Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix-7211

CMakeLists.txt

@@ -20,8 +20,10 @@ set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
 include(system)
 
 project(paddle CXX C Go)
-message(STATUS "CXX compiler: " ${CMAKE_CXX_COMPILER} ", version: " ${CMAKE_CXX_COMPILER_VERSION})
-message(STATUS "C compiler: " ${CMAKE_C_COMPILER} ", version: " ${CMAKE_C_COMPILER_VERSION})
+message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
+        "${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
+message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
+        "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
 
 find_package(Sphinx)
 if(NOT CMAKE_CROSSCOMPILING)

cmake/external/eigen.cmake

@@ -19,7 +19,7 @@ ExternalProject_Add(
 
 if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
     set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/eigen3_dummy.c)
-    file(WRITE ${dummyfile} "const char * dummy_eigen3 = \"${dummyfile}\";")
+    file(WRITE ${dummyfile} "const char *dummy_eigen3 = \"${dummyfile}\";")
     add_library(eigen3 STATIC ${dummyfile})
 else()
     add_library(eigen3 INTERFACE)

cmake/external/mkldnn.cmake

@@ -63,9 +63,30 @@ ExternalProject_Add(
                         -DMKLROOT:PATH=${MKLML_ROOT}
 )
 
-ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
-SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
-ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
+ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
+SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
+ADD_DEPENDENCIES(shared_mkldnn ${MKLDNN_PROJECT})
 MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
 add_definitions(-DPADDLE_WITH_MKLDNN)
-LIST(APPEND external_project_dependencies mkldnn)
+LIST(APPEND external_project_dependencies shared_mkldnn)
+
+# generate a static dummy target to track mkldnn dependencies
+# for cc_library(xxx SRCS xxx.c DEPS mkldnn)
+SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/mkldnn_dummy.c)
+FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
+ADD_LIBRARY(mkldnn STATIC ${dummyfile})
+TARGET_LINK_LIBRARIES(mkldnn ${MKLDNN_LIB} ${MKLML_LIB} ${MKLML_IOMP_LIB})
+ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
+
+# copy the real so.0 lib to install dir
+# it can be directly contained in wheel or capi
+SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
+ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}
+    COMMAND cp ${MKLDNN_LIB} ${MKLDNN_SHARED_LIB}
+    DEPENDS mkldnn)
+ADD_CUSTOM_TARGET(mkldnn_shared_lib ALL DEPENDS ${MKLDNN_SHARED_LIB})
+
+IF(WITH_C_API)
+  INSTALL(FILES ${MKLDNN_SHARED_LIB} DESTINATION lib)
+ENDIF()
+

cmake/external/mklml.cmake

@@ -66,3 +66,7 @@ ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
 ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
 LIST(APPEND external_project_dependencies mklml)
+
+IF(WITH_C_API)
+  INSTALL(FILES ${MKLML_LIB} ${MKLML_IOMP_LIB} DESTINATION lib)
+ENDIF()

cmake/external/openblas.cmake

@@ -30,23 +30,21 @@ IF(NOT ${CBLAS_FOUND})
         CACHE FILEPATH "openblas library." FORCE)
 
     SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
+    SET(OPENBLAS_COMMIT "v0.2.20")
 
     IF(CMAKE_CROSSCOMPILING)
         SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER})
         GET_FILENAME_COMPONENT(CROSS_SUFFIX ${CMAKE_C_COMPILER} DIRECTORY)
         SET(CROSS_SUFFIX ${CROSS_SUFFIX}/)
         IF(ANDROID)
-            # arm_soft_fp_abi branch of OpenBLAS to support softfp
-            #   https://github.com/xianyi/OpenBLAS/tree/arm_soft_fp_abi
-            SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5")
             IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
+                # use softfp
                 SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0)
             ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
                 SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0)
             ENDIF()
         ELSEIF(IOS)
             IF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
-                SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5")
                 SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}")
                 SET(OPENBLAS_CC "${OPENBLAS_CC} -arch arm64")
                 SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0 CROSS_SUFFIX=${CROSS_SUFFIX})
@@ -56,14 +54,12 @@ IF(NOT ${CBLAS_FOUND})
             ENDIF()
         ELSEIF(RPI)
             # use hardfp
-            SET(OPENBLAS_COMMIT "v0.2.20")
             SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 USE_THREAD=0)
         ENDIF()
     ELSE()
         IF(APPLE)
             SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
         ENDIF()
-        SET(OPENBLAS_COMMIT "v0.2.20")
         SET(OPTIONAL_ARGS "")
         IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$")
             SET(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64)
@@ -113,7 +109,7 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
 # FIXME(gangliao): generate cblas target to track all high performance
 # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
 SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
-FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
+FILE(WRITE ${dummyfile} "const char *dummy_cblas = \"${dummyfile}\";")
 ADD_LIBRARY(cblas STATIC ${dummyfile})
 TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
 

cmake/generic.cmake

@@ -120,7 +120,7 @@ function(merge_static_libs TARGET_NAME)
       DEPENDS ${libs})
 
     # Generate dummy staic lib
-    file(WRITE ${target_SRCS} "const char *dummy = \"${target_SRCS}\";")
+    file(WRITE ${target_SRCS} "const char *dummy_${TARGET_NAME} = \"${target_SRCS}\";")
     add_library(${TARGET_NAME} STATIC ${target_SRCS})
     target_link_libraries(${TARGET_NAME} ${libs_deps})
 
@@ -160,7 +160,7 @@ function(merge_static_libs TARGET_NAME)
       DEPENDS ${libs} ${target_OBJS})
 
     # Generate dummy staic lib
-    file(WRITE ${target_SRCS} "const char *dummy = \"${target_SRCS}\";")
+    file(WRITE ${target_SRCS} "const char *dummy_${TARGET_NAME} = \"${target_SRCS}\";")
     add_library(${TARGET_NAME} STATIC ${target_SRCS})
     target_link_libraries(${TARGET_NAME} ${libs_deps})
 
@@ -324,7 +324,7 @@ function(go_library TARGET_NAME)
     )
 
   # Add dummy code to support `make target_name` under Terminal Command
-  file(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
+  file(WRITE ${dummyfile} "const char *dummy_${TARGET_NAME} = \"${dummyfile}\";")
   if (go_library_SHARED OR go_library_shared)
     add_library(${TARGET_NAME} SHARED ${dummyfile})
   else()

doc/api/v2/fluid/layers.rst

@@ -307,6 +307,12 @@ sequence_expand
     :noindex:
 
 
+gru_unit
+--------
+..  autofunction:: paddle.v2.fluid.layers.gru_unit
+    :noindex:
+
+
 lstm_unit
 ---------
 ..  autofunction:: paddle.v2.fluid.layers.lstm_unit

doc/design/backward.md

@@ -106,9 +106,11 @@ See function `_addup_repetitive_outputs_` in `backward.py` for implementation de
 
 In our framework, variables can be marked as *no_gradient*, it means that the gradient of this variable is unnecessary and can be considered as zero in model training. Apparently, when all the outputs of some `grad_op` are marked as *no_gradient*, the `grad_op` itself can be skipped in backward pass. 
 
-But these unnecessary gradients still need to be creating and initialized by something, otherwise following `grad_op`s who take these gradients as inputs take the risk of using uninitialized memory. In our code, we employ `fill_zeros_like_op` to initialize them as all zeros. 
+Another situation is all the gradient inputs of some `grad_op` are marked as *no_gradient*, which means all of them can be considered as zeros. For `grad_op`s are in essence the propagation of gradients, all the outputs are definitely zeros when all gradient inputs are zeros. Therefore the `grad_op` can also be skipped.
 
-This features are implemented in function `_remove_no_grad_branch_`. It checks new created `grad_op`s one-by-one, removes whose outputs are all in `no_grad_set` or inserts `fill_zeros_like_op` when its necessary. We can get the `no_grad_set` from the `_append_backward_ops_` argument `no_grad_dict` or generate it on the fly by scanning all variables' `no_gradient` attribute(True or False). 
+It should be noted that all these zero gradients still need to be creating and initialized by something, otherwise following `grad_op`s who take these gradients as inputs take the risk of using uninitialized memory. In our code, we employ `fill_zeros_like_op` to initialize them as all zeros. 
+
+This features are implemented in function `_remove_no_grad_branch_`. It checks new created `grad_op`s one-by-one, removes who can be skipped and inserts `fill_zeros_like_op` when its necessary. We can get the `no_grad_set` from the `_append_backward_ops_` argument `no_grad_dict` or generate it on the fly by scanning all variables' `no_gradient` attribute(True or False). 
 
 ### Creating Backward Variables
 

doc/design/concurrent_programming.md

+# Design Doc: Concurrent Programming with Fluid
+
+With PaddlePaddle Fluid, users describe a program other than a model.  The program is a [`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto) protobuf message. TensorFlow/MxNet/Caffe2 applications generate protobuf messages too, but their protobuf messages represent the model, a graph of operators, but not the program that trains/uses the model.   
+
+Many know that when we program TensorFlow, we can specify the device on which each operator runs.  This allows us to create a concurrent/parallel AI application.   An interesting questions is **how does a `ProgramDesc` represents a concurrent program?**  
+
+The answer relies on the fact that a `ProgramDesc` is similar to an abstract syntax tree (AST) that describes a program.  So users just program a concurrent program that they do with any concurrent programming language, e.g., [Go](https://golang.org).
+
+## An Analogy
+
+The following table compares concepts in Fluid and Go
+
+| Go | Fluid |
+|----|-------|
+|user-defined functions | [layers](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/v2/fluid) |
+| control-flow and built-in functions | [intrinsics/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators) |
+| goroutines, channels | [class ThreadPool](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework/thread_pool.h) |
+| runtime | [class Executor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h) |
+
+## An Example Concurrent Program
+
+To review all above concepts in an example, let us take a simple program and writes its distributed version.
+
+Suppose that we want to parallelize a naive Fluid program (written in Go and calling Fluid's Go binding) that multiplies two tensors.
+
+```go
+import "fluid"
+
+func paddlepaddle() {
+  X = fluid.read(...)
+  W = fluid.Tensor(...)
+  Y = fluid.mult(X, W)
+}
+```
+
+Please be aware that the Fluid's Go binding provides the default `main` function, which calls the `paddlepaddle` function, which, in this case, is defined in above program and creates the following `ProgramDesc` message.
+
+```protobuf
+message ProgramDesc {
+  block[0] = Block {
+    vars = [X, W, Y],
+    ops = [
+      read(output = X)
+      assign(input = ..., output = W)
+      mult(input = {X, W}, output = Y)
+    ],
+  }
+}
+```
+
+Then, the default `main` function calls `fluid.run()`, which creates an instance of the [`class Executor`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h) and calls `Executor.Run(block[0])`, where `block[0]` is the first and only block defined in above `ProgramDesc` message.
+
+The default `main` function is defined as follows:
+
+```go
+func main() {
+  paddlepaddle()
+  fluid.run()
+}
+```
+
+## The Concurrent Version
+
+By parallelizing the above program, we could support very big tensor X by splitting into small pieces {x_1, x_2, ...} and sent each piece to worker process/node for parallel multiplication.
+
+In this case, we can write a transpiler that takes a `ProgramDesc` message that represents the above example program and outputs two `ProgramDesc` messages, one for running on the master process/node, and the other one for worker processes/nodes.
+
+### The Master Program
+
+The master program could look like the following:
+
+```protobuf
+message ProgramDesc {
+  block[0] = Block {
+    vars = [X, L, Y],
+    ops = [
+      read(output = X)
+      kube_get_workers_addrs(output = L)
+      Y = tensor_array(len(L))
+      parallel_for(input = X, output = Y, 
+                   attrs = {L, block_id(1)}) # referring to block 1
+    ]
+  }
+  
+  block[1] = Block {
+    parent = 0,
+    vars = [x, y, index],
+    ops = [
+      slice(input = [X, index], output = x) # index is initialized by parallel_for
+      send(input = x, attrs = L[index])
+      recv(outputs = y, attrs = L[index])
+      assign(input = y, output = Y[index])
+    ]
+  }
+}
+```
+
+The equivalent Fluid program (calling the Go binding) is:
+
+```go
+func main() {  //// block 0
+  X = fluid.read(...)
+  L = fluid.k8s.get_worker_addrs()
+  Y = fluid.tensor_array(len(L))
+  fluid.parallel_for(X, L, 
+                     func(index int) {  //// block 1
+                       x = X[index]
+                       fluid.send(L[index], x)
+                       y = fluid.recv(L[index])
+                       Y[index] = y
+                     })
+}
+```
+
+An explanation of the above program:
+
+- `fluid.k8s` is a package that provides access to Kubernetes API.  
+- `fluid.k8s.get_worker_addrs` returns the list of IP and ports of all pods of the current job except for the current one (the master pod).  
+- `fluid.tensor_array` creates a [tensor array](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor_array.h).  `fluid.parallel_for` creates a `ParallelFor` intrinsic, which, when executed, 
+
+  1. creates `len(L)` scopes, each for the concurrent running of the sub-block (block 1 in this case), and initializes a variable named "index" in the scope to an integer value in the range `[0, len(L)-1]`, and
+  2. creates `len(L)` threads by calling into the `ThreadPool` singleton, each thread  
+     1. creates an Executor instance, and
+     2. calls `Executor.Run(block)`, where `block` is block 1 as explained above.
+1. Please be aware that block 1 is a sub-block of block 0, so ops in block 1 could refer to variables defined in block 0.
+
+### The Worker Program
+
+The worker program looks like
+
+```go
+func main() {
+  W = Tensor(...)
+  x = fluid.listen_and_do(
+        fluid.k8s.self_addr(),
+        func(input Tensor) {
+          output = fluid.mult(input, W)
+        })
+}
+```
+
+where
+
+- `fluid.listen_and_do` creates a `ListenAndDo` intrinsic, which, when executed,
+  1. listens on the current pod's IP address, as returned by `fliud.k8s.self_addr()`,
+  2. once a connection is established,
+     1. creates a scope of two parameters, "input" and "output",
+     2. reads a [Fluid variable](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h) and saves it into "input",
+     3. creates an Executor instance and calls `Executor.Run(block)`, where the block is generated by running the lambda specified as the second parameter of `fluid.listen_and_do`.
+
+## Summarization
+
+From the above example, we see that:
+
+1. Fluid enables the imperative programming paradigm by:
+   1. letting users describe a program, but not a model (a sequence of layers, or a graph of operators), and
+   2. call the `fluid.run` function that runs the program implicitly.
+1. The program is described as a `ProgramDesc` protobuf message.
+2. Function `Executor.Run` takes a block, instead of a `ProgramDesc`, as its parameter.
+3. `fluid.run` calls `Executor.Run` to run the first block in the `ProgramDesc` message.
+4. `Executor.Run`'s implementation is extremely simple -- it doesn't plan the execution nor create threads; instead, it runs on the current thread and execute intrinsics/operators' `Run` method sequentially as they appear in the `Block.ops` array.
+5. Intrinsics/operators' `Run` method might create threads.  For example, the `ListenAndDo` operator creates a thread to handle each incoming request.
+6. Threads are not necessarily OS thread; instead, they could be [green threads](https://en.wikipedia.org/wiki/Green_threads) managed by ThreadPool.  Multiple green threads might run on the same OS thread.  An example green threads is Go's [goroutines](https://tour.golang.org/concurrency/1).

