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

Dockerfile

@@ -22,7 +22,7 @@ COPY ./paddle/scripts/docker/root/ /root/
 
 RUN apt-get update && \
     apt-get install -y \
-    git python-pip python-dev openssh-server bison  \
+    git python-pip python-dev openssh-server bison libnccl-dev \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
     python-matplotlib gcc-4.8 g++-4.8 \

doc/design/block.md

@@ -189,7 +189,7 @@ OpDesc {
   inputs = {0} // the index of x in vars of BlockDesc above
   outputs = {5, 3} // indices of act and hidden_out in vars of BlockDesc above
   attrs {
-    "memories" : {1} // the index of h
+    "states" : {1} // the index of h
     "step_net" : <above step net>
   }
 };

doc/design/register_grad_op.md

@@ -3,17 +3,17 @@
 
 ## The Problem Posed
 
-Currently, for each C++ operator class definition, there registers a *gradient operator creator* function, which takes a C++ operator instance and returns the corresponding gradient operator instance.
+Currently, for each C++ operator class definition, a *gradient operator creator* function is registered, which takes as input a C++ operator instance and returns the corresponding gradient operator instance.
 
-However, we noticed two problems with the current deisgn:
+However, we noticed two problems with the current design:
 
-1. As we decided to separate the *compilation* and *execution* phases, we need to change the creator to take an `OpDesc` protobuf message in a `ProgramDesc` and inserts corresponding `OpDesc` messages into the `ProgramDesc` message.
+1. As we decided to separate the *compilation* and the *execution* phases, we need to change the creator to take an `OpDesc` protobuf message in a `ProgramDesc` and inserts corresponding `OpDesc` messages into the `ProgramDesc` message.
 
-1. Some operator's gradient computation requires more than one gradient operators.  For example, the gradient of *minus* consists of two operators -- an identity operaotr and a scale operator.  So we need to make the registration mechanism to support the mapping from an operator to a set of operators for gradient computation.
+1. For some operators, the gradient computation can be written in terms of existing operators.  For example, the gradient of *minus* operator consists of two operators -- an *identity* operator followed by a *scale* operator.  Hence the registration mechanism needs to support mapping from an operator to a set of operators for the gradient computation.
 
 ## The Current Implementation
 
-The C++ class `OpInfos` store in a association map which key is the operator type. The `grad_op_type` indicate associated gradient operator type. Operator can create gradient operator by `OpInfo::creator_` of gradient. The pseudo code is
+Instances of the C++ class `OpInfo` are stored an associative map whose key is the operator type. The `grad_op_type` indicates the associated gradient operator type. An operator can create the gradient operator by invoking `OpInfo::creator_` of the gradient operator. The pseudo code is as follows
 
 ```cpp
 struct OpInfo {
@@ -31,16 +31,16 @@ OperatorBase* CreateGradientOperator(const OperatorBase& op) {
 
 ## Proposed Solution
 
-The mapping relationship between an operator and its gradient operators is a function. The interface of that function is:
+The mapping relationship between an operator and its gradient operators is a function. The interface of this function is:
 
 ```cpp
 // (OpDesc) --> vector<OpDesc>
 std::function<std::vector<OpDescBind>(const OpDescBind&)>;
 ```
 
-The function takes an `OpDescBind` of the forward operator and returns one or many gradient operator descriptions. `OpDescBind` is a C++ wrapper for protobuf message `OpDesc` to manipulate `OpDesc` fast.
+The function takes an `OpDescBind` of the forward operator and returns one or many gradient operator descriptions. `OpDescBind` is a C++ wrapper for  the protobuf message `OpDesc` for rapid manipulation of `OpDesc`.
 
-The `GradOpDescMaker` will be registered in `OpInfo`, to replace `grad_op_type_` field. The `OpInfo` should be
+The `GradOpDescMaker` will be registered in `OpInfo` and will replace the `grad_op_type_` field. The `OpInfo` should look like 
 
