriptors.";
     std::vector<int> dims;
     std::vector<int> strides;
-    if (tensor_format == TensorFormat::NCHW) {
+    if (data_layout == DataLayout::kNCHW) {
       dims = {N, C, H, W, D};
       strides = {C * H * W * D, H * W * D, W * D, D, 1};
     } else {
@@ -180,9 +181,9 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
                    "It must use GPUPlace.");
     double epsilon = static_cast<double>(ctx.Attr<float>("epsilon"));
-    const std::string tensor_format_str =
-        ctx.Attr<std::string>("tensor_format");
-    const TensorFormat tensor_format = StringToTensorFormat(tensor_format_str);
+    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
+    const DataLayout data_layout =
+        framework::StringToDataLayout(data_layout_str);
     const auto *x = ctx.Input<Tensor>("X");
     const auto *d_y = ctx.Input<Tensor>(framework::GradVarName("Y"));
     const auto *scale = ctx.Input<Tensor>("Scale");
@@ -192,7 +193,7 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
     PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
                    "The Input dim size should be between 2 and 5");
     int N, C, H, W, D;
-    ExtractNCWHD(x_dims, tensor_format, &N, &C, &H, &W, &D);
+    ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D);
 
     PADDLE_ENFORCE_EQ(scale->dims().size(), 1UL);
     PADDLE_ENFORCE_EQ(scale->dims()[0], C);
@@ -219,7 +220,7 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
 
     std::vector<int> dims;
     std::vector<int> strides;
-    if (tensor_format == TensorFormat::NCHW) {
+    if (data_layout == DataLayout::kNCHW) {
       dims = {N, C, H, W, D};
       strides = {C * H * W * D, H * W * D, W * D, D, 1};
     } else {

paddle/operators/batch_norm_op.h

@@ -19,21 +19,6 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-enum TensorFormat {
-  NHWC = 0,
-  NCHW = 1,
-};
-
-inline TensorFormat StringToTensorFormat(const std::string& str) {
-  if (str == "NHWC" || str == "nhwc") {
-    return TensorFormat::NHWC;
-  } else if (str == "NCHW" || str == "nchw") {
-    return TensorFormat::NCHW;
-  } else {
-    PADDLE_THROW("Unknown storage order string: %s", str);
-  }
-}
-
 template <typename DeviceContext, typename T>
 class BatchNormKernel : public framework::OpKernel<T> {
  public:

paddle/operators/beam_search_decode_op.cc

@@ -83,9 +83,8 @@ class BeamSearchDecodeOp : public framework::OperatorBase {
 
 class BeamSearchDecodeOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BeamSearchDecodeOpProtoMaker(framework::OpProto* proto,
-                               framework::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
+  BeamSearchDecodeOpProtoMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("Ids",
              "(LodTensorArray)"
              "score of the candidate words in each step");
@@ -120,13 +119,13 @@ class BeamSearchDecodeInferShape : public framework::InferShapeBase {
 
 class BeamSearchDecodeInferVarType : public framework::VarTypeInference {
  public:
-  void operator()(const framework::OpDescBind& op_desc,
-                  framework::BlockDescBind* block) const override {
+  void operator()(const framework::OpDesc& op_desc,
+                  framework::BlockDesc* block) const override {
     for (auto& o : op_desc.Output("SentenceIds")) {
-      block->Var(o)->SetType(framework::VarDesc::LOD_TENSOR);
+      block->Var(o)->SetType(framework::proto::VarDesc::LOD_TENSOR);
     }
     for (auto& o : op_desc.Output("SentenceScores")) {
-      block->Var(o)->SetType(framework::VarDesc::LOD_TENSOR);
+      block->Var(o)->SetType(framework::proto::VarDesc::LOD_TENSOR);
     }
   }
 };

paddle/operators/beam_search_op.cc

@@ -153,8 +153,7 @@ bool BeamSearch::NextItemSet(std::vector<BeamSearch::Item> *items) {
 class BeamSearchProtoAndCheckerMaker
     : public framework::OpProtoAndCheckerMaker {
  public:
-  BeamSearchProtoAndCheckerMaker(framework::OpProto *proto,
-                                 framework::OpAttrChecker *op_checker)
+  BeamSearchProtoAndCheckerMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     // inputs and outputs stored in proto
     AddInput("pre_ids", "ids in previous step");

paddle/operators/bilinear_tensor_product_op.cc

@@ -65,8 +65,7 @@ class BilinearTensorProductOp : public framework::OperatorWithKernel {
 