doc/design/memory_optimization.md

+# Memory Optimization
+
+
+## Problem
+
+In a lecture from Andrew Ng, he attributes the recent sucess of AI due to a combination of these:
+
+- availability of Big Data
+- supercomputing power to process this Big Data over very large neural networks
+- modern algorithms
+
+Following graph shows the details:
+
+![](images/deep_learning.png)
+
+Larger model usually brings better performance. However, GPU memory is certain limited. For example, the memory size of a GTX TITAN X is only 12GB. To train complex and large model, we have to take care of memory using. Besides, memory optimization is also necessary in both online/mobile inference. 
+
+## Solution
+
+### Basic Strategy
+
+There are some basic strategies to make memory optimization, including in-place operation and memory sharing.
+
+#### In-place Operation
+In a relu activation operator： 
+
+$y = \max(x, 0)$
+
+If the variable x is not used in any other operator, we can make an in-place operation. In other words, the memory block of variable y and variable x are the same. In-place operation will save 50% memory occupancy immediately.
+
+#### Memory Sharing
+
+Not all operators support in-place operations. Memory sharing is a more general strategy.
+
+Following is an example:
+
+```
+a = op1(b, c);
+d = op2(a)
+e = op3(d, f)
+```
+
+In this case, variable a is no longer used, and op2 does not support in-place operation. After op2 finished, we can put the memory of variable a to a memory pool. Then, variable e can share the memory of variable a from the pool.
+
+
+### Live Variable Analysis
+
+It's not enough to only have some basic strategies. The prerequisite of memory optimization is to know if a variable is still "live" after an operation.
+
+In our design, the neural network topology is defined as a program. Luckily, [live variable analysis](https://en.wikipedia.org/wiki/Live_variable_analysis) is a classic problem in compilers which can be used in many stages, such as register allocation. 
+
+In compilers, the front end of the compilers translates programs into an intermediate language with an unbounded number of temporaries. This program must run on a machine with a bounded number of registers. Two temporaries a and b can fit into the same register, if a and b are never "in use" at the same time. Thus, many temporaries can fit in few registers; if they don't all fit, the excess temporaries can be kept in memory.
+
+Therefore, the compiler needs to analyze the intermediate-representation program to determine which temporaries are in use at the same time. We say a variable is "live" if it holds a value that may be needed in the future, so this analysis is called liveness analysis. 
+
+We can leran these techniques from compilers. There are mainly two stages to make live variable analysis:
+
+- construct a control flow graph
+- solve the dataflow equations
+
+
+#### Control Flow Graph
+To preform analyses on a program, it is often useful to make a control flow graph. A [control flow graph](https://en.wikipedia.org/wiki/Control_flow_graph) (CFG) in computer science is a representation, using graph notation, of all paths that might be traversed through a program during its execution. Each statement in the program is a node in the flow graph; if statemment x can be followed by statement y, there is an egde from x to y.
+
+Following is the flow graph for a simple loop.
+
+![](images/control_flow_graph.png)
+
+#### Dataflow Analysis
+
+liveness of variable "flows" around the edges of the control flow graph; determining the live range of each variable is an example of a dataflow problem. [Dataflow analysis](https://en.wikipedia.org/wiki/Data-flow_analysis) is a technique for gathering information about the possible set of values calculated at various points in a computer program.
+
+A simple way to perform data-flow analysis of programs is to set up dataflow equations for each node of the control flow graph and solve them by repeatedly calculating the output from the input locally at each node until the whole system stabilizes.
+
+- Flow Graph Terminology
+
+A flow graph node has out-edges that lead to sucessor nodes, and in-edges that come from presucessor nodes. The set *pred[n]* is all the predecessors of node n, and *succ[n]* is the set of sucessors.
+In former control flow graph, the out-edges of node 5 are 5 --> 6 and 5 --> 2, and *succ[5]* = {2, 6}. The in-edges of 2 are 5 --> 2 and 1 --> 2, and *pred[2]* = {1, 5}.
+
+- Uses and Defs
+
+An assignmemt to a variable or temporary defines that variable. An occurence of a variable on the right-hand side of an assginment(or in other expressions) uses the variable. We can speak the *def* of a variable as the set of graph nodes that define it; or the *def* of a graph node as the set of variables that it defines; and the similarly for the *use* of a variable or graph node. In former control flow graph, *def(3)* = {c}, *use(3)* = {b, c}.
+
+- Liveness
+
+A variable is *live* on an edge if there is a directed path from that edge to a *use* of the variable that does not go through any *def*. A variable is *live-in* at a node if it is live on any of the in-edges of that node; it is *live-out* at a node if it is live on any of the out-edges of the node.
+
+
+The calcution of liveness can be solved by iteration until a fixed pointer is reached. Following is the recursive formula:
+
+![](images/dataflow_equations.png)
+
+### Memory optimization transpiler
+
+At last, we take basic strategy and liveness analysis techniques learning from compilers to implement our memory optimization transpiler.
+
+#### add in-place attribute
+
+In-place is a built-in attribute of an operator. Since we treat in-place and other operators differently, we have to add an in-place attribute for every operator.
+
+
+#### contruct control flow graph
+
+Following is the ProgramDesc protobuf of [machine translation](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/book/test_machine_translation.py) example.
+
+- Block0:
+
+```
+lookup_table
+mul
+...
+while(sub-block idx 1)
+...
+array_to_lod_tensor
+cross_entropy
+...
+while_grad(sub-block idx 2)
+read_from_array
+array_to_lod_tensor
+...
+```
+
+- Block1
+
+```
+read_from_array
+read_from_array
+...
+write_to_array
+increment
+write_to_array
+less_than
+```
+
+- Block2
+
+```
+read_from_array
+increment
+...
+write_to_array
+write_to_array
+```
+
+We can transfer all the operators and variables in ProgramDesc to build a control flow graph.
+
+```python
+class ControlFlowGraph(object):
+    def __init__(self, Program):
+        self._sucessors = defaultdict(set)
+        self._presucessors = defaultdict(set)
+        self._uses = defaultdict(set)
+        self._defs = defaultdict(set)
+        self._live_in = defaultdict(set)
+        self._live_out = defaultdict(set)
+        self._program = Program
+    
+    def build(self):
+        pass
+    
+    def dataflow_analysis(self):
+        pass
+        
+    def memory_optimization(self):
+        pass
+        
+    def get_program(self):
+        return self._program
+```
+
+#### make dataflow analysis
+
+We follow guide from compilers and try to solve the dataflow equation to get liveness of every variable. If the live-in of an operator node is different from the live-out, then we can make memory sharing. 
+
+For example:
+
+```
+a = op1(b, c);
+d = op2(a)
+e = op3(d, f)
+```
+
+The dataflow analysis result is:
+
+```
+live_in(op1) = {b, c, f}
+live_out(op1) = {a, f}
+
+live_in(op2) = {a, f}
+live_out(op2) = {d, f}
+
+live_in(op3) = {d, f}
+live_out(op3) = {}
+```
+
+After op1, we can process variable b and variable c; After op2, we can process variable a. After op3, we can process variable d and variable f.
+
+#### memory sharing policy
+
+A memory pool will be mantained in the stage of memory optimization. Each operator node will be scanned to determine memory optimization is done or not. If an operator satifies the requirement, following policy will be taken to handle input/output variables.
+
+```
+if op.support_inplace():
+    i --> pool
+    pool --> o
+else:
+    pool --> o
+    i --> pool
+```
+
+
+
+## Reference
+
+- [Lecture Notes From Artificial Intelligence Is The New Electricity By Andrew Ng](https://manavsehgal.com/lecture-notes-from-artificial-intelligence-is-the-new-electricity-by-andrew-ng-4712dcbf26e5)
+- Modern compiler implementation in ML, by Andrew W. Appel
+- [Optimizing Memory Consumption in Deep learning](https://mxnet.incubator.apache.org/architecture/note_memory.html)

doc/design/profiler.md

+## Introduction
+
+There are many performance analysis tools for [different programming languages and different software frameworks](https://en.wikipedia.org/wiki/List_of_performance_analysis_tools). For most popular deep learning frameworks, they use several programming languages and adapt to heterogeneous platforms. Similar to most of the deep learning frameworks, PaddlePaddle also uses C++, CUDA and Python as the basic programming languages to adapt to run on CPU and GPU devices.  The [`nvprof` tools](http://docs.nvidia.com/cuda/profiler-users-guide/index.html#nvprof-overview) is usually used to analyse the CUDA program.  We have [a document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/optimization/cpu_profiling.md) to profile CPU and Python program by [yep](https://pypi.python.org/pypi/yep) and [Google's perftools](https://github.com/google/pprof) to profile only the CPU and Python program. But for [PaddlePaddle fluid](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/fluid.md), the operator is the basic computing unit. The developers usually want to collect the time of each operator and locate bottlenecks.  The `nvprof` usually collect the timeline of CUDA-related activities on both CPU and GPU, including kernel execution, memory transfers, memory set and CUDA API calls and events or metrics for CUDA kernels. And the `yep` and `Google's perftools` can't collect the timeline for CUDA program. All these tools can't collect time in the operator level. So we design this profiling tool.
+
+## Architecture
+
+The work flow for most task is as follows. Each operator will run many times in the all iterations. So the profiler must collect the total time of each operator during the iteration. For more, sometimes, the developers may want to collect more detailed time span inside the operator or record time span for elsewhere, this requires that the profiler must support to record the nested time span. And in order to speedup training, all the deep learning frameworks support parallel computing, including multiple threads on CPU and multiple GPUs. So the profiler must be able to collect the timeline for each thread. In addition, the profiler also occupies certain resources. It must can be easily to be enabled or disabled by the developers. At last, the profiler should present a human-readable report.  
+
+```python
+for i in xrange(M):  # M is  the iteration number
+  for op in operator_lists: # The `operator_lists` contains all the operators in the network.
+    op.run();
+```
+
+In summary, the proflier should have following features:
+
+- records time span in loop.
+- supports nested time span.
+- supports multiple threads/multiple GPUs.
+- supports to be enabled and disabled by users.
+
+But how to record the time for the mixed C++ and CUDA program?  There many C++ APIs to get the current calendar time in host program. But for GPU, the CUDA kernels may be executed concurrently if they are in different [streams](http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#streams) and the CUDA kernels is asynchronous with the host program if there is no the synchronous aftern the CUDA kernels. CUDA provides [event](http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#events) to monitor the device and perform accurate timing. Inspired by PyTorch and CUDA event, we also design and apply the events to record the timeline. Then summarize and present statistics based on these events.  
+
+The overall flow is shown as the following figure.
+
+<img src="./images/profiler.png" align="center"/><br/>
+
+### Event
+
+In above work flow, a pair of events are needed before and after the piece of code to collect time. So the event has a flag to mark whether it is a starting event or an ending event. Except this two kinds of event, sometime, a only marker with a text message is needed, for example, a marker to specify the profiling start or end. There are three kinds of event:
+
+```c++
+enum EventKind {
+  kMark,
+  kPushRange,
+  kPopRange};
+```
+- kMark: only a marker without time range.
+- kPushRange: mark the starting event for time range. 
+- kPopRange: mark the ending event for time range.
+
+For the CPU code, the events only need to record the current time. For the CUDA code, the [event management functions of CUDA](http://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__EVENT.html#group__CUDART__EVENT) are used.  For many pieces of code, an event lists are used to record each piece. 
+
+```c++
+class Event {
+ public:
+  // The DeviceContext is used to get current  CUDA stream.
+  Event(EventKind kind, std::string name, uint32_t thread_id,
+        const platform::DeviceContext* dev_ctx = nullptr);
+  double CpuElapsedUs(const Event& e) const;
+  double CudaElapsedUs(const Event& e) const;
+
+ private:
+  EventKind kind_;
+  std::string name_;
+  uint32_t thread_id_;
+  int64_t cpu_ns_;
+#ifdef PADDLE_WITH_CUDA
+  cudaEvent_t event_ = nullptr;
+  int device_ = -1;
+#endif
+};
+
+struct EventList {
+  std::forward_list<std::vector<Event>> event_blocks;
+};
+```
+
+As mentioned above, there is no need to record the timeline when disabling the profiler. So there is a global state to enable or disable the profiler. 
+
+```c++
+enum ProfilerState {
+  kDisabled, 
+  kCPU,
+  kCUDA
+};
+ProfilerState g_state;
+```
+- kDisabled: the disabled state.
+- kCPU: CPU profiling state.
+- kCUDA: GPU profiling state.
+
+A pair of starting and ending events are pushed to event lists in constructor and destructor of `RecordEvent`. So the timeline is recorded for the code in the lifecycle of an object of `RecordEvent`.
+
+```c++
+struct RecordEvent {
+  explicit RecordEvent(const std::string name,
+                       platform::DeviceContext* dev_ctx = nullptr) {
+    if (kState == ProfilerState::kDisabled) return;
+    // push the starting event to the event lists.
+  }
+  ~RecordEvent() {
+    if (kState == ProfilerState::kDisabled) return;
+    // push the ending event to the event lists.
+  }
+};
+```

doc/design/releasing_process.md

@@ -7,11 +7,9 @@ PaddlePaddle每次发新的版本，遵循以下流程:
 1. 从`develop`分支派生出新的分支，分支名为`release/版本号`。例如，`release/0.10.0`
 1. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。
 1. 对这个版本的提交，做如下几个操作:
+  * 使用Regression Test List作为检查列表，测试本次release的正确性。
+	  * 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，到第二步
 	* 修改`python/setup.py.in`中的版本信息,并将`istaged`字段设为`True`。
-	* 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步
-	* 编译这个版本的Ubuntu Deb包。如果失败，修复Ubuntu Deb包编译问题，Patch号加一，返回第二步。
-	* 使用Regression Test List作为检查列表，测试Docker镜像/ubuntu安装包的功能正确性
-		* 如果失败，记录下所有失败的例子，在这个`release/版本号`分支中，修复所有bug后，Patch号加一，返回第二步
 	* 编译这个版本的python wheel包，并发布到pypi。
 		* 由于pypi.python.org目前遵循[严格的命名规范PEP 513](https://www.python.org/dev/peps/pep-0513)，在使用twine上传之前，需要重命名wheel包中platform相关的后缀，比如将`linux_x86_64`修改成`manylinux1_x86_64`。
 		* pypi上的package名称为paddlepaddle和paddlepaddle_gpu，如果要上传GPU版本的包，需要修改build/python/setup.py中，name: "paddlepaddle_gpu"并重新打包wheel包：`python setup.py bdist_wheel`。
@@ -21,8 +19,8 @@ PaddlePaddle每次发新的版本，遵循以下流程:
 			pip install twine
 			twine upload dist/[package to upload]
 			```
+		* 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步
 1. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。
-1. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面
 1. 协同完成Release Note的书写
 
 
@@ -31,6 +29,30 @@ PaddlePaddle每次发新的版本，遵循以下流程:
 * `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试PaddlePaddle的行为。
 * 在`release/版本号`分支存在的时候，如果有bugfix的行为，需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。
 
+## 发布wheel包到pypi
+
+使用[PaddlePaddle CI](https://paddleci.ngrok.io/project.html?projectId=Manylinux1&tab=projectOverview)
+完成自动化二进制编译，参考下图，选择需要发布的版本（通常包含一个CPU版本和一个GPU版本），点击"run"右侧的"..."按钮，可以
+弹出下面的选择框，在第二个tab (Changes)里选择需要发布的分支，这里选择0.11.0，然后点击"Run Build"按钮。等待编译完成后
+可以在此页面的"Artifacts"下拉框中找到生成的3个二进制文件，分别对应CAPI，`cp27m`和`cp27mu`的版本。然后按照上述的方法
+使用`twine`工具上传即可。
+
+<img src="ci_build_whl.png">
+
+* 注：CI环境使用 https://github.com/PaddlePaddle/buildtools 这里的DockerImage作为编译环境以支持更多的Linux
+  发型版，如果需要手动编译，也可以使用这些镜像。这些镜像也可以从 https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/ 下载得到。
+* pypi不支持覆盖上传，所以一个版本号的wheel包发布之后，不可以更改。下一个wheel包需要更新版本号才可以上传。
+
+## 发布Docker镜像
+
+上述PaddlePaddle CI编译wheel完成后会自动将Docker镜像push到DockerHub，所以，发布Docker镜像只需要对自动push的镜像打上
+版本号对应的tag即可：
+
+1. 进入 https://hub.docker.com/r/paddlepaddle/paddle/tags/ 查看latest tag的更新时间是否在上述编译wheel包完成后是否最新。
+1. 执行 `docker pull paddlepaddle/paddle:[latest tag]`，latest tag可以是latest或latest-gpu等。
+1. 执行 `docker tag paddlepaddle/paddle:[latest tag] paddlepaddle/paddle:[version]`
+1. 执行 `docker push paddlepaddle/paddle:[version]`
+
 ## PaddlePaddle 分支规范
 
 PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。

doc/design/support_new_device.md

@@ -48,8 +48,8 @@ Fluid uses class [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/dev
 
 
 ```
-                /->  CPUDeviceContext   --> MKLDeviceContext
-DeviceContext ---->  CUDADeviceContext  --> CUDNNDeviceContext
+                /->  CPUDeviceContext   
+DeviceContext ---->  CUDADeviceContext  
                 \->  FPGADeviceContext
 ```
 