 ```cpp
 struct OpInfo {
@@ -49,7 +49,7 @@ struct OpInfo {
 };
 ```
 
-The `grad_op_maker_ ` is `nullptr` if the operator does not have associated gradient operators.
+The `grad_op_maker_ ` is a `nullptr` if the operator does not have any associated gradient operators.
 
 We propose a base class called `GradOpDescMakerBase` to let operator developers generate `Gradient Operators` easily. The public interface of that class is
 
@@ -74,7 +74,7 @@ func = [] (const OpDescBind& fwd_op) {
 
 We can write many helper functions since the `GradOpDescMakerBase` is a class now. The basic helper functions get the variables of `Input`, `Output`, `InputGradient` and `OutputGradient` in the forwarding operator.
 
-We should chagne register macros at the same time. In the current solution, there is no difference between forwarding operators and backward operators. So `REGISTER_OP` just register one operator. If the `REGISTER_OPERATOR ` contains `OpProtoAndCheckerMaker` and `GradOpDescMaker`, we just list them in the same macro. It can be done by a macro contains `__VA_ARGS__`.
+We should change register macros at the same time. In the current solution, there is no difference between forwarding operators and backward operators. So `REGISTER_OP` just register one operator. If the `REGISTER_OPERATOR ` contains `OpProtoAndCheckerMaker` and `GradOpDescMaker`, we just list them in the same macro. It can be done by a macro contains `__VA_ARGS__`.
 
 The user interface should be
 

doc/faq/local/index_cn.rst

@@ -174,7 +174,7 @@ decoder_inputs = paddle.layer.fc(
 1. 两者都是对梯度的截断，但截断时机不同，前者在 :code:`optimzier` 更新网络参数时应用；后者在激活函数反向计算时被调用；
 2. 截断对象不同：前者截断可学习参数的梯度，后者截断回传给前层的梯度;
 
-除此之外，还可以通过减小学习律或者对数据进行归一化处理来解决这类问题。
+除此之外，还可以通过减小学习率或者对数据进行归一化处理来解决这类问题。
 
 5.  如何调用 infer 接口输出多个layer的预测结果
 -----------------------------------------------

doc/howto/usage/cluster/src/word2vec/api_train_v2.py

+import gzip
+import math
+
+import paddle.v2 as paddle
+
+embsize = 32
+hiddensize = 256
+N = 5
+
+
+def wordemb(inlayer):
+    wordemb = paddle.layer.embedding(
+        input=inlayer,
+        size=embsize,
+        param_attr=paddle.attr.Param(
+            name="_proj",
+            initial_std=0.001,
+            learning_rate=1,
+            l2_rate=0,
+            sparse_update=True))
+    return wordemb
+
+
+def main():
+    # for local training
+    cluster_train = False
+
+    if not cluster_train:
+        paddle.init(use_gpu=False, trainer_count=1)
+    else:
+        paddle.init(
+            use_gpu=False,
+            trainer_count=2,
+            port=7164,
+            ports_num=1,
+            ports_num_for_sparse=1,
+            num_gradient_servers=1)
+    word_dict = paddle.dataset.imikolov.build_dict()
+    dict_size = len(word_dict)
+    firstword = paddle.layer.data(
+        name="firstw", type=paddle.data_type.integer_value(dict_size))
+    secondword = paddle.layer.data(
+        name="secondw", type=paddle.data_type.integer_value(dict_size))
+    thirdword = paddle.layer.data(
+        name="thirdw", type=paddle.data_type.integer_value(dict_size))
+    fourthword = paddle.layer.data(
+        name="fourthw", type=paddle.data_type.integer_value(dict_size))
+    nextword = paddle.layer.data(
+        name="fifthw", type=paddle.data_type.integer_value(dict_size))
+
+    Efirst = wordemb(firstword)
+    Esecond = wordemb(secondword)
+    Ethird = wordemb(thirdword)
+    Efourth = wordemb(fourthword)
+
+    contextemb = paddle.layer.concat(input=[Efirst, Esecond, Ethird, Efourth])
+    hidden1 = paddle.layer.fc(input=contextemb,
+                              size=hiddensize,
+                              act=paddle.activation.Sigmoid(),
+                              layer_attr=paddle.attr.Extra(drop_rate=0.5),
+                              bias_attr=paddle.attr.Param(learning_rate=2),
+                              param_attr=paddle.attr.Param(
+                                  initial_std=1. / math.sqrt(embsize * 8),
+                                  learning_rate=1))
+    predictword = paddle.layer.fc(input=hidden1,
+                                  size=dict_size,
+                                  bias_attr=paddle.attr.Param(learning_rate=2),
+                                  act=paddle.activation.Softmax())
+
+    def event_handler(event):
+        if isinstance(event, paddle.event.EndIteration):
+            if event.batch_id % 100 == 0:
+                with gzip.open("batch-" + str(event.batch_id) + ".tar.gz",
+                               'w') as f:
+                    trainer.save_parameter_to_tar(f)
+                result = trainer.test(
+                    paddle.batch(
+                        paddle.dataset.imikolov.test(word_dict, N), 32))
+                print "Pass %d, Batch %d, Cost %f, %s, Testing metrics %s" % (
+                    event.pass_id, event.batch_id, event.cost, event.metrics,
+                    result.metrics)
+
+    cost = paddle.layer.classification_cost(input=predictword, label=nextword)
+
+    parameters = paddle.parameters.create(cost)
+    adagrad = paddle.optimizer.AdaGrad(
+        learning_rate=3e-3,
+        regularization=paddle.optimizer.L2Regularization(8e-4))
+    trainer = paddle.trainer.SGD(cost,
+                                 parameters,
+                                 adagrad,
+                                 is_local=not cluster_train)
+    trainer.train(
+        paddle.batch(paddle.dataset.imikolov.train(word_dict, N), 32),
+        num_passes=30,
+        event_handler=event_handler)
+
+
+if __name__ == '__main__':
+    main()

doc/howto/usage/cluster/src/word2vec/api_train_v2_cluster.py

+import math
+import os
+import paddle.v2 as paddle
+import pickle
+
+embsize = 32
+hiddensize = 256
+N = 5
+cluster_train_file = "./train_data_dir/train/train.txt"
+cluster_test_file = "./test_data_dir/test/test.txt"
+node_id = os.getenv("OMPI_COMM_WORLD_RANK")
+if not node_id:
+    raise EnvironmentError("must provied OMPI_COMM_WORLD_RANK")
+
+
+def wordemb(inlayer):
+    wordemb = paddle.layer.embedding(
+        input=inlayer,
+        size=embsize,
+        param_attr=paddle.attr.Param(
+            name="_proj",
+            initial_std=0.001,
+            learning_rate=1,
+            l2_rate=0,
+            sparse_update=True))
+    return wordemb
+
+
+def cluster_reader_cluster(filename, node_id):
+    def cluster_reader():
+        with open("-".join([filename, "%05d" % int(node_id)]), "r") as f:
+            for l in f:
+                csv_data = [int(cell) for cell in l.split(",")]
+                yield tuple(csv_data)
+
+    return cluster_reader
+
+
+def main():
+    # get arguments from env
+
+    # for local training
+    TRUTH = ["true", "True", "TRUE", "1", "yes", "Yes", "YES"]
+    cluster_train = os.getenv('PADDLE_CLUSTER_TRAIN', "False") in TRUTH
+    use_gpu = os.getenv('PADDLE_INIT_USE_GPU', "False")
+
+    if not cluster_train:
+        paddle.init(
+            use_gpu=use_gpu,
+            trainer_count=int(os.getenv("PADDLE_INIT_TRAINER_COUNT", "1")))
+    else:
+        paddle.init(
+            use_gpu=use_gpu,
+            trainer_count=int(os.getenv("PADDLE_INIT_TRAINER_COUNT", "1")),
+            port=int(os.getenv("PADDLE_INIT_PORT", "7164")),
+            ports_num=int(os.getenv("PADDLE_INIT_PORTS_NUM", "1")),
+            ports_num_for_sparse=int(
+                os.getenv("PADDLE_INIT_PORTS_NUM_FOR_SPARSE", "1")),
+            num_gradient_servers=int(
+                os.getenv("PADDLE_INIT_NUM_GRADIENT_SERVERS", "1")),
+            trainer_id=int(os.getenv("PADDLE_INIT_TRAINER_ID", "0")),
+            pservers=os.getenv("PADDLE_INIT_PSERVERS", "127.0.0.1"))
+    fn = open("thirdparty/wuyi_train_thdpty/word_dict.pickle", "r")
+    word_dict = pickle.load(fn)
+    fn.close()
+    dict_size = len(word_dict)
+    firstword = paddle.layer.data(
+        name="firstw", type=paddle.data_type.integer_value(dict_size))
+    secondword = paddle.layer.data(
+        name="secondw", type=paddle.data_type.integer_value(dict_size))
+    thirdword = paddle.layer.data(
+        name="thirdw", type=paddle.data_type.integer_value(dict_size))
+    fourthword = paddle.layer.data(
+        name="fourthw", type=paddle.data_type.integer_value(dict_size))
+    nextword = paddle.layer.data(
+        name="fifthw", type=paddle.data_type.integer_value(dict_size))
+
+    Efirst = wordemb(firstword)
+    Esecond = wordemb(secondword)
+    Ethird = wordemb(thirdword)
+    Efourth = wordemb(fourthword)
+
+    contextemb = paddle.layer.concat(input=[Efirst, Esecond, Ethird, Efourth])
+    hidden1 = paddle.layer.fc(input=contextemb,
+                              size=hiddensize,
+                              act=paddle.activation.Sigmoid(),
+                              layer_attr=paddle.attr.Extra(drop_rate=0.5),
+                              bias_attr=paddle.attr.Param(learning_rate=2),
+                              param_attr=paddle.attr.Param(
+                                  initial_std=1. / math.sqrt(embsize * 8),
+                                  learning_rate=1))
+    predictword = paddle.layer.fc(input=hidden1,
+                                  size=dict_size,
+                                  bias_attr=paddle.attr.Param(learning_rate=2),
+                                  act=paddle.activation.Softmax())
+
+    def event_handler(event):
+        if isinstance(event, paddle.event.EndIteration):
+            if event.batch_id % 100 == 0:
+                result = trainer.test(
+                    paddle.batch(
+                        cluster_reader_cluster(cluster_test_file, node_id), 32))
+                print "Pass %d, Batch %d, Cost %f, %s, Testing metrics %s" % (
+                    event.pass_id, event.batch_id, event.cost, event.metrics,
+                    result.metrics)
+
+    cost = paddle.layer.classification_cost(input=predictword, label=nextword)
+    parameters = paddle.parameters.create(cost)
+    adagrad = paddle.optimizer.AdaGrad(
+        learning_rate=3e-3,
+        regularization=paddle.optimizer.L2Regularization(8e-4))
+    trainer = paddle.trainer.SGD(cost,
+                                 parameters,
+                                 adagrad,
+                                 is_local=not cluster_train)
+    trainer.train(
+        paddle.batch(cluster_reader_cluster(cluster_train_file, node_id), 32),
+        num_passes=30,
+        event_handler=event_handler)
+
+
+if __name__ == '__main__':
+    main()

doc/howto/usage/cluster/src/word2vec/prepare.py

+import paddle.v2 as paddle
+import tarfile
+import os
+import pickle
+
+SPLIT_COUNT = 3
+N = 5
+
+
+def file_len(fd):
+    for i, l in enumerate(fd):
+        pass
+    return i + 1
+
+
+def split_from_reader_by_line(filename, reader, split_count):
+    fn = open(filename, "w")
+    for batch_id, batch_data in enumerate(reader()):
+        batch_data_str = [str(d) for d in batch_data]
+        fn.write(",".join(batch_data_str))
+        fn.write("\n")
+    fn.close()
+
+    fn = open(filename, "r")
+    total_line_count = file_len(fn)
+    fn.close()
+    per_file_lines = total_line_count / split_count + 1
+    cmd = "split -d -a 5 -l %d %s %s-" % (per_file_lines, filename, filename)
+    os.system(cmd)
+
+
+word_dict = paddle.dataset.imikolov.build_dict()
+with open("word_dict.pickle", "w") as dict_f:
+    pickle.dump(word_dict, dict_f)
+
+split_from_reader_by_line("train.txt",
+                          paddle.dataset.imikolov.train(word_dict, N),
+                          SPLIT_COUNT)
+split_from_reader_by_line("test.txt",
+                          paddle.dataset.imikolov.test(word_dict, N),
+                          SPLIT_COUNT)

paddle/capi/CMakeLists.txt

@@ -28,23 +28,37 @@ add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER}
 
 add_dependencies(paddle_capi paddle_proto)
 
-# combine all paddle static libraries together, into libpaddle_capi_whole.a
-# user should use PaddleCAPI as -lpaddle_capi_whole
-set(PADDLE_CAPI_INFER_LIBS
-    paddle_utils
-    paddle_parameter
-    paddle_math
-    paddle_cuda
-    paddle_function
-    paddle_gserver
-    paddle_proto)
-
+# TODO: paddle_capi_whole will be removed.
+if(MOBILE_INFERENCE)
+    set(PADDLE_CAPI_INFER_LIBS
+        paddle_utils
+        paddle_parameter
+        paddle_math
+        paddle_cuda
+        paddle_function
+        paddle_gserver
+        paddle_proto)
+else()
+    set(PADDLE_CAPI_INFER_LIBS
+        paddle_utils
+        paddle_parameter
+        paddle_math
+        paddle_cuda
+        paddle_function
+        paddle_gserver
+        paddle_proto
+        paddle_pserver
+        paddle_network)
+endif()
 cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
 
-# No shared library for iOS
+# Link the static library for inference
+cc_library(paddle_capi_engine DEPS paddle_capi paddle_utils paddle_parameter paddle_math paddle_cuda paddle_proto)
+cc_library(paddle_capi_layers DEPS paddle_function paddle_gserver)
+
+# Link the shared library for inference
 if(NOT IOS)
-  set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
-  # TODO: merge mkl into paddle_capi_shared
+  set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_capi.map")
   add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
   set_target_properties(paddle_capi_shared	PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
   target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
@@ -53,9 +67,10 @@ endif()
 
 # install library & headers.
 install(FILES ${CAPI_HEADERS} DESTINATION include/paddle)
+install(FILES paddle_capi.map DESTINATION include/paddle)
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/config.h DESTINATION include/paddle)
 if(ANDROID)
-  install(TARGETS paddle_capi_whole paddle_capi_shared
+  install(TARGETS paddle_capi_whole paddle_capi_engine paddle_capi_layers paddle_capi_shared
           ARCHIVE DESTINATION lib/${ANDROID_ABI}
           LIBRARY DESTINATION lib/${ANDROID_ABI})
   execute_process(
@@ -80,7 +95,7 @@ if(ANDROID)
       )"
   )
 else(ANDROID)
-  install(TARGETS paddle_capi_whole ARCHIVE DESTINATION lib)
+  install(TARGETS paddle_capi_whole paddle_capi_engine paddle_capi_layers ARCHIVE DESTINATION lib)
   if(NOT IOS)
     install(TARGETS paddle_capi_shared DESTINATION lib)
   endif()

paddle/framework/CMakeLists.txt

@@ -19,15 +19,15 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope)
 proto_library(framework_proto SRCS framework.proto)
 
 cc_library(attribute SRCS attribute.cc DEPS framework_proto)
-cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS attribute ddim op_info)
 cc_test(program_desc_test SRCS program_desc_test.cc DEPS proto_desc)
 cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
-cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope proto_desc glog)
+cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
+cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS attribute ddim op_info operator)
 
-cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog)
+cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 
 py_proto_compile(framework_py_proto SRCS framework.proto)
@@ -43,7 +43,7 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
 cc_library(backward SRCS backward.cc DEPS net_op)
 cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context)
 
-cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto backward)
+cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto backward glog)
 
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)

paddle/framework/backward.cc

@@ -21,6 +21,7 @@
 
 #include "paddle/framework/block_desc.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/dynamic_recurrent_op.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
 
@@ -220,8 +221,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
     // process recurrent gradient op as a special operator.
     if (forwardOp.Type() == "recurrent") {
       // NOTE clean up cycle call somewhere (RNN's stepnet constains itself),
-      // or
-      // this will result in infinite loop.
+      // or this will result in infinite loop.
       const auto& rnnop =
           *static_cast<const operators::RecurrentOp*>(&forwardOp);
       auto rnn_grad_op =
@@ -231,6 +231,18 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
       // create stepnet's gradient op
       rnn_grad_op->set_stepnet(
           BackwardRecursive(stepnet_op, no_grad_names, grad_to_var, uniq_id));
+    } else if (forwardOp.Type() == "dynamic_recurrent") {
+      // NOTE clean up cycle call somewhere (RNN's stepnet constains itself),
+      // or this will result in infinite loop.
+      const auto& rnnop =
+          *static_cast<const operators::DynamicRecurrentOp*>(&forwardOp);
+      auto rnn_grad_op =
+          static_cast<operators::DynamicRecurrentGradientOp*>(grad_op.get());
+      const auto& stepnet_op =
+          *static_cast<const OperatorBase*>(&rnnop.rnn.GetStepUnit());
+      // create stepnet's gradient op
+      rnn_grad_op->rnn.SetStepUnit(
+          BackwardRecursive(stepnet_op, no_grad_names, grad_to_var, uniq_id));
     }
 
     if (net->ops_.empty()) {  // Current no aux op is added to network

paddle/framework/block_desc.cc

@@ -41,6 +41,19 @@ bool BlockDescBind::HasVar(const std::string &name) const {
   return vars_.find(name) != vars_.end();
 }
 
+VarDescBind *BlockDescBind::FindVarRecursive(const std::string &name) const {
+  auto it = vars_.find(name);
+  if (it == vars_.end()) {
+    return Parent() == kNoneBlockIndex ? nullptr
+                                       : ParentBlock()->FindVarRecursive(name);
+  }
+  return it->second.get();
+}
+
+bool BlockDescBind::HasVarRecursive(const std::string &name) const {
+  return FindVarRecursive(name) != nullptr;
+}
+
 std::vector<VarDescBind *> BlockDescBind::AllVars() const {
   std::vector<VarDescBind *> res;
   for (const auto &p : vars_) {
@@ -97,7 +110,7 @@ void BlockDescBind::Flush() {
 }
 
 BlockDescBind *BlockDescBind::ParentBlock() const {
-  if (this->desc_->parent_idx() == -1) {
+  if (this->desc_->parent_idx() == kNoneBlockIndex) {
     return nullptr;
   }
   return prog_->Block(static_cast<size_t>(this->desc_->parent_idx()));

paddle/framework/block_desc.h

@@ -21,6 +21,7 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/framework/op_desc.h"
+#include "paddle/framework/proto_desc.h"
 #include "paddle/framework/var_desc.h"
 #include "paddle/platform/macros.h"
 
@@ -56,6 +57,10 @@ class BlockDescBind {
 
   bool HasVar(const std::string &var_name) const;
 
+  VarDescBind *FindVarRecursive(const std::string &name_bytes) const;
+
+  bool HasVarRecursive(const std::string &var_name) const;
+
   std::set<std::string> LocalVarNames() const {
     std::set<std::string> var_names;
     for (auto &var : vars_) {

paddle/framework/data_type.h

@@ -26,6 +26,8 @@ inline DataType ToDataType(std::type_index type) {
     return DataType::FP64;
   } else if (typeid(int).hash_code() == type.hash_code()) {
     return DataType::INT32;
+  } else if (typeid(int64_t).hash_code() == type.hash_code()) {
+    return DataType::INT64;
   } else {
     PADDLE_THROW("Not supported");
   }

paddle/framework/executor.cc

@@ -68,9 +68,13 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id) {
 
   for (auto& var : block.vars()) {
     if (var.persistable()) {
-      scope->Var(var.name());
+      auto* ptr = scope->Var(var.name());
+      VLOG(3) << "Create Variable " << var.name()
+              << " global, which pointer is " << ptr;
     } else {
-      local_scope.Var(var.name());
+      auto* ptr = local_scope.Var(var.name());
+      VLOG(3) << "Create Variable " << var.name()
+              << " locally, which pointer is " << ptr;
     }
   }
 
@@ -80,8 +84,7 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id) {
     op->Run(local_scope, *device);
   }
 
-  // TODO(tonyyang-svail):
-  //  - Destroy local_scope
+  scope->DeleteScope(&local_scope);
 }
 
 }  // namespace framework

paddle/framework/feed_fetch_method.h

@@ -13,37 +13,45 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include "glog/logging.h"
+#include "paddle/framework/feed_fetch_type.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/variable.h"
 
 namespace paddle {
 namespace framework {
 
-template <typename T>
-void SetFeedVariable(const LoDTensor& input, const std::string& var_name,
-                     size_t index) {
+void SetFeedVariable(Scope* scope, const LoDTensor& input,
+                     const std::string& var_name, size_t index) {
   // If var_name Variable is not found in GlobalScope, a new variable will
   // be created.
-  Variable* g_feed_value = GetGlobalScope().Var(var_name);
+  VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
+  Variable* g_feed_value = scope->Var(var_name);
   auto& feed_inputs =
       *(g_feed_value->GetMutable<std::vector<paddle::framework::LoDTensor>>());
   if (index >= feed_inputs.size()) {
     feed_inputs.resize(index + 1);
   }
   // shared data with input tensor
-  feed_inputs[index].ShareDataWith<T>(input);
+  feed_inputs[index].ShareDataWith(input);
   // set lod
   feed_inputs[index].set_lod(input.lod());
 }
 
-LoDTensor& GetFetchVariable(const std::string& var_name, size_t index) {
+LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
+                            size_t index) {
   // Since we want to fetch LodTensor from a variable, the variable must
   // be created alreadly.
-  Variable* g_fetch_value = GetGlobalScope().FindVar(var_name);
-  auto& fetch_outputs =
-      *(g_fetch_value->GetMutable<std::vector<paddle::framework::LoDTensor>>());
+  Variable* g_fetch_value = scope.FindVar(var_name);
+  PADDLE_ENFORCE(g_fetch_value->IsType<FeedFetchList>(),
+                 "Only %s can be invoked by GetFetchVariable",
+                 typeid(FeedFetchList).name());
+  auto& fetch_outputs = *g_fetch_value->GetMutable<FeedFetchList>();
+  auto& tensor = fetch_outputs[index];
+  VLOG(3) << "Fetch " << var_name << " with index " << index
+          << " shape= " << tensor.dims();
   PADDLE_ENFORCE_LT(index, fetch_outputs.size());
-  return fetch_outputs[index];
+  return tensor;
 }
 
 }  // namespace framework

paddle/framework/framework.proto

@@ -68,6 +68,7 @@ message OpProto {
 
     optional bool duplicable = 3 [ default = false ];
     optional bool intermediate = 4 [ default = false ];
+    optional bool dispensable = 5 [ default = false ];
   }
 
   // AttrProto describes the C++ type Attribute.
@@ -112,6 +113,8 @@ message VarDesc {
   enum VarType {
     LOD_TENSOR = 1;
     SELECTED_ROWS = 2;
+    FEED_MINIBATCH = 3;
+    FETCH_LIST = 4;
   }
   required string name = 1;
   required VarType type = 2;

paddle/framework/lod_tensor.cc

@@ -25,31 +25,50 @@ LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) {
   for (size_t i = level_begin; i < level_end; i++) {
     new_lod.emplace_back(in.at(i));
   }
+  // transform the lowest level to absolute offset.
+  LoD abs_offset_lod = ToAbsOffset(in);
+  new_lod.back() = abs_offset_lod[level_end - 1];
   return new_lod;
 }
 
 LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
                  size_t elem_end) {
-  // slice the lod.
-  LoD new_lod;
-  new_lod.reserve(in.size() - level);
-  auto start = in.at(level)[elem_begin];
-  auto end = in.at(level)[elem_end];
-
-  for (auto it = in.begin() + level; it != in.end(); it++) {