 class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  BilinearTensorProductOpMaker(framework::OpProto* proto,
-                               framework::OpAttrChecker* op_checker)
+  BilinearTensorProductOpMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The first input of bilinear_tensor_product operator.");
     AddInput("Y", "The second input of bilinear_tensor_product operator.");

paddle/operators/cast_op.cc

@@ -20,8 +20,7 @@ namespace operators {
 
 class CastOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CastOpProtoMaker(framework::OpProto *proto,
-                   framework::OpAttrChecker *op_checker)
+  CastOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input tensor of cast op");
     AddOutput("Out", "The output tensor of cast op");
@@ -53,14 +52,14 @@ class CastOpGradMaker : public framework::SingleGradOpDescMaker {
   using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
 
  protected:
-  std::unique_ptr<framework::OpDescBind> Apply() const override {
-    auto grad = new framework::OpDescBind();
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    auto grad = new framework::OpDesc();
     grad->SetType("cast");
     grad->SetInput("X", OutputGrad("Out"));
     grad->SetOutput("Out", InputGrad("X"));
     grad->SetAttr("out_dtype", GetAttr("in_dtype"));
     grad->SetAttr("in_dtype", GetAttr("out_dtype"));
-    return std::unique_ptr<framework::OpDescBind>(grad);
+    return std::unique_ptr<framework::OpDesc>(grad);
   }
 };
 

paddle/operators/cast_op.h

@@ -55,7 +55,7 @@ class CastOpKernel : public framework::OpKernel<InT> {
     auto* in = context.Input<framework::Tensor>("X");
     auto* out = context.Output<framework::Tensor>("Out");
     framework::VisitDataType(
-        static_cast<framework::DataType>(context.Attr<int>("out_dtype")),
+        static_cast<framework::proto::DataType>(context.Attr<int>("out_dtype")),
         CastOpFunctor<DeviceContext, InT>(
             in, out, context.template device_context<DeviceContext>()));
   }

paddle/operators/chunk_eval_op.cc

@@ -57,15 +57,14 @@ class ChunkEvalOp : public framework::OperatorWithKernel {
  protected:
   framework::OpKernelType GetKernelType(
       const framework::ExecutionContext &ctx) const override {
-    return framework::OpKernelType(framework::DataType::FP32,
+    return framework::OpKernelType(framework::proto::DataType::FP32,
                                    ctx.device_context());
   }
 };
 
 class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ChunkEvalOpMaker(framework::OpProto *proto,
-                   framework::OpAttrChecker *op_checker)
+  ChunkEvalOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("Inference",
              "(Tensor, default: Tensor<int64_t>). "

paddle/operators/clip_by_norm_op.cc

@@ -37,8 +37,7 @@ class ClipByNormOp : public framework::OperatorWithKernel {
 
 class ClipByNormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ClipByNormOpMaker(framework::OpProto* proto,
-                    framework::OpAttrChecker* op_checker)
+  ClipByNormOpMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X",
              "(Tensor) The input of clip_by_norm op."

paddle/operators/clip_op.cc

@@ -38,7 +38,7 @@ class ClipOp : public framework::OperatorWithKernel {
 template <typename AttrType>
 class ClipOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ClipOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+  ClipOpMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X",
              "(Tensor)The input of clip op."

paddle/operators/compare_op.cc

@@ -20,8 +20,7 @@ namespace operators {
 template <typename OpComment>
 class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CompareOpProtoMaker(framework::OpProto *proto,
-                      framework::OpAttrChecker *op_checker)
+  CompareOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     OpComment comment;
     AddInput("X",

paddle/operators/concat_op.cc

@@ -58,7 +58,7 @@ class ConcatOp : public framework::OperatorWithKernel {
 
 class ConcatOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ConcatOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  ConcatOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "Input tensors of concat operator.").AsDuplicable();
     AddOutput("Out", "Output tensor of concat operator.");
@@ -98,8 +98,8 @@ class ConcatOpGrad : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad,
-            ops::ConcatOpGrad)
+REGISTER_OP_EX(concat, ops::ConcatOp, ops::ConcatOpMaker, concat_grad,
+               ops::ConcatOpGrad, false)
 REGISTER_OP_CPU_KERNEL(concat,
                        ops::ConcatKernel<paddle::platform::CPUPlace, float>)
 REGISTER_OP_CPU_KERNEL(concat_grad,

paddle/operators/cond_op.cc

@@ -205,8 +205,7 @@ void CondOp::Run(const Scope& scope,
 
 class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CondOpProtoAndCheckerMaker(framework::OpProto* proto,
-                             framework::OpAttrChecker* op_checker)
+  CondOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("Cond", "The condition, which is a bool vector");
     AddInput("Xs", "Inputs of Subnets").AsDuplicable();