@@ -79,16 +79,6 @@ private:
 };
 ```
 
-- CUDNNDeviceContext
-
-```
-class CUDNNDeviceContext : public CUDADeviceContext {
-  private:
-    cudnnHandle_t cudnn_handle_;
-};
-```
-
-
 ### Memory and Tensor
 
 

doc/mobile/cross_compiling_for_android_cn.md

 # Android平台编译指南
 
 用户可通过如下两种方式，交叉编译Android平台上适用的PaddlePaddle库：
-- 基于Docker容器的编译方式
-- 基于Linux交叉编译环境的编译方式
+
+- [基于Docker容器的编译方式](#基于docker容器的编译方式)
+- [基于Linux交叉编译环境的编译方式](#基于linux交叉编译环境的编译方式)
 
 ## 基于Docker容器的编译方式
 Docker能在所有主要操作系统（包括Linux，Mac OS X和Windows）上运行，因此，使用基于Docker容器的编译方式，用户可在自己熟悉的开发平台上编译Android平台上适用的PaddlePaddle库。
@@ -16,6 +17,12 @@ $ cd Paddle
 $ docker build -t username/paddle-android:dev . -f Dockerfile.android
 ```
 
+用户也可以使用PaddlePaddle提供的官方开发镜像：
+
+```bash
+$ docker pull paddlepaddle/paddle:latest-dev-android
+```
+
 ### 编译PaddlePaddle C-API库
 构建好开发镜像后，即可使用开发镜像来编译Android版PaddlePaddle C-API库。
 Android的Docker开发镜像向用户提供两个可配置的参数：
@@ -41,23 +48,25 @@ Android的Docker开发镜像向用户提供两个可配置的参数：
 </tr>
 <tr class="row-odd">
   <td>ANDROID_API</td>
-  <td>>= 21</td>
+  <td>>= 16</td>
   <td>21</td>
 </tr>
 </tbody>
 </table>
 
 - 编译`armeabi-v7a`，`Android API 21`的PaddlePaddle库
+
   ```bash
   $ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=armeabi-v7a" -e "ANDROID_API=21" username/paddle-android:dev
   ```
 
 - 编译`arm64-v8a`，`Android API 21`的PaddlePaddle库
+
   ```bash
   $ docker run -it --rm -v $PWD:/paddle -e "ANDROID_ABI=arm64-v8a" -e "ANDROID_API=21" username/paddle-android:dev
   ```
 
-执行上述`docker run`命令时，容器默认执行[paddle/scripts/docker/build_android.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文**配置交叉编译参数**章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。
+执行上述`docker run`命令时，容器默认执行[paddle/scripts/docker/build_android.sh](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build_android.sh)脚本。该脚本中记录了交叉编译Android版PaddlePaddle库常用的CMake配置，并且会根据`ANDROID_ABI`和`ANDROID_API`自动构建独立工具链、进行编译和安装。由于arm64架构要求Android API不小于21。因此当`ANDROID_ABI=arm64-v8a`，`ANDROID_API<21`时，Docker容器中将默认使用`Android API 21`的编译工具链。用户可以参考下文[配置交叉编译参数](#配置交叉编译参数)章节，根据个人的需求修改定制Docker容器所执行的脚本。编译安装结束之后，PaddlePaddle的C-API库将被安装到`$PWD/install_android`目录，所依赖的第三方库同时也被安装到`$PWD/install_android/third_party`目录。
 
 ## 基于Linux交叉编译环境的编译方式
 本文档将以Linux x86-64平台为例，介绍交叉编译Android平台上适用的PaddlePaddle库的方法和步骤。
@@ -83,6 +92,7 @@ your/path/to/android-ndk-r14b-linux-x86_64/build/tools/make-standalone-toolchain
 此命令将在`your/path/to/arm_standalone_toolchain`目录生成一套独立编译工具链，面向架构为32位ARM架构，支持的最小的Android API级别为21，支持编译器`arm-linux-androideabi-gcc (GCC) 4.9`和`clang 3.8`。
 
 - 构建`arm64-v8a`、 `Android API 21`的独立工具链：
+
 ```bash
 your/path/to/android-ndk-r14b-linux-x86_64/build/tools/make-standalone-toolchain.sh \
         --arch=arm64 --platform=android-21 --install-dir=your/path/to/arm64_standalone_toolchain
@@ -90,14 +100,12 @@ your/path/to/android-ndk-r14b-linux-x86_64/build/tools/make-standalone-toolchain
 
 此命令将在`your/path/to/arm64_standalone_toolchain`目录生成一套独立编译工具链，面向架构为64位ARM64架构，支持的最小Android API级别为21，支持编译器`arm-linux-androideabi-gcc (GCC) 4.9`和`clang 3.8`。
 
-注意：**PaddlePaddle要求使用的编译工具链所支持的Android API级别不小于21**。
-
 ### 配置交叉编译参数
 
 CMake系统对交叉编译提供了支持[cmake-toolchains](https://cmake.org/cmake/help/v3.0/manual/cmake-toolchains.7.html#cross-compiling)。为了简化cmake配置，PaddlePaddle为交叉编译提供了工具链配置文档[cmake/cross_compiling/android.cmake](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/android.cmake)，以提供一些默认的编译器和编译参数相关配置。注意，从CMake 3.7版本开始，CMake官方对Android平台的交叉编译提供了通用的支持。PaddlePaddle若检测到用户使用的CMake版本不低于3.7时，将会将用户传进来的配置参数传递CMake系统，交由CMake系统本身来处理。有关参数配置的详细说明见[cmake-toolchains](https://cmake.org/cmake/help/v3.7/manual/cmake-toolchains.7.html#cross-compiling)。
 
 交叉编译Android版本的PaddlePaddle库时，有一些必须配置的参数：
-- `CMAKE_SYSTEM_NAME`，CMake编译的目标平台，必须设置为`Android`。在设置`CMAKE_SYSTEM_NAME=Android`后，PaddlePaddle的CMake系统才认为是在交叉编译Android系统的版本，并自动编译宿主机版protoc可执行文件、目标机版protobuf库、以及Android所需`arm_soft_fp_abi`分支的目标机版OpenBLAS库。此外，还会强制设置一些PaddlePaddle参数的值（`WITH_GPU=OFF`、`WITH_AVX=OFF`、`WITH_PYTHON=OFF`、`WITH_RDMA=OFF`）。
+- `CMAKE_SYSTEM_NAME`，CMake编译的目标平台，必须设置为`Android`。在设置`CMAKE_SYSTEM_NAME=Android`后，PaddlePaddle的CMake系统才认为是在交叉编译Android系统的版本，并自动编译PaddlePaddle所需的所有第三方库。此外，还会强制设置一些PaddlePaddle参数的值（`WITH_GPU=OFF`、`WITH_AVX=OFF`、`WITH_PYTHON=OFF`、`WITH_RDMA=OFF`、`WITH_MKL=OFF`、`WITH_GOLANG=OFF`）。
 - `WITH_C_API`，必须设置为`ON`。在Android平台上只支持使用C-API来预测。
 - `WITH_SWIG_PY`，必须设置为`OFF`。在Android平台上不支持通过swig调用来训练或者预测。
 
@@ -119,7 +127,7 @@ Android平台可选配置参数：
 其他配置参数：
 
 - `USE_EIGEN_FOR_BLAS`，是否使用Eigen库进行矩阵计算。可设置`ON/OFF`，默认值为`OFF`。
-- `HOST_C/CXX_COMPILER`，宿主机的C/C++编译器。在编译宿主机版protoc可执行文件和目标机版OpenBLAS库时需要用到。默认设置成环境变量`CC`的值；若环境变量`CC`没有设置，则设置成`cc`编译器。
+- `HOST_C/CXX_COMPILER`，宿主机的C/C++编译器。在编译宿主机版protoc可执行文件和目标机版OpenBLAS库时需要用到。默认设置成环境变量`CC/CXX`的值；若环境变量`CC/CXX`没有设置，则设置成`cc/c++`编译器。
 
 常用的cmake配置如下：
 
@@ -147,9 +155,10 @@ cmake -DCMAKE_SYSTEM_NAME=Android \
       ..
 ```
 
-用户还可根据自己的需求设置其他编译参数。比如希望最小化生成的库的大小，可以设置`CMAKE_BUILD_TYPE`为`MinSizeRel`；若希望最快的执行速度，则可设置`CMAKE_BUILD_TYPE`为`Release`。亦可以通过手动设置`CMAKE_C/CXX_FLAGS_MINSIZEREL/RELEASE`来影响PaddlePaddle的编译过程。
+用户还可根据自己的需求设置其他编译参数。比如希望最小化生成的库的大小，可以设置`CMAKE_BUILD_TYPE`为`MinSizeRel`；若希望最快的执行速度，则可设置`CMAKE_BUILD_TYPE`为`Release`。亦可以通过手动设置`CMAKE_C/CXX_FLAGS`来影响PaddlePaddle的编译过程。
 
 **性能TIPS**，为了达到最快的计算速度，在CMake参数配置上，有以下建议：
+
 - 设置`CMAKE_BUILD_TYPE`为`Release`
 - 使用`clang`编译工具链
 - `armeabi-v7a`时，设置`USE_EIGEN_BLAS=ON`，使用Eigen进行矩阵计算；`arm64-v8a`时，设置`USE_EIGEN_FOR_BLAS=OFF`，使用OpenBLAS进行矩阵计算

doc/mobile/cross_compiling_for_android_en.md

 # Build PaddlePaddle for Android
 
-There are two approaches to build PaddlePaddle for Android: using Docker and on Linux without Docker. 
+There are two approaches to build PaddlePaddle for Android: 
+
+- [Cross-Compiling Using Docker](#cross-compiling-using-docker)
+- [Cross-Compiling on Linux](#cross-compiling-on-linux) 
 