-    auto it_begin = std::find(it->begin(), it->end(), start);
-    auto it_end = std::find(it_begin, it->end(), end);
-    PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
-    PADDLE_ENFORCE(it_end != it->end(), "error in parsing lod info");
-    new_lod.emplace_back(it_begin, it_end + 1);
-    // reset offset if tensor is copyed and sliced.
-    std::transform(new_lod.back().begin(), new_lod.back().end(),
-                   new_lod.back().begin(),
-                   [start](int v) { return v - start; });
-    PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LoD");
+  PADDLE_ENFORCE_LT(level, in.size());
+  PADDLE_ENFORCE_LT(elem_end, in[level].size());
+
+  LoD res;
+  res.resize(in.size() - level);
+  // copy the first level
+  res[0].assign(in[level].begin() + elem_begin,
+                in[level].begin() + elem_end + 1);
+  for (size_t lvl = 1; lvl < res.size(); lvl++) {
+    const auto& in_level = in[level + lvl];
+    const auto& above_level = res[lvl - 1];
+    auto& out_level = res[lvl];
+    out_level.assign(in_level.begin() + above_level.front(),
+                     in_level.begin() + above_level.back() + 1);
   }
-  PADDLE_ENFORCE_LE(new_lod.size(), in.size());
-  return new_lod;
+  for (size_t lvl = 0; lvl < res.size(); lvl++) {
+    // to make the first offset equals 0, all the elements minus the first
+    // element
+    size_t front = res[lvl].front();
+    for (auto& ele : res[lvl]) {
+      ele -= front;
+    }
+  }
+  return res;
+}
+
+LoD ToAbsOffset(const LoD& in) {
+  // the lowest level stores relative offsets
+  if (in.empty() || in.size() == 1) return in;
+  LoD result = in;
+  for (int level = result.size() - 2; level >= 0; level--) {
+    for (auto& ele : result[level]) {
+      ele = result[level + 1][ele];
+    }
+  }
+  return result;
 }
 
 bool operator==(const LoD& a, const LoD& b) {
@@ -75,17 +94,7 @@ bool operator==(const LoD& a, const LoD& b) {
 size_t LoDTensor::NumElements(size_t level, size_t idx) const {
   PADDLE_ENFORCE_LT(level, NumLevels());
   PADDLE_ENFORCE_LT(idx, NumElements(level));
-  // the last level of LoD, just return number of records in Tensor
-  if (level == NumLevels() - 1) {
-    return lod_[level][idx + 1] - lod_[level][idx];
-  }
-  // high level of LoD, and there is another lower level, return number of
-  // lower-level elements
-  auto tmp = SliceInLevel(lod_, level, idx, idx + 1);
-  PADDLE_ENFORCE_GE(tmp.size(), 2);
-  // there is a 0 as a placeholder stored in LoD, so the number of elements
-  // equals lod.size() - 1
-  return tmp[1].size() - 1;
+  return lod_[level][idx + 1] - lod_[level][idx];
 }
 
 void LoDTensor::ShrinkLevels(size_t level_begin, size_t level_end) {

paddle/framework/lod_tensor.h

@@ -39,23 +39,36 @@ using Vector = thrust::host_vector<
 #endif
 
 /*
- * 3-level LoD stores
+ * LoD is short for Level of Details.
  *
- * 0 10 20
- * 0 5 10 15 20
- * 0 2 5 7 10 12 15 20
- *
- * - in a level, each element indicates offset in the underlying Tensor
+ * - in a level, each element indicates relative offset of the lower level
  * - the first element should be 0 and that indicates that this sequence start
  * from 0
  * - each sequence's begin and end(no-inclusive) is level[id, id+1]
+ *
+ * For example:
+ *    3-level LoD stores
+ *
+ *    0 2 3
+ *    0 2 4 7
+ *    0 2 5 7 10 12 15 20
  */
 using LoD = std::vector<Vector<size_t>>;
 
+/*
+ * Slice levels from a LoD.
+ * NOTE the lowest level should always be the absolute offsets of the underlying
+ * tensor instances. So if higher layers are sliced without the lowest level,
+ * the lower level of the sliced LoD will be transformed to the absolute offset.
+ */
 LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end);
 
 LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
                  size_t elem_end);
+/*
+ * Transform an LoD from relative offsets to absolute offsets.
+ */
+LoD ToAbsOffset(const LoD& in);
 
 bool operator==(const LoD& a, const LoD& b);
 

paddle/framework/lod_tensor_test.cc

@@ -30,8 +30,8 @@ class LoDTensorTester : public ::testing::Test {
     // 0 5 10 15 20
     // 0 2 5 7 10 12 15 20
     LoD lod;
-    lod.push_back(std::vector<size_t>{0, 10, 20});
-    lod.push_back(std::vector<size_t>{0, 5, 10, 15, 20});
+    lod.push_back(std::vector<size_t>{0, 2, 3});
+    lod.push_back(std::vector<size_t>{0, 2, 5, 8});
     lod.push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
 
     ASSERT_EQ(lod.size(), 3UL);
@@ -52,14 +52,14 @@ TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor_.NumLevels(), 3UL); }
 
 TEST_F(LoDTensorTester, NumElements) {
   ASSERT_EQ(lod_tensor_.NumElements(0), 2UL);
-  ASSERT_EQ(lod_tensor_.NumElements(1), 4UL);
+  ASSERT_EQ(lod_tensor_.NumElements(1), 3UL);
   ASSERT_EQ(lod_tensor_.NumElements(2), 8UL);
 }
 
 TEST_F(LoDTensorTester, NumElements2) {
   ASSERT_EQ(lod_tensor_.NumElements(0, 0), 2UL);
-  ASSERT_EQ(lod_tensor_.NumElements(0, 1), 2UL);
-  ASSERT_EQ(lod_tensor_.NumElements(1, 1), 2UL);
+  ASSERT_EQ(lod_tensor_.NumElements(0, 1), 1UL);
+  ASSERT_EQ(lod_tensor_.NumElements(1, 1), 3UL);
 }
 
 TEST_F(LoDTensorTester, ShrinkLevels) {
@@ -68,17 +68,16 @@ TEST_F(LoDTensorTester, ShrinkLevels) {
     LoDTensor new_lod_tensor = lod_tensor_;
     new_lod_tensor.ShrinkLevels(level, level + 1);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
     ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
   // shrink 2 level
   for (size_t level = 0; level < 2UL; ++level) {
     LoDTensor new_lod_tensor = lod_tensor_;
     new_lod_tensor.ShrinkLevels(level, level + 2);
+    // the lowest level's last element should be the tensor's batch_size.
+    ASSERT_EQ(new_lod_tensor.lod().back().back(),
+              lod_tensor_.lod().back().back());
     ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
-    ASSERT_EQ(new_lod_tensor.NumElements(1),
-              lod_tensor_.NumElements(level + 1));
     ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
 }
@@ -86,19 +85,19 @@ TEST_F(LoDTensorTester, ShrinkLevels) {
 TEST_F(LoDTensorTester, ShrinkInLevel) {
   size_t level = 0;
   LoDTensor new_lod_tensor = lod_tensor_;
-  new_lod_tensor.ShrinkInLevel(level, 0, 2);
+  new_lod_tensor.ShrinkInLevel(level, 0, 1);
   EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
-  EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
-  EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
+  EXPECT_EQ(new_lod_tensor.NumElements(0), 1UL);
+  EXPECT_EQ(new_lod_tensor.NumElements(1), 2UL);
+  EXPECT_EQ(new_lod_tensor.NumElements(2), 5UL);
   ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 
   level = 1;
   new_lod_tensor = lod_tensor_;
-  new_lod_tensor.ShrinkInLevel(level, 0, 2);
+  new_lod_tensor.ShrinkInLevel(level, 1, 2);
   ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
+  ASSERT_EQ(new_lod_tensor.NumElements(0), 1UL);
+  ASSERT_EQ(new_lod_tensor.NumElements(1), 3UL);
   ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 }
 

paddle/framework/op_proto_maker.h

@@ -44,6 +44,11 @@ class OpProtoAndCheckerMaker {
       var_->set_intermediate(true);
       return *this;
     }
+
+    VariableBuilder& AsDispensable() {
+      var_->set_dispensable(true);
+      return *this;
+    }
   };
 
   VariableBuilder AddInput(const std::string& name, const std::string& comment);

paddle/framework/operator.cc

@@ -252,5 +252,20 @@ std::ostream& operator<<(std::ostream& os,
   return os;
 }
 
+bool OpSupportGPU(const std::string& op_type) {
+  auto& all_kernels = OperatorWithKernel::AllOpKernels();
+  auto it = all_kernels.find(op_type);
+  if (it == all_kernels.end()) {
+    // All control operator must support GPU
+    return true;
+  }
+  for (auto& kern_pair : it->second) {
+    if (platform::is_gpu_place(kern_pair.first.place_)) {
+      return true;
+    }
+  }
+  return false;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/operator.h

@@ -327,37 +327,47 @@ class CompileTimeInferShapeContext : public InferShapeContext {
   bool HasInput(const std::string& name) const override {
     const std::vector<std::string>& input_names = op_.Input(name);
     auto length = input_names.size();
+    if (length == 0) {
+      return false;
+    }
     PADDLE_ENFORCE_EQ(length, 1UL,
                       "Input(%s) should have only one value, "
                       "but it have %d now",
                       name, length);
-    return block_.HasVar(input_names[0]);
+    return block_.HasVarRecursive(input_names[0]);
   }
 
   bool HasOutput(const std::string& name) const override {
     const std::vector<std::string>& output_names = op_.Output(name);
     auto length = output_names.size();
+    if (length == 0) {
+      return false;
+    }
     PADDLE_ENFORCE_EQ(length, 1UL,
                       "Output(%s) should have only one value, "
                       "but it have %d now",
                       name, length);
-    return block_.HasVar(output_names[0]);
+    return block_.HasVarRecursive(output_names[0]);
   }
 
   bool HasInputs(const std::string& name) const override {
     const std::vector<std::string>& input_names = op_.Input(name);
-    PADDLE_ENFORCE(!input_names.empty(), "Inputs(%s) length is 0", name);
+    if (input_names.empty()) {
+      return false;
+    }
     for (auto& input : input_names) {
-      if (!block_.HasVar(input)) return false;
+      if (!block_.HasVarRecursive(input)) return false;
     }
     return true;
   }
 
   bool HasOutputs(const std::string& name) const override {
     const std::vector<std::string>& output_names = op_.Output(name);
-    PADDLE_ENFORCE(!output_names.empty(), "Inputs(%s) length is 0", name);
+    if (output_names.empty()) {
+      return false;
+    }
     for (auto& output : output_names) {
-      if (!block_.HasVar(output)) return false;
+      if (!block_.HasVarRecursive(output)) return false;
     }
     return true;
   }
@@ -404,11 +414,11 @@ class CompileTimeInferShapeContext : public InferShapeContext {
 
  private:
   DDim GetDim(const std::string& name) const override {
-    return framework::make_ddim(block_.FindVar(name)->Shape());
+    return framework::make_ddim(block_.FindVarRecursive(name)->Shape());
   }
 
   void SetDim(const std::string& name, const DDim& dim) override {
-    block_.FindVar(name)->SetShape(framework::vectorize(dim));
+    block_.FindVarRecursive(name)->SetShape(framework::vectorize(dim));
   }
 
   const OpDescBind& op_;
@@ -421,13 +431,27 @@ class RuntimeInferShapeContext : public InferShapeContext {
       : op_(op), scope_(scope) {}
 
   bool HasInput(const std::string& name) const override {
-    auto ipt = op_.Input(name);
+    auto& ins = Inputs(name);
+    size_t length = ins.size();
+    if (length == 0) {
+      return false;
+    }
+    PADDLE_ENFORCE_EQ(length, 1UL, "Input %s should have more than one inputs",
+                      name);
+    auto ipt = ins[0];
     auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
     return var != nullptr;
   }
 
   bool HasOutput(const std::string& name) const override {
-    auto ipt = op_.Output(name);
+    auto& outs = Outputs(name);
+    size_t length = outs.size();
+    if (length == 0) {
+      return false;
+    }
+    PADDLE_ENFORCE_EQ(length, 1UL, "Output %s should have more than one inputs",
+                      name);
+    auto ipt = outs[0];
     auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
     return var != nullptr;
   }
@@ -649,5 +673,7 @@ class OperatorWithKernel : public OperatorBase {
 std::ostream& operator<<(std::ostream& os,
                          const OperatorWithKernel::OpKernelKey& kernel_key);
 
+extern bool OpSupportGPU(const std::string& op_type);
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/program_desc.cc

@@ -35,8 +35,8 @@ ProgramDesc *ProgramDescBind::Proto() {
 
 ProgramDescBind::ProgramDescBind() {
   auto *block = prog_.mutable_blocks()->Add();
-  block->set_idx(0);
-  block->set_parent_idx(-1);
+  block->set_idx(kRootBlockIndex);
+  block->set_parent_idx(kNoneBlockIndex);
   blocks_.emplace_back(new BlockDescBind(this, block));
 }
 

paddle/framework/program_desc.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <memory>
 #include <vector>
 #include "paddle/framework/framework.pb.h"
+#include "paddle/framework/proto_desc.h"
 #include "paddle/platform/macros.h"
 
 namespace paddle {

paddle/framework/program_desc_test.cc

@@ -80,4 +80,4 @@ TEST(ProgramDesc, copy_ctor) {
   // different and it is correct.
 }
 }  // namespace framework
-}  // namespace paddle
+}  // namespace paddle
\ No newline at end of file

paddle/framework/proto_desc.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 {
+
+// The Index of first Block in Program. also called root block.
+constexpr int kRootBlockIndex = 0;
+// The Parent Index of root Block, this block does not exist.
+constexpr int kNoneBlockIndex = -1;
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/scope.cc

@@ -65,12 +65,11 @@ void Scope::DropKids() {
   kids_.clear();
 }
 
-framework::Scope& GetGlobalScope() {
-  static framework::Scope* g_scope = nullptr;
-  if (g_scope == nullptr) {
-    g_scope = new framework::Scope();
-  }
-  return *g_scope;
+void Scope::DeleteScope(Scope* scope) {
+  auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
+  PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
+  this->kids_.erase(it);
+  delete scope;
 }
 
 }  // namespace framework

paddle/framework/scope.h

@@ -59,6 +59,8 @@ class Scope {
   /// Find the scope or an ancestor scope that contains the given variable.
   const Scope* FindScope(const Variable* var) const;
 
+  void DeleteScope(Scope* scope);
+
   /// Drop all kids scopes belonged to this scope.
   void DropKids();
 
@@ -72,8 +74,5 @@ class Scope {
 
   DISABLE_COPY_AND_ASSIGN(Scope);
 };
-
-framework::Scope& GetGlobalScope();
-
 }  // namespace framework
 }  // namespace paddle

paddle/framework/tensor.h

@@ -60,6 +60,10 @@ class Tensor {
   template <typename T>
   inline T* mutable_data(platform::Place place);
 
+  inline void* mutable_data(platform::Place place, std::type_index type);
+
+  inline void* mutable_data(platform::Place place);
+
   /**
    * @brief     Return a pointer to mutable memory block.
    *
@@ -81,7 +85,6 @@ class Tensor {
   inline Tensor& Resize(const DDim& dims);
 
   /*! The internal of two tensors share the same memory block. */
-  template <typename T>
   inline Tensor& ShareDataWith(const Tensor& src);
 
   /**
@@ -96,26 +99,9 @@ class Tensor {
   // TODO(qijun): https://github.com/PaddlePaddle/Paddle/issues/4647
   // Remove `CopyFrom` and `CopyFromVector` from Tensor interface
   // and make them global functions
-  template <typename T>
   inline void CopyFrom(const Tensor& src, const platform::Place& dst_place,
                        const platform::DeviceContext& ctx);
 
-  // FIXME(yuyang18): CopyFrom should without template T, use the replace
-  // `CopyFrom` with `CopyFromTensor`
-  inline void CopyFromTensor(const Tensor& src,
-                             const platform::Place& dst_place,
-                             const platform::DeviceContext& ctx) {
-    // NOLINTNEXTLINES_8 cpplint.py will recognize below lines as functions.
-    // That is a bug of cpplint.py. Just ignore lint these lines.
-    if (src.type() == std::type_index(typeid(double))) {
-      CopyFrom<double>(src, dst_place, ctx);
-    } else if (src.type() == std::type_index(typeid(float))) {
-      CopyFrom<float>(src, dst_place, ctx);
-    } else if (src.type() == std::type_index(typeid(int))) {
-      CopyFrom<int>(src, dst_place, ctx);
-    }
-  }
-
   /**
    * @brief   Copy the content of an external vector to a tensor.
    *
@@ -135,7 +121,6 @@ class Tensor {
    * @param[in] begin_idx   The begin index of the slice.
    * @param[in] end_idx     The end index of the slice.
    */
-  template <typename T>
   inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
 
   platform::Place place() const {
@@ -146,7 +131,6 @@ class Tensor {
   std::type_index type() const { return holder_->type(); }
 
  private:
-  template <typename T>
   inline void check_memory_size() const;
 
  private:
@@ -155,20 +139,22 @@ class Tensor {
    *          parameter of Variable.
    */
   struct Placeholder {
-    virtual ~Placeholder() {}
+    virtual ~Placeholder() = default;
     virtual void* ptr() const = 0;
     virtual size_t size() const = 0;
     virtual std::type_index type() const = 0;
     virtual platform::Place place() const = 0;
+    virtual void set_type(std::type_index type) = 0;
   };
 
-  template <typename T, typename Place>
+  template <typename Place>
   struct PlaceholderImpl : public Placeholder {
-    PlaceholderImpl(Place place, size_t size)
-        : ptr_(static_cast<T*>(memory::Alloc(place, size)),
-               memory::PODDeleter<T, Place>(place)),
+    PlaceholderImpl(Place place, size_t size, std::type_index type)
+        : ptr_(static_cast<uint8_t*>(memory::Alloc(place, size)),
+               memory::PODDeleter<uint8_t, Place>(place)),
           place_(place),
-          size_(size) {
+          size_(size),
+          type_(type) {
       PADDLE_ENFORCE_NOT_NULL(ptr_, "Insufficient %s memory to allocation.",
                               (is_cpu_place(place_) ? "CPU" : "GPU"));
     }
@@ -176,16 +162,20 @@ class Tensor {
     virtual size_t size() const { return size_; }
     virtual platform::Place place() const { return place_; }
     virtual void* ptr() const { return static_cast<void*>(ptr_.get()); }
-    virtual std::type_index type() const { return std::type_index(typeid(T)); }
+    virtual std::type_index type() const { return type_; }
+    virtual void set_type(std::type_index type) { type_ = type; }
 
     /*! the pointer of memory block. */
-    std::unique_ptr<T, memory::PODDeleter<T, Place>> ptr_;
+    std::unique_ptr<uint8_t, memory::PODDeleter<uint8_t, Place>> ptr_;
 
     /*! the place of memory block. */
     platform::Place place_;
 
     /*! the size of memory block. */
     size_t size_;
+
+    /* the current type of memory */
+    std::type_index type_;
   };
 
   /*! holds the memory block if allocated. */

paddle/framework/tensor_array.cc

@@ -106,8 +106,8 @@ void TensorArray::Write(size_t index, const LoDTensor& value) {
 
   values_[index].Resize(value.dims());
   values_[index].mutable_data<value_type>(platform::CPUPlace());
-  values_[index].CopyFrom<value_type>(value, platform::CPUPlace(),
-                                      platform::CPUDeviceContext());
+  values_[index].CopyFrom(value, platform::CPUPlace(),
+                          platform::CPUDeviceContext());
 }
 
 void TensorArray::WriteShared(size_t index, const LoDTensor& value) {
@@ -116,7 +116,7 @@ void TensorArray::WriteShared(size_t index, const LoDTensor& value) {
     values_.resize(index + 1);
   }
 
-  values_[index].ShareDataWith<value_type>(value);
+  values_[index].ShareDataWith(value);
 }
 
 LoDTensor TensorArray::Pack(size_t level, const std::vector<DySeqMeta>& meta,
@@ -163,9 +163,9 @@ LoDTensor TensorArray::Stack() const {
   result.mutable_data<value_type>(platform::CPUPlace());
 
   for (size_t idx = 0; idx < size(); idx++) {
-    result.Slice<value_type>(idx, idx + 1)
-        .CopyFrom<value_type>(Read(idx), platform::CPUPlace(),
-                              platform::CPUDeviceContext());
+    result.Slice(idx, idx + 1)
+        .CopyFrom(Read(idx), platform::CPUPlace(),
+                  platform::CPUDeviceContext());
   }
   return result;
 }
@@ -191,13 +191,12 @@ void TensorArray::Unstack(const LoDTensor& source, bool data_shared) const {
     auto& value = values_[elem];
     if (data_shared) {
       // share memory
-      value.ShareDataWith<value_type>(source.Slice<value_type>(elem, elem + 1));
+      value.ShareDataWith(source.Slice(elem, elem + 1));
     } else {
       // copy
       value.Resize(value_dims);
-      value.CopyFrom<value_type>(source.Slice<value_type>(elem, elem + 1),
-                                 platform::CPUPlace(),
-                                 platform::CPUDeviceContext());
+      value.CopyFrom(source.Slice(elem, elem + 1), platform::CPUPlace(),
+                     platform::CPUDeviceContext());
     }
   }
 }
@@ -242,11 +241,10 @@ LoDTensor DynamicBatchUnpacker::GetBatch(size_t index) {
 
   for (size_t i = 0; i < indice.size(); i++) {
     auto index = indice[i];
-    auto target = result.Slice<value_type>(i, i + 1);
-    auto slice = source->Slice<value_type>(index, index + 1);
+    auto target = result.Slice(i, i + 1);
+    auto slice = source->Slice(index, index + 1);
 
-    target.CopyFrom<value_type>(slice, platform::CPUPlace(),
-                                platform::CPUDeviceContext());
+    target.CopyFrom(slice, platform::CPUPlace(), platform::CPUDeviceContext());
   }
 
   return result;
@@ -277,10 +275,10 @@ LoDTensor PackDynamicBatch(const std::vector<LoDTensor>& source,
       // target is result[index]
       auto index = seq_meta.begin + batch_id;
       if (index >= seq_meta.end) break;
-      auto source_ = source[batch_id].Slice<float>(seq_id, seq_id + 1);
-      auto target = result.Slice<float>(index, index + 1);
-      target.CopyFrom<float>(source_, platform::CPUPlace(),
-                             platform::CPUDeviceContext());
+      auto source_ = source[batch_id].Slice(seq_id, seq_id + 1);
+      auto target = result.Slice(index, index + 1);
+      target.CopyFrom(source_, platform::CPUPlace(),
+                      platform::CPUDeviceContext());
     }
   }
 

paddle/framework/tensor_array_test.cc

@@ -91,7 +91,7 @@ class TensorArrayPackTester : public ::testing::Test {
       size_t begin = level[i];
       size_t end = level[i + 1];
       for (size_t j = begin; j < end; j++) {
-        auto record = source.Slice<int>(j, j + 1);
+        auto record = source.Slice(j, j + 1);
         for (int dim = 0; dim < 128; dim++) {
           record.mutable_data<int>(platform::CPUPlace())[dim] = j - begin;
         }

paddle/framework/tensor_impl.h

@@ -19,12 +19,50 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+template <typename... T>
+struct SizeOfTypeFunctor;
+
 template <typename T>
+struct SizeOfTypeFunctor<T> {
+  size_t operator()(std::type_index type) const {
+    if (typeid(T).hash_code() == type.hash_code()) {
+      return sizeof(T);
+    } else {
+      return 0UL;
+    }
+  }
+};
+
+template <>
+struct SizeOfTypeFunctor<> {
+  size_t operator()(std::type_index type) const { return 0UL; }
+};
+
+template <typename HEAD, typename... TAIL>
+struct SizeOfTypeFunctor<HEAD, TAIL...> {
+  size_t operator()(std::type_index type) const {
+    SizeOfTypeFunctor<HEAD> head;
+    size_t head_size = head(type);
+    if (head_size != 0) {
+      return head_size;
+    }
+    SizeOfTypeFunctor<TAIL...> tail;
+    return tail(type);
+  }
+};
+
+static inline size_t SizeOfType(std::type_index type) {
+  SizeOfTypeFunctor<int, float, double, int16_t, int64_t> functor;
+  size_t size = functor(type);
+  PADDLE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name());
+  return size;
+}
+
 inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
       holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
   PADDLE_ENFORCE_GE(
-      holder_->size(), numel() * sizeof(T) + offset_,
+      holder_->size(), numel() * SizeOfType(type()) + offset_,
       "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