 ## Cross-Compiling Using Docker
 
@@ -16,6 +19,12 @@ $ cd Paddle
 $ docker build -t paddle:dev-android . -f Dockerfile.android
 ```
 
+Users can directly use the published Docker image.
+
+```bash
+$ docker pull paddlepaddle/paddle:latest-dev-android
+```
+
 ### Build the Inference Library
 
 We can run the Docker image we just created to build the inference library of PaddlePaddle for Android using the command below:
@@ -47,7 +56,7 @@ The Docker image accepts two arguments `ANDROID_ABI` and `ANDROID_API`:
 </tr>
 <tr class="row-odd">
   <td>ANDROID_API</td>
-  <td>>= 21</td>
+  <td>>= 16</td>
   <td>21</td>
 </tr>
 </tbody>
@@ -93,15 +102,13 @@ Android NDK includes everything we need to build the [*standalone toolchain*](ht
 
   The generated standalone toolchain will be in `your/path/to/arm64_standalone_toolchain`.
 
-**Please be aware that the minimum level of Android API required by PaddlePaddle is 21.**
-
 ### Cross-Compiling Arguments
 
 CMake supports [choosing the toolchain](https://cmake.org/cmake/help/v3.0/manual/cmake-toolchains.7.html#cross-compiling).  PaddlePaddle provides [`android.cmake`](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/android.cmake), which configures the Android cross-compiling toolchain for CMake.  `android.cmake` is not required for CMake >= 3.7, which support Android cross-compiling. PaddlePaddle detects the CMake version, for those newer than 3.7, it uses [the official version](https://cmake.org/cmake/help/v3.7/manual/cmake-toolchains.7.html#cross-compiling).
 
 Some other CMake arguments you need to know:
 
-- `CMAKE_SYSTEM_NAME` must be `Android`.  This tells PaddlePaddle's CMake system to cross-compile third-party dependencies.  This also changes some other CMake arguments like `WITH_GPU=OFF`, `WITH_AVX=OFF`, `WITH_PYTHON=OFF`, and `WITH_RDMA=OFF`.
+- `CMAKE_SYSTEM_NAME` must be `Android`.  This tells PaddlePaddle's CMake system to cross-compile third-party dependencies. This also changes some other CMake arguments like `WITH_GPU=OFF`, `WITH_AVX=OFF`, `WITH_PYTHON=OFF`, `WITH_RDMA=OFF`, `WITH_MKL=OFF` and `WITH_GOLANG=OFF`.
 - `WITH_C_API` must be `ON`, to build the C-based inference library for Android.
 - `WITH_SWIG_PY` must be `OFF` because the Android platform doesn't support SWIG-based API.
 
@@ -123,7 +130,7 @@ Some Android-specific arguments:
 Other useful arguments:
 
 - `USE_EIGEN_FOR_BLAS`: indicates if using Eigen.  Could be `ON` or `OFF`, defaults to `OFF`.
-- `HOST_C/CXX_COMPILER`: specifies the host compiler, which is used to build the host-specific protoc and target-specific OpenBLAS.  It defaults to the value of the environment variable `CC`, or `cc`.
+- `HOST_C/CXX_COMPILER`: specifies the host compiler, which is used to build the host-specific protoc and target-specific OpenBLAS.  It defaults to the value of the environment variable `CC/C++`, or `cc/c++`.
 
 Some frequent configurations for your reference:
 
@@ -158,6 +165,7 @@ There are some other arguments you might want to configure.
 - `CMAKE_BUILD_TYPE-Release` optimizes the runtime performance.
 
 Our own tip for performance optimization to use clang and Eigen or OpenBLAS:
+
 - `CMAKE_BUILD_TYPE=Release`
 - `ANDROID_TOOLCHAIN=clang`
 - `USE_EIGEN_BLAS=ON` for `armeabi-v7a`, or `USE_EIGEN_FOR_BLAS=OFF` for `arm64-v8a`.

doc/mobile/cross_compiling_for_ios_en.md

-# PaddlePaddle Compiling Guide for iOS
+# Build PaddlePaddle for iOS
 
 This tutorial will walk you through cross compiling the PaddlePaddle library for iOS from the source in MacOS.
 
@@ -98,7 +98,7 @@ You can set other compiling parameters for your own need. I.E. if you are trying
 - set `CMAKE_BUILD_TYPE` with `Release`
 - set `IOS_USE_VECLIB_FOR_BLAS` with `ON`
 
-## Compile and install
+## Build and install
 
 After CMake, run following commands, PaddlePaddle will download the compile 3rd party dependencies, compile and install PaddlePaddle inference library.
 
@@ -109,7 +109,7 @@ $ make install
 
 Please Note: if you compiled PaddlePaddle in the source directory for other platforms, do remove `third_party` and `build` directory within the source with `rm -rf` to ensure that all the 3rd party libraries dependencies and PaddlePaddle is newly compiled with current CMake configuration.
 
-`your/path/to/install` directory will have following directories after `compile` and `install`:
+`your/path/to/install` directory will have following directories after `make install`:
 
 - `include`, contains all the C-API header files.
 - `lib`, contains PaddlePaddle C-API static library.

paddle/CMakeLists.txt

@@ -24,6 +24,7 @@ else()
     add_subdirectory(framework)
     add_subdirectory(operators)
     add_subdirectory(pybind)
+    add_subdirectory(inference)
   endif()
 
   if(WITH_SWIG_PY)

paddle/framework/CMakeLists.txt

@@ -26,10 +26,15 @@ nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
 
 cc_test(variable_test SRCS variable_test.cc)
 
-cc_library(scope SRCS scope.cc DEPS glog)
+cc_library(threadpool SRCS threadpool.cc)
+cc_test(threadpool_test SRCS threadpool_test.cc DEPS threadpool)
+
+cc_library(scope SRCS scope.cc DEPS glog threadpool)
 cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
-cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor framework_proto)
+cc_library(device_data_transform SRCS device_data_transform.cc DEPS tensor)
+
+cc_library(data_transform SRCS data_transform.cc DEPS math_function tensor framework_proto selected_rows device_data_transform)
 cc_test(data_transform_test SRCS data_transform_test.cc DEPS data_transform device_context)
 
 cc_library(attribute SRCS attribute.cc DEPS framework_proto)
@@ -38,7 +43,7 @@ device_context)
 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(shape_inference SRCS shape_inference.cc DEPS ddim attribute)
+cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
     shape_inference data_transform)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init)
@@ -70,9 +75,10 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
-cc_library(threadpool SRCS threadpool.cc)
-cc_test(threadpool_test SRCS threadpool_test.cc DEPS threadpool)
-cc_library(init SRCS init.cc DEPS gflags device_context place stringpiece)
+cc_library(init SRCS init.cc DEPS gflags device_context place stringpiece operator)
 cc_test(init_test SRCS init_test.cc DEPS init)
 
 cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
+
+nv_test(device_data_transform_test SRCS device_data_transform_test.cu
+        DEPS operator op_registry init math_function)

paddle/framework/backward.cc

@@ -427,7 +427,8 @@ std::vector<std::unique_ptr<OpDesc>> MakeBlockBackward(
     VLOG(5) << "Making backward " << (*it)->Type() << " op";
     std::vector<std::unique_ptr<OpDesc>> op_grads;
 
-    if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") {
+    if ((*it)->Type() == "recurrent" || (*it)->Type() == "while" ||
+        (*it)->Type() == "parallel_do") {
       int step_block_idx = (*it)->GetBlockAttr("sub_block");
       BlockDesc* backward_block = CreateStepBlock(program_desc, no_grad_vars,
                                                   grad_to_var, step_block_idx);

paddle/framework/data_transform.cc

@@ -11,9 +11,12 @@ 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 <functional>
 
 #include "paddle/framework/data_transform.h"
+#include "paddle/framework/device_data_transform.h"
 #include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/selected_rows.h"
 #include "paddle/platform/device_context.h"
 
 namespace paddle {
@@ -24,6 +27,37 @@ DataTransformFnMap& DataTransformFnMap::Instance() {
   return data_transform_map;
 }
 
+Tensor* DataTransform(const OpKernelType& expected_kernel_type,
+                      const OpKernelType& kernel_type_for_var,
+                      const Tensor& input_tensor) {
+  Tensor* out = nullptr;
+  if (!platform::is_same_place(kernel_type_for_var.place_,
+                               expected_kernel_type.place_)) {
+    out = DeviceTransform(input_tensor, expected_kernel_type.place_);
+  }
+  PADDLE_ENFORCE_NOT_NULL(out, "out should not be null");
+  return out;
+}
+
+void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
+                            Variable& out_var) {
+  if (in_var.IsType<LoDTensor>()) {
+    auto& in_lod_tensor = in_var.Get<LoDTensor>();
+    auto* tran_lod_tensor = out_var.GetMutable<LoDTensor>();
+    tran_lod_tensor->set_lod(in_lod_tensor.lod());
+    tran_lod_tensor->set_layout(in_lod_tensor.layout());
+    tran_lod_tensor->ShareDataWith(tensor);
+  } else if (in_var.IsType<SelectedRows>()) {
+    auto& in_selected_rows = in_var.Get<SelectedRows>();
+    auto* trans_selected_rows = out_var.GetMutable<SelectedRows>();
+    trans_selected_rows->set_height(in_selected_rows.height());
+    trans_selected_rows->set_rows(in_selected_rows.rows());
+    trans_selected_rows->mutable_value()->ShareDataWith(tensor);
+  } else {
+    PADDLE_THROW("unknown var type");
+  }
+}
+
 auto KernelFP32 = OpKernelType(proto::DataType::FP32, platform::CPUPlace(),
                                DataLayout::kNHWC, LibraryType::kPlain);
 
@@ -36,6 +70,28 @@ auto KernelNHWC = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
 auto KernelNCHW = OpKernelType(proto::DataType::FP64, platform::CPUPlace(),
                                DataLayout::kNCHW, LibraryType::kPlain);
 
+// TODO(dzhwinter): Only for testing multiple op kernel.
+// Dummy transform function for library_type
+// should be removed.
+auto KernelPlain = OpKernelType(proto::DataType::FP32, platform::CUDAPlace(0),
+                                DataLayout::kAnyLayout, LibraryType::kPlain);
+
+auto KernelCUDNN = OpKernelType(proto::DataType::FP32, platform::CUDAPlace(0),
+                                DataLayout::kAnyLayout, LibraryType::kCUDNN);
+
+void DummyTrans(const platform::DeviceContext* ctx,
+                const KernelTypePair& kernel_pair, const Variable& in,
+                Variable* out) {
+  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only Support Tensor transform!.");
+  PADDLE_ENFORCE(
+      platform::places_are_same_class(kernel_pair.first.place_,
+                                      kernel_pair.second.place_),
+      "TransDataType Only Support DataType transform on same place!");
+  auto src = in.Get<Tensor>();
+  auto* dst = out->GetMutable<Tensor>();
+  *dst = src;
+}
+
 void TransDataType(const platform::DeviceContext* ctx,
                    const KernelTypePair& kernel_pair, const Variable& in,
                    Variable* out) {
@@ -74,26 +130,36 @@ void TransDataType(const platform::DeviceContext* ctx,
   }
 }
 
-void TransDataLayout(const platform::DeviceContext* ctx,
+void TransDataLayout(const std::vector<int>& axis,
+                     const platform::DeviceContext* ctx,
                      const KernelTypePair& kernel_pair, const Variable& in,
                      Variable* out) {
-  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only Support Tensor transform!.");
+  PADDLE_ENFORCE(in.IsType<Tensor>(), "Only support Tensor transform!.");
   PADDLE_ENFORCE(
       platform::places_are_same_class(kernel_pair.first.place_,
                                       kernel_pair.second.place_),
-      "TransDataType Only Support DataType transform on same place!");
+      "TransDataLayout only support DataLayout transform on same place!");
+  PADDLE_ENFORCE(kernel_pair.first.data_type_ == kernel_pair.second.data_type_,
+                 "TransDataLayout only support Datatype are same!");
 
   auto src = in.Get<Tensor>();
   auto* dst = out->GetMutable<Tensor>();
   PADDLE_ENFORCE(arity(src.dims()) == 4, "Input Arity Only Suppport 4!");
 
-  dst->Resize(src.dims());
+  auto src_dim = src.dims();
+  std::vector<int64_t> dst_dim;
+
+  dst_dim.resize(axis.size());
+  for (size_t i = 0; i < axis.size(); i++) {
+    dst_dim[i] = src_dim[axis[i]];
+  }
+
+  dst->Resize(make_ddim(dst_dim));
   auto place = kernel_pair.second.place_;
-  CopyFrom(src, place, *ctx, dst);
-  const std::vector<int> axis = {0, 2, 3, 1};
+  dst->mutable_data(place, src.type());
 
   auto src_type = kernel_pair.first.data_type_;
-  framework::VisitDataType(src_type, CastDataLayout(src, dst, ctx, axis));
+  framework::VisitDataType(src_type, CastDataLayout(ctx, axis, src, dst));
 
   dst->set_layout(kernel_pair.second.data_layout_);
 }
@@ -102,5 +168,24 @@ void TransDataLayout(const platform::DeviceContext* ctx,
 }  // namespace paddle
 
 namespace f = paddle::framework;
+
+namespace {
+std::vector<int> NHWC2NCHW = {0, 3, 1, 2};
+std::vector<int> NCHW2NHWC = {0, 2, 3, 1};
+}
+
 REGISTER_DATA_TRANSFORM_FN(f::KernelFP32, f::KernelFP64, f::TransDataType);
-REGISTER_DATA_TRANSFORM_FN(f::KernelNHWC, f::KernelNCHW, f::TransDataLayout);
+REGISTER_DATA_TRANSFORM_FN(f::KernelPlain, f::KernelCUDNN, f::DummyTrans);
+REGISTER_DATA_TRANSFORM_FN(f::KernelCUDNN, f::KernelPlain, f::DummyTrans);
+REGISTER_DATA_TRANSFORM_FN(f::KernelNHWC, f::KernelNCHW,
+                           std::bind(f::TransDataLayout, NHWC2NCHW,
+                                     std::placeholders::_1,
+                                     std::placeholders::_2,
+                                     std::placeholders::_3,
+                                     std::placeholders::_4));
+REGISTER_DATA_TRANSFORM_FN(f::KernelNCHW, f::KernelNHWC,
+                           std::bind(f::TransDataLayout, NCHW2NHWC,
+                                     std::placeholders::_1,
+                                     std::placeholders::_2,
+                                     std::placeholders::_3,
+                                     std::placeholders::_4));

paddle/framework/data_transform.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/framework/op_kernel_type.h"
+#include "paddle/framework/selected_rows.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/framework/variable.h"
 #include "paddle/operators/math/math_function.h"
@@ -49,6 +50,13 @@ struct KernelTypePairHash {
   }
 };
 
+Tensor* DataTransform(const OpKernelType& expected_kernel_type,
+                      const OpKernelType& kernel_type_for_var,
+                      const Tensor& input_tensor);
+
+void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
+                            Variable& out_var);
+
 template <typename InType, typename OutType>
 struct CastDataTypeFunctor {
   HOSTDEVICE inline OutType operator()(InType in) const {
@@ -73,6 +81,7 @@ struct CastDataType {
     auto numel = in_.numel();
     auto* in_end = in_begin + numel;
     auto* out_begin = out_->mutable_data<OutType>(place);
+
     if (platform::is_cpu_place(place)) {
       platform::Transform<platform::CPUDeviceContext> trans;
       auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);
@@ -86,9 +95,9 @@ struct CastDataType {
 };
 
 struct CastDataLayout {
-  CastDataLayout(const framework::Tensor& in, framework::Tensor* out,
-                 const platform::DeviceContext* ctx,
-                 const std::vector<int>& axis)
+  CastDataLayout(const platform::DeviceContext* ctx,
+                 const std::vector<int>& axis, const framework::Tensor& in,
+                 framework::Tensor* out)
       : in_(in), out_(out), ctx_(ctx), axis_(axis) {}
   const framework::Tensor in_;
   framework::Tensor* out_;
@@ -98,6 +107,7 @@ struct CastDataLayout {
   template <typename T>
   void operator()() {
     auto place = ctx_->GetPlace();
+
     if (platform::is_cpu_place(place)) {
       operators::math::Transpose<platform::CPUDeviceContext, T, 4> trans4;
       auto* context = static_cast<const platform::CPUDeviceContext*>(ctx_);

paddle/framework/data_transform_test.cc

@@ -32,18 +32,18 @@ using namespace platform;
  *       1111 -> FP64, GPUPlace, kNCHW, kMKLDNN
  */
 
-std::array<proto::DataType, 2> kDataType = {proto::DataType::FP32,
-                                            proto::DataType::FP64};
+std::array<proto::DataType, 2> kDataType = {
+    {proto::DataType::FP32, proto::DataType::FP64}};
 
-std::array<Place, 2> kPlace = {CPUPlace(), CUDAPlace(0)};
+std::array<Place, 2> kPlace = {{CPUPlace(), CUDAPlace(0)}};
 
-std::array<DataLayout, 2> kDataLayout = {
+std::array<DataLayout, 2> kDataLayout = {{
     DataLayout::kNHWC, DataLayout::kNCHW,
-};
+}};
 
-std::array<LibraryType, 2> kLibraryType = {
+std::array<LibraryType, 2> kLibraryType = {{
     LibraryType::kPlain, LibraryType::kMKLDNN,
-};
+}};
 
 OpKernelType GenFromBit(const std::vector<bool> bits) {
   return OpKernelType(kDataType[bits[0]], kPlace[bits[1]], kDataLayout[bits[2]],