       "first to re-allocate memory.\n"
       "or maybe the required data-type mismatches the data already stored.");
@@ -32,14 +70,23 @@ inline void Tensor::check_memory_size() const {
 
 template <typename T>
 inline const T* Tensor::data() const {
-  check_memory_size<T>();
+  check_memory_size();
+  PADDLE_ENFORCE(std::is_same<T, void>::value ||
+                     holder_->type().hash_code() == typeid(T).hash_code(),
+                 "Tensor holds the wrong type, it holds %s",
+                 this->holder_->type().name());
+
   return reinterpret_cast<const T*>(
       reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
 }
 
 template <typename T>
 inline T* Tensor::data() {
-  check_memory_size<T>();
+  check_memory_size();
+  PADDLE_ENFORCE(std::is_same<T, void>::value ||
+                     holder_->type().hash_code() == typeid(T).hash_code(),
+                 "Tensor holds the wrong type, it holds %s",
+                 this->holder_->type().name());
   return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
                               offset_);
 }
@@ -54,51 +101,62 @@ inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
 template <typename T>
 inline T* Tensor::mutable_data(platform::Place place) {
   static_assert(std::is_pod<T>::value, "T must be POD");
+  return reinterpret_cast<T*>(mutable_data(place, typeid(T)));
+}
+
+inline void* Tensor::mutable_data(platform::Place place, std::type_index type) {
+  if (holder_ != nullptr) {
+    holder_->set_type(type);
+  }
   PADDLE_ENFORCE_GT(numel(), 0,
                     "Tensor's numel must be larger than zero to call "
                     "Tensor::mutable_data. Call Tensor::set_dim first.");
+  int64_t size = numel() * SizeOfType(type);
   /* some versions of boost::variant don't have operator!= */
-  int64_t size = numel() * sizeof(T);
   if (holder_ == nullptr || !(holder_->place() == place) ||
       holder_->size() < size + offset_) {
     if (platform::is_cpu_place(place)) {
-      holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
-          boost::get<platform::CPUPlace>(place), size));
+      holder_.reset(new PlaceholderImpl<platform::CPUPlace>(
+          boost::get<platform::CPUPlace>(place), size, type));
     } else if (platform::is_gpu_place(place)) {
 #ifndef PADDLE_WITH_CUDA
       PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
     }
 #else
-      holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
-          boost::get<platform::GPUPlace>(place), size));
+      holder_.reset(new PlaceholderImpl<platform::GPUPlace>(
+          boost::get<platform::GPUPlace>(place), size, type));
     }
 #endif
     offset_ = 0;
   }
-  return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
-                              offset_);
+  return reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
+                                 offset_);
+}
+
+inline void* Tensor::mutable_data(platform::Place place) {
+  PADDLE_ENFORCE(this->holder_ != nullptr,
+                 "Cannot invoke mutable data if current hold nothing");
+  return mutable_data(place, holder_->type());
 }
 
-template <typename T>
 inline Tensor& Tensor::ShareDataWith(const Tensor& src) {
-  src.check_memory_size<T>();
+  src.check_memory_size();
   *this = src;
   return *this;
 }
 
-template <typename T>
 inline void Tensor::CopyFrom(const Tensor& src,
                              const platform::Place& dst_place,
                              const platform::DeviceContext& ctx) {
-  src.check_memory_size<T>();
+  src.check_memory_size();
   Resize(src.dims());
 
   auto src_place = src.holder_->place();
-  auto src_ptr = static_cast<const void*>(src.data<T>());
+  auto src_ptr = src.data<void>();
 
-  auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
+  auto dst_ptr = mutable_data(dst_place, src.type());
 
-  auto size = src.numel() * sizeof(T);
+  auto size = src.numel() * SizeOfType(src.type());
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
@@ -165,9 +223,8 @@ inline void Tensor::CopyFromVector(const std::vector<T>& src,
 #endif
 }
 
-template <typename T>
 inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
-  check_memory_size<T>();
+  check_memory_size();
   PADDLE_ENFORCE_GE(begin_idx, 0, "Slice begin index is less than zero.");
   PADDLE_ENFORCE_LE(end_idx, dims_[0], "Slice end index is out of bound.");
   PADDLE_ENFORCE_LT(begin_idx, end_idx,
@@ -182,7 +239,7 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
     DDim dst_dims = dims_;
     dst_dims[0] = end_idx - begin_idx;
     dst.Resize(dst_dims);
-    dst.offset_ = offset_ + begin_idx * base * sizeof(T);
+    dst.offset_ = offset_ + begin_idx * base * SizeOfType(type());
     return dst;
   }
 }
@@ -196,10 +253,9 @@ inline const DDim& Tensor::dims() const { return dims_; }
 
 inline int64_t Tensor::numel() const { return product(dims_); }
 
-template <typename T>
 inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
   Tensor res;
-  res.ShareDataWith<T>(src);
+  res.ShareDataWith(src);
   res.Resize(flatten_to_2d(src.dims(), num_col_dims));
   return res;
 }

paddle/framework/tensor_test.cc

@@ -108,7 +108,7 @@ TEST(Tensor, ShareDataWith) {
     // Try to share data form uninitialized tensor
     bool caught = false;
     try {
-      dst_tensor.ShareDataWith<float>(src_tensor);
+      dst_tensor.ShareDataWith(src_tensor);
     } catch (paddle::platform::EnforceNotMet err) {
       caught = true;
       std::string msg =
@@ -122,7 +122,7 @@ TEST(Tensor, ShareDataWith) {
     ASSERT_TRUE(caught);
 
     src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), CPUPlace());
-    dst_tensor.ShareDataWith<int>(src_tensor);
+    dst_tensor.ShareDataWith(src_tensor);
     ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
   }
 
@@ -131,7 +131,7 @@ TEST(Tensor, ShareDataWith) {
     Tensor src_tensor;
     Tensor dst_tensor;
     src_tensor.mutable_data<int>(make_ddim({2, 3, 4}), GPUPlace());
-    dst_tensor.ShareDataWith<int>(src_tensor);
+    dst_tensor.ShareDataWith(src_tensor);
     ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
   }
 #endif
@@ -143,7 +143,7 @@ TEST(Tensor, Slice) {
   {
     Tensor src_tensor;
     src_tensor.mutable_data<int>(make_ddim({5, 3, 4}), CPUPlace());
-    Tensor slice_tensor = src_tensor.Slice<int>(1, 3);
+    Tensor slice_tensor = src_tensor.Slice(1, 3);
     DDim slice_dims = slice_tensor.dims();
     ASSERT_EQ(arity(slice_dims), 3);
     EXPECT_EQ(slice_dims[0], 2);
@@ -167,7 +167,7 @@ TEST(Tensor, Slice) {
   {
     Tensor src_tensor;
     src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
-    Tensor slice_tensor = src_tensor.Slice<double>(2, 6);
+    Tensor slice_tensor = src_tensor.Slice(2, 6);
     DDim slice_dims = slice_tensor.dims();
     ASSERT_EQ(arity(slice_dims), 2);
     EXPECT_EQ(slice_dims[0], 4);
@@ -202,7 +202,7 @@ TEST(Tensor, CopyFrom) {
     memcpy(src_ptr, arr, 9 * sizeof(int));
 
     auto cpu_place = new paddle::platform::CPUPlace();
-    dst_tensor.CopyFrom<int>(src_tensor, *cpu_place, cpu_ctx);
+    dst_tensor.CopyFrom(src_tensor, *cpu_place, cpu_ctx);
 
     const int* dst_ptr = dst_tensor.data<int>();
     ASSERT_NE(src_ptr, dst_ptr);
@@ -210,8 +210,8 @@ TEST(Tensor, CopyFrom) {
       EXPECT_EQ(src_ptr[i], dst_ptr[i]);
     }
 
-    Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
-    dst_tensor.CopyFrom<int>(slice_tensor, *cpu_place, cpu_ctx);
+    Tensor slice_tensor = src_tensor.Slice(1, 2);
+    dst_tensor.CopyFrom(slice_tensor, *cpu_place, cpu_ctx);
     const int* slice_ptr = slice_tensor.data<int>();
     dst_ptr = dst_tensor.data<int>();
     ASSERT_NE(dst_ptr, slice_ptr);
@@ -233,11 +233,11 @@ TEST(Tensor, CopyFrom) {
     // CPU Tensor to GPU Tensor
     auto gpu_place = new paddle::platform::GPUPlace(0);
     CUDADeviceContext gpu_ctx(*gpu_place);
-    gpu_tensor.CopyFrom<int>(src_tensor, *gpu_place, gpu_ctx);
+    gpu_tensor.CopyFrom(src_tensor, *gpu_place, gpu_ctx);
 
     // GPU Tensor to CPU Tensor
     auto cpu_place = new paddle::platform::CPUPlace();
-    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place, gpu_ctx);
+    dst_tensor.CopyFrom(gpu_tensor, *cpu_place, gpu_ctx);
 
     // Sync before Compare Tensors
     gpu_ctx.Wait();
@@ -247,13 +247,13 @@ TEST(Tensor, CopyFrom) {
       EXPECT_EQ(src_ptr[i], dst_ptr[i]);
     }
 
-    Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
+    Tensor slice_tensor = src_tensor.Slice(1, 2);
 
     // CPU Slice Tensor to GPU Tensor
-    gpu_tensor.CopyFrom<int>(slice_tensor, *gpu_place, gpu_ctx);
+    gpu_tensor.CopyFrom(slice_tensor, *gpu_place, gpu_ctx);
 
     // GPU Tensor to CPU Tensor
-    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place, gpu_ctx);
+    dst_tensor.CopyFrom(gpu_tensor, *cpu_place, gpu_ctx);
 
     // Sync before Compare Slice Tensors
     gpu_ctx.Wait();
@@ -320,7 +320,7 @@ TEST(Tensor, CopyFromVector) {
     CUDADeviceContext gpu_ctx(*gpu_place);
     gpu_tensor.CopyFromVector<int>(src_vec, gpu_ctx);
     // Copy from GPU to CPU tensor for comparison
-    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place, gpu_ctx);
+    dst_tensor.CopyFrom(gpu_tensor, *cpu_place, gpu_ctx);
 
     // Sync before Compare Tensors
     gpu_ctx.Wait();
@@ -340,7 +340,7 @@ TEST(Tensor, CopyFromVector) {
     cpu_tensor.CopyFromVector<int>(src_vec, cpu_ctx);
     gpu_tensor.Resize(make_ddim({2, 2}));
     gpu_tensor.CopyFromVector<int>(src_vec, gpu_ctx);
-    dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_place, gpu_ctx);
+    dst_tensor.CopyFrom(gpu_tensor, *cpu_place, gpu_ctx);
 
     // Sync before Compare Tensors
     gpu_ctx.Wait();
@@ -368,7 +368,7 @@ TEST(Tensor, ReshapeToMatrix) {
   for (int i = 0; i < 2 * 3 * 4 * 9; ++i) {
     src_ptr[i] = i;
   }
-  Tensor res = ReshapeToMatrix<int>(src, 2);
+  Tensor res = ReshapeToMatrix(src, 2);
   ASSERT_EQ(res.dims()[0], 2 * 3);
   ASSERT_EQ(res.dims()[1], 4 * 9);
 }

paddle/framework/variable.h

@@ -25,7 +25,10 @@ class Variable {
  public:
   template <typename T>
   const T& Get() const {
-    PADDLE_ENFORCE(IsType<T>(), "Variable must be type %s", typeid(T).name());
+    PADDLE_ENFORCE(holder_ != nullptr, "Variable must hold some thing");
+    PADDLE_ENFORCE(IsType<T>(),
+                   "Variable must be type %s, the holding type is %s",
+                   typeid(T).name(), holder_->Type().name());
     return *static_cast<const T*>(holder_->Ptr());
   }
 

paddle/gserver/activations/MKLDNNActivation.cpp

@@ -126,7 +126,7 @@ void MKLDNNEltwiseActivation::resetFwd(Argument& act) {
   copyInVal_ = nullptr;
   if (act.grad && algo == algorithm::eltwise_tanh) {
     // tanh need save src input for backward
-    inVal_ = MKLDNNMatrix::create(nullptr, val_->getPrimitiveDesc());
+    inVal_ = MKLDNNMatrix::create(val_->getPrimitiveDesc());
     copyInVal_ = std::make_shared<mkldnn::reorder>(*val_, *inVal_);
     CHECK(copyInVal_) << "should not be emptry";
     pipelineFwd_.push_back(*copyInVal_);
@@ -145,7 +145,7 @@ void MKLDNNEltwiseActivation::resetBwd(Argument& act) {
   algorithm algo = getAlgo(this->getName());
   float alpha = getBwdAlpha();
   float beta = getBeta();
-  grad_ = MKLDNNMatrix::create(act.grad, val_->getPrimitiveDesc());
+  grad_ = MKLDNNMatrix::create(val_->getPrimitiveDesc(), act.grad);
   auto eng = CPUEngine::Instance().getEngine();
   auto bwdDesc = eltwise_bwd::desc(
       algo, grad_->getMemoryDesc(), val_->getMemoryDesc(), alpha, beta);
@@ -230,7 +230,7 @@ void MKLDNNActivation::resetFwd(Argument& act) {
     int ic = cnt_ / bs / ih / iw;
     CHECK_EQ(cnt_, (size_t)bs * ic * ih * iw);
     val_ = MKLDNNMatrix::create(
-        act.value, {bs, ic, ih, iw}, mkldnn::memory::format::nchw, *engine_);
+        {bs, ic, ih, iw}, mkldnn::memory::format::nchw, *engine_, act.value);
     CHECK(val_);
     val_->downSpatial();
   }

paddle/gserver/layers/MKLDNNBase.h

@@ -21,8 +21,8 @@ namespace paddle {
 typedef enum {
   MKLDNN_BASE = 1,   // basical info of MKLDNN
   MKLDNN_TESTS = 1,  // gtest info of MKLDNN
-  MKLDNN_SIZES = 2,  // size info of MKLDNN
-  MKLDNN_FMTS = 3,   // format info of MKLDNN
+  MKLDNN_FMTS = 2,   // format info of MKLDNN
+  MKLDNN_SIZES = 3,  // size info of MKLDNN
   MKLDNN_ALL = 4,    // show all info of MKLDNN
 } MKLDNN_LOG_LEVEL;
 

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -116,8 +116,6 @@ void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,
   resetFwdBuffers(fwdPD_, in, wgt, bias, out);
 
   resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
-
-  printValueFormatFlow();
 }
 
 void MKLDNNConvLayer::resetBwd(std::vector<primitive>& pipeline,
@@ -135,12 +133,6 @@ void MKLDNNConvLayer::resetBwd(std::vector<primitive>& pipeline,
   resetBwdBuffers(bwdWgtPD, bwdDataPD, in, wgt, bias, out);
 
   resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
-
-  printGradFormatFlow();
-}
-
-void MKLDNNConvLayer::updateInputData() {
-  cpuInVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
 }
 
 void MKLDNNConvLayer::updateWeights(const UpdateCallback& callback) {
@@ -211,11 +203,18 @@ void MKLDNNConvLayer::resetFwdBuffers(
     MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
   CHECK(pd);
-  resetInValue(pd, in);
+  resetInValue(
+      in, std::make_shared<memory::primitive_desc>(pd->src_primitive_desc()));
+
+  resetOutValue(out, pd->dst_primitive_desc());
 
-  resetWgtBiasValue(pd, wgt, bias);
+  resetWithMatrix(wgt, weight_->getW(), pd->weights_primitive_desc());
 
-  resetOutValue(pd, out);
+  if (biases_ && biases_->getW()) {
+    resetWithMatrix(bias, biases_->getW(), pd->bias_primitive_desc());
+  } else {
+    bias = nullptr;
+  }
 }
 
 void MKLDNNConvLayer::resetFwdPipeline(
@@ -225,104 +224,12 @@ void MKLDNNConvLayer::resetFwdPipeline(
     MKLDNNMatrixPtr& wgt,
     MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
-  if (cvtInVal_) {
-    pipeline.push_back(*cvtInVal_);
-  }
-
   if (bias) {
     fwd_.reset(new conv_fwd(*pd, *in, *wgt, *bias, *out));
   } else {
     fwd_.reset(new conv_fwd(*pd, *in, *wgt, *out));
   }
   pipeline.push_back(*fwd_);
-
-  if (cvtOutVal_) {
-    pipeline.push_back(*cvtOutVal_);
-  }
-}
-
-void MKLDNNConvLayer::resetInValue(
-    std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& in) {
-  const MatrixPtr& inMat = inputLayers_[0]->getOutputValue();
-  in = MKLDNNMatrix::create(inMat, pd->src_primitive_desc());
-
-  // create buffer and reorder if input value do not match
-  cpuInVal_ = nullptr;
-  cvtInVal_ = nullptr;
-
-  MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
-  CHECK_EQ(inputIsOnlyMKLDNN(), dnnIn != nullptr);
-  if (dnnIn != nullptr && dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc()) {
-    in = dnnIn;
-    return;
-  }
-  if (dnnIn) {
-    if (dnnIn->getFormat() == format::nc) {
-      CHECK(ih_ == 1 && iw_ == 1) << "when input is nc format";
-      // create a new one with nchw format and same data
-      memory::dims inDims = memory::dims{bs_, ic_, 1, 1};
-      dnnIn = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
-    }
-    if (dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc()) {
-      in = dnnIn;
-      return;
-    }
-    cpuInVal_ = dnnIn;
-    in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());
-    cvtInVal_ = MKLDNNMatrix::createReorder(cpuInVal_, in);
-    CHECK(cvtInVal_) << "should not be emptry";
-  } else {
-    memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
-    cpuInVal_ = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
-    if (cpuInVal_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
-      // create new mkldnn matrix
-      in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());
-      cvtInVal_ = MKLDNNMatrix::createReorder(cpuInVal_, in);
-      CHECK(cvtInVal_) << "should not be emptry";
-    } else {
-      in = cpuInVal_;
-    }
-  }
-}
-
-void MKLDNNConvLayer::resetWgtBiasValue(
-    std::shared_ptr<conv_fwd::primitive_desc>& pd,
-    MKLDNNMatrixPtr& wgt,
-    MKLDNNMatrixPtr& bias) {
-  wgt = MKLDNNMatrix::create(weight_->getW(), pd->weights_primitive_desc());
-  VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
-
-  bias = (biases_ && biases_->getW())
-             ? MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc())
-             : nullptr;
-}
-
-void MKLDNNConvLayer::resetOutValue(
-    std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& out) {
-  out = MKLDNNMatrix::create(output_.value, pd->dst_primitive_desc());
-
-  // create reorder if output value has cpu device and pd do not match
-  cpuOutVal_ = nullptr;
-  cvtOutVal_ = nullptr;
-  if (!outputIsOnlyMKLDNN()) {
-    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
-    memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
-    cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
-    if (cpuOutVal_->getPrimitiveDesc() != pd->dst_primitive_desc()) {
-      out = MKLDNNMatrix::create(nullptr, pd->dst_primitive_desc());
-      cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
-      CHECK(cvtOutVal_) << "should not be empty";
-    } else {
-      cpuOut->setData(output_.value->getData());
-      cpuOutVal_ = out;
-    }
-    // when output is cpu device, change the mkldnn output value and make them
-    // share the same data. Then if next layer use inputlayer->getOuputValue()
-    // to achieve the input value, it will get the right data.
-    output_.value = std::dynamic_pointer_cast<Matrix>(cpuOutVal_);
-    return;
-  }
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
 }
 
 void MKLDNNConvLayer::resetBwdWgtPD(
@@ -331,8 +238,8 @@ void MKLDNNConvLayer::resetBwdWgtPD(
   loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
 
   // create backward weight using input, output and weight value memory desc
-  CHECK(inVal_) << "Should have input value";
-  CHECK(outVal_) << "Should have output value";
+  CHECK(inVal_) << "Should have internal input value";
+  CHECK(outVal_) << "Should have internal output value";
   CHECK(wgtVal_) << "Should have weight value";
   algorithm algo = algorithm::convolution_direct;
   padding_kind padKind = padding_kind::zero;
@@ -372,8 +279,8 @@ void MKLDNNConvLayer::resetBwdDataPD(
 
   memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
   loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
-  CHECK(inVal_) << "Should have input value";
-  CHECK(outVal_) << "Should have output value";
+  CHECK(inVal_) << "Should have internal input value";
+  CHECK(outVal_) << "Should have internal output value";
   // create backward data using input and output value memory desc
   // but using weight memory desc with any format
   auto bwdDataDesc = conv_bwdData::desc(algorithm::convolution_direct,
@@ -399,12 +306,27 @@ void MKLDNNConvLayer::resetBwdBuffers(
     MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
   CHECK(wgtPD);
-  resetOutGrad(wgtPD, out);
+  resetOutGrad(out, wgtPD->diff_dst_primitive_desc());
 
-  resetWgtBiasGrad(wgtPD, wgt, bias);
+  resetWithMatrix(
+      wgt, weight_->getWGrad(), wgtPD->diff_weights_primitive_desc());
+  CHECK(wgtVal_ != nullptr &&
+        wgt->getPrimitiveDesc() == wgtVal_->getPrimitiveDesc())
+      << "primitive desc of weight grad and value should be equal";
 
-  resetInGrad(dataPD, in);
+  bias = nullptr;
+  if (biases_ && biases_->getWGrad()) {
+    resetWithMatrix(
+        bias, biases_->getWGrad(), wgtPD->diff_bias_primitive_desc());
+    CHECK(bias && biasVal_ &&
+          bias->getPrimitiveDesc() == biasVal_->getPrimitiveDesc())
+        << "primitive desc of bias grad should equal the bias value";
+  }
 
+  if (dataPD == nullptr) {
+    return;
+  }
+  resetInGrad(in, dataPD->diff_src_primitive_desc());
   resetWgtValBwdData(dataPD, wgtValBwdData_);
 }
 
@@ -416,10 +338,7 @@ void MKLDNNConvLayer::resetBwdPipeline(
     MKLDNNMatrixPtr& wgt,
     MKLDNNMatrixPtr& bias,
     MKLDNNMatrixPtr& out) {
-  if (cvtOutGrad_) {
-    pipeline.push_back(*cvtOutGrad_);
-  }
-
+  CHECK(inVal_);
   // add bwdWgt handle
   if (bias) {
     bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt, *bias));
@@ -431,99 +350,13 @@ void MKLDNNConvLayer::resetBwdPipeline(
   if (dataPD == nullptr) {
     return;
   }
-
   if (cvtWgtVal_) {
     pipeline.push_back(*cvtWgtVal_);
   }
-
   // add bwdData handle
   CHECK(wgtValBwdData_) << "Should have weight memory";
   bwdData_.reset(new conv_bwdData(*dataPD, *out, *wgtValBwdData_, *in));
   pipeline.push_back(*bwdData_);
-
-  if (cvtInGrad_) {
-    pipeline.push_back(*cvtInGrad_);
-  }
-}
-
-void MKLDNNConvLayer::resetOutGrad(
-    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD, MKLDNNMatrixPtr& out) {
-  cpuOutGrad_ = nullptr;
-  cvtOutGrad_ = nullptr;
-  CHECK(outVal_ != nullptr &&
-        outVal_->getPrimitiveDesc() == wgtPD->diff_dst_primitive_desc())
-      << "primitive desc of out grad and value should be equal";
-  if (outputIsOnlyMKLDNN()) {
-    MKLDNNLayer::resetOutGrad(out, outVal_->getPrimitiveDesc());
-  } else {
-    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
-    // always share the same grad data of CPU output
-    // then the activation can get the right grad from output_.grad
-    output_.grad->setData(cpuOut->getData());
-    // same PrimitiveDesc with cpuInVal_
-    CHECK(cpuOutVal_);
-    cpuOutGrad_ = MKLDNNMatrix::create(cpuOut, cpuOutVal_->getPrimitiveDesc());
-    // create reorder if primitive desc does not match
-    if (cpuOutGrad_->getPrimitiveDesc() != outVal_->getPrimitiveDesc()) {
-      out = MKLDNNMatrix::create(nullptr, outVal_->getPrimitiveDesc());
-      cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
-      CHECK(cvtOutGrad_);
-    } else {
-      out = cpuOutGrad_;
-    }
-  }
-}
-
-void MKLDNNConvLayer::resetWgtBiasGrad(
-    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
-    MKLDNNMatrixPtr& wgt,
-    MKLDNNMatrixPtr& bias) {
-  wgt = MKLDNNMatrix::create(weight_->getWGrad(),
-                             wgtPD->diff_weights_primitive_desc());
-  CHECK(nullptr != wgtVal_ &&
-        wgt->getPrimitiveDesc() == wgtVal_->getPrimitiveDesc())
-      << "primitive desc of weight grad and value should be equal";
-  VLOG(MKLDNN_FMTS) << "weight grad format: " << wgt->getFormat();
-
-  bias = nullptr;
-  if (biasVal_ == nullptr) {
-    return;
-  }
-  bias = MKLDNNMatrix::create(biases_->getWGrad(),
-                              wgtPD->diff_bias_primitive_desc());
-  CHECK(bias->getPrimitiveDesc() == biasVal_->getPrimitiveDesc())
-      << "primitive desc of bias grad should equal the bias value";
-}
-
-void MKLDNNConvLayer::resetInGrad(
-    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
-    MKLDNNMatrixPtr& in) {
-  in = nullptr;
-  cpuInGrad_ = nullptr;
-  cvtInGrad_ = nullptr;
-  if (dataPD == nullptr) {
-    return;
-  }
-
-  if (inputIsOnlyMKLDNN()) {
-    MKLDNNLayer::resetInGrad(in, dataPD->diff_src_primitive_desc());
-    CHECK(nullptr != inVal_ &&
-          in->getPrimitiveDesc() == inVal_->getPrimitiveDesc())
-        << "primitive desc of input grad and value should be equal";
-  } else {
-    const MatrixPtr& cpuIn = getInputGrad(0, CPU_DEVICE);
-    // same PrimitiveDesc with cpuInVal_
-    CHECK(cpuInVal_);
-    cpuInGrad_ = MKLDNNMatrix::create(cpuIn, cpuInVal_->getPrimitiveDesc());
-    in = cpuInGrad_;
-    // create reorder if PrimitiveDesc does not match
-    if (cpuInGrad_->getPrimitiveDesc() != dataPD->diff_src_primitive_desc()) {
-      in = MKLDNNMatrix::create(getInputGrad(0, MKLDNN_DEVICE),
-                                dataPD->diff_src_primitive_desc());
-      cvtInGrad_ = MKLDNNMatrix::createReorder(in, cpuInGrad_);
-      CHECK(cvtInGrad_);
-    }
-  }
 }
 
 void MKLDNNConvLayer::resetWgtValBwdData(
@@ -537,8 +370,7 @@ void MKLDNNConvLayer::resetWgtValBwdData(
   // since the primitive_desc would be different with wgtVal_
   CHECK(wgtVal_) << "should have weight value";
   if (dataPD->weights_primitive_desc() != wgtVal_->getPrimitiveDesc()) {
-    wgtValBwdData_ =
-        MKLDNNMatrix::create(nullptr, dataPD->weights_primitive_desc());
+    wgtValBwdData_ = MKLDNNMatrix::create(dataPD->weights_primitive_desc());
     cvtWgtVal_ = MKLDNNMatrix::createReorder(wgtVal_, wgtValBwdData_);
     CHECK(cvtWgtVal_);
   } else {

paddle/gserver/layers/MKLDNNConvLayer.h

@@ -48,17 +48,6 @@ protected:
   // save forward primitive_desc, which can be used backward
   std::shared_ptr<conv_fwd::primitive_desc> fwdPD_;
 
-  // MKLDNNMatrixPtr which should be created from CPU Device
-  MKLDNNMatrixPtr cpuInVal_;
-  MKLDNNMatrixPtr cpuInGrad_;
-  MKLDNNMatrixPtr cpuOutVal_;
-  MKLDNNMatrixPtr cpuOutGrad_;
-  // convert handle between CPU device and MKLDNN device
-  std::shared_ptr<mkldnn::reorder> cvtInVal_;
-  std::shared_ptr<mkldnn::reorder> cvtInGrad_;
-  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
-  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
-
   // whether the weight has been init
   bool hasInitedWgt_;
 
@@ -94,8 +83,6 @@ public:
                 MKLDNNMatrixPtr& bias,
                 MKLDNNMatrixPtr& out) override;
 
-  void updateInputData() override;
-
   void updateWeights(const UpdateCallback& callback) override;
 
   void convertWeightsFromPaddle() override;
@@ -109,26 +96,6 @@ public:
                        << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
   }
 
-  void printValueFormatFlow() override {
-    if (cpuInVal_) {
-      VLOG(MKLDNN_FMTS) << cpuInVal_->getFormat() << " >>>";
-    }
-    MKLDNNLayer::printValueFormatFlow();
-    if (cpuOutVal_) {
-      VLOG(MKLDNN_FMTS) << " >>> " << cpuOutVal_->getFormat();
-    }
-  }
-
-  void printGradFormatFlow() override {
-    if (cpuInGrad_) {
-      VLOG(MKLDNN_FMTS) << cpuInGrad_->getFormat() << " <<<";
-    }
-    MKLDNNLayer::printGradFormatFlow();
-    if (cpuOutGrad_) {
-      VLOG(MKLDNN_FMTS) << " <<< " << cpuOutGrad_->getFormat();
-    }
-  }
-
 protected:
   /**
    * load the dims settings of this conv
@@ -162,23 +129,6 @@ protected:
                         MKLDNNMatrixPtr& bias,
                         MKLDNNMatrixPtr& out);
 
-  /**
-   * reset MKLDNNMatrix of input value
-   */
-  void resetInValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
-                    MKLDNNMatrixPtr& in);
-  /**
-   * reset MKLDNNMatrix of weight and bias value
-   */
-  void resetWgtBiasValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
-                         MKLDNNMatrixPtr& wgt,
-                         MKLDNNMatrixPtr& bias);
-  /**
-   * reset MKLDNNMatrix of output value
-   */
-  void resetOutValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
-                     MKLDNNMatrixPtr& out);
-
   /**
    * reset the backward weight primitive descriptor.
    */
@@ -207,22 +157,6 @@ protected:
                         MKLDNNMatrixPtr& bias,
                         MKLDNNMatrixPtr& out);
 
-  /**
-   * reset MKLDNNMatrix of output grad
-   */
-  void resetOutGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
-                    MKLDNNMatrixPtr& out);
-  /**
-   * reset MKLDNNMatrix of weight and bias grad
-   */
-  void resetWgtBiasGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
-                        MKLDNNMatrixPtr& wgt,
-                        MKLDNNMatrixPtr& bias);
-  /**
-   * reset MKLDNNMatrix of input grad
-   */
-  void resetInGrad(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
-                   MKLDNNMatrixPtr& in);
   /**
    * reset MKLDNNMatrix of weight value for backward data
    * since the primitive_desc would be different with wgtVal_

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -62,7 +62,7 @@ void MKLDNNFcLayer::convertWeightsFromPaddle() {
   CHECK(wgtVal_) << "should have been initialized";
   bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
   auto targetDim = wgtVal_->getDims();
-  auto srcFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
+  auto srcFmt = hasNoSpatial_ ? format::io : format::ihwo;
   wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
   hasInitedWgt_ = true;
 }
@@ -71,7 +71,7 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
   CHECK(wgtVal_) << "should have been initialized";
   bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
   auto targetDim = wgtVal_->getDims();
-  auto dstFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
+  auto dstFmt = hasNoSpatial_ ? format::io : format::ihwo;
   wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
 }
 
@@ -100,8 +100,6 @@ void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
   resetFwdPD(fwdPD_, in, wgt, bias, out);
 
   resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
-
-  printValueFormatFlow();
 }
 
 void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
@@ -119,12 +117,6 @@ void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
   resetBwdDataPD(bwdDataPD, in, out);
 
   resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
-
-  printGradFormatFlow();
-}
-
-void MKLDNNFcLayer::updateInputData() {
-  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
 }
 
 void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
@@ -139,51 +131,30 @@ void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
                                     MKLDNNMatrixPtr& bias,
                                     MKLDNNMatrixPtr& out) {
   resetInValue(in);
-
-  resetWgtBiasValue(wgt, bias);
-
-  resetOutValue(out);
-}
-
-void MKLDNNFcLayer::resetInValue(MKLDNNMatrixPtr& in) {
-  if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& dnnIn = getInputValue(0);
-    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
-    CHECK(in) << "Input should be MKLDNNMatrix";
-  } else {
-    CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
-    in = MKLDNNMatrix::create(
-        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
-  }
+  CHECK(in);
   in->downSpatial();
-}
 
-void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
-                                      MKLDNNMatrixPtr& bias) {
+  auto outPD =
+      MKLDNNMatrix::createPrimitiveDesc({bs_, oc_}, format::nc, engine_);
+  resetOutValue(out, outPD);
+
   format wgtFmt = format::oihw;
-  if (inVal_->getFormat() == format::nChw8c) {
+  if (in->getFormat() == format::nChw8c) {
     wgtFmt = format::oIhw8i;
-  } else if (inVal_->getFormat() == format::nChw16c) {
+  } else if (in->getFormat() == format::nChw16c) {
     wgtFmt = format::oIhw16i;
   }
-  wgt = MKLDNNMatrix::create(
-      weight_->getW(), {oc_, ic_, ih_, iw_}, wgtFmt, engine_);
+  auto wgtPD =
+      MKLDNNMatrix::createPrimitiveDesc({oc_, ic_, ih_, iw_}, wgtFmt, engine_);
+  resetWithMatrix(wgt, weight_->getW(), wgtPD);
   wgt->downSpatial();
-  VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
-
-  bias = (biases_ && biases_->getW())
-             ? MKLDNNMatrix::create(biases_->getW(), {oc_}, format::x, engine_)
-             : nullptr;
-}
 
-void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
-  out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
-  if (!outputIsOnlyMKLDNN()) {
-    // fc cpu output value do not need create convert, just share data
-    getOutput(CPU_DEVICE).value->setData(out->getData());
+  if (biases_ && biases_->getW()) {
+    auto biasPD = MKLDNNMatrix::createPrimitiveDesc({oc_}, format::x, engine_);
+    resetWithMatrix(bias, biases_->getW(), biasPD);
+  } else {
+    bias = nullptr;
   }
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
 }
 
 void MKLDNNFcLayer::resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
@@ -219,7 +190,6 @@ void MKLDNNFcLayer::resetFwdPipeline(
   } else {
     fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out));
   }
-
   pipeline.push_back(*fwd_);
 }
 
@@ -227,44 +197,18 @@ void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
                                     MKLDNNMatrixPtr& wgt,
                                     MKLDNNMatrixPtr& bias,
                                     MKLDNNMatrixPtr& out) {
-  resetOutGrad(out);
-
-  resetWgtBiasGrad(wgt, bias);
-
-  resetInGrad(in);
-}
-
-void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
-  CHECK(outVal_);
-  if (outputIsOnlyMKLDNN()) {
-    MKLDNNLayer::resetOutGrad(out, outVal_->getPrimitiveDesc());
-  } else {
-    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
-    output_.grad->setData(cpuOut->getData());
-    out = MKLDNNMatrix::create(cpuOut, outVal_->getPrimitiveDesc());
-  }
-}
+  CHECK(inVal_ && outVal_);
+  resetOutGrad(out, outVal_->getPrimitiveDesc());
+  resetInGrad(in, inVal_->getPrimitiveDesc());
 
-void MKLDNNFcLayer::resetWgtBiasGrad(MKLDNNMatrixPtr& wgt,
-                                     MKLDNNMatrixPtr& bias) {
   CHECK(wgtVal_);
-  wgt = MKLDNNMatrix::create(weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
+  resetWithMatrix(wgt, weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
 
-  bias = nullptr;
-  if (biasVal_ == nullptr) {
-    return;
-  }
-  bias =
-      MKLDNNMatrix::create(biases_->getWGrad(), biasVal_->getPrimitiveDesc());
-}
-
-void MKLDNNFcLayer::resetInGrad(MKLDNNMatrixPtr& in) {
-  in = nullptr;
-  if (inputLayers_[0]->getOutput().grad == nullptr) {
-    return;
+  if (biasVal_) {
+    resetWithMatrix(bias, biases_->getWGrad(), biasVal_->getPrimitiveDesc());
+  } else {
+    bias = nullptr;
   }
-  CHECK(inVal_);
-  MKLDNNLayer::resetInGrad(in, inVal_->getPrimitiveDesc());
 }
 
 void MKLDNNFcLayer::resetBwdWgtPD(

paddle/gserver/layers/MKLDNNFcLayer.h

@@ -66,8 +66,6 @@ public:
                 MKLDNNMatrixPtr& bias,
                 MKLDNNMatrixPtr& out) override;
 
-  void updateInputData() override;
-
   void updateWeights(const UpdateCallback& callback) override;
 
   void convertWeightsFromPaddle() override;
@@ -84,9 +82,6 @@ protected:
                        MKLDNNMatrixPtr& wgt,
                        MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);
-  void resetInValue(MKLDNNMatrixPtr& in);
-  void resetWgtBiasValue(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
-  void resetOutValue(MKLDNNMatrixPtr& out);
   void resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
                   MKLDNNMatrixPtr in,
                   MKLDNNMatrixPtr wgt,
@@ -109,9 +104,6 @@ protected:
                        MKLDNNMatrixPtr& wgt,
                        MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);
-  void resetOutGrad(MKLDNNMatrixPtr& out);
-  void resetWgtBiasGrad(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
-  void resetInGrad(MKLDNNMatrixPtr& in);
   void resetBwdWgtPD(std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
                      MKLDNNMatrixPtr& wgt,
                      MKLDNNMatrixPtr& bias,

paddle/gserver/layers/MKLDNNLayer.cpp

+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "MKLDNNLayer.h"
+
+using namespace mkldnn;  // NOLINT
+typedef memory::format format;
+
+namespace paddle {
+
+bool MKLDNNLayer::init(const LayerMap& layerMap,
+                       const ParameterMap& parameterMap) {
+  CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
+                          << "Please set WITH_MKLDNN=ON "
+                          << "and set use_mkldnn=True";
+  CHECK(!useGpu_) << "Do not support GPU yet";
+
+  // set device id before Layer::init
+  setDevice(MKLDNN_DEVICE);
+  // change param device to MKLDNN device
+  setParamsDevice(MKLDNN_DEVICE, parameterMap);
+  if (!Layer::init(layerMap, parameterMap)) {
+    return false;
+  }
+  setOutputMap();
+  checkCPUOutputsNumber();
+
+  stream_.reset(new MKLDNNStream());
+  engine_ = CPUEngine::Instance().getEngine();
+  return true;
+}
+
+void MKLDNNLayer::forward(PassType passType) {
+  passType_ = passType;
+
+  {
+    REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
+    CHECK(!inputLayers_.empty());
+    copySeqInfoToOutputs();
+    size_t elemenCnt = inputLayers_[0]->getOutputValue()->getElementCnt();
+    if (inputElemenCnt_ != elemenCnt) {
+      VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward";
+      // reset when input total sizes changed, not only the batchsize
+      inputElemenCnt_ = elemenCnt;
+      pipelineFwd_.clear();
+      reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
+      // all cpu device output grad or value share output's
+      shareCPUDevice();
+      resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
+      // MKLDNNLayer output value should be MKLDNNMatrix
+      // so external output value is necessary.
+      // Then external input value is not necessary,
+      // since input may be mkldnn internal buffer.
+      CHECK(extOutVal_) << "external output value is necessary";
+      output_.value = std::dynamic_pointer_cast<Matrix>(extOutVal_);
+      CHECK(inVal_ && outVal_) << "internal memories are necessary";
+      if (cvtInVal_) {
+        pipelineFwd_.insert(pipelineFwd_.begin(), *cvtInVal_);
+      }
+      if (cvtOutVal_) {
+        pipelineFwd_.push_back(*cvtOutVal_);
+      }
+      convertWeightsFromPaddle();
+      printSizeInfo();
+      printValueFormat();
+      needResetBwd_ = true;
+    }
+
+    if (inputLayers_[0]->getType() == "data") {
+      // Update input value data when input layer is "data" type,
+      // since the input value data address might be changed.
+      CHECK(extInVal_);
+      extInVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+    }
+
+    if (!outputOnlyMKLDNN_) {
+      clearGrads();
+    }
+    stream_->submit(pipelineFwd_);
+  }
+  {
+    REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
+    forwardActivation();
+  }
+}
+
+void MKLDNNLayer::backward(const UpdateCallback& callback) {
+  if (needResetBwd_) {
+    VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward";
+    pipelineBwd_.clear();
+    pipelineMergeGrad_.clear();
+    mergeGrad_ = nullptr;
+    resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
+    // external output grad is not necessary
+    // since output may be mkldnn internal buffer or merge them directly.
+    CHECK(outGrad_) << "internal output grad is necessary";
+    if (extOutGrad_) {
+      CHECK_EQ(extOutGrad_->getData(), output_.grad->getData())
+          << "the external buffer should share the same data with output_.grad";
+    }
+    if (cvtOutGrad_) {
+      pipelineBwd_.insert(pipelineBwd_.begin(), *cvtOutGrad_);
+    }
+    if (cvtInGrad_) {
+      pipelineBwd_.push_back(*cvtInGrad_);
+    }
+    printGradFormat();
+    needResetBwd_ = false;
+  }
+
+  // merge grad must before backward activation
+  if (mergeGrad_) {
+    REGISTER_TIMER_INFO("MergeBpGrad", getName().c_str());
+    stream_->submit(pipelineMergeGrad_);
+  }
+  {
+    REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
+    backwardActivation();
+  }
+  {
+    REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
+    stream_->submit(pipelineBwd_);
+  }
+  {
+    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+    updateWeights(callback);
+  }
+}
+
+void MKLDNNLayer::reshapeInput(int& batchsize, int& height, int& width) {
+  const Argument& input = inputLayers_[0]->getOutput();
+  batchsize = input.getBatchSize();
+  int h = input.getFrameHeight();
+  int w = input.getFrameWidth();
+  if (h != 0) {
+    height = h;
+  }
+  if (w != 0) {
+    width = w;
+  }
+}
+
+void MKLDNNLayer::reshapeOutput(size_t height, size_t width) {
+  output_.setFrameHeight(height);
+  output_.setFrameWidth(width);
+  for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+    outputOtherDevice_[i].setFrameHeight(height);
+    outputOtherDevice_[i].setFrameWidth(width);
+  }
+}
+
+void MKLDNNLayer::resetWithMatrix(MKLDNNMatrixPtr& dnn,
+                                  const MatrixPtr& mat,
+                                  memory::primitive_desc pd) {
+  dnn = nullptr;
+  if (mat == nullptr) {
+    return;
+  }
+  dnn = MKLDNNMatrix::create(pd, mat);
+}
+
+void MKLDNNLayer::resetInValue(
+    MKLDNNMatrixPtr& in, const std::shared_ptr<memory::primitive_desc>& intPD) {
+  cvtInVal_ = nullptr;
+  extInVal_ = nullptr;
+  in = nullptr;
+  CHECK_GT(bs_ * ic_ * ih_ * iw_, 0);
+  auto extPD = MKLDNNMatrix::createPrimitiveDesc(
+      {bs_, ic_, ih_, iw_}, format::nchw, engine_);
+  const MatrixPtr& inMat = inputLayers_[0]->getOutputValue();
+  in = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
+  CHECK_EQ(inputIsOnlyMKLDNN(), in != nullptr);
+  if (in == nullptr || in->getFormat() == format::nc) {
+    in = MKLDNNMatrix::create(extPD, inMat);
+  }
+  extInVal_ = isPaddleFormat(in->getFormat()) ? in : nullptr;
+  if (in->getFormat() == format::nc) {
+    CHECK(ih_ == 1 && iw_ == 1);
+  }
+  if (nullptr == intPD || in->getPrimitiveDesc() == *intPD) {
+    return;
+  }
+  // need create reorder
+  in = MKLDNNMatrix::create(*intPD);
+  extInVal_ = extInVal_ ? extInVal_ : MKLDNNMatrix::create(extPD, inMat);
+  cvtInVal_ = MKLDNNMatrix::createReorder(extInVal_, in);
+  CHECK(cvtInVal_) << "should not be emptry";
+}
+
+void MKLDNNLayer::resetOutValue(MKLDNNMatrixPtr& out,
+                                memory::primitive_desc intPD) {
+  cvtOutVal_ = nullptr;
+  out = MKLDNNMatrix::create(intPD, output_.value);
+  extOutVal_ = out;
+  if (outputIsOnlyMKLDNN() || isPaddleFormat(extOutVal_->getFormat())) {
+    return;
+  }
+  // need create reorder
+  CHECK_GT(bs_ * oc_ * oh_ * ow_, 0);
+  extOutVal_ = MKLDNNMatrix::create(
+      memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_, output_.value);
+  out = MKLDNNMatrix::create(intPD);
+  cvtOutVal_ = MKLDNNMatrix::createReorder(out, extOutVal_);
+  CHECK(cvtOutVal_) << "should not be empty";
+}
+
+void MKLDNNLayer::resetInGrad(MKLDNNMatrixPtr& in,
+                              memory::primitive_desc intPD) {
+  cvtInGrad_ = nullptr;
+  extInGrad_ = nullptr;
+  in = nullptr;
+  LayerPtr& input = inputLayers_[0];
+  if (input->getOutputGrad() == nullptr) {
+    // no need input grad
+    return;
+  }
+  CHECK(inputIsOnlyMKLDNN() || input->getOutputMapSize() <= 1)
+      << "only support input is MKLDNN layer or only have one output layer";
+  // when input is a mkldnn branch node,
+  // this layer will save input grad to a internal buffer,
+  // and the mkldnn input layer will merge them to actual prev->output_.grad
+  const MatrixPtr& inMat =
+      input->getOutputMapSize() <= 1 ? input->getOutputGrad() : nullptr;
+  in = MKLDNNMatrix::create(intPD, inMat);
+  Argument& arg = input->getOutput(this->getName());
+  arg.grad = std::dynamic_pointer_cast<Matrix>(in);
+  CHECK(inVal_);
+  CHECK(inVal_->getPrimitiveDesc() == intPD) << "the primitive desc must equal";
+  if (inputIsOnlyMKLDNN()) {
+    return;
+  }
+
+  extInGrad_ = in;
+  if (isPaddleFormat(extInGrad_->getFormat())) {
+    return;
+  }
+  // need create reorder
+  // TODO(TJ): add macro definition to simplify it
+  CHECK(extInVal_ != nullptr && isPaddleFormat(extInVal_->getFormat()))
+      << "should have external input value and the format must be nchw(nc)";
+  extInGrad_ = MKLDNNMatrix::create(extInVal_->getPrimitiveDesc(), inMat);
+  CHECK(inVal_ != nullptr && inVal_->getPrimitiveDesc() == intPD)
+      << "should have internal input value and primitive desc must equal";
+  in = MKLDNNMatrix::create(intPD);
+  cvtInGrad_ = MKLDNNMatrix::createReorder(in, extInGrad_);
+  CHECK(cvtInGrad_);
+}
+
+void MKLDNNLayer::resetOutGrad(MKLDNNMatrixPtr& out,
+                               memory::primitive_desc intPD) {
+  cvtOutGrad_ = nullptr;
+  extOutGrad_ = nullptr;
+  out = nullptr;
+  MatrixPtr& outMat = output_.grad;
+  out = MKLDNNMatrix::create(intPD, outMat);
+  resetMergeGrad(out);
+  if (outputIsOnlyMKLDNN()) {
+    return;
+  }
+  CHECK_LE(outputMap_.size(), 1U) << "do not support mixed with cpu device";
+  extOutGrad_ = out;
+  if (isPaddleFormat(extOutGrad_->getFormat())) {
+    return;
+  }
+  // need create reorder
+  CHECK(extOutVal_ != nullptr && isPaddleFormat(extOutVal_->getFormat()))
+      << "should have external output value and the format must be nchw(nc)";
+  extOutGrad_ = MKLDNNMatrix::create(extOutVal_->getPrimitiveDesc(), outMat);
+  CHECK(outVal_ != nullptr && outVal_->getPrimitiveDesc() == intPD)
+      << "should have internal output value and primitive desc must equal";
+  out = MKLDNNMatrix::create(intPD);
+  cvtOutGrad_ = MKLDNNMatrix::createReorder(extOutGrad_, out);
+  CHECK(cvtOutGrad_);
+}
+
+void MKLDNNLayer::resetMergeGrad(MKLDNNMatrixPtr& out) {
+  mergeGrad_ = nullptr;
+  pipelineMergeGrad_.clear();
+  if (outputMap_.size() <= 1 || !outputIsOnlyMKLDNN()) {
+    // do not merge when output is not all MKLDNN or only one output
+    return;
+  }
+  CHECK(out) << "should have reset internal ouput grad";
+  std::vector<double> scales(outputMap_.size(), 1.0);
+  std::vector<memory::primitive_desc> srcPDs;
+  std::vector<primitive::at> srcs;
+  for (auto it = outputMap_.begin(); it != outputMap_.end(); ++it) {
+    MKLDNNMatrixPtr src =
+        std::dynamic_pointer_cast<MKLDNNMatrix>(it->second->grad);
+    CHECK(src) << "should be MKLDNNMatrix";
+    auto srcDims = src->getDims();
+    auto dstDims = out->getDims();
+    CHECK_EQ(srcDims.size(), dstDims.size());
+    for (size_t i = 0; i < srcDims.size(); ++i) {
+      CHECK_EQ(srcDims[i], dstDims[i]);
+    }
+    VLOG(MKLDNN_BASE) << getName() << " has output grad " << it->first
+                      << ", format " << src->getFormat();
+    srcPDs.push_back(src->getPrimitiveDesc());
+    srcs.push_back(*src);
+  }
+
+  // TODO(TJ): remove me when mkldnn sum support different formats
+  for (size_t i = 1; i < srcPDs.size(); ++i) {
+    CHECK(srcPDs[0] == srcPDs[i]);
+  }
+  tmpOutGrad_ = out;
+  tmpCvt_ = nullptr;
+  if (out->getPrimitiveDesc() != srcPDs[0]) {
+    tmpOutGrad_ = MKLDNNMatrix::create(srcPDs[0]);
+    tmpCvt_ = MKLDNNMatrix::createReorder(tmpOutGrad_, out);
+    CHECK(tmpCvt_);
+    pipelineMergeGrad_.push_back(*tmpCvt_);
+  }
+
+  auto sumPD =
+      sum::primitive_desc(tmpOutGrad_->getMemoryDesc(), scales, srcPDs);
+  mergeGrad_.reset(new sum(sumPD, srcs, *tmpOutGrad_));
+  pipelineMergeGrad_.insert(pipelineMergeGrad_.begin(), *mergeGrad_);
+}
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNLayer.h

@@ -58,11 +58,31 @@ protected:
   std::vector<mkldnn::primitive> pipelineFwd_;
   std::vector<mkldnn::primitive> pipelineBwd_;
 
-  // MKLDNNMatrixPtr with internal format
+  /* Value and grad are seperated as internal and external buffers.
+   * Each MKLDNNLayer must init or reset internal buffer at least,
+   * and the external buffer format is always nchw of nc(when h==w==1),
+   * which is the same format as paddle.
+   * The output_.value and output_.grad always save the external data,
+   * when mixed with cpu device.
+   * When all layers are mkldnn layers, they could save internal data.
+   */
+  // below MKLDNNMatrix buffers are all internal buffers
   MKLDNNMatrixPtr inVal_;