@@ -106,7 +106,7 @@ TEST(DataTransform, Register) {
   ASSERT_EQ(test_value, 2);
 }
 
-TEST(DataTransform, Layout) {
+TEST(DataTransform, DataLayout) {
   using namespace paddle::framework;
   using namespace paddle::platform;
 
@@ -127,7 +127,19 @@ TEST(DataTransform, Layout) {
   }
 
   Tensor dst = out.Get<Tensor>();
-  EXPECT_TRUE(dst.layout() != src->layout());
+
+  EXPECT_TRUE(dst.layout() == DataLayout::kNCHW);
+  EXPECT_TRUE(dst.dims() == make_ddim({2, 2, 3, 1}));
+
+  {
+    auto kernel1 = GenFromBit({1, 0, 1, 0});
+    auto kernel2 = GenFromBit({1, 0, 0, 0});
+    auto pair0 = std::make_pair(kernel1, kernel2);
+    instance.Get(pair0)(ctx, pair0, out, &in);
+  }
+
+  EXPECT_TRUE(src->layout() == DataLayout::kNHWC);
+  EXPECT_TRUE(src->dims() == make_ddim({2, 3, 1, 2}));
 }
 
 TEST(DataTransform, DataType) {

paddle/framework/device_data_transform.cc

+/* 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. */
+
+#include "paddle/framework/device_data_transform.h"
+
+namespace paddle {
+namespace framework {
+
+static const platform::DeviceContext* GetDeviceContext(
+    const platform::Place& src_place, const platform::Place& dst_place) {
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+
+  if (platform::is_gpu_place(src_place) && platform::is_cpu_place(dst_place)) {
+    return pool.Get(src_place);
+  } else if (platform::is_cpu_place(src_place) &&
+             platform::is_gpu_place(dst_place)) {
+    return pool.Get(dst_place);
+  } else {
+    PADDLE_THROW(
+        "Currently, model parallelism is only supported between CPU and CUDA");
+  }
+}
+
+Tensor* DeviceTransform(const Tensor& in, const platform::Place& dst_place) {
+  VLOG(3) << "DeviceTransform in, src_place " << in.place()
+          << " dst_place: " << dst_place;
+  Tensor* out = new Tensor();
+  auto* dev_ctx = GetDeviceContext(in.place(), dst_place);
+  dev_ctx->Wait();
+  CopyFrom(in, dst_place, *dev_ctx, out);
+  dev_ctx->Wait();
+  return out;
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/device_data_transform.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
+
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/framework/tensor_util.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+
+Tensor* DeviceTransform(const Tensor& in, const platform::Place& dst_place);
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/device_data_transform_test.cu

+/* 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. */
+
+#include "gtest/gtest.h"
+
+#include "paddle/framework/init.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/op_info.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/elementwise_op_function.h"
+#include "paddle/operators/math/math_function.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+
+template <typename T>
+struct AddFunctor {
+  inline HOSTDEVICE T operator()(T a, T b) const { return a + b; }
+};
+
+class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
+ public:
+  OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("input", "input1 of test op");
+    AddOutput("output", "output of test op");
+    AddAttr<bool>("use_gpu", "force to use gpu kernel").SetDefault(false);
+    AddComment("This is test op");
+  }
+};
+
+class TestOpWithKernel : public OperatorWithKernel {
+ public:
+  using OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {}
+  OpKernelType GetExpectedKernelType(
+      const ExecutionContext& ctx) const override {
+    if (Attr<bool>("use_gpu")) {
+      VLOG(3) << "force use gpu kernel";
+      return OpKernelType(proto::DataType::FP32, platform::CUDAPlace(0));
+    } else {
+      VLOG(3) << "use default kernel";
+      return OpKernelType(proto::DataType::FP32,
+                          ctx.Input<Tensor>("input")->place());
+    }
+  }
+};
+
+template <typename DeviceContext, typename T>
+class TestKernel : public OpKernel<float> {
+ public:
+  void Compute(const ExecutionContext& ctx) const {
+    std::cout << ctx.op().DebugString() << std::endl;
+
+    const Tensor* input = ctx.Input<Tensor>("input");
+
+    std::cout << "input place:" << input->place() << std::endl;
+    auto* output = ctx.Output<framework::LoDTensor>("output");
+    output->Resize(input->dims());
+    output->mutable_data<T>(ctx.GetPlace());
+
+    operators::TransformFunctor<AddFunctor<T>, T, DeviceContext> functor(
+        input, input, output, ctx.template device_context<DeviceContext>(),
+        AddFunctor<T>());
+    functor.Run();
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(
+    test_op, paddle::framework::TestOpWithKernel,
+    paddle::framework::OpKernelTestProtoAndCheckerMaker);
+REGISTER_OP_CPU_KERNEL(
+    test_op,
+    paddle::framework::TestKernel<paddle::platform::CPUDeviceContext, float>);
+REGISTER_OP_CUDA_KERNEL(
+    test_op,
+    paddle::framework::TestKernel<paddle::platform::CUDADeviceContext, float>);
+
+static void BuildVar(const std::string& param_name,
+                     std::initializer_list<const char*> arguments,
+                     paddle::framework::proto::OpDesc::Var* var) {
+  var->set_parameter(param_name);
+  for (auto& arg_name : arguments) {
+    *var->mutable_arguments()->Add() = arg_name;
+  }
+}
+
+TEST(Operator, CPUtoGPU) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+
+  ASSERT_EQ(InitDevices({"CPU", "GPU:0"}), true);
+
+  paddle::framework::Scope scope;
+  paddle::platform::CPUPlace cpu_place;
+
+  // create an op to run on CPU
+  paddle::framework::proto::OpDesc cpu_op_desc;
+  cpu_op_desc.set_type("test_op");
+  BuildVar("input", {"IN1"}, cpu_op_desc.add_inputs());
+  BuildVar("output", {"OUT1"}, cpu_op_desc.add_outputs());
+
+  auto cpu_op = paddle::framework::OpRegistry::CreateOp(cpu_op_desc);
+  // prepare input
+  auto* in_t = scope.Var("IN1")->GetMutable<LoDTensor>();
+  auto* src_ptr = in_t->mutable_data<float>({2, 3}, CPUPlace());
+  for (int i = 0; i < 2 * 3; ++i) {
+    src_ptr[i] = static_cast<float>(i);
+  }
+
+  // get output
+  auto* output = scope.Var("OUT1");
+  cpu_op->Run(scope, cpu_place);
+
+  auto* output_ptr = output->Get<LoDTensor>().data<float>();
+  for (int i = 0; i < 2 * 3; ++i) {
+    ASSERT_EQ(output_ptr[i], static_cast<float>(i) * 2);
+  }
+
+  // create an op to run on GPU
+  paddle::framework::proto::OpDesc gpu_op_desc;
+  gpu_op_desc.set_type("test_op");
+  BuildVar("input", {"OUT1"}, gpu_op_desc.add_inputs());
+  BuildVar("output", {"OUT2"}, gpu_op_desc.add_outputs());
+
+  auto attr = gpu_op_desc.mutable_attrs()->Add();
+  attr->set_name("use_gpu");
+  attr->set_type(paddle::framework::proto::AttrType::BOOLEAN);
+  attr->set_b(true);
+
+  auto gpu_op = paddle::framework::OpRegistry::CreateOp(gpu_op_desc);
+
+  paddle::platform::CUDAPlace cuda_place(0);
+  // get output
+  auto* output2 = scope.Var("OUT2");
+  gpu_op->Run(scope, cuda_place);
+
+  // auto* output2_ptr = output2->Get<LoDTensor>().data<float>();
+  DeviceContextPool& pool = DeviceContextPool::Instance();
+  auto dev_ctx = pool.Get(cuda_place);
+
+  paddle::framework::Tensor output_tensor;
+  CopyFrom(output2->Get<LoDTensor>(), paddle::platform::CPUPlace(), *dev_ctx,
+           &output_tensor);
+
+  dev_ctx->Wait();
+  float* output2_ptr = output_tensor.data<float>();
+  for (int i = 0; i < 2 * 3; ++i) {
+    ASSERT_EQ(output2_ptr[i], static_cast<float>(i) * 4);
+  }
+}

paddle/framework/executor.cc

@@ -111,7 +111,7 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
 
   for (auto& op_desc : block.AllOps()) {
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
-    VLOG(3) << op->DebugString();
+    VLOG(3) << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);
     if (FLAGS_check_nan_inf) {
       for (auto& vname : op->OutputVars(true)) {

paddle/framework/init.cc

@@ -15,6 +15,7 @@ limitations under the License. */
 #include <string>
 
 #include "paddle/framework/init.h"
+#include "paddle/framework/operator.h"
 #include "paddle/platform/device_context.h"
 #include "paddle/platform/place.h"
 #include "paddle/string/piece.h"
@@ -24,7 +25,6 @@ namespace framework {
 
 std::once_flag gflags_init_flag;
 
-// TODO(qijun) move init gflags to init.cc
 void InitGflags(std::vector<std::string> &argv) {
   std::call_once(gflags_init_flag, [&]() {
     int argc = argv.size();
@@ -72,8 +72,14 @@ bool InitDevices(const std::vector<std::string> &devices) {
     LOG(WARNING) << "Not specified CPU device, create CPU by Default.";
   }
   platform::DeviceContextPool::Init(places);
+  framework::UseALL();
   return true;
 }
 
+void InitGLOG(const std::string &prog_name) {
+  google::InitGoogleLogging(prog_name.c_str());
+  google::InstallFailureSignalHandler();
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/init.h

@@ -22,6 +22,8 @@ namespace framework {
 
 void InitGflags(std::vector<std::string> &argv);
 
+void InitGLOG(const std::string &prog_name);
+
 bool InitDevices(const std::vector<std::string> &devices);
 
 }  // namespace framework

paddle/framework/library_type.h

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include <cctype>
 
 namespace paddle {
 namespace framework {
@@ -41,6 +42,9 @@ inline std::string LibraryTypeToString(const LibraryType& library_type) {
 
 inline LibraryType StringToLibraryType(const char* ctype) {
   std::string s(ctype);
+  for (size_t i = 0; i < s.size(); ++i) {
+    s[i] = toupper(s[i]);
+  }
   if (s == std::string("PLAIN")) {
     return LibraryType::kPlain;
   } else if (s == std::string("MKLDNN")) {

paddle/framework/lod_tensor.cc

@@ -43,6 +43,22 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) {
   return os;
 }
 
+std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
+  PADDLE_ENFORCE(platform::is_cpu_place(t.place()));
+  PADDLE_ENFORCE(t.type().hash_code() == typeid(float).hash_code());
+
+  os << "dim: " << t.dims() << "\n";
+  os << "lod: " << t.lod() << "\n";
+
+  // only print first ten elements
+  int64_t size = t.numel() < 10 ? t.numel() : 10;
+  for (int64_t i = 0; i < size; ++i) {
+    os << t.data<float>()[i] << " ";
+  }
+
+  return os;
+}
+
 LoD SliceLevels(const LoD &in, size_t level_begin, size_t level_end) {
   LoD new_lod;
   new_lod.reserve(level_end - level_begin);
@@ -177,6 +193,9 @@ void AppendLoD(LoD *lod, const LoD &lod_length) {
       lod->empty() || lod->size() == lod_length.size(),
       "The lod_length should has the same size with the appended lod.");
   if (lod->empty()) {
+    for (size_t i = 0; i < lod_length.size(); ++i) {
+      lod->emplace_back(1, 0);  // size = 1, value = 0;
+    }
     *lod = LoD(lod_length.size(), std::vector<size_t>({0}));
   }
   for (size_t i = 0; i < lod->size(); ++i) {
@@ -214,9 +233,10 @@ void SerializeToStream(std::ostream &os, const LoDTensor &tensor,
   SerializeToStream(os, static_cast<Tensor>(tensor), dev_ctx);
 }
 
-void DeserializeFromStream(std::istream &is, LoDTensor *tensor) {
+void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
+                           const platform::DeviceContext &dev_ctx) {
   {
-    // the 1st field, unit32_t version for SelectedRows
+    // the 1st field, unit32_t version for LoDTensor
     uint32_t version;
     is.read(reinterpret_cast<char *>(&version), sizeof(version));
     PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
@@ -237,7 +257,71 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor) {
     }
   }
   // the 3st filed, Tensor
-  DeserializeFromStream(is, static_cast<Tensor *>(tensor));
+  DeserializeFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
+}
+
+std::vector<LoDTensor> LoDTensor::SplitLoDTensor(
+    const std::vector<platform::Place> places) const {
+  check_memory_size();
+  //  PADDLE_ENFORCE(lod().empty() || (lod().size() == 1 && lod()[0].empty())
+  //                 , "Disable parallel lod for now");
+  PADDLE_ENFORCE(lod().empty(), "Disable parallel lod for now");
+  PADDLE_ENFORCE(dims()[0] % places.size() == 0,
+                 "Batch size should be divided by places size");
+
+  std::vector<LoDTensor> lods;
+  for (size_t place_idx = 0; place_idx < places.size(); ++place_idx) {
+    size_t begin = place_idx * dims()[0] / places.size();
+    size_t end = (place_idx + 1) * dims()[0] / places.size();
+    auto src = Slice(static_cast<int>(begin), static_cast<int>(end));
+
+    LoDTensor dst;
+    dst.Resize(src.dims());
+    auto &dst_place = places[place_idx];
+    auto dst_ptr = dst.mutable_data(dst_place, src.type());
+
+    // TODO(tonyyang-svail):
+    //   change the following to framework::CopyFrom
+    auto src_place = src.place();
+    auto src_ptr = src.data<void>();
+    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,
+                   boost::get<platform::CPUPlace>(src_place), src_ptr, size);
+    } else {
+      PADDLE_THROW("Not Implemented");
+    }
+
+    lods.emplace_back(dst);
+  }
+
+  return lods;
+}
+
+void LoDTensor::MergeLoDTensor(
+    const std::vector<const LoDTensor *> &lod_tensors, platform::Place place) {
+  PADDLE_ENFORCE(platform::is_cpu_place(place));
+  PADDLE_ENFORCE(!lod_tensors.empty());
+
+  framework::DDim new_dim = lod_tensors[0]->dims();
+  std::type_index new_type = lod_tensors[0]->type();
+  for (auto *lod : lod_tensors) {
+    PADDLE_ENFORCE(new_dim == lod->dims());
+    PADDLE_ENFORCE(new_type == lod->type());
+    PADDLE_ENFORCE(platform::is_cpu_place(lod->place()));
+  }
+  new_dim[0] *= lod_tensors.size();
+  Resize(new_dim);
+
+  auto *dst_ptr = reinterpret_cast<uint8_t *>(mutable_data(place, new_type));
+  for (auto *src : lod_tensors) {
+    auto size = src->numel() * SizeOfType(src->type());
+    memory::Copy(boost::get<platform::CPUPlace>(place), dst_ptr,
+                 boost::get<platform::CPUPlace>(src->place()),
+                 src->data<void>(), size);
+    dst_ptr += size;
+  }
 }
 
 }  // namespace framework

paddle/framework/lod_tensor.h

@@ -58,6 +58,7 @@ using Vector = thrust::host_vector<
 using LoD = std::vector<Vector<size_t>>;
 
 std::ostream& operator<<(std::ostream& os, const LoD& lod);
+std::ostream& operator<<(std::ostream& os, const LoDTensor& t);
 
 /*
  * Slice levels from a LoD.
@@ -144,6 +145,12 @@ class LoDTensor : public Tensor {
    */
   void ShrinkInLevel(size_t level, size_t elem_begin, size_t elem_end);
 
+  std::vector<LoDTensor> SplitLoDTensor(
+      const std::vector<platform::Place> places) const;
+
+  void MergeLoDTensor(const std::vector<const LoDTensor*>& lod_tensors,
+                      platform::Place place);
+
  private:
   LoD lod_;
 };
@@ -208,7 +215,8 @@ void AppendLoD(LoD* lod, const LoD& lod_length);
  */
 void SerializeToStream(std::ostream& os, const LoDTensor& tensor,
                        const platform::DeviceContext& dev_ctx);
-void DeserializeFromStream(std::istream& is, LoDTensor* tensor);
+void DeserializeFromStream(std::istream& is, LoDTensor* tensor,
+                           const platform::DeviceContext& dev_ctx);
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor_test.cc

@@ -132,7 +132,7 @@ TEST_F(LoDTensorTester, SerializeAndDeserialize) {
   std::ostringstream oss;
   SerializeToStream(oss, lod_tensor_, cpu_ctx);
   std::istringstream iss(oss.str());
-  DeserializeFromStream(iss, &dst_tensor);
+  DeserializeFromStream(iss, &dst_tensor, cpu_ctx);
   float* dst_ptr = dst_tensor.mutable_data<float>(platform::CPUPlace());
   for (int i = 0; i < kLodTensorSize; ++i) {
     EXPECT_EQ(dst_ptr[i], i);

paddle/framework/op_desc.cc

@@ -64,8 +64,9 @@ class CompileTimeInferShapeContext : public InferShapeContext {
     PADDLE_ENFORCE_EQ(in_var->GetType(), proto::VarDesc::LOD_TENSOR,
                       "The %d-th output of Output(%s) must be LoDTensor.", j,
                       out);
-    out_var->SetLoDLevel(in_var->GetLodLevel());
+    out_var->SetLoDLevel(in_var->GetLoDLevel());
   }
+
   bool IsRuntime() const override;
 
  protected:
@@ -260,7 +261,13 @@ struct SetAttrDescVisitor : public boost::static_visitor<void> {
   void operator()(int v) const { attr_->set_i(v); }
   void operator()(float v) const { attr_->set_f(v); }
   void operator()(const std::string &v) const { attr_->set_s(v); }
-  void operator()(bool b) const { attr_->set_b(b); }
+
+  // Please refer to https://github.com/PaddlePaddle/Paddle/issues/7162
+  template <class T,
+            class = typename std::enable_if<std::is_same<bool, T>::value>::type>
+  void operator()(T b) const {
+    attr_->set_b(b);
+  }
 
   void operator()(const std::vector<int> &v) const {
     VectorToRepeated(v, attr_->mutable_ints());