   MKLDNNMatrixPtr inGrad_;
   MKLDNNMatrixPtr outVal_;
   MKLDNNMatrixPtr outGrad_;
+  // below are external value and grad
+  MKLDNNMatrixPtr extInVal_;
+  MKLDNNMatrixPtr extInGrad_;
+  MKLDNNMatrixPtr extOutVal_;
+  MKLDNNMatrixPtr extOutGrad_;
+  // convert handle between external and internal buffers
+  std::shared_ptr<mkldnn::reorder> cvtInVal_;
+  std::shared_ptr<mkldnn::reorder> cvtInGrad_;
+  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
+  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
+
+  // weight and bias are always internal buffers
   MKLDNNMatrixPtr wgtVal_;
   MKLDNNMatrixPtr wgtGrad_;
   MKLDNNMatrixPtr biasVal_;
@@ -91,6 +111,7 @@ public:
         oh_(0),
         ow_(0),
         needResetBwd_(true),
+        outputOnlyMKLDNN_(false),
         engine_(mkldnn::engine::cpu, 0),
         stream_(nullptr),
         fwd_(nullptr),
@@ -99,92 +120,9 @@ public:
 
   ~MKLDNNLayer() {}
 
-  virtual bool init(const LayerMap& layerMap,
-                    const ParameterMap& parameterMap) {
-    CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
-                            << "Please set WITH_MKLDNN=ON "
-                            << "and set use_mkldnn=True";
-    CHECK(!useGpu_) << "Do not support GPU yet";
-
-    // set device id before Layer::init
-    setDevice(MKLDNN_DEVICE);
-    // change param device to MKLDNN device
-    setParamsDevice(MKLDNN_DEVICE, parameterMap);
-    if (!Layer::init(layerMap, parameterMap)) {
-      return false;
-    }
-    setOutputMap();
-    checkCPUOutputsNumber();
-
-    stream_.reset(new MKLDNNStream());
-    engine_ = CPUEngine::Instance().getEngine();
-    return true;
-  }
-
-  void forward(PassType passType) override {
-    passType_ = passType;
-
-    {
-      REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
-      CHECK(!inputLayers_.empty());
-      copySeqInfoToOutputs();
-      size_t elemenCnt = inputLayers_[0]->getOutput().value->getElementCnt();
-      if (inputElemenCnt_ != elemenCnt) {
-        VLOG(MKLDNN_BASE) << getName() << " reset mkldnn forward";
-        // reset when input total sizes changed, not only the batchsize
-        inputElemenCnt_ = elemenCnt;
-        pipelineFwd_.clear();
-        reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
-        resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
-        convertWeightsFromPaddle();
-        needResetBwd_ = true;
-      }
-
-      if (inputLayers_[0]->getType() == "data") {
-        updateInputData();
-      }
-
-      if (!outputOnlyMKLDNN_) {
-        clearGrads();
-      }
-      stream_->submit(pipelineFwd_);
-    }
-
-    /* activation */ {
-      REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
-      forwardActivation();
-    }
-  }
-
-  void backward(const UpdateCallback& callback) override {
-    if (needResetBwd_) {
-      VLOG(MKLDNN_BASE) << getName() << " reset mkldnn backward";
-      pipelineBwd_.clear();
-      pipelineMergeGrad_.clear();
-      mergeGrad_ = nullptr;
-      resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
-      needResetBwd_ = false;
-    }
-
-    // merge grad must before backward activation
-    if (mergeGrad_) {
-      REGISTER_TIMER_INFO("MergeBpGrad", getName().c_str());
-      stream_->submit(pipelineMergeGrad_);
-    }
-    {
-      REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
-      backwardActivation();
-    }
-    {
-      REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
-      stream_->submit(pipelineBwd_);
-    }
-
-    {
-      REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
-      updateWeights(callback);
-    }
-  }
+  virtual bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+  virtual void forward(PassType passType);
+  virtual void backward(const UpdateCallback& callback);
 
   /**
    * reshape the input image sizes
@@ -195,7 +133,7 @@ public:
       int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) = 0;
 
   /**
-   * reset the mkldnn forward primitve and memory
+   * reset the mkldnn forward primitve and memories
    * only would be called when input size changes
    */
   virtual void resetFwd(std::vector<mkldnn::primitive>& pipeline,
@@ -205,7 +143,7 @@ public:
                         MKLDNNMatrixPtr& out) = 0;
 
   /**
-   * reset the mkldnn backward primitve and memory for mkldnn fc
+   * reset the mkldnn backward primitve and memories
    * only would be called when needed
    */
   virtual void resetBwd(std::vector<mkldnn::primitive>& pipeline,
@@ -214,12 +152,6 @@ public:
                         MKLDNNMatrixPtr& bias,
                         MKLDNNMatrixPtr& out) = 0;
 
-  /**
-   * Update input value data when input layer is "data" type.
-   * Since the input value data address might be changed.
-   */
-  virtual void updateInputData() {}
-
   /**
    * Update weights and biases if necessary.
    */
@@ -246,131 +178,78 @@ protected:
   /**
    * reshape the input image sizes and input batchsize
    */
-  virtual void reshapeInput(int& batchsize, int& height, int& width) {
-    const Argument& input = inputLayers_[0]->getOutput();
-    batchsize = input.getBatchSize();
-    int h = input.getFrameHeight();
-    int w = input.getFrameWidth();
-    if (h != 0) {
-      height = h;
-    }
-    if (w != 0) {
-      width = w;
-    }
-  }
+  void reshapeInput(int& batchsize, int& height, int& width);
 
   /**
    * reshape output image sizes
    */
-  virtual void reshapeOutput(size_t height, size_t width) {
-    output_.setFrameHeight(height);
-    output_.setFrameWidth(width);
-    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
-      outputOtherDevice_[i].setFrameHeight(height);
-      outputOtherDevice_[i].setFrameWidth(width);
-    }
-  }
+  void reshapeOutput(size_t height, size_t width);
 
   /**
-   * reset the output grad matrix from primitive desc.
-   * and reset the merge grad primitive if needed.
-   * note: when this layer has serval outputs,
-   *       it could not be mixed with cpu device,
-   *       since it can not get memory desc from cpu device.
+   * reset MKLDNNMatrix from Matrix and internal primitive desc.
+   * reset nullptr if matrix or primitive desc is empty
    */
-  virtual void resetOutGrad(MKLDNNMatrixPtr& out,
-                            mkldnn::memory::primitive_desc pd) {
-    CHECK(outputIsOnlyMKLDNN()) << "do not support mixed with other device yet";
-    mergeGrad_ = nullptr;
-    pipelineMergeGrad_.clear();
-    out = MKLDNNMatrix::create(output_.grad, pd);
-    if (outputMap_.size() <= 1) {
-      return;
-    }
-    std::vector<double> scales(outputMap_.size(), 1.0);
-    std::vector<mkldnn::memory::primitive_desc> srcPDs;
-    std::vector<mkldnn::primitive::at> srcs;
-    for (auto it = outputMap_.begin(); it != outputMap_.end(); ++it) {
-      MKLDNNMatrixPtr src =
-          std::dynamic_pointer_cast<MKLDNNMatrix>(it->second->grad);
-      VLOG(MKLDNN_BASE) << getName() << " has output grad " << it->first;
-      CHECK(src) << "should be MKLDNNMatrix";
-      auto srcDims = src->getDims();
-      auto dstDims = out->getDims();
-      CHECK_EQ(srcDims.size(), dstDims.size());
-      for (size_t i = 0; i < srcDims.size(); ++i) {
-        CHECK_EQ(srcDims[i], dstDims[i]);
-      }
-      srcPDs.push_back(src->getPrimitiveDesc());
-      srcs.push_back(*src);
-    }
+  void resetWithMatrix(MKLDNNMatrixPtr& dnn,
+                       const MatrixPtr& mat,
+                       mkldnn::memory::primitive_desc pd);
 
-    // TODO(TJ): remove me when mkldnn sum support different formats
-    for (size_t i = 1; i < srcPDs.size(); ++i) {
-      CHECK(srcPDs[0] == srcPDs[i]);
-    }
-    tmpOutGrad_ = nullptr;
-    tmpCvt_ = nullptr;
-    if (out->getPrimitiveDesc() != srcPDs[0]) {
-      tmpOutGrad_ = MKLDNNMatrix::create(nullptr, srcPDs[0]);
-      tmpCvt_ = MKLDNNMatrix::createReorder(tmpOutGrad_, out);
-      CHECK(tmpCvt_);
-      pipelineMergeGrad_.push_back(*tmpCvt_);
-    } else {
-      tmpOutGrad_ = out;
-    }
+  /**
+   * reset input value from input MKLDNNMatrix and internal primitive desc.
+   * reset both internal and external buffer and create reorder if necessary.
+   */
+  void resetInValue(
+      MKLDNNMatrixPtr& in,
+      const std::shared_ptr<mkldnn::memory::primitive_desc>& intPD = nullptr);
 
-    auto sumPD = mkldnn::sum::primitive_desc(
-        tmpOutGrad_->getMemoryDesc(), scales, srcPDs);
-    mergeGrad_.reset(new mkldnn::sum(sumPD, srcs, *tmpOutGrad_));
-    pipelineMergeGrad_.insert(pipelineMergeGrad_.begin(), *mergeGrad_);
-  }
+  /**
+   * reset output value from internal primitive desc.
+   * reset both internal and external buffer and create reorder if necessary.
+   */
+  void resetOutValue(MKLDNNMatrixPtr& out,
+                     mkldnn::memory::primitive_desc intPD);
 
   /**
-   * reset input grad from primitive desc.
-   * this function is avaiable for input is only mkldnn
-   * or input do not care cpu device
+   * reset input grad from internal primitive desc.
+   * reset both internal and external buffer and create reorder if necessary.
    */
-  virtual void resetInGrad(MKLDNNMatrixPtr& in,
-                           mkldnn::memory::primitive_desc pd) {
-    LayerPtr& input = inputLayers_[0];
-    const MatrixPtr& grad =
-        input->getOutputMapSize() > 1 ? nullptr : input->getOutput().grad;
-    in = MKLDNNMatrix::create(grad, pd);
-    Argument& arg = input->getOutput(this->getName());
-    arg.grad = std::dynamic_pointer_cast<Matrix>(in);
-  }
+  void resetInGrad(MKLDNNMatrixPtr& in, mkldnn::memory::primitive_desc intPD);
 
   /**
-   * print info about sizes
+   * reset output grad from internal primitive desc.
+   * merge grad if necessary.
+   * reset both internal and external buffer and create reorder if necessary.
+   * note: about merge grad, when this layer has several outputs,
+   *       it could not be mixed with cpu device,
+   *       since it can not get memory desc from cpu device.
    */
-  virtual void printSizeInfo() {
-    VLOG(MKLDNN_SIZES) << getName() << ": bs: " << bs_ << ", ic: " << ic_
-                       << ", ih: " << ih_ << ", iw: " << iw_ << ", oc: " << oc_
-                       << ", oh: " << oh_ << ", ow: " << ow_;
-  }
+  void resetOutGrad(MKLDNNMatrixPtr& out, mkldnn::memory::primitive_desc intPD);
 
   /**
-   * Print the mkldnn memory format flow of value
+   * reset the merge grad primitive if necessary.
+   * note: do not support the grads mixed with cpu device,
+   *       since it can not get memory desc from cpu device.
    */
-  virtual void printValueFormatFlow() {
-    if (inVal_ && outVal_) {
-      VLOG(MKLDNN_FMTS) << inVal_->getFormat() << " >>> "
-                        << outVal_->getFormat();
-    }
-  }
+  void resetMergeGrad(MKLDNNMatrixPtr& out);
 
+protected:
   /**
-   * Print the mkldnn memory format flow of grad
+   * Set deviceId of this layer.
    */
-  virtual void printGradFormatFlow() {
-    if (inGrad_ && outGrad_) {
-      VLOG(MKLDNN_FMTS) << inGrad_->getFormat() << " <<< "
-                        << outGrad_->getFormat();
+  void setDevice(int id) { deviceId_ = id; }
+
+  /**
+   * check the format is nchw or nc,
+   * which is supported by Paddle default memory layout
+   */
+  bool isPaddleFormat(mkldnn::memory::format fmt) {
+    if (fmt == mkldnn::memory::format::nchw ||
+        fmt == mkldnn::memory::format::nc) {
+      return true;
+    } else {
+      return false;
     }
   }
 
-protected:
   /**
    * If input only has MKLDNN device.
    * Otherwise, only support the previous layer using CPU device.
@@ -380,7 +259,6 @@ protected:
     if (prevDevice == MKLDNN_DEVICE) {
       return true;
     } else {
-      // do not support GPU yet
       CHECK_EQ(prevDevice, CPU_DEVICE) << "Only support CPU yet";
       return false;
     }
@@ -400,9 +278,61 @@ protected:
   }
 
   /**
-   * Set deviceId of this layer.
+   * print info about sizes
    */
-  void setDevice(int id) { deviceId_ = id; }
+  virtual void printSizeInfo() {
+    VLOG(MKLDNN_SIZES) << getName() << ": bs: " << bs_ << ", ic: " << ic_
+                       << ", ih: " << ih_ << ", iw: " << iw_ << ", oc: " << oc_
+                       << ", oh: " << oh_ << ", ow: " << ow_;
+  }
+
+  /**
+   * print the mkldnn memory format of value
+   */
+  virtual void printValueFormat() {
+    if (extInVal_) {
+      VLOG(MKLDNN_FMTS) << extInVal_->getFormat() << " >>> ";
+    }
+    if (inVal_) {
+      VLOG(MKLDNN_FMTS) << inVal_->getFormat() << " >>>";
+    }
+    if (outVal_) {
+      VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> ";
+    }
+    if (extOutVal_) {
+      VLOG(MKLDNN_FMTS) << extOutVal_->getFormat();
+    }
+    if (wgtVal_) {
+      VLOG(MKLDNN_FMTS) << "Weight value format: " << wgtVal_->getFormat();
+    }
+    if (biasVal_) {
+      VLOG(MKLDNN_FMTS) << "Bias value format: " << biasVal_->getFormat();
+    }
+  }
+
+  /**
+   * print the mkldnn memory format of grad
+   */
+  virtual void printGradFormat() {
+    if (extOutGrad_) {
+      VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat();
+    }
+    if (outGrad_) {
+      VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< ";
+    }
+    if (inGrad_) {
+      VLOG(MKLDNN_FMTS) << inGrad_->getFormat() << " <<<";
+    }
+    if (extInGrad_) {
+      VLOG(MKLDNN_FMTS) << extInGrad_->getFormat() << " <<< ";
+    }
+    if (wgtGrad_) {
+      VLOG(MKLDNN_FMTS) << "Weight grad format: " << wgtGrad_->getFormat();
+    }
+    if (biasGrad_) {
+      VLOG(MKLDNN_FMTS) << "Bias grad format: " << biasGrad_->getFormat();
+    }
+  }
 
 private:
   /**
@@ -449,6 +379,19 @@ private:
     }
   }
 
+  /**
+   * if have cpu device, share value and grad data with output_
+   */
+  void shareCPUDevice() {
+    if (outputIsOnlyMKLDNN()) {
+      return;
+    }
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].value = output_.value;
+      outputOtherDevice_[i].grad = output_.grad;
+    }
+  }
+
   /**
    * Check the cpu device number of outputOtherDevice_.
    * should have only one at most.

paddle/gserver/layers/MKLDNNPoolLayer.cpp

@@ -85,8 +85,6 @@ void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,
   resetFwdPD(fwdPD_, in, out);
 
   resetFwdPipeline(pipeline, fwdPD_, in, out);
-
-  printValueFormatFlow();
 }
 
 void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
@@ -101,65 +99,22 @@ void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
   resetBwdPD(pd, in, out);
 
   resetBwdPipeline(pipeline, pd, in, out);
-
-  printGradFormatFlow();
-}
-
-void MKLDNNPoolLayer::updateInputData() {
-  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
 }
 
 void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
                                       MKLDNNMatrixPtr& out) {
   resetInValue(in);
 
-  resetOutValue(out);
-}
-
-void MKLDNNPoolLayer::resetInValue(MKLDNNMatrixPtr& in) {
-  if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& dnnIn = getInputValue(0);
-    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
-    CHECK(in) << "Input should be MKLDNNMatrix";
-  } else {
-    CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
-    in = MKLDNNMatrix::create(
-        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
-  }
-}
-
-void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
-  CHECK(inVal_) << "Should reset input value first";
   memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
-  out = MKLDNNMatrix::create(
-      output_.value, outDims, inVal_->getFormat(), engine_);
-
-  // create reorder if output value has cpu device and pd do not match
-  cpuOutVal_ = nullptr;
-  cvtOutVal_ = nullptr;
-  if (!outputIsOnlyMKLDNN()) {
-    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
-    cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
-    if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
-      out = MKLDNNMatrix::create(nullptr, out->getPrimitiveDesc());
-      cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
-      CHECK(cvtOutVal_) << "should not be emptry";
-    } else {
-      cpuOut->setData(output_.value->getData());
-      cpuOutVal_ = out;
-    }
-    output_.value = std::dynamic_pointer_cast<Matrix>(cpuOutVal_);
-    return;
-  }
-  output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
+  CHECK(in);
+  auto outPD =
+      MKLDNNMatrix::createPrimitiveDesc(outDims, in->getFormat(), engine_);
+  resetOutValue(out, outPD);
 }
 
 void MKLDNNPoolLayer::resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
                                  MKLDNNMatrixPtr in,
                                  MKLDNNMatrixPtr out) {
-  memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
-  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
   memory::dims kernels = memory::dims{fh_, fw_};
   memory::dims strides = memory::dims{sh_, sw_};
   memory::dims padL = memory::dims{ph_, pw_};
@@ -194,58 +149,26 @@ void MKLDNNPoolLayer::resetFwdPipeline(
              ? std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out, *workspace_))
              : std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out));
   pipeline.push_back(*fwd_);
-
-  if (cvtOutVal_) {
-    pipeline.push_back(*cvtOutVal_);
-  }
 }
 
 void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
                                       MKLDNNMatrixPtr& out) {
-  resetOutGrad(out);
-
-  resetInGrad(in);
-}
-void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
-  cpuOutGrad_ = nullptr;
-  cvtOutGrad_ = nullptr;
-  CHECK(outVal_);
-  if (outputIsOnlyMKLDNN()) {
-    MKLDNNLayer::resetOutGrad(out, outVal_->getPrimitiveDesc());
-  } else {
-    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
-    // always share the same grad data of CPU output
-    // then the activation can get the right grad from output_.grad
-    output_.grad->setData(cpuOut->getData());
-    cpuOutGrad_ = MKLDNNMatrix::create(
-        cpuOut, memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_);
-    if (cpuOutGrad_->getPrimitiveDesc() != outVal_->getPrimitiveDesc()) {
-      out = MKLDNNMatrix::create(nullptr, outVal_->getPrimitiveDesc());
-      cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
-      CHECK(cvtOutGrad_) << "should not be emptry";
-    } else {
-      out = cpuOutGrad_;
-    }
-  }
-}
-
-void MKLDNNPoolLayer::resetInGrad(MKLDNNMatrixPtr& in) {
-  in = nullptr;
-  if (inputLayers_[0]->getOutput().grad == nullptr) {
-    return;
-  }
-  CHECK(inVal_);
-  MKLDNNLayer::resetInGrad(in, inVal_->getPrimitiveDesc());
+  CHECK(inVal_ && outVal_);
+  resetOutGrad(out, outVal_->getPrimitiveDesc());
+  resetInGrad(in, inVal_->getPrimitiveDesc());
 }
 
 void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
                                  MKLDNNMatrixPtr& in,
                                  MKLDNNMatrixPtr& out) {
+  pd = nullptr;
+  if (in == nullptr) {
+    return;
+  }
   memory::dims kernels = memory::dims{fh_, fw_};
   memory::dims strides = memory::dims{sh_, sw_};
   memory::dims padL = memory::dims{ph_, pw_};
   memory::dims padR = getPaddingR();
-  CHECK(in);
   CHECK(out);
   auto bwdDesc = pool_bwd::desc(poolAlgo_,
                                 in->getMemoryDesc(),
@@ -263,8 +186,8 @@ void MKLDNNPoolLayer::resetBwdPipeline(
     std::shared_ptr<pool_bwd::primitive_desc>& pd,
     MKLDNNMatrixPtr& in,
     MKLDNNMatrixPtr& out) {
-  if (cvtOutGrad_) {
-    pipeline.push_back(*cvtOutGrad_);
+  if (pd == nullptr) {
+    return;
   }
 
   bwdData_ =

paddle/gserver/layers/MKLDNNPoolLayer.h

@@ -38,13 +38,6 @@ protected:
   // pooling_avg or pooling_max
   mkldnn::algorithm poolAlgo_;
 
-  // MKLDNNMatrixPtr which should be created from CPU Device
-  MKLDNNMatrixPtr cpuOutVal_;
-  MKLDNNMatrixPtr cpuOutGrad_;
-  // convert handle between CPU device and MKLDNN device
-  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
-  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
-
   // save forward primitive_desc, which can be used backward
   std::shared_ptr<pool_fwd::primitive_desc> fwdPD_;
   // according to https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
@@ -74,8 +67,6 @@ public:
                 MKLDNNMatrixPtr& bias,
                 MKLDNNMatrixPtr& out) override;
 
-  void updateInputData() override;
-
   void printSizeInfo() override {
     MKLDNNLayer::printSizeInfo();
     VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
@@ -90,8 +81,6 @@ protected:
    *                    reset pipeline.
    */
   void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
-  void resetInValue(MKLDNNMatrixPtr& in);
-  void resetOutValue(MKLDNNMatrixPtr& out);
   void resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
                   MKLDNNMatrixPtr in,
                   MKLDNNMatrixPtr out);
@@ -106,8 +95,6 @@ protected:
    *                     reset pipeline.
    */
   void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
-  void resetOutGrad(MKLDNNMatrixPtr& out);
-  void resetInGrad(MKLDNNMatrixPtr& in);
   void resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
                   MKLDNNMatrixPtr& in,
                   MKLDNNMatrixPtr& out);

paddle/gserver/tests/MKLDNNTester.cpp

@@ -97,7 +97,7 @@ void MKLDNNTester::randomWgtDatas() {
     parameters_[REF][i]->randomize();
     dnnValue->copyFrom(*refValue);
 
-    VLOG(lvl_) << "Random weight data " << parameters_[DNN][i]->getName();
+    VLOG(MKLDNN_TESTS) << "Random weight " << parameters_[DNN][i]->getName();
     printVector(dnnValue);
   }
 }
@@ -109,7 +109,7 @@ void MKLDNNTester::randomBotDatas() {
     dataLayers_[REF][i]->getOutputValue()->randomizeUniform();
     dataLayers_[DNN][i]->getOutputValue()->copyFrom(
         *(dataLayers_[REF][i]->getOutputValue()));
-    VLOG(lvl_) << "Input " << i << " data:";
+    VLOG(MKLDNN_TESTS) << "Random Foward, InputValue " << i;
     printMatrix(dataLayers_[REF][i]->getOutputValue());
   }
 }
@@ -118,12 +118,12 @@ void MKLDNNTester::randomTopDiffs() {
   refLayer_->getOutputGrad()->randomizeUniform();
   dnnLayer_->getOutput(CPU_DEVICE)
       .grad->copyFrom(*(refLayer_->getOutputGrad()));
-  VLOG(lvl_) << "Random Backward Input, TopDiff: ";
+  VLOG(MKLDNN_TESTS) << "Random Backward, OutputGrad";
   printMatrix(refLayer_->getOutputGrad());
 }
 
 void MKLDNNTester::checkForward() {
-  VLOG(MKLDNN_ALL) << "Check Forward";
+  VLOG(MKLDNN_TESTS) << "Check Forward";
   printTopDatas();
   double delta =
       compareMatrix(dnnLayer_->getOutputValue(), refLayer_->getOutputValue());
@@ -131,15 +131,15 @@ void MKLDNNTester::checkForward() {
 }
 
 void MKLDNNTester::checkBackwardData() {
-  VLOG(MKLDNN_ALL) << "Check Backward Data";
+  VLOG(MKLDNN_TESTS) << "Check Backward Data";
   // TODO(TJ): uncomment me when batch norm ready
   // const bool isBN = dnnLayer_->getType() == "mkldnn_batch_norm";
   for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) {
     const MatrixPtr& dnnDiff = dataLayers_[DNN][i]->getOutputGrad();
     const MatrixPtr& refDiff = dataLayers_[REF][i]->getOutputGrad();
-    VLOG(lvl_) << "Mkldnn Backward Output BotDiff " << i;
+    VLOG(MKLDNN_ALL) << "MKLDNN Backward Result: InputGrad " << i;
     printMatrix(dnnDiff);
-    VLOG(lvl_) << "Reference Backward Output BotDiff " << i;
+    VLOG(MKLDNN_ALL) << "Reference Backward Result: InputGrad " << i;
     printMatrix(refDiff);
 
     double delta = compareMatrix(dnnDiff, refDiff);
@@ -153,7 +153,7 @@ void MKLDNNTester::checkBackwardData() {
 }
 
 void MKLDNNTester::checkBackwardWgts() {
-  VLOG(MKLDNN_ALL) << "Check Backward Weight";
+  VLOG(MKLDNN_TESTS) << "Check Backward Weight";
   CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
   vector<VectorPtr> dnnWgts;  // used to temply save mkldnn weights
   saveWgt(parameters_[DNN], dnnWgts);
@@ -165,9 +165,11 @@ void MKLDNNTester::checkBackwardWgts() {
   for (size_t i = 0; i < parameters_[DNN].size(); ++i) {
     const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE);
     const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE);
-    VLOG(lvl_) << "Mkldnn Output weight " << parameters_[DNN][i]->getName();
+    VLOG(MKLDNN_ALL) << "MKLDNN Result: weight value"
+                     << parameters_[DNN][i]->getName();
     printVector(dnn);
-    VLOG(lvl_) << "Reference Output weight " << parameters_[REF][i]->getName();
+    VLOG(MKLDNN_ALL) << "Reference Result: weight value "
+                     << parameters_[REF][i]->getName();
     printVector(ref);
 
     double delta = compareVector(dnn, ref);
@@ -240,7 +242,8 @@ void MKLDNNTester::printTopDatas() {
   }
 
   for (int n = 0; n < NUM; ++n) {
-    VLOG(lvl_) << testLayers_[n]->getType() << " forward output TopData: ";
+    VLOG(MKLDNN_ALL) << testLayers_[n]->getType()
+                     << " Forward Result: OutputValue";
     printMatrix(testLayers_[n]->getOutputValue());
   }
 }
@@ -252,7 +255,7 @@ void MKLDNNTester::printMatrix(const MatrixPtr& m) {
 
   std::ostringstream ostr;
   m->print(ostr);
-  VLOG(lvl_) << std::endl << ostr.str();
+  VLOG(MKLDNN_ALL) << std::endl << ostr.str();
 }
 
 void MKLDNNTester::printVector(const VectorPtr& v) {
@@ -262,7 +265,7 @@ void MKLDNNTester::printVector(const VectorPtr& v) {
 
   std::ostringstream ostr;
   v->print(ostr, v->getSize());
-  VLOG(lvl_) << std::endl << ostr.str();
+  VLOG(MKLDNN_ALL) << std::endl << ostr.str();
 }
 
 double MKLDNNTester::getDelta(const real* d1,
@@ -314,7 +317,7 @@ void MKLDNNTester::runOnce() {
   UpdateCallback updateCallback = [](Parameter* para) {
     auto& grad = para->getBuf(PARAMETER_GRADIENT);
     auto& value = para->getBuf(PARAMETER_VALUE);
-    real lr = 1e-3;
+    real lr = 1e-2;
     value->add(*grad, lr);
     grad->zeroMem();
   };
@@ -340,10 +343,9 @@ void MKLDNNTester::run(const TestConfig& dnn,
                        size_t batchSize,
                        size_t inputImgH,
                        size_t inputImgW,
+                       bool printDetails,
                        size_t iter,
-                       float epsilon,
-                       bool log,
-                       int level) {
+                       float epsilon) {
   CHECK(dnn.layerConfig.type().compare(0, 7, "mkldnn_") == 0 ||
         dnn.layerConfig.active_type().compare(0, 7, "mkldnn_") == 0)
       << "should be MKLDNN layer or MKLDNN activation";
@@ -359,10 +361,9 @@ void MKLDNNTester::run(const TestConfig& dnn,
 
   ih_ = inputImgH;
   iw_ = inputImgW;
+  log_ = printDetails;
   iter_ = iter;
   eps_ = epsilon;
-  log_ = log;
-  lvl_ = level;
 
   // Firstly test mkldnn init from PARAM_FORMAT_ORIGINAL weight
   reset(dnn, ref, batchSize);
@@ -531,9 +532,11 @@ void MKLDNNTester::getOutResult(const std::string& configPath,
 void MKLDNNTester::compareResult(DataOut& ref, DataOut& dnn, float eps) {
   CHECK_EQ(ref.outValues.size(), dnn.outValues.size());
   CHECK_EQ(ref.paraValues.size(), dnn.paraValues.size());
+  VLOG(MKLDNN_TESTS) << "compare value size: " << ref.outValues.size();
   for (size_t i = 0; i < ref.outValues.size(); i++) {
     EXPECT_LE(fabs(compareMatrix(ref.outValues[i], dnn.outValues[i])), eps);
   }
+  VLOG(MKLDNN_TESTS) << "compare param size: " << ref.outValues.size();
   for (size_t i = 0; i < ref.paraValues.size(); i++) {
     EXPECT_LE(fabs(compareVector(ref.paraValues[i], dnn.paraValues[i])), eps);
   }
@@ -544,9 +547,10 @@ void MKLDNNTester::runBranchesTest(const std::string& configPath,
                                    float eps) {
   DataIn in;
   initArgument(in, configPath, iter);
-
   DataOut outCpu, outDnn;
+  VLOG(MKLDNN_TESTS) << "runing cpu network";
   getOutResult(configPath, in, outCpu, false, iter);
+  VLOG(MKLDNN_TESTS) << "runing mkldnn network";
   getOutResult(configPath, in, outDnn, true, iter);
 
   compareResult(outCpu, outDnn, eps);

paddle/gserver/tests/MKLDNNTester.h

@@ -58,8 +58,6 @@ protected:
   size_t iter_;
   /// whether to print out the details
   bool log_;
-  /// vlog level to print the matrix details datas
-  int lvl_;
   /// epsilon
   float eps_;
   /// input image size, default 1
@@ -70,7 +68,6 @@ public:
     iter_ = iter;
     eps_ = epsilon;
     log_ = false;
-    lvl_ = MKLDNN_ALL;
   }
 
   ~MKLDNNTester() {}
@@ -81,10 +78,9 @@ public:
            size_t batchSize,
            size_t inputImgH = 1,
            size_t inputImgW = 1,
+           bool printDetails = false,
            size_t iter = 3,
-           float epsilon = 1e-4,
-           bool log = false,
-           int level = MKLDNN_ALL);
+           float epsilon = 1e-4);
   static void runBranchesTest(const std::string& configPath,
                               size_t iter = 3,
                               float eps = 1e-4);

paddle/gserver/tests/mkldnn_branches_fc.conf

+# Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from paddle.trainer_config_helpers import *
+
+settings(batch_size=16)
+channels = get_config_arg("channels", int, 2)
+
+def two_fc(input, group_name):
+  out1 = fc_layer(input=input,
+            name=group_name+'_fc1',
+            size=channels,
+            bias_attr=False,
+            act=LinearActivation())
+
+  out2 = fc_layer(input=input,
+            name=group_name+'_fc2',
+            size=channels,
+            bias_attr=False,
+            act=LinearActivation())
+  return out1, out2
+
+data = data_layer(name ="input", size=channels*16*16)
+
+conv = img_conv_layer(input=data,
+            num_channels=channels,
+            filter_size=3,
+            num_filters=channels,
+            padding=1,
+            shared_biases=True,
+            act=LinearActivation())
+
+pool = img_pool_layer(input=conv,
+            pool_size=3,
+            stride=2,
+            padding=1,
+            pool_type=AvgPooling())
+
+a1, a2 = two_fc(input=pool, group_name='a')
+
+concat = concat_layer(input=[a1, a2])
+
+b1, b2 = two_fc(input=pool, group_name='b')
+
+addto = addto_layer(input=[b1, b2])
+
+outputs([concat, addto])

paddle/gserver/tests/mkldnn_branches_pool.conf

+# Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from paddle.trainer_config_helpers import *
+
+settings(batch_size=16)
+channels = get_config_arg("channels", int, 2)
+
+def two_pool(input, group_name):
+  out1 = img_pool_layer(input=input,
+            name=group_name+'_pool1',
+            pool_size=3,
+            stride=2,
+            padding=0,
+            pool_type=MaxPooling())
+
+  out2 = img_pool_layer(input=input,
+            name=group_name+'_pool2',
+            pool_size=5,
+            stride=2,
+            padding=1,
+            pool_type=MaxPooling())
+  return out1, out2
+
+data = data_layer(name ="input", size=channels*16*16)
+
+conv = img_conv_layer(input=data,
+            num_channels=channels,
+            filter_size=3,
+            num_filters=channels,
+            padding=1,
+            shared_biases=True,
+            act=LinearActivation())
+
+pool = img_pool_layer(input=conv,
+            pool_size=3,
+            stride=1,
+            padding=1,
+            pool_type=AvgPooling())
+
+a1, a2 = two_pool(input=pool, group_name='a')
+
+concat = concat_layer(input=[a1, a2])
+
+b1, b2 = two_pool(input=pool, group_name='b')
+
+addto = addto_layer(input=[b1, b2])
+
+outputs([concat, addto])

paddle/gserver/tests/test_MKLDNN.cpp

@@ -250,7 +250,7 @@ TEST(MKLDNNActivation, Activations) {
 
 DECLARE_string(config_args);
 TEST(MKLDNNLayer, branches) {
-  std::vector<std::string> cases = {"conv"};
+  std::vector<std::string> cases = {"conv", "pool", "fc"};
   for (auto name : cases) {
     std::string config = "./gserver/tests/mkldnn_branches_" + name + ".conf";
     for (auto channels : {2, 32}) {

paddle/gserver/tests/test_PyDataProvider2.py

@@ -51,7 +51,10 @@ def test_sparse_non_value_no_seq(setting, filename):
         yield [(i + 1) * (j + 1) for j in xrange(10)]
 
 
-@provider(input_types=[sparse_vector(30000, seq_type=SequenceType.NO_SEQUENCE)])
+@provider(input_types=[
+    sparse_float_vector(
+        30000, seq_type=SequenceType.NO_SEQUENCE)
+])
 def test_sparse_value_no_seq(setting, filename):
     for i in xrange(200):
         yield [((i + 1) * (j + 1), float(j) / float(i + 1)) for j in xrange(10)]

paddle/math/MKLDNNMatrix.cpp

@@ -18,7 +18,7 @@ using namespace mkldnn;  // NOLINT
 
 namespace paddle {
 
-MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
+MKLDNNMatrixPtr MKLDNNMatrix::create(memory::primitive_desc pd, MatrixPtr m) {
   memory::desc md = pd.desc();
   size_t ndims = md.data.ndims;
   int* dims = md.data.dims;
@@ -41,12 +41,12 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
   return std::make_shared<MKLDNNMatrix>(cpuMatrix, pd);
 }
 
-MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
-                                     memory::dims dims,
+MKLDNNMatrixPtr MKLDNNMatrix::create(memory::dims dims,
                                      memory::format fmt,
                                      engine& eg,
+                                     MatrixPtr m,
                                      mkldnn::memory::data_type dtype) {
-  return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg));
+  return create(createPrimitiveDesc(dims, fmt, eg, dtype), m);
 }
 
 std::shared_ptr<reorder> MKLDNNMatrix::createReorder(const MKLDNNMatrixPtr& src,

paddle/math/MKLDNNMatrix.h

@@ -40,24 +40,37 @@ public:
   /**
    * Create MKLDNNMatrix from a MatrixPtr and memory primitive_desc
    */
-  static MKLDNNMatrixPtr create(MatrixPtr m, mkldnn::memory::primitive_desc pd);
+  static MKLDNNMatrixPtr create(mkldnn::memory::primitive_desc pd,
+                                MatrixPtr m = nullptr);
 
   /**
    * Create MKLDNNMatrix from a MatrixPtr and memory details info
    */
   static MKLDNNMatrixPtr create(
-      MatrixPtr m,
       mkldnn::memory::dims dims,
       mkldnn::memory::format fmt,
       mkldnn::engine& eg,
+      MatrixPtr m = nullptr,
       mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32);
 
+  /**
+   * Create primitive descriptor.
+   * default with f32 dtype
+   */
+  static mkldnn::memory::primitive_desc createPrimitiveDesc(
+      const mkldnn::memory::dims dims,
+      const mkldnn::memory::format& fmt,
+      const mkldnn::engine& eg,
+      const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) {
+    return mkldnn::memory::primitive_desc(memory::desc(dims, dtype, fmt), eg);
+  }
+
   /**
    * Create Memory descriptor.
    * default with any format and f32 dtype
    */
   static mkldnn::memory::desc createMemoryDesc(
-      const mkldnn::memory::dims& dims,
+      const mkldnn::memory::dims dims,
       const mkldnn::memory::format& fmt = mkldnn::memory::format::any,
       const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) {
     return mkldnn::memory::desc(dims, dtype, fmt);

paddle/operators/accuracy_op.cc

@@ -69,5 +69,8 @@ information, or not. But the output only shares the LoD with input `Inference`.
 
 namespace ops = paddle::operators;
 REGISTER_OP_WITHOUT_GRADIENT(accuracy, ops::AccuracyOp, ops::AccuracyOpMaker);
-REGISTER_OP_CPU_KERNEL(accuracy,
-                       ops::AccuracyKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    accuracy, ops::AccuracyKernel<paddle::platform::CPUPlace, float>,
+    ops::AccuracyKernel<paddle::platform::CPUPlace, int>,
+    ops::AccuracyKernel<paddle::platform::CPUPlace, double>,
+    ops::AccuracyKernel<paddle::platform::CPUPlace, int64_t>);

paddle/operators/accuracy_op.cu

@@ -21,9 +21,9 @@ namespace paddle {
 namespace operators {
 using platform::PADDLE_CUDA_NUM_THREADS;
 
-template <int BlockSize>
-__global__ void AccuracyCudaKernel(const int N, const int D, const int* Xdata,
-                                   const int* labeldata, float* accuracy) {
+template <typename T, int BlockSize>
+__global__ void AccuracyCudaKernel(const int N, const int D, const T* Xdata,
+                                   const T* labeldata, float* accuracy) {
   int count = 0;
   __shared__ int total[BlockSize];
 
@@ -57,8 +57,8 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
     auto* accuracy = ctx.Output<Tensor>("Accuracy");
     // FIXME(typhoonzero): only support indices currently
     // if add support for output values, how to detect the data type?
-    const int* inference_data = inference->data<int>();
-    const int* label_data = label->data<int>();
+    const T* inference_data = inference->data<T>();
+    const T* label_data = label->data<T>();
     float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
 
     size_t num_samples = inference->dims()[0];
@@ -69,7 +69,7 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
       return;
     }
 
-    AccuracyCudaKernel<PADDLE_CUDA_NUM_THREADS><<<
+    AccuracyCudaKernel<T, PADDLE_CUDA_NUM_THREADS><<<
         1, PADDLE_CUDA_NUM_THREADS, 0,
         reinterpret_cast<const platform::CUDADeviceContext&>(
             ctx.device_context())
@@ -81,5 +81,7 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_GPU_KERNEL(accuracy,
-                       paddle::operators::AccuracyOpCUDAKernel<float>);
+REGISTER_OP_GPU_KERNEL(accuracy, paddle::operators::AccuracyOpCUDAKernel<float>,
+                       paddle::operators::AccuracyOpCUDAKernel<double>,
+                       paddle::operators::AccuracyOpCUDAKernel<int>,
+                       paddle::operators::AccuracyOpCUDAKernel<int64_t>);

paddle/operators/adam_op.cc

@@ -43,10 +43,6 @@ class AdamOp : public framework::OperatorWithKernel {
                    "Output(Moment1Out) of AdamOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Moment2Out"),
                    "Output(Moment2Out) of AdamOp should not be null.");
-    PADDLE_ENFORCE(ctx->HasOutput("Beta1PowOut"),
-                   "Output(Beta1PowOut) of AdamOp should not be null.");
-    PADDLE_ENFORCE(ctx->HasOutput("Beta2PowOut"),
-                   "Output(Beta2PowOut) of AdamOp should not be null.");
 
     auto lr_dims = ctx->GetInputDim("LearningRate");
     PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1,
@@ -72,8 +68,6 @@ class AdamOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("ParamOut", param_dims);
     ctx->SetOutputDim("Moment1Out", param_dims);
     ctx->SetOutputDim("Moment2Out", param_dims);
-    ctx->SetOutputDim("Beta1PowOut", beta1_pow_dims);
-    ctx->SetOutputDim("Beta2PowOut", beta2_pow_dims);
   }
 };
 
@@ -92,8 +86,6 @@ class AdamOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("ParamOut", "(Tensor) Output parameter");
     AddOutput("Moment1Out", "(Tensor) Output first moment");
     AddOutput("Moment2Out", "(Tensor) Output second moment");
-    AddOutput("Beta1PowOut", "(Tensor) Output beta1 power accumulator");
-    AddOutput("Beta2PowOut", "(Tensor) Output beta2 power accumulator");
 
     AddAttr<float>("beta1",
                    "(float, default 0.9) "
@@ -121,10 +113,8 @@ Adam updates:
 
 moment1_out = beta1 * moment1 + (1 − beta1) * grad
 moment2_out = beta2 * moment2 + (1 − beta2) * grad * grad
-beta1_pow_out = beta1_pow * beta1
-beta2_pow_out = beta2_pow * beta2
 learning_rate_t = learning_rate_t *
-                  sqrt(1 - beta2_pow_out) / (1 - beta1_pow_out)
+                  sqrt(1 - beta2_pow) / (1 - beta1_pow)
 param_out = param - learning_rate_t * moment1/ (sqrt(moment2) + epsilon)
 
 References:

paddle/operators/adam_op.h

@@ -26,14 +26,10 @@ class AdamOpKernel : public framework::OpKernel<T> {
     auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
     auto moment1_out_tensor = ctx.Output<framework::Tensor>("Moment1Out");
     auto moment2_out_tensor = ctx.Output<framework::Tensor>("Moment2Out");
-    auto beta1_pow_out_tensor = ctx.Output<framework::Tensor>("Beta1PowOut");
-    auto beta2_pow_out_tensor = ctx.Output<framework::Tensor>("Beta2PowOut");
 
     param_out_tensor->mutable_data<T>(ctx.GetPlace());
     moment1_out_tensor->mutable_data<T>(ctx.GetPlace());
     moment2_out_tensor->mutable_data<T>(ctx.GetPlace());
-    beta1_pow_out_tensor->mutable_data<T>(ctx.GetPlace());
-    beta2_pow_out_tensor->mutable_data<T>(ctx.GetPlace());
 
     float beta1 = ctx.Attr<float>("beta1");
     float beta2 = ctx.Attr<float>("beta2");
@@ -56,18 +52,13 @@ class AdamOpKernel : public framework::OpKernel<T> {
     auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
     auto moment1_out = framework::EigenVector<T>::Flatten(*moment1_out_tensor);
     auto moment2_out = framework::EigenVector<T>::Flatten(*moment2_out_tensor);
-    auto beta1_pow_out =
-        framework::EigenVector<T>::Flatten(*beta1_pow_out_tensor);
-    auto beta2_pow_out =
-        framework::EigenVector<T>::Flatten(*beta2_pow_out_tensor);
     auto place = ctx.GetEigenDevice<Place>();
 
     moment1_out.device(place) = beta1 * moment1 + (1 - beta1) * grad;
     moment2_out.device(place) = beta2 * moment2 + (1 - beta2) * grad.square();
-    beta1_pow_out.device(place) = beta1_pow * beta1;
-    beta2_pow_out.device(place) = beta2_pow * beta2;
+
     // All of these are tensors of 1 element
-    auto lr_t = lr * (1 - beta2_pow_out).sqrt() / (1 - beta1_pow_out);
+    auto lr_t = lr * (1 - beta2_pow).sqrt() / (1 - beta1_pow);
     // Eigen does not support automatic broadcast
     // Get dimensions of moment vector to broadcast lr_t
     Eigen::DSizes<int, 1> m_dsize(moment1_out_tensor->numel());

paddle/operators/adamax_op.cc

@@ -41,8 +41,6 @@ class AdamaxOp : public framework::OperatorWithKernel {
                    "Output(MomentOut) of AdamaxOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("InfNormOut"),
                    "Output(InfNormOut) of AdamaxOp should not be null.");
-    PADDLE_ENFORCE(ctx->HasOutput("Beta1PowOut"),
-                   "Output(Beta1PowOut) of AdamaxOp should not be null.");
 
     auto lr_dims = ctx->GetInputDim("LearningRate");
     PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1,
@@ -64,7 +62,6 @@ class AdamaxOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("ParamOut", param_dims);
     ctx->SetOutputDim("MomentOut", param_dims);
     ctx->SetOutputDim("InfNormOut", param_dims);
-    ctx->SetOutputDim("Beta1PowOut", beta1_pow_dims);
   }
 };
 
@@ -86,7 +83,6 @@ class AdamaxOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("InfNormOut",
               "(Tensor) "
               "Output exponentially weighted infinity norm");
-    AddOutput("Beta1PowOut", "(Tensor) Output beta1 power accumulator");
 
     AddAttr<float>("beta1",
                    "(float, default 0.9) "
@@ -113,8 +109,7 @@ Adamax updates:
 
 moment_out = beta1 * moment + (1 - beta1) * grad
 inf_norm_out = max(beta2 * inf_norm + epsilon, abs(grad))
-beta1_pow_out = beta1_pow * beta1
-learning_rate_t = learning_rate/(1 - beta1_pow_out)
+learning_rate_t = learning_rate/(1 - beta1_pow)
 param_out = param - learning_rate_t * moment_out/inf_norm_out
 
 The original paper does not have an epsilon attribute.

paddle/operators/adamax_op.h

@@ -26,12 +26,10 @@ class AdamaxOpKernel : public framework::OpKernel<T> {
     auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
     auto moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");
     auto inf_norm_out_tensor = ctx.Output<framework::Tensor>("InfNormOut");
-    auto beta1_pow_out_tensor = ctx.Output<framework::Tensor>("Beta1PowOut");
 
     param_out_tensor->mutable_data<T>(ctx.GetPlace());
     moment_out_tensor->mutable_data<T>(ctx.GetPlace());
     inf_norm_out_tensor->mutable_data<T>(ctx.GetPlace());
-    beta1_pow_out_tensor->mutable_data<T>(ctx.GetPlace());
 
     float beta1 = ctx.Attr<float>("beta1");
     float beta2 = ctx.Attr<float>("beta2");
@@ -53,15 +51,12 @@ class AdamaxOpKernel : public framework::OpKernel<T> {
     auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
     auto inf_norm_out =
         framework::EigenVector<T>::Flatten(*inf_norm_out_tensor);
-    auto beta1_pow_out =
-        framework::EigenVector<T>::Flatten(*beta1_pow_out_tensor);
     auto place = ctx.GetEigenDevice<Place>();
 
     moment_out.device(place) = beta1 * moment + (1 - beta1) * grad;
     inf_norm_out.device(place) =
         grad.abs().cwiseMax((beta2 * inf_norm) + epsilon);
-    beta1_pow_out.device(place) = beta1_pow * beta1;
-    auto lr_t = lr / (1 - beta1_pow_out);
+    auto lr_t = lr / (1 - beta1_pow);
     Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
     param_out.device(place) =
         param - lr_t.broadcast(m_dsize) * (moment_out / inf_norm_out);

paddle/operators/clip_op.cc

@@ -27,8 +27,8 @@ class ClipOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of ClipOp should not be null.");
     auto x_dims = ctx->GetInputDim("X");
-    auto max = Attr<float>("max");
-    auto min = Attr<float>("min");
+    auto max = ctx->Attrs().Get<float>("max");
+    auto min = ctx->Attrs().Get<float>("min");
     PADDLE_ENFORCE_LT(min, max, "max should be greater than min.");
     ctx->SetOutputDim("Out", x_dims);
     ctx->ShareLoD("X", /*->*/ "Out");

paddle/operators/conv2d_op.h

@@ -108,17 +108,17 @@ class GemmConv2DKernel : public framework::OpKernel<T> {
     int in_step = input_channels / groups;
     int out_step = output_channels / groups;
     for (int i = 0; i < batch_size; i++) {
-      Tensor in_batch = input->Slice<T>(i, i + 1).Resize(input_shape);
-      Tensor out_batch = output->Slice<T>(i, i + 1).Resize(output_matrix_shape);
+      Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
+      Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
       for (int g = 0; g < groups; g++) {
         // im2col
-        Tensor in_slice = in_batch.Slice<T>(g * in_step, (g + 1) * in_step);
+        Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
         im2col(context.device_context(), in_slice, col, strides[0], strides[1],
                paddings[0], paddings[1]);
 
         // gemm
-        Tensor out_slice = out_batch.Slice<T>(g * out_step, (g + 1) * out_step);
-        Tensor filter_slice = filter.Slice<T>(g * out_step, (g + 1) * out_step);
+        Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
+        Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
         math::matmul<Place, T>(context.device_context(), filter_slice, false,
                                col_matrix, false, T(1.0), &out_slice, T(0.0));
       }
@@ -198,22 +198,20 @@ class GemmConvGrad2DKernel : public framework::OpKernel<T> {
 
       for (int i = 0; i < batch_size; i++) {
         Tensor out_grad_batch =
-            output_grad->Slice<T>(i, i + 1).Resize(output_matrix_shape);
-        Tensor in_grad_batch =
-            input_grad->Slice<T>(i, i + 1).Resize(input_shape);
+            output_grad->Slice(i, i + 1).Resize(output_matrix_shape);
+        Tensor in_grad_batch = input_grad->Slice(i, i + 1).Resize(input_shape);
         for (int g = 0; g < groups; g++) {
           // gemm
           Tensor out_grad_slice =