@@ -274,9 +281,7 @@ struct SetAttrDescVisitor : public boost::static_visitor<void> {
   void operator()(const std::vector<bool> &v) const {
     VectorToRepeated(v, attr_->mutable_bools());
   }
-  void operator()(proto::BlockDesc *desc) const {
-    attr_->set_block_idx(desc->idx());
-  }
+  void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); }
   void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }
 };
 

paddle/framework/op_kernel_type.h

@@ -26,13 +26,12 @@ namespace framework {
 struct OpKernelType {
   struct Hash {
     size_t operator()(const OpKernelType& key) const {
-      int place = key.place_.which() + (1 << LEFT_SHIFT);
-      int data_type =
-          static_cast<int>(key.data_type_) + (1 << (LEFT_SHIFT + 1));
-      int data_layout =
-          static_cast<int>(key.data_layout_) + (1 << (LEFT_SHIFT + 2));
-      int library_type =
-          static_cast<int>(key.library_type_) + (1 << (LEFT_SHIFT + 3));
+      int place = key.place_.which();
+      int data_type = static_cast<int>(key.data_type_) << LEFT_SHIFT;
+      int data_layout = static_cast<int>(key.data_layout_) << (LEFT_SHIFT * 2);
+      int library_type = static_cast<int>(key.library_type_)
+                         << (LEFT_SHIFT * 3);
+
       std::hash<int> hasher;
       return hasher(place + data_type + data_layout + library_type);
     }

paddle/framework/op_registry.h

@@ -37,8 +37,8 @@ class Registrar {
  public:
   // In our design, various kinds of classes, e.g., operators and kernels,
   // have their corresponding registry and registrar. The action of
-  // registration is in the constructor of a global registrar variable, which,
-  // however, are not used in the code that calls package framework, and would
+  // registration is in the constructor of a global registrar variable, which
+  // are not used in the code that calls package framework, and would
   // be removed from the generated binary file by the linker. To avoid such
   // removal, we add Touch to all registrar classes and make USE_OP macros to
   // call this method. So, as long as the callee code calls USE_OP, the global

paddle/framework/op_registry_test.cc

@@ -12,13 +12,16 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "paddle/framework/op_registry.h"
+#include <glog/logging.h>
 #include <gtest/gtest.h>
 
+#include "paddle/framework/op_registry.h"
+
 namespace pd = paddle::framework;
 
 namespace paddle {
 namespace framework {
+
 class CosineOp : public OperatorBase {
  public:
   using OperatorBase::OperatorBase;
@@ -215,7 +218,7 @@ class OpWithKernelTest : public OperatorWithKernel {
  protected:
   void InferShape(InferShapeContext* ctx) const override {}
 
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(proto::DataType::FP32, ctx.device_context());
   }
@@ -252,7 +255,6 @@ TEST(OperatorRegistrar, CPU) {
   op->Run(scope, cpu_place);
 }
 
-#ifdef PADDLE_WITH_CUDA
 TEST(OperatorRegistrar, CUDA) {
   paddle::framework::proto::OpDesc op_desc;
   paddle::platform::CUDAPlace cuda_place(0);
@@ -263,4 +265,127 @@ TEST(OperatorRegistrar, CUDA) {
 
   op->Run(scope, cuda_place);
 }
-#endif
+
+static int op_test_value = 0;
+
+using paddle::platform::DeviceContext;
+using paddle::platform::CPUDeviceContext;
+using paddle::platform::CUDADeviceContext;
+
+namespace paddle {
+namespace framework {
+
+class OpWithMultiKernelTest : public OperatorWithKernel {
+ public:
+  using OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(InferShapeContext* ctx) const override {}
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::OpKernelType(
+        proto::DataType::FP32, platform::CUDAPlace(0), DataLayout::kAnyLayout,
+        framework::LibraryType::kCUDNN);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class OpMultiKernelTest : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const;
+};
+
+template <typename T>
+class OpMultiKernelTest<CPUDeviceContext, T>
+    : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const {
+    ++op_test_value;
+  }
+};
+
+template <typename T>
+class OpMultiKernelTest<CUDADeviceContext, T>
+    : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const {
+    --op_test_value;
+  }
+};
+
+template <typename DeviceContext, typename T>
+class OpMultiKernelTest2 : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const;
+};
+
+template <typename T>
+class OpMultiKernelTest2<CPUDeviceContext, T>
+    : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const {
+    op_test_value += 10;
+  }
+};
+
+template <typename T>
+class OpMultiKernelTest2<CUDADeviceContext, T>
+    : public paddle::framework::OpKernel<T> {
+ public:
+  void Compute(const paddle::framework::ExecutionContext& ctx) const {
+    op_test_value -= 10;
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(op_with_multi_kernel,
+                             paddle::framework::OpWithMultiKernelTest,
+                             paddle::framework::OpKernelTestMaker);
+REGISTER_OP_KERNEL(
+    op_with_multi_kernel, CPU, paddle::platform::CPUPlace,
+    paddle::framework::OpMultiKernelTest<CPUDeviceContext, float>);
+REGISTER_OP_KERNEL(
+    op_with_multi_kernel, MKLDNN, paddle::platform::CPUPlace,
+    paddle::framework::OpMultiKernelTest2<CPUDeviceContext, float>);
+REGISTER_OP_KERNEL(
+    op_with_multi_kernel, CUDA, paddle::platform::CUDAPlace,
+    paddle::framework::OpMultiKernelTest<CUDADeviceContext, float>);
+REGISTER_OP_KERNEL(
+    op_with_multi_kernel, CUDNN, paddle::platform::CUDAPlace,
+    paddle::framework::OpMultiKernelTest2<CUDADeviceContext, float>);
+
+TEST(OperatorRegistrar, OpWithMultiKernel) {
+  paddle::framework::proto::OpDesc op_desc;
+  paddle::platform::CUDAPlace cuda_place(0);
+  paddle::platform::CPUPlace cpu_place;
+  paddle::framework::Scope scope;
+
+  op_desc.set_type("op_with_multi_kernel");
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
+
+  // TODO(qiao) add priority back
+  // use all available kernels
+  paddle::framework::UseALL();
+  op->Run(scope, cuda_place);
+  EXPECT_EQ(op_test_value, -10);
+
+  // remove cuda kernels
+  paddle::framework::UseCPU();
+  op->Run(scope, cpu_place);
+
+  EXPECT_EQ(op_test_value, -20);
+
+  // add cuda kernels
+  paddle::framework::UseCUDA();
+  op->Run(scope, cuda_place);
+
+  EXPECT_EQ(op_test_value, -30);
+
+  // use cudnn kernel
+  paddle::framework::UseCUDNN();
+  op->Run(scope, cuda_place);
+  EXPECT_EQ(op_test_value, -40);
+}

paddle/framework/operator.cc

@@ -11,13 +11,13 @@ 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 <glog/logging.h>
 
 #include <algorithm>
-#include <atomic>
 
 #include "paddle/framework/data_transform.h"
+#include "paddle/framework/device_data_transform.h"
 #include "paddle/framework/executor.h"
-#include "paddle/framework/lod_tensor_array.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/shape_inference.h"
 #include "paddle/framework/var_type.h"
@@ -25,6 +25,64 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority;
+
+void UseCPU() {
+  kKernelPriority.clear();
+  /*Plain CPU*/
+  auto pair0 = std::make_tuple(platform::CPUPlace(), LibraryType::kPlain);
+  kKernelPriority.insert(kKernelPriority.begin(), pair0);
+}
+
+void UseMKLDNN() {
+  UseCPU();
+#if PADDLE_WITH_MKLML
+  {
+    /*MKLDNN Kernel*/
+    auto pair0 = std::make_tuple(platform::CPUPlace(), LibraryType::kMKLDNN);
+    kKernelPriority.insert(kKernelPriority.begin(), pair0);
+  }
+#endif
+}
+
+void UseCUDA() {
+  UseMKLDNN();
+#if PADDLE_WITH_CUDA
+  /*Plain GPU*/
+  auto pair0 = std::make_tuple(platform::CUDAPlace(0), LibraryType::kPlain);
+  kKernelPriority.insert(kKernelPriority.begin(), pair0);
+#endif
+}
+
+void UseCUDNN() {
+  UseCUDA();
+#if PADDLE_WITH_CUDA
+  if (platform::dynload::HasCUDNN()) {
+    /*CUDNN Kernel*/
+    auto pair0 = std::make_tuple(platform::CUDAPlace(0), LibraryType::kCUDNN);
+    kKernelPriority.insert(kKernelPriority.begin(), pair0);
+  }
+#endif
+}
+
+void UseALL() {
+  UseCPU();
+  UseMKLDNN();
+  UseCUDA();
+  UseCUDNN();
+}
+
+static DDim GetDims(const Scope& scope, const std::string& name) {
+  Variable* var = scope.FindVar(name);
+  if (var->IsType<LoDTensor>()) {
+    return var->Get<LoDTensor>().dims();
+  } else if (var->IsType<SelectedRows>()) {
+    return var->Get<SelectedRows>().GetCompleteDims();
+  } else {
+    return DDim({-1});
+  }
+}
+
 std::string OperatorBase::Input(const std::string& name) const {
   auto& ins = Inputs(name);
   PADDLE_ENFORCE_LE(ins.size(), 1UL,
@@ -57,7 +115,7 @@ const std::vector<std::string>& OperatorBase::Outputs(
   return it->second;
 }
 
-std::string OperatorBase::DebugString() const {
+std::string OperatorBase::DebugStringEx(const Scope* scope) const {
   std::stringstream ss;
   ss << "Op(" << type_ << "), inputs:{";
   for (auto it = inputs_.begin(); it != inputs_.end();) {
@@ -65,6 +123,9 @@ std::string OperatorBase::DebugString() const {
     ss << input.first << "[";
     for (size_t i = 0; i < input.second.size(); ++i) {
       ss << input.second[i];
+      if (scope) {
+        ss << "(" << GetDims(*scope, input.second[i]) << ")";
+      }
       if (i != input.second.size() - 1) {
         ss << ", ";
       }
@@ -81,6 +142,9 @@ std::string OperatorBase::DebugString() const {
     ss << output.first << "[";
     for (size_t i = 0; i < output.second.size(); ++i) {
       ss << output.second[i];
+      if (scope) {
+        ss << "(" << GetDims(*scope, output.second[i]) << ")";
+      }
       if (i != output.second.size() - 1) {
         ss << ", ";
       }
@@ -178,6 +242,10 @@ void OperatorBase::GenerateTemporaryNames() {
   }
 }
 
+static bool VarIsTensor(const Variable* var) {
+  return var->IsType<LoDTensor>() || var->IsType<SelectedRows>();
+}
+
 static const Tensor* GetTensorFromVar(const Variable* var) {
   const Tensor* t = nullptr;
   if (var->IsType<LoDTensor>()) {
@@ -185,7 +253,8 @@ static const Tensor* GetTensorFromVar(const Variable* var) {
   } else if (var->IsType<SelectedRows>()) {
     t = &(var->Get<SelectedRows>().value());
   } else {
-    PADDLE_THROW("Variable type must be LoDTensor/SelectedRows.");
+    PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
+                 var->Type().name());
   }
   return t;
 }
@@ -197,7 +266,8 @@ static Tensor* GetMutableTensorFromVar(Variable* var) {
   } else if (var->IsType<SelectedRows>()) {
     t = var->GetMutable<SelectedRows>()->mutable_value();
   } else {
-    PADDLE_THROW("Variable type must be LoDTensor/SelectedRows.");
+    PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
+                 var->Type().name());
   }
   return t;
 }
@@ -347,6 +417,25 @@ class RuntimeInferShapeContext : public InferShapeContext {
     auto in_tensor = in_var->Get<LoDTensor>();
     auto* out_tensor = out_var->GetMutable<LoDTensor>();
     out_tensor->set_lod(in_tensor.lod());
+
+    // TODO(dzhwinter) : reuse ShareLoD in most operators.
+    // Need to call ShareLayout explicitly in sequence related ops.
+    // Shall we have a better method to shared info between in/out Tensor?
+    out_tensor->set_layout(in_tensor.layout());
+  }
+
+  void ShareLayout(const std::string& in, const std::string& out, size_t i = 0,
+                   size_t j = 0) const {
+    PADDLE_ENFORCE_LT(i, Inputs(in).size());
+    PADDLE_ENFORCE_LT(j, Outputs(out).size());
+    Variable* in_var = scope_.FindVar(Inputs(in)[i]);
+    Variable* out_var = scope_.FindVar(Outputs(out)[j]);
+    if (!in_var->IsType<LoDTensor>()) return;
+    PADDLE_ENFORCE(out_var->IsType<LoDTensor>(),
+                   "The %d-th output of Output(%s) must be LoDTensor.", j, out);
+    auto in_tensor = in_var->Get<LoDTensor>();
+    auto* out_tensor = out_var->GetMutable<LoDTensor>();
+    out_tensor->set_layout(in_tensor.layout());
   }
 
   bool IsRuntime() const override { return true; }
@@ -359,7 +448,8 @@ class RuntimeInferShapeContext : public InferShapeContext {
     } else if (var->IsType<SelectedRows>()) {
       return var->Get<SelectedRows>().GetCompleteDims();
     } else {
-      PADDLE_THROW("Variable type must be LoDTensor/SelectedRows.");
+      PADDLE_THROW("Variable %s type_id %s, expect LoDTensor/SelectedRows.",
+                   name, var->Type().name());
     }
   }
 
@@ -370,7 +460,8 @@ class RuntimeInferShapeContext : public InferShapeContext {
     } else if (var->IsType<SelectedRows>()) {
       var->GetMutable<SelectedRows>()->set_height(dim[0]);
     } else {
-      PADDLE_THROW("Variable type must be LoDTensor/SelectedRows.");
+      PADDLE_THROW("Variable %s type_id %s, expect LoDTensor/SelectedRows.",
+                   name, var->Type().name());
     }
   }
 
@@ -384,24 +475,6 @@ class RuntimeInferShapeContext : public InferShapeContext {
   const Scope& scope_;
 };
 
-const platform::DeviceContext* GetDeviceContext(
-    framework::KernelTypePair& kernel_pair) {
-  auto& actual_kernel_key = kernel_pair.first;
-  auto& expected_kernel_key = kernel_pair.second;
-  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-
-  if (platform::is_gpu_place(actual_kernel_key.place_) &&
-      platform::is_cpu_place(expected_kernel_key.place_)) {
-    return pool.Get(actual_kernel_key.place_);
-  } else if (platform::is_cpu_place(actual_kernel_key.place_) &&
-             platform::is_gpu_place(expected_kernel_key.place_)) {
-    return pool.Get(expected_kernel_key.place_);
-  } else {
-    PADDLE_THROW(
-        "Currently, model parallelism is only supported between CPU and CUDA");
-  }
-}
-
 void OperatorWithKernel::Run(const Scope& scope,
                              const platform::Place& place) const {
   RuntimeInferShapeContext infer_shape_ctx(*this, scope);
@@ -417,71 +490,43 @@ void OperatorWithKernel::Run(const Scope& scope,
         "There are no kernels which are registered in the %s operator.", type_);
   }
 
-  // check if op[type] have kernel for kernel_key
-  OpKernelMap& kernels = kernels_iter->second;
-
   ExecutionContext ctx(*this, scope, *dev_ctx);
-  auto actual_kernel_key = GetActualKernelType(ctx);
-  auto expected_kernel_key = GetExpectedKernelType(actual_kernel_key);
-  auto kernel_iter = kernels.find(expected_kernel_key);
-
-  if (kernel_iter == kernels.end()) {
-    PADDLE_THROW("The operator %s does not support %s", type_,
-                 expected_kernel_key);
-  }
-
-  if (actual_kernel_key == expected_kernel_key) {
-    PADDLE_ENFORCE_EQ(actual_kernel_key.place_, expected_kernel_key.place_,
-                      "Currently, model parallelism is only supported between "
-                      "CPU and other devices. For example, multi-GPU model "
-                      "parallelism will failed.");
-  } else {
-    auto kernel_pair = std::make_pair(actual_kernel_key, expected_kernel_key);
-    const DataTransformFn* trans_fun =
-        DataTransformFnMap::Instance().GetNullable(kernel_pair);
-    if (trans_fun) {
-      auto input_vars = this->InputVars();
-      // TODO(qijun) filter the input vars that do not need to be transformed
-
-      // filter vars that has been transformed
-      std::vector<std::string> need_trans;
-      for (auto var_name : input_vars) {
-        auto var_name_trans =
-            var_name + framework::KernelTypeToString(expected_kernel_key);
-        if (!scope.FindVar(var_name_trans)) {
-          const_cast<Scope&>(scope).Var(var_name_trans);
-          need_trans.push_back(var_name);
-        }
-      }
-
-      if (!need_trans.empty()) {
-        auto trans_dev_ctx = GetDeviceContext(kernel_pair);
-
-        // Wait for transform starting
-        dev_ctx->Wait();
-
-        for (auto var_name : need_trans) {
-          (*trans_fun)(trans_dev_ctx, kernel_pair, *(scope.FindVar(var_name)),
-                       scope.FindVar(var_name + framework::KernelTypeToString(
-                                                    expected_kernel_key)));
+  auto expected_kernel_key = this->GetExpectedKernelType(ctx);
+
+  Scope& new_scope = scope.NewScope();
+
+  for (auto& var_name_item : this->Inputs()) {
+    for (auto& var_name : var_name_item.second) {
+      auto* var = scope.FindVar(var_name);
+      if (var && VarIsTensor(var)) {
+        auto* tensor_in = GetTensorFromVar(var);
+        if (tensor_in->IsInitialized()) {
+          auto kernel_type_for_var = this->GetKernelTypeForVar(
+              var_name_item.first, *tensor_in, expected_kernel_key);
+          if (kernel_type_for_var != expected_kernel_key) {
+            auto out_var_names = OutputVars(true);