-              out_grad_batch.Slice<T>(g * out_step, (g + 1) * out_step);
-          Tensor filter_slice =
-              filter.Slice<T>(g * out_step, (g + 1) * out_step);
+              out_grad_batch.Slice(g * out_step, (g + 1) * out_step);
+          Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
           math::matmul<Place, T>(context.device_context(), filter_slice, true,
                                  out_grad_slice, false, T(1.0), &col_matrix,
                                  T(0.0));
 
           // col2im
           Tensor in_grad_slice =
-              in_grad_batch.Slice<T>(g * in_step, (g + 1) * in_step);
+              in_grad_batch.Slice(g * in_step, (g + 1) * in_step);
           col2im(context.device_context(), in_grad_slice, col, strides[0],
                  strides[1], paddings[0], paddings[1]);
         }
@@ -229,19 +227,19 @@ class GemmConvGrad2DKernel : public framework::OpKernel<T> {
 
       for (int i = 0; i < batch_size; i++) {
         Tensor out_grad_batch =
-            output_grad->Slice<T>(i, i + 1).Resize(output_matrix_shape);
-        Tensor in_batch = input->Slice<T>(i, i + 1).Resize(input_shape);
+            output_grad->Slice(i, i + 1).Resize(output_matrix_shape);
+        Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
         for (int g = 0; g < groups; g++) {
           // im2col
           Tensor out_grad_slice =
-              out_grad_batch.Slice<T>(g * out_step, (g + 1) * out_step);
-          Tensor in_slice = in_batch.Slice<T>(g * in_step, (g + 1) * in_step);
+              out_grad_batch.Slice(g * out_step, (g + 1) * out_step);
+          Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
           im2col(context.device_context(), in_slice, col, strides[0],
                  strides[1], paddings[0], paddings[1]);
 
           // gemm
           Tensor filter_grad_slice =
-              filter_grad_.Slice<T>(g * out_step, (g + 1) * out_step);
+              filter_grad_.Slice(g * out_step, (g + 1) * out_step);
           math::matmul<Place, T>(context.device_context(), out_grad_slice,
                                  false, col_matrix, true, T(1.0),
                                  &filter_grad_slice, T(1.0));

paddle/operators/dynamic_recurrent_op.cc

@@ -23,6 +23,7 @@ using framework::Scope;
 using framework::TensorArray;
 using framework::LoDTensor;
 using framework::Variable;
+using framework::OperatorBase;
 using framework::DySeqMetaBatch;
 
 namespace detail {
@@ -43,72 +44,72 @@ inline void CreateVariables(Scope& scope,
  * be reordered, but the RNN op should not change the `boot_state` as an input
  * variable's content.
  */
-template <typename T>
-inline void ReorderBootState(const DySeqMetaBatch& metas,
-                             const LoDTensor& boot_state, LoDTensor* tensor,
-                             const platform::Place& dst_place) {
+inline void ReorderInitialState(const DySeqMetaBatch& metas,
+                                const LoDTensor& boot_state, LoDTensor* tensor,
+                                const platform::Place& dst_place) {
   for (size_t seq_id = 0; seq_id < metas.size(); seq_id++) {
-    auto slice = tensor->Slice<T>(seq_id, seq_id + 1);
+    auto slice = tensor->Slice(seq_id, seq_id + 1);
     auto boot_slice =
-        boot_state.Slice<T>(metas[seq_id].ori_idx, metas[seq_id].ori_idx + 1);
+        boot_state.Slice(metas[seq_id].ori_idx, metas[seq_id].ori_idx + 1);
     // TODO(superjom) pass in device context as an argument
-    slice.template CopyFrom<T>(boot_slice, dst_place,
-                               platform::CPUDeviceContext());
+    slice.CopyFrom(boot_slice, dst_place, platform::CPUDeviceContext());
   }
 }
 
-}  // namespace detail
-
-class DynamicRecurrentOpProtoAndCheckerMaker
-    : public framework::OpProtoAndCheckerMaker {
- public:
-  DynamicRecurrentOpProtoAndCheckerMaker(framework::OpProto* proto,
-                                         framework::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    const auto& name = DynamicRecurrentOp::kArgName;
-    // inputs and outputs stored in proto
-    AddInput(name.inlinks,
-             "the inputs that need to be segmented for each step.")
-        .AsDuplicable();
-    AddInput(name.boot_memories, "variables to initialize memories.")
-        .AsDuplicable();
-
-    AddOutput(name.outlinks, "the outputs that need to concated for all steps.")
-        .AsDuplicable();
-    AddOutput(name.step_scopes, "step scopes");
-
-    // Attributes stored in AttributeMap
-    AddAttr<std::vector<std::string>>(name.pre_memories,
-                                      "names of pre-memories");
-    AddAttr<std::vector<std::string>>(name.memories, "names of memories");
-
-    AddComment("This is a RNN operator for varience-length sequences.");
+inline void RestoreInitialState(const DySeqMetaBatch& metas,
+                                const LoDTensor& tensor, LoDTensor* boot_state,
+                                const platform::Place& dst_place) {
+  for (size_t seq_id = 0; seq_id < metas.size(); seq_id++) {
+    auto slice = tensor.Slice(seq_id, seq_id + 1);
+    auto boot_slice =
+        boot_state->Slice(metas[seq_id].ori_idx, metas[seq_id].ori_idx + 1);
+    boot_slice.CopyFrom(slice, dst_place, platform::CPUDeviceContext());
   }
-};
+}
 
-void DynamicRecurrentOp::Run(const Scope& scope,
-                             const platform::DeviceContext& dev_ctx) const {
-  cache_.Init(kArgName, *this, scope, &arg_);
+}  // namespace detail
+
+// Implementation for forward propagation.
+template <>
+void RNNAlgorithm::Run<RNNAlgorithm::ComputeMode::kForward>(
+    const framework::Scope& scope, const framework::OperatorBase& op,
+    const platform::DeviceContext& dev_ctx) {
+  SetComputeMode(ComputeMode::kForward);
+  cache_.Init(kArgNames[mode_], op, scope, &dev_ctx, &arg_);
   SplitInputs();
   CreateScopes();
   WriteStepInputs();
   InitStates();
   WriteStepOutputs();
+  RunSteps();
+  ConcatOutputs();
+}
 
-  // call stepnet in all the time steps
-  for (size_t step = 0; step < cache_.num_steps; step++) {
-    auto& step_scope = cache_.GetScope(step);
-    stepnet_->Run(step_scope, dev_ctx);
+// Implementation for backward propagation.
+template <>
+void RNNAlgorithm::Run<RNNAlgorithm::ComputeMode::kBackward>(
+    const framework::Scope& scope, const framework::OperatorBase& op,
+    const platform::DeviceContext& dev_ctx) {
+  SetComputeMode(ComputeMode::kBackward);
+  cache_.Init(kArgNames[mode_], op, scope, &dev_ctx, &arg_);
+  SplitInputs();
+  WriteStepInputs();
+  InitStates();
+  WriteStepOutputs();
+  RunSteps();
+  // copy boot-states' gradients back.
+  for (const auto& state : arg_.states) {
+    ExportInitialStateGradient(state);
   }
 
   ConcatOutputs();
 }
 
-void DynamicRecurrentOp::SplitInputs() const {
+void RNNAlgorithm::SplitInputs() {
   // TODO(superjom) make level a config
   // TODO(superjom) check all the inputs has the same LoD
   int level = 0;
-  for (const auto& item : cache_.inlinks) {
+  for (const auto& item : cache_.inputs) {
     const auto& var = item.second;
     const auto& tensor = var->Get<LoDTensor>();
     TensorArray& ta = step_inputs_[item.first];
@@ -125,8 +126,8 @@ void DynamicRecurrentOp::SplitInputs() const {
   }
 }
 
-void DynamicRecurrentOp::WriteStepInputs() const {
-  for (const auto& item : cache_.inlinks) {
+void RNNAlgorithm::WriteStepInputs() {
+  for (const auto& item : cache_.inputs) {
     auto ta_it = step_inputs_.find(item.first);
     PADDLE_ENFORCE(ta_it != step_inputs_.end(),
                    "step_inputs_ not compatible with memory set");
@@ -138,20 +139,20 @@ void DynamicRecurrentOp::WriteStepInputs() const {
       if (var == nullptr) {
         var = step_scope.Var(item.first);
       }
-      var->GetMutable<LoDTensor>()->ShareDataWith<value_type>(tensor);
+      var->GetMutable<LoDTensor>()->ShareDataWith(tensor);
     }
   }
 }
 
-void DynamicRecurrentOp::WriteStepOutputs() const {
+void RNNAlgorithm::WriteStepOutputs() {
   // initialize step outputs
-  for (const auto& item : cache_.outlinks) {
+  for (const auto& item : cache_.outputs) {
     step_outputs_.emplace(item.first, TensorArray());
   }
   PADDLE_ENFORCE_GT(step_outputs_.size(), 0UL);
 }
 
-void DynamicRecurrentOp::CreateScopes() const {
+void RNNAlgorithm::CreateScopes() {
   PADDLE_ENFORCE_GT(cache_.num_steps, 0);
   // resize scopes
   size_t num_scopes_need_create = cache_.num_steps - cache_.scopes->size();
@@ -160,19 +161,19 @@ void DynamicRecurrentOp::CreateScopes() const {
   }
 
   // init temporary inputs
-  PADDLE_ENFORCE_NOT_NULL(stepnet_, "stepnet should be set first");
-  std::vector<std::string> memories;
-  std::vector<std::string> pre_memories;
-  std::vector<std::string> stepnet_outputs;
-  std::transform(arg_.memories.begin(), arg_.memories.end(),
-                 std::back_inserter(memories),
-                 [](const rnn::MemoryAttr& m) { return m.var; });
-  std::transform(arg_.memories.begin(), arg_.memories.end(),
-                 std::back_inserter(pre_memories),
-                 [](const rnn::MemoryAttr& m) { return m.pre_var; });
-  for (const auto& item : stepnet_->Outputs()) {
+  PADDLE_ENFORCE_NOT_NULL(step_unit_, "stepnet should be set first");
+  std::vector<std::string> states;
+  std::vector<std::string> ex_states;
+  std::vector<std::string> step_unit_outputs;
+  std::transform(arg_.states.begin(), arg_.states.end(),
+                 std::back_inserter(states),
+                 [](const rnn::StateAttr& m) { return m.var; });
+  std::transform(arg_.states.begin(), arg_.states.end(),
+                 std::back_inserter(ex_states),
+                 [](const rnn::StateAttr& m) { return m.pre_var; });
+  for (const auto& item : step_unit_->Outputs()) {
     for (const auto& var : item.second) {
-      stepnet_outputs.push_back(var);
+      step_unit_outputs.push_back(var);
     }
   }
 
@@ -180,13 +181,13 @@ void DynamicRecurrentOp::CreateScopes() const {
     auto& scope = cache_.GetScope(step);
     detail::CreateVariables(scope, arg_.inlinks);
     detail::CreateVariables(scope, arg_.outlinks);
-    detail::CreateVariables(scope, memories);
-    detail::CreateVariables(scope, pre_memories);
-    detail::CreateVariables(scope, stepnet_outputs);
+    detail::CreateVariables(scope, states);
+    detail::CreateVariables(scope, ex_states);
+    detail::CreateVariables(scope, step_unit_outputs);
   }
 }
 
-void DynamicRecurrentOp::ConcatOutputs() const {
+void RNNAlgorithm::ConcatOutputs() {
   // TODO(superjom) transform this to a config
   int level = 0;
   for (size_t step = 0; step < cache_.num_steps; step++) {
@@ -199,31 +200,45 @@ void DynamicRecurrentOp::ConcatOutputs() const {
       item.second.WriteShared(step, *tensor);
     }
   }
-  // the inlinks' lods should be the same, so randomly get one lod.
+  // the inputs' lods should be the same, so randomly get one lod.
   const auto& some_lod =
       cache_.scope->FindVar(arg_.inlinks.front())->Get<LoDTensor>().lod();
   const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()];
   for (auto& item : step_outputs_) {
     auto tensor = item.second.Pack(level, some_meta, some_lod);
-    auto* output = cache_.outlinks[item.first]->GetMutable<LoDTensor>();
-    const_cast<LoDTensor*>(output)->ShareDataWith<value_type>(tensor);
+    auto* output = cache_.outputs[item.first]->GetMutable<LoDTensor>();
+    const_cast<LoDTensor*>(output)->ShareDataWith(tensor);
+  }
+}
+
+void RNNAlgorithm::RunSteps() {
+  if (IsBackward()) {
+    // call stepnet in all the time steps reversely
+    for (int step = cache_.num_steps - 1; step >= 0; step--) {
+      auto& step_scope = cache_.GetScope(step);
+      step_unit_->Run(step_scope, *cache_.dev_ctx);
+    }
+  } else {
+    for (size_t step = 0; step < cache_.num_steps; step++) {
+      auto& step_scope = cache_.GetScope(step);
+      step_unit_->Run(step_scope, *cache_.dev_ctx);
+    }
   }
 }
 
-void DynamicRecurrentOp::InitStates() const {
+void RNNAlgorithm::InitStates() {
   for (size_t step = 0; step < cache_.num_steps; step++) {
-    for (const auto& memory : arg_.memories) {
-      CreateState(memory, step);
-      LinkState(memory, step);
+    for (const auto& state : arg_.states) {
+      CreateState(state, step);
+      LinkState(state, step);
     }
   }
 }
 
-void DynamicRecurrentOp::CreateState(const rnn::MemoryAttr& memory,
-                                     size_t step) const {
+void RNNAlgorithm::CreateState(const rnn::StateAttr& state_attr, size_t step) {
   auto& scope = cache_.GetScope(step);
-  auto& state = *cache_.GetTensor(scope, memory.var);
-  auto& boot_state = *cache_.GetTensor(*cache_.scope, memory.boot_var);
+  auto& state = *cache_.GetTensor(scope, state_attr.var);
+  auto& boot_state = *cache_.GetTensor(*cache_.scope, state_attr.boot_var);
 
   size_t num_instances =
       step_inputs_[arg_.inlinks.front()].Read(step).dims()[0];
@@ -232,56 +247,79 @@ void DynamicRecurrentOp::CreateState(const rnn::MemoryAttr& memory,
 
   state.Resize(dims);
   state.mutable_data<value_type>(platform::CPUPlace());
-  states_[memory.var].WriteShared(step, state);
+  states_[state_attr.var].WriteShared(step, state);
 }
 
-void DynamicRecurrentOp::LinkState(const rnn::MemoryAttr& memory,
-                                   size_t step) const {
+void RNNAlgorithm::LinkState(const rnn::StateAttr& state, size_t step) {
   auto& scope = cache_.GetScope(step);
-  auto& state_pre = *cache_.GetTensor(scope, memory.pre_var);
+  auto& state_pre = *cache_.GetTensor(scope, state.pre_var);
+
+  // process the first state's boot-state(the 0-step in forward mode or the
+  // last step in backward mode)
+  // Only forward mode need to link the boot-state to the `pre-state` in first
+  // time step. In backward mode, need to copy the gradient of `pre-state` in
+  // first time step to the gradient of `boot-state`.
+  if (step == 0 && IsForward()) {
+    LinkInitialState(state);
+  } else {
+    size_t num_instances =
+        step_inputs_[arg_.inlinks.front()].Read(step).dims()[0];
+    auto* pre_state = cache_.GetTensor(cache_.GetScope(step - 1), state.var);
+    // shink and share from previous state
+    auto shrinked_pre_state = pre_state->Slice(0, num_instances);
+    state_pre.ShareDataWith(shrinked_pre_state);
+  }
+}
 
+void RNNAlgorithm::LinkInitialState(const rnn::StateAttr& state) {
   // all the step_inputs' metas should be the same, just randomly select one
   // and get the dyseq meta.
   const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()];
-  size_t num_instances =
-      step_inputs_[arg_.inlinks.front()].Read(step).dims()[0];
+  auto& scope = cache_.GetScope(0);
+  auto& state_pre = *cache_.GetTensor(scope, state.pre_var);
+  auto* pre_state = cache_.GetTensor(*cache_.scope, state.boot_var);
+  pre_state->mutable_data<float>(platform::CPUPlace());
+  // allocate state
+  state_pre.Resize(pre_state->dims());
+  state_pre.mutable_data<value_type>(platform::CPUPlace());
+  detail::ReorderInitialState(some_meta, *pre_state, &state_pre,
+                              pre_state->place());
+}
 
-  LoDTensor* pre_state{nullptr};
-  if (step == 0) {
-    pre_state = cache_.GetTensor(*cache_.scope, memory.boot_var);
-    pre_state->mutable_data<float>(platform::CPUPlace());
-    // allocate memory
-    state_pre.Resize(pre_state->dims());
-    state_pre.mutable_data<value_type>(platform::CPUPlace());
-    detail::ReorderBootState<value_type>(some_meta, *pre_state, &state_pre,
-                                         pre_state->place());
-  } else {
-    pre_state = cache_.GetTensor(cache_.GetScope(step - 1), memory.var);
-  }
+void RNNAlgorithm::ExportInitialStateGradient(const rnn::StateAttr& state) {
+  // all the step_inputs' metas should be the same, just randomly select one
+  // and get the dyseq meta.
+  const auto& some_meta = dy_seq_metas_[arg_.inlinks.front()];
+  auto& scope = cache_.GetScope(0);
 
-  // shink and share from previous state
-  auto shrinked_pre_state = pre_state->Slice<value_type>(0, num_instances);
-  state_pre.ShareDataWith<value_type>(shrinked_pre_state);
+  auto& state_pre = *cache_.GetTensor(scope, state.pre_var);
+  auto& pre_state = *cache_.GetTensor(*cache_.scope, state.boot_var);
+  pre_state.Resize(state_pre.dims());
+  detail::RestoreInitialState(some_meta, state_pre, &pre_state,
+                              pre_state.place());
 }
 
-void DynamicRecurrentOp::ArgCache::Init(
-    const rnn::ArgumentName& name, const paddle::framework::OperatorBase& op,
-    const paddle::framework::Scope& scope, rnn::Argument* arg) {
+void RNNAlgorithm::ArgCache::Init(const rnn::ArgumentName& name,
+                                  const paddle::framework::OperatorBase& op,
+                                  const paddle::framework::Scope& scope,
+                                  platform::DeviceContext const* dev_ctx,
+                                  rnn::Argument* arg) {
   this->scope = &scope;
   InitArgument(name, op, arg);
   CacheScopes(scope, *arg);
   CacheInlinks(scope, arg->inlinks);
   CacheOutlinks(scope, arg->outlinks);
+  this->dev_ctx = dev_ctx;
 }
 
-void DynamicRecurrentOp::ArgCache::InitArgument(const rnn::ArgumentName& name,
-                                                const OperatorBase& op,
-                                                rnn::Argument* arg) {
+void RNNAlgorithm::ArgCache::InitArgument(const rnn::ArgumentName& name,
+                                          const OperatorBase& op,
+                                          rnn::Argument* arg) {
   rnn::InitArgument(name, arg, op, false /*is_grad*/);
 }
 
-void DynamicRecurrentOp::ArgCache::CacheScopes(const Scope& scope,
-                                               const rnn::Argument& arg) {
+void RNNAlgorithm::ArgCache::CacheScopes(const Scope& scope,
+                                         const rnn::Argument& arg) {
   auto scopes_var = scope.FindVar(arg.step_scopes);
   PADDLE_ENFORCE(scopes_var != nullptr,
                  "the step_scopes output argument [%s] should be created first "
@@ -290,45 +328,85 @@ void DynamicRecurrentOp::ArgCache::CacheScopes(const Scope& scope,
   this->scopes = scopes_var->GetMutable<std::vector<Scope*>>();
 }
 
-void DynamicRecurrentOp::ArgCache::CacheInlinks(
+void RNNAlgorithm::ArgCache::CacheInlinks(
     const Scope& scope, const std::vector<std::string>& names) {
   for (auto name : names) {
     auto* var = GetVariable(scope, name);
-    inlinks[name] = var;
+    inputs[name] = var;
   }
 }
 
-void DynamicRecurrentOp::ArgCache::CacheOutlinks(
+void RNNAlgorithm::ArgCache::CacheOutlinks(
     const Scope& scope, const std::vector<std::string>& names) {
   for (auto name : names) {
     auto* var = GetVariable(scope, name);
-    outlinks[name] = var;
+    outputs[name] = var;
   }
 }
 
-Variable* DynamicRecurrentOp::ArgCache::GetVariable(const Scope& scope,
-                                                    const std::string& name) {
+Variable* RNNAlgorithm::ArgCache::GetVariable(const Scope& scope,
+                                              const std::string& name) {
   auto* var = scope.FindVar(name);
   PADDLE_ENFORCE_NOT_NULL(var, "variable [%s] not exist in scope", name);
   return var;
 }
 
-LoDTensor* DynamicRecurrentOp::ArgCache::GetTensor(
-    const framework::Scope& scope, const std::string& name) {
+LoDTensor* RNNAlgorithm::ArgCache::GetTensor(const framework::Scope& scope,
+                                             const std::string& name) {
   auto* var = GetVariable(scope, name);
   return var->GetMutable<LoDTensor>();
 }
 
-const rnn::ArgumentName DynamicRecurrentOp::kArgName{
-    "step_net", "step_scopes",  "inlinks",      "outlinks",
-    "memories", "pre_memories", "boot_memories"};
+const std::array<rnn::ArgumentName, 2> RNNAlgorithm::kArgNames{
+    {rnn::ArgumentName{"step_unit", "step_scopes", "inputs", "outputs",
+                       "states", "ex_states", "initial_states"},
+     rnn::ArgumentName{"step_unit", "step_scopes@GRAD", "outputs@GRAD",
+                       "inputs@GRAD", "states", "ex_states",
+                       "initial_states@GRAD"}}};
+
+void DynamicRecurrentOp::Run(const framework::Scope& scope,
+                             const platform::DeviceContext& dev_ctx) const {
+  rnn.Run<RNNAlgorithm::ComputeMode::kForward>(
+      scope, *dynamic_cast<const OperatorBase*>(this), dev_ctx);
+}
 
 void DynamicRecurrentGradientOp::Run(
-    const Scope& scope, const platform::DeviceContext& dev_ctx) const {}
+    const Scope& scope, const platform::DeviceContext& dev_ctx) const {
+  rnn.Run<RNNAlgorithm::ComputeMode::kBackward>(
+      scope, *dynamic_cast<const OperatorBase*>(this), dev_ctx);
+}
+
+class DynamicRecurrentOpProtoAndCheckerMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  DynamicRecurrentOpProtoAndCheckerMaker(framework::OpProto* proto,
+                                         framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    const auto& name =
+        RNNAlgorithm::kArgNames[RNNAlgorithm::ComputeMode::kForward];
+    // inputs and outputs stored in proto
+    AddInput(name.inlinks,
+             "the inputs that need to be segmented for each step.")
+        .AsDuplicable();
+    AddInput(name.initial_states, "variables to initialize states.")
+        .AsDuplicable();
+
+    AddOutput(name.outlinks, "the outputs that need to concated for all steps.")
+        .AsDuplicable();
+    AddOutput(name.step_scopes, "step scopes");
+
+    // Attributes stored in AttributeMap
+    AddAttr<std::vector<std::string>>(name.ex_states, "names of ex_states");
+    AddAttr<std::vector<std::string>>(name.states, "names of states");
+
+    AddComment("This is a RNN operator for varience-length sequences.");
+  }
+};
 
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_WITHOUT_GRADIENT(
-    dynamic_recurrent, paddle::operators::DynamicRecurrentOp,
-    paddle::operators::DynamicRecurrentOpProtoAndCheckerMaker);
+REGISTER_OP(dynamic_recurrent, paddle::operators::DynamicRecurrentOp,
+            paddle::operators::DynamicRecurrentOpProtoAndCheckerMaker,
+            dynamic_recurrent_grad,
+            paddle::operators::DynamicRecurrentGradientOp);

paddle/operators/elementwise_op_function.h

@@ -108,7 +108,7 @@ void ElementwiseCompute(const framework::ExecutionContext& ctx) {
   PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
                     "Rank of first input must >= rank of second input.")
 
-  if (x_dims == y_dims || product(y_dims) == 1) {
+  if (x_dims == y_dims) {
     functor f;
     f.template Run<Place, T>(x, y, z, ctx);
     return;
@@ -174,12 +174,6 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
     return;
   }
 
-  if (product(y_dims) == 1) {
-    functor1 f;
-    f(place, x, y, out, dx, dy, dout);
-    return;
-  }
-
   int axis = ctx.Attr<int>("axis");
   axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
 

paddle/operators/gaussian_random_op.cc

@@ -59,7 +59,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
  protected:
   framework::DataType IndicateDataType(
       const framework::ExecutionContext& ctx) const override {
-    return static_cast<framework::DataType>(Attr<int>("data_type"));
+    return static_cast<framework::DataType>(ctx.Attr<int>("data_type"));
   }
 };