+            if (std::find(out_var_names.begin(), out_var_names.end(),
+                          var_name) != out_var_names.end()) {
+              PADDLE_THROW(
+                  "var %s is both input and output, "
+                  "does not support transform",
+                  var_name);
+            }
+            VLOG(3) << "need to do transform for var " << var_name;
+            auto* trans_var = new_scope.Var(var_name);
+            auto* out = DataTransform(expected_kernel_key, kernel_type_for_var,
+                                      *tensor_in);
+            CopyVariableWithTensor(*var, *out, *trans_var);
+          }
         }
-        // Wait for data transform finishing
-        trans_dev_ctx->Wait();
       }
     }
   }
 
-  kernel_iter->second->Compute(ctx);
-}
+  OpKernelMap& kernels = kernels_iter->second;
+  auto kernel_iter = kernels.find(expected_kernel_key);
 
-OpKernelType OperatorWithKernel::GetActualKernelType(
-    const ExecutionContext& ctx) const {
-  return OpKernelType(IndicateDataType(ctx), ctx.GetPlace());
-}
-
-OpKernelType OperatorWithKernel::GetExpectedKernelType(
-    const OpKernelType& actual_kernel_type) const {
-  return actual_kernel_type;
+  kernel_iter->second->Compute(ExecutionContext(*this, new_scope, *dev_ctx));
 }
 
 proto::DataType OperatorWithKernel::IndicateDataType(
@@ -513,5 +558,16 @@ proto::DataType OperatorWithKernel::IndicateDataType(
   return static_cast<proto::DataType>(data_type);
 }
 
+OpKernelType OperatorWithKernel::GetExpectedKernelType(
+    const ExecutionContext& ctx) const {
+  return OpKernelType(IndicateDataType(ctx), ctx.GetPlace());
+}
+
+OpKernelType OperatorWithKernel::GetKernelTypeForVar(
+    const std::string& var_name, const Tensor& tensor,
+    const OpKernelType& expected_kernel_type) const {
+  return OpKernelType(expected_kernel_type.data_type_, tensor.place());
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/operator.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <algorithm>
 #include <atomic>
 #include <string>
+#include <tuple>
 #include <unordered_map>
 #include <vector>
 
@@ -52,10 +53,33 @@ constexpr char kGradVarSuffix[] = "@GRAD";
 /// Variables with this suffix are supposed to be filled up with zeros.
 constexpr char kZeroVarSuffix[] = "@ZERO";
 
-// define some kernel hint
-const std::string kUseCPU = "use_cpu";
-const std::string kUseCUDNN = "use_cudnn";
-const std::string kUseMKLDNN = "use_mkldnn";
+// define some kernel priority
+extern std::vector<std::tuple<platform::Place, LibraryType>> kKernelPriority;
+
+/**
+ * @brief Use cpu kernel only
+ */
+void UseCPU();
+
+/**
+ * @brief Perfer MKLDNN kernel than Plain CPU kernel
+ */
+void UseMKLDNN();
+
+/**
+ * @brief Perfer CUDA kernel than Plain CPU kernel
+ */
+void UseCUDA();
+
+/**
+ * @brief Perfer cudnn kernel than Plain CUDA kernel
+ */
+void UseCUDNN();
+
+/**
+ * @brief Use all available kernels
+ */
+void UseALL();
 
 inline std::string GradVarName(const std::string& var_name) {
   return var_name + kGradVarSuffix;
@@ -84,7 +108,10 @@ class OperatorBase {
     return boost::get<T>(attrs_.at(name));
   }
 
-  virtual std::string DebugString() const;
+  /// if scope is not null, also show dimensions of arguments
+  virtual std::string DebugStringEx(const Scope* scope) const;
+
+  std::string DebugString() const { return DebugStringEx(nullptr); }
 
   /// Net will call this function to Run an op.
   virtual void Run(const Scope& scope, const platform::Place& place) const = 0;
@@ -381,9 +408,10 @@ class OperatorWithKernel : public OperatorBase {
   }
 
  protected:
-  virtual OpKernelType GetActualKernelType(const ExecutionContext& ctx) const;
-  virtual OpKernelType GetExpectedKernelType(
-      const OpKernelType& actual_kernel_type) const;
+  virtual OpKernelType GetExpectedKernelType(const ExecutionContext& ctx) const;
+  virtual OpKernelType GetKernelTypeForVar(
+      const std::string& var_name, const Tensor& tensor,
+      const OpKernelType& expected_kernel_type) const;
 
  private:
   // indicate kernel DataType by input data. Defaultly all input data must be

paddle/framework/operator_test.cc

@@ -114,7 +114,8 @@ class OpWithKernelTest : public OperatorWithKernel {
 
  protected:
   void InferShape(framework::InferShapeContext* ctx) const override {}
-  OpKernelType GetActualKernelType(const ExecutionContext& ctx) const override {
+  OpKernelType GetExpectedKernelType(
+      const ExecutionContext& ctx) const override {
     return OpKernelType(proto::DataType::FP32, ctx.GetPlace());
   }
 };

paddle/framework/scope.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <memory>  // for unique_ptr
 #include <mutex>   // for call_once
 #include "glog/logging.h"
+#include "paddle/framework/threadpool.h"
 #include "paddle/string/printf.h"
 
 namespace paddle {
@@ -87,7 +88,8 @@ 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;
+  // Make delete async.
+  Async([scope] { delete scope; });
 }
 
 void Scope::Rename(const std::string& origin_name,
@@ -107,6 +109,7 @@ std::string Scope::Rename(const std::string& origin_name) const {
   Rename(origin_name, var_name);
   return var_name;
 }
+
 Variable* Scope::FindVarLocally(const std::string& name) const {
   auto it = vars_.find(name);
   if (it != vars_.end()) return it->second;

paddle/framework/scope.h

@@ -75,9 +75,9 @@ class Scope {
   // Rename variable to a new name and return the new name
   std::string Rename(const std::string& origin_name) const;
 
- private:
   Variable* FindVarLocally(const std::string& name) const;
 
+ private:
   // Call Scope::NewScope for a sub-scope.
   explicit Scope(Scope const* parent) : parent_(parent) {}
 

paddle/framework/selected_rows.cc

@@ -37,8 +37,8 @@ void SerializeToStream(std::ostream& os, const SelectedRows& selected_rows,
   SerializeToStream(os, selected_rows.value(), dev_ctx);
 }
 
-void DeserializeFromStream(std::istream& is, SelectedRows* selected_rows) {
-  auto tensor = *selected_rows->mutable_value();
+void DeserializeFromStream(std::istream& is, SelectedRows* selected_rows,
+                           const platform::DeviceContext& dev_ctx) {
   {
     // the 1st field, unit32_t version for SelectedRows
     uint32_t version;
@@ -62,7 +62,7 @@ void DeserializeFromStream(std::istream& is, SelectedRows* selected_rows) {
     selected_rows->set_height(height);
   }
   // the 4st field, tensor which contains the data
-  DeserializeFromStream(is, &tensor);
+  DeserializeFromStream(is, selected_rows->mutable_value(), dev_ctx);
 }
 
 }  // namespace framework

paddle/framework/selected_rows.h

@@ -66,7 +66,8 @@ class SelectedRows {
  */
 void SerializeToStream(std::ostream& os, const SelectedRows& selected_rows,
                        const platform::DeviceContext& dev_ctx);
-void DeserializeFromStream(std::istream& is, SelectedRows* selected_rows);
+void DeserializeFromStream(std::istream& is, SelectedRows* selected_rows,
+                           const platform::DeviceContext& dev_ctx);
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/selected_rows_test.cc

@@ -51,10 +51,12 @@ TEST_F(SelectedRowsTester, SerializeAndDeseralize) {
   SerializeToStream(oss, *selected_rows_, cpu_ctx);
 
   std::istringstream iss(oss.str());
-  DeserializeFromStream(iss, &dst_tensor);
+  DeserializeFromStream(iss, &dst_tensor, cpu_ctx);
 
   ASSERT_EQ(selected_rows_->rows(), dst_tensor.rows());
   ASSERT_EQ(selected_rows_->height(), dst_tensor.height());
+  ASSERT_EQ(selected_rows_->value().dims(), dst_tensor.value().dims());
+  ASSERT_EQ(selected_rows_->GetCompleteDims(), dst_tensor.GetCompleteDims());
 }
 
 }  // namespace framework

paddle/framework/tensor.h

@@ -55,6 +55,10 @@ class Tensor {
   template <typename T>
   inline const T* data() const;
 
+  inline bool IsInitialized() const;
+
+  inline void switch_place(platform::Place new_place);
+
   /**
    * @brief   Return a pointer to mutable memory block.
    * @note    If not exist, then allocation.
@@ -200,6 +204,15 @@ class Tensor {
   size_t offset_;
 };
 
+inline void Tensor::switch_place(platform::Place new_place) {
+  if (holder_->place() == new_place) {
+    return;
+  }
+
+  // TODO(tonyyang-svail): do memcpy here.
+  PADDLE_THROW("Not Implemented");
+}
+
 }  // namespace framework
 }  // namespace paddle
 

paddle/framework/tensor_impl.h

@@ -84,6 +84,8 @@ inline const T* Tensor::data() const {
       reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
 }
 
+inline bool Tensor::IsInitialized() const { return holder_ != nullptr; }
+
 template <typename T>
 inline T* Tensor::data() {
   check_memory_size();

paddle/framework/tensor_util.h

@@ -270,7 +270,23 @@ inline void SerializeToStream(std::ostream& os, const Tensor& tensor,
   }
 }
 
-inline void DeserializeFromStream(std::istream& is, Tensor* tensor) {
+struct DeserializedDataFunctor {
+  DeserializedDataFunctor(void** buf, Tensor* tensor,
+                          const platform::Place& place)
+      : buf_(buf), tensor_(tensor), place_(place) {}
+
+  template <typename T>
+  void operator()() {
+    *buf_ = tensor_->mutable_data<T>(place_);
+  }
+
+  void** buf_;
+  Tensor* tensor_;
+  platform::Place place_;
+};
+
+inline void DeserializeFromStream(std::istream& is, Tensor* tensor,
+                                  const platform::DeviceContext& dev_ctx) {
   uint32_t version;
   is.read(reinterpret_cast<char*>(&version), sizeof(version));
   PADDLE_ENFORCE_EQ(version, 0U, "Only version 0 is supported");
@@ -289,27 +305,28 @@ inline void DeserializeFromStream(std::istream& is, Tensor* tensor) {
     dims.reserve(static_cast<size_t>(desc.dims().size()));
     std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims));
     tensor->Resize(framework::make_ddim(dims));
-
     void* buf;
-    platform::Place cpu = platform::CPUPlace();
-    // TODO(Yancey1989): use VisiterDataType instead of DataType switch
-    switch (desc.data_type()) {
-      case proto::FP32:
-        buf = tensor->mutable_data<float>(cpu);
-        break;
-      case proto::FP64:
-        buf = tensor->mutable_data<double>(cpu);
-        break;
-      case proto::INT32:
-        buf = tensor->mutable_data<int>(cpu);
-        break;
-      case proto::INT64:
-        buf = tensor->mutable_data<int64_t>(cpu);
-        break;
-      default:
-        PADDLE_THROW("DataType %d not supported", desc.data_type());
+    auto ctx = platform::CPUDeviceContext();
+    if (platform::is_gpu_place(dev_ctx.GetPlace())) {
+#ifdef PADDLE_WITH_CUDA
+      Tensor cpu_tensor;
+      cpu_tensor.Resize(framework::make_ddim(dims));
+      framework::VisitDataType(
+          desc.data_type(),
+          DeserializedDataFunctor(&buf, &cpu_tensor, ctx.GetPlace()));
+      is.read(static_cast<char*>(buf), cpu_tensor.memory_size());
+      auto cpu_place = new platform::CPUPlace();
+      framework::CopyFrom(cpu_tensor, *cpu_place, dev_ctx, tensor);
+      delete cpu_place;
+#else
+      PADDLE_THROW("Unexpected branch");
+#endif
+    } else {
+      framework::VisitDataType(
+          desc.data_type(),
+          DeserializedDataFunctor(&buf, tensor, ctx.GetPlace()));
+      is.read(static_cast<char*>(buf), tensor->memory_size());
     }
-    is.read(static_cast<char*>(buf), tensor->memory_size());
   }
 }
 

paddle/framework/tensor_util_test.cc

@@ -270,11 +270,12 @@ TEST(Tensor, SerializeAndDeserialize) {
     SerializeToStream(oss, src_tensor, cpu_ctx);
 
     std::istringstream iss(oss.str());
-    DeserializeFromStream(iss, &dst_tensor);
+    DeserializeFromStream(iss, &dst_tensor, cpu_ctx);
     int* dst_ptr = dst_tensor.mutable_data<int>(platform::CPUPlace());
     for (int i = 0; i < 5; ++i) {
       ASSERT_EQ(dst_ptr[i], array[i]);
     }
+    ASSERT_EQ(dst_tensor.dims(), src_tensor.dims());
     delete place;
   }
 #ifdef PADDLE_WITH_CUDA
@@ -292,13 +293,12 @@ TEST(Tensor, SerializeAndDeserialize) {
     SerializeToStream(oss, gpu_tensor, gpu_ctx);
 
     std::istringstream iss(oss.str());
-    DeserializeFromStream(iss, &dst_tensor);
+    DeserializeFromStream(iss, &dst_tensor, gpu_ctx);
 
     int* dst_ptr = dst_tensor.mutable_data<int>(platform::CPUPlace());
     for (int i = 0; i < 6; ++i) {
       ASSERT_EQ(dst_ptr[i], array[i]);
     }
-
     delete gpu_place;
   }
 #endif

paddle/framework/threadpool.h

@@ -29,7 +29,6 @@ namespace framework {
 class ThreadPool {
  public:
   typedef std::packaged_task<void()> Task;
-  typedef std::function<void()> Fun;
 
   /**
    * @brief   Get a instance of threadpool, the thread number will
@@ -67,7 +66,8 @@ class ThreadPool {
    * @return    std::future<void>, we could wait for the task finished by
    *            f.wait().
    */
-  std::future<void> Run(const Fun& fn) {
+  template <typename Callback>
+  std::future<void> Run(Callback fn) {
     std::unique_lock<std::mutex> lock(mutex_);
     Task task(std::bind(fn));
     std::future<void> f = task.get_future();
@@ -159,5 +159,13 @@ class ThreadPool {
   std::condition_variable completed_;
 };
 
+// Run a function asynchronously.
+// NOTE: The function must return void. If the function need to return a value,
+// you can use lambda to capture a value pointer.
+template <typename Callback>
+std::future<void> Async(Callback callback) {
+  return ThreadPool::GetInstance()->Run(callback);
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/var_desc.cc

@@ -52,7 +52,7 @@ void VarDesc::SetLoDLevel(int32_t lod_level) {
   }
 }
 
-int32_t VarDesc::GetLodLevel() const {
+int32_t VarDesc::GetLoDLevel() const {
   switch (desc_.type()) {
     case proto::VarDesc::LOD_TENSOR:
       return desc_.lod_tensor().lod_level();

paddle/framework/var_desc.h

@@ -76,7 +76,7 @@ class VarDesc {
 
   void SetLoDLevel(int32_t lod_level);
 
-  int32_t GetLodLevel() const;
+  int32_t GetLoDLevel() const;
 
   proto::VarDesc::VarType GetType() const;
 

paddle/framework/var_type.h

@@ -17,6 +17,8 @@ limitations under the License. */
 #include "paddle/framework/lod_rank_table.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/lod_tensor_array.h"
+#include "paddle/framework/selected_rows.h"
+#include "paddle/framework/variable.h"
 
 namespace paddle {
 namespace framework {
@@ -35,7 +37,7 @@ inline proto::VarDesc::VarType ToVarType(std::type_index type) {
 }
 
 template <typename Visitor>
-inline void VisitVarType(const Variable& var, Visitor visitor) {
+inline void VisitVarType(const framework::Variable& var, Visitor visitor) {
   switch (ToVarType(var.Type())) {
     case proto::VarDesc_VarType_LOD_TENSOR:
       visitor(var.Get<framework::LoDTensor>());

paddle/framework/variable.h

@@ -32,6 +32,8 @@ class Variable {
     return *static_cast<const T*>(holder_->Ptr());
   }
 
+  bool IsInitialized() const { return holder_ != nullptr; }
+
   template <typename T>
   T* GetMutable() {
     if (!IsType<T>()) {

paddle/gserver/CMakeLists.txt

@@ -34,6 +34,16 @@ else()
     message(STATUS "Compile with MKLDNNLayers and MKLDNNActivations")
 endif()
 
+if(NOT WITH_MKLML)
+    file(GLOB_RECURSE MKL_HEADER RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "MKLPacked*.h")
+    file(GLOB_RECURSE MKL_SOURCES RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "MKLPacked*.cpp")
+    list(REMOVE_ITEM GSERVER_HEADER ${MKL_HEADER})
+    list(REMOVE_ITEM GSERVER_SOURCES ${MKL_SOURCES})
+    message(STATUS "Skip compiling with MKLPackedLayers")
+else()
+    message(STATUS "Compile with MKLPackedLayers")
+endif()
+
 if(NOT WITH_GPU)
     list(REMOVE_ITEM GSERVER_HEADER
         layers/CudnnConvBaseLayer.h

paddle/gserver/layers/MKLPackedRecurrentLayer.cpp

+/* 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. */
+
+#include "MKLPackedRecurrentLayer.h"
+
+namespace paddle {
+
+REGISTER_LAYER(mkl_packed_recurrent, MKLPackedRecurrentLayer);
+
+bool MKLPackedRecurrentLayer::init(const LayerMap& layerMap,
+                                   const ParameterMap& parameterMap) {
+  if (!RecurrentLayer::init(layerMap, parameterMap)) return false;
+  packed_weight_.reset(new MKLPackedWeight(weight_->getW()));
+  packed_weight_->pack();
+  if (needGradient_) {
+    packed_weightT_.reset(new MKLPackedWeight(weight_->getW(), true));
+    packed_weightT_->pack();
+  }
+  return true;
+}
+
+void MKLPackedRecurrentLayer::backward(const UpdateCallback& callback) {
+  RecurrentLayer::backward(callback);
+  packed_weight_->pack();
+  if (needGradient_) {
+    packed_weightT_->pack();
+  }
+}
+
+void MKLPackedRecurrentLayer::forwardBatch(int batchSize,
+                                           size_t numSequences,
+                                           const int* starts) {
+  if (!batchValue_) {
+    batchValue_.reset(new SequenceToBatch(useGpu_));
+  }
+
+  batchValue_->resizeOrCreateBatch(batchSize, numSequences, starts, reversed_);
+
+  batchValue_->copyFromSeq(*output_.value);
+
+  {
+    REGISTER_TIMER_INFO("RecurrentFwBatch", getName().c_str());
+    /* forward one batch */
+    for (size_t n = 0; n < batchValue_->getNumBatch(); n++) {
+      MatrixPtr batchValue = batchValue_->getBatchValue(n);
+
+      if (n != 0) {
+        MatrixPtr preBatchValue =
+            batchValue_->getBatchValue(n - 1, batchValue->getHeight());
+
+        packed_weight_->gemm_compute(preBatchValue, batchValue);
+      }
+      Argument arg;
+      arg.value = batchValue;
+      activation_->forward(arg).check();
+    }
+  }
+  batchValue_->copyBackSeq(*output_.value);
+}
+
+void MKLPackedRecurrentLayer::backwardBatch(int batchSize,
+                                            size_t numSequences,
+                                            const int* starts) {
+  if (!batchGrad_) {
+    batchGrad_.reset(new SequenceToBatch(useGpu_));
+  }
+  batchGrad_->shareIndexWith(*batchValue_);
+
+  size_t numBatch = batchGrad_->getNumBatch();
+  bool backwardByBatch = numBatch < numSequences;
+
+  batchGrad_->copyFromSeq(*output_.grad);
+  {
+    REGISTER_TIMER_INFO("RecurrentBwData", getName().c_str());
+    /* backward one batch */
+    for (int n = (int)numBatch - 1; n >= 0; n--) {
+      MatrixPtr batchGrad = batchGrad_->getBatchValue(n);
+      MatrixPtr batchValue =
+          batchValue_->getBatchValue(n, batchGrad->getHeight());
+
+      Argument arg;
+      arg.value = batchValue;
+      arg.grad = batchGrad;
+      activation_->backward(arg).check();
+
+      if (n != 0) {
+        batchValue = batchGrad_->getBatchValue(n - 1, batchGrad->getHeight());
+        packed_weightT_->gemm_compute(batchGrad, batchValue);
+      }
+
+      if (backwardByBatch && weight_->getWGrad()) {
+        if (n != 0) {
+          /* backward weight */
+          batchValue =
+              batchValue_->getBatchValue(n - 1, batchGrad->getHeight());
+          weight_->getWGrad()->mul(
+              *batchValue->getTranspose(), *batchGrad, 1, 1);
+        }
+      }
+    }
+  }
+
+  batchGrad_->copyBackSeq(*output_.grad);
+
+  if (!backwardByBatch && weight_->getWGrad()) {
+    REGISTER_TIMER_INFO("RecurrentBwWeight", getName().c_str());
+    for (size_t seq = 0; seq < numSequences; ++seq) {
+      int len = starts[seq + 1] - starts[seq];
+      weight_->getWGrad()->mul(
+          *output_.value
+               ->subMatrix(reversed_ ? starts[seq] + 1 : starts[seq], len - 1)
+               ->getTranspose(),
+          *output_.grad->subMatrix(reversed_ ? starts[seq] : starts[seq] + 1,
+                                   len - 1),
+          1,
+          1);
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/MKLPackedRecurrentLayer.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
+
+#include "MKLPackedWeight.h"
+#include "RecurrentLayer.h"
+
+DECLARE_bool(rnn_use_batch);
+
+namespace paddle {
+
+/**
+ * @brief MKLPackedRecurrentLayer is almost the same with RecurrentLayer
+ * but is optimized with MKL cblas packed gemm.
+ * More details:
+ * https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/mkl/mkl_packed.md
+ */
+
+class MKLPackedRecurrentLayer : public RecurrentLayer {
+public:
+  explicit MKLPackedRecurrentLayer(const LayerConfig& config)
+      : RecurrentLayer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void backward(const UpdateCallback& callback) override;
+
+protected:
+  void forwardBatch(int batchSize,
+                    size_t numSequences,
+                    const int* starts) override;
+
+  void backwardBatch(int batchSize,
+                     size_t numSequences,
+                     const int* starts) override;
+
+protected:
+  /// packed_weight_ contains same data with
+  /// RecurrentLayer::weight_ but is packed
+  std::unique_ptr<MKLPackedWeight> packed_weight_;
+  /// packed_weightT_ is the transposition matrix of packed_weight_
+  std::unique_ptr<MKLPackedWeight> packed_weightT_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/MKLPackedWeight.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
+
+#include "paddle/math/MathFunctions.h"
+#include "paddle/parameter/Parameter.h"
+#include "paddle/parameter/Weight.h"
+
+namespace paddle {
+
+class MKLPackedWeight {
+protected:
+  /// The pointer of weight
+  real *weight_;
+  /// The pointer of cblas packed gemm to weight
+  real *packedWeight_;
+  size_t height_;
+  size_t width_;
+  bool transW_;
+
+public:
+  explicit MKLPackedWeight(MatrixPtr weight, bool transW = false) {
+    packedWeight_ = nullptr;
+    weight_ = weight->getData();
+    height_ = weight->getHeight();
+    width_ = weight->getWidth();
+    transW_ = transW;
+  }
+
+  ~MKLPackedWeight() { free_(); }
+
+  void pack() { pack_(weight_); }
+
+  void gemm_compute(const MatrixPtr src, MatrixPtr dst) {
+    cblas_sgemm_compute(CblasRowMajor,
+                        CblasNoTrans,
+                        CblasPacked,
+                        src->getHeight(),
+                        transW_ ? height_ : width_,
+                        transW_ ? width_ : height_,
+                        src->getData(),
+                        src->getWidth(),
+                        packedWeight_,
+                        width_,
+                        1.0,
+                        dst->getData(),
+                        dst->getWidth());
+  }
+
+protected:
+  void pack_(real *src) {
+    if (!packedWeight_) {
+      packedWeight_ = cblas_sgemm_alloc(CblasBMatrix, 1, width_, height_);
+    }
+    cblas_sgemm_pack(CblasRowMajor,
+                     CblasBMatrix,
+                     transW_ ? CblasTrans : CblasNoTrans,
+                     1,
+                     transW_ ? height_ : width_,
+                     transW_ ? width_ : height_,
+                     1.0,
+                     src,
+                     width_,
+                     packedWeight_);
+  }
+
+  void free_() {
+    if (packedWeight_) {
+      cblas_sgemm_free(packedWeight_);
+    }
+  }
+};
+
+}  // namespace paddle

paddle/gserver/layers/RecurrentLayer.cpp

@@ -12,119 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include <gflags/gflags.h>
-#include "Layer.h"
-#include "SequenceToBatch.h"
-#include "paddle/utils/Stat.h"
+#include "RecurrentLayer.h"
 
 DEFINE_bool(rnn_use_batch, false, "Using the batch method for calculation.");
 
 namespace paddle {
 
-/**
- * @brief RecurrentLayer takes 1 input layer. The output size is the same with
- * input layer.
- * For each sequence [start, end] it performs the following computation:
- * \f[
- *    out_{i} = act(in_{i})     \      \      \text{for} \ i = start \\
- *    out_{i} = act(in_{i} + out_{i-1} * W) \ \ \text{for} \ start < i <= end
- *
- * \f]
- * If reversed is true, the order is reversed:
- * \f[
- *   out_{i} = act(in_{i})           \    \   \text{for} \ i = end  \\
- *   out_{i} = act(in_{i} + out_{i+1} * W) \ \ \text{for} \ start <= i < end
- * \f]
- * There are two methods to calculate rnn. One way is to compute rnn one
- * sequence by one sequence. The other way is to reorganize the input
- * into batches, then compute rnn one batch by one batch. Users can select
- * them by rnn_use_batch flag.
- */
-
-class RecurrentLayer : public Layer {
-public:
-  explicit RecurrentLayer(const LayerConfig& config) : Layer(config) {}
-
-  bool init(const LayerMap& layerMap,
-            const ParameterMap& parameterMap) override;
-
-  void forward(PassType passType) override;
-
-  void backward(const UpdateCallback& callback) override;
-
-  void resetState() override;
-
-  void setState(LayerStatePtr state) override;
-
-  LayerStatePtr getState() override;
-
-protected:
-  /**
-   * @brief If user do not set --rnn_use_batch=true, it will
-   * compute rnn forward one sequence by one sequence in default.
-   * @param batchSize Total words number of all samples in this batch.
-   * @param numSequences The sample number.
-   * @param starts Each start position of each samples.
-   */
-  void forwardSequence(int batchSize, size_t numSequences, const int* starts);
-  /**
-   * @brief Compute rnn forward by one sequence.
-   * @param start The start position of this sequence (or sample).
-   * @param length The length of this sequence (or sample), namely the words
-   * number of this sequence.
-   */
-  void forwardOneSequence(int start, int length);
-  /**
-   * @brief Compute rnn backward one sequence by onesequence.
-   * @param batchSize Total words number of all samples in this batch.
-   * @param numSequences The sample number.
-   * @param starts Each start position of each samples.
-   */
-  void backwardSequence(int batchSize, size_t numSequences, const int* starts);
-  /**
-   * @brief Compute rnn backward by one sequence.
-   * @param start The start position of this sequence (or sample).
-   * @param length The length of this sequence (or sample), namely the words
-   * number of this sequence.
-   */
-  void backwardOneSequence(int start, int length);
-
-  /**
-   * @brief Reorganize input into batches and compute rnn forward batch
-   * by batch. It will convert batch shape to sequence after finishing forward.
-   * The batch info can refer to SequenceToBatch class.
-   * @param batchSize Total words number of all samples in this batch.
-   * @param numSequences The sample number.
-   * @param starts Each start position of each samples.
-   */
-  void forwardBatch(int batchSize, size_t numSequences, const int* starts);
-
-  /**
-   * @brief Reorganize input into batches and compute rnn forward batch
-   * by batch.
-   * @param batchSize Total words number of all samples in this batch.
-   * @param numSequences The sample number.
-   * @param starts Each start position of each samples.
-   */
-  void backwardBatch(int batchSize, size_t numSequences, const int* starts);
-
-protected:
-  std::unique_ptr<Weight> weight_;
-  std::unique_ptr<Weight> bias_;
-
-  /// frameOutput_[i] is used to hold the i-th sample of output_
-  std::vector<Argument> frameOutput_;
-  MatrixPtr prevOutput_;
-  /// Whether compute rnn by reverse.
-  bool reversed_;
-  /// If compute batch by batch, batchValue_ will be used to save the
-  /// reorganized input value.
-  std::unique_ptr<SequenceToBatch> batchValue_;
-  /// If compute batch by batch, batchGrad_ will be used to save the
-  /// gradient with respect to reorganized input value.
-  std::unique_ptr<SequenceToBatch> batchGrad_;
-};
-
 REGISTER_LAYER(recurrent, RecurrentLayer);
 
 bool RecurrentLayer::init(const LayerMap& layerMap,
@@ -260,7 +153,6 @@ void RecurrentLayer::backward(const UpdateCallback& callback) {
     bias_->getWGrad()->collectBias(*output_.grad, 1);
     bias_->getParameterPtr()->incUpdate(callback);
   }
-
   weight_->getParameterPtr()->incUpdate(callback);
 }
 

paddle/gserver/layers/RecurrentLayer.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
+#include <gflags/gflags.h>
+#include "Layer.h"
+#include "SequenceToBatch.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+/**
+ * @brief RecurrentLayer takes 1 input layer. The output size is the same with
+ * input layer.
+ * For each sequence [start, end] it performs the following computation:
+ * \f[
+ *    out_{i} = act(in_{i})     \      \      \text{for} \ i = start \\
+ *    out_{i} = act(in_{i} + out_{i-1} * W) \ \ \text{for} \ start < i <= end
+ *
+ * \f]
+ * If reversed is true, the order is reversed:
+ * \f[
+ *   out_{i} = act(in_{i})           \    \   \text{for} \ i = end  \\
+ *   out_{i} = act(in_{i} + out_{i+1} * W) \ \ \text{for} \ start <= i < end
+ * \f]
+ * There are two methods to calculate rnn. One way is to compute rnn one
+ * sequence by one sequence. The other way is to reorganize the input
+ * into batches, then compute rnn one batch by one batch. Users can select
+ * them by rnn_use_batch flag.
+ */
+
+class RecurrentLayer : public Layer {
+public:
+  explicit RecurrentLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+
+  void backward(const UpdateCallback& callback) override;
+
+  void resetState() override;
+
+  void setState(LayerStatePtr state) override;
+
+  LayerStatePtr getState() override;
+
+protected:
+  /**
+   * @brief If user do not set --rnn_use_batch=true, it will
+   * compute rnn forward one sequence by one sequence in default.
+   * @param batchSize Total words number of all samples in this batch.
+   * @param numSequences The sample number.
+   * @param starts Each start position of each samples.
+   */
+  void forwardSequence(int batchSize, size_t numSequences, const int* starts);
+  /**
+   * @brief Compute rnn forward by one sequence.
+   * @param start The start position of this sequence (or sample).
+   * @param length The length of this sequence (or sample), namely the words
+   * number of this sequence.
+   */
+  void forwardOneSequence(int start, int length);
+  /**
+   * @brief Compute rnn backward one sequence by onesequence.
+   * @param batchSize Total words number of all samples in this batch.
+   * @param numSequences The sample number.
+   * @param starts Each start position of each samples.
+   */
+  void backwardSequence(int batchSize, size_t numSequences, const int* starts);
+  /**
+   * @brief Compute rnn backward by one sequence.
+   * @param start The start position of this sequence (or sample).
+   * @param length The length of this sequence (or sample), namely the words
+   * number of this sequence.
+   */
+  void backwardOneSequence(int start, int length);
+
+  /**
+   * @brief Reorganize input into batches and compute rnn forward batch
+   * by batch. It will convert batch shape to sequence after finishing forward.
+   * The batch info can refer to SequenceToBatch class.
+   * @param batchSize Total words number of all samples in this batch.
+   * @param numSequences The sample number.
+   * @param starts Each start position of each samples.
+   */
+  virtual void forwardBatch(int batchSize,
+                            size_t numSequences,
+                            const int* starts);
+
+  /**
+   * @brief Reorganize input into batches and compute rnn forward batch
+   * by batch.
+   * @param batchSize Total words number of all samples in this batch.
+   * @param numSequences The sample number.
+   * @param starts Each start position of each samples.
+   */
+  virtual void backwardBatch(int batchSize,
+                             size_t numSequences,
+                             const int* starts);
+
+protected:
+  std::unique_ptr<Weight> weight_;
+  std::unique_ptr<Weight> bias_;
+
+  /// frameOutput_[i] is used to hold the i-th sample of output_
+  std::vector<Argument> frameOutput_;
+  MatrixPtr prevOutput_;
+  /// Whether compute rnn by reverse.
+  bool reversed_;
+  /// If compute batch by batch, batchValue_ will be used to save the
+  /// reorganized input value.
+  std::unique_ptr<SequenceToBatch> batchValue_;
+  /// If compute batch by batch, batchGrad_ will be used to save the
+  /// gradient with respect to reorganized input value.
+  std::unique_ptr<SequenceToBatch> batchGrad_;
+};
+
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1472,7 +1472,8 @@ TEST(Layer, RecurrentLayer) {
     for (auto reversed : {false, true}) {
       config.layerConfig.set_reversed(reversed);
       config.testState = !reversed;
-      testLayerGrad(config, "recurrent", 50, /* trans= */ false, useGpu);
+      testLayerGrad(
+          config, "recurrent", 50, /* trans= */ false, useGpu, false, 1.0);
     }
   }
 }
@@ -1494,7 +1495,8 @@ TEST(Layer, LstmLayer) {
     for (auto reversed : {false, true}) {
       config.layerConfig.set_reversed(reversed);
       config.testState = !reversed;
-      testLayerGrad(config, "lstmemory", 100, /* trans= */ false, useGpu);
+      testLayerGrad(
+          config, "lstmemory", 100, /* trans= */ false, useGpu, false, 0.02);
     }
   }
   for (auto useGpu : {true}) {

paddle/gserver/tests/test_RecurrentLayer.cpp

@@ -222,6 +222,7 @@ TEST(Layer, RecurrentLayer) {
 #define protected public
 #include "paddle/gserver/layers/GatedRecurrentLayer.h"
 #include "paddle/gserver/layers/LstmLayer.h"
+#include "paddle/gserver/layers/RecurrentLayer.h"
 template <class T>
 class TestRecurrentLayer {
 public:
@@ -420,12 +421,151 @@ TEST(Layer, LstmLayer) {
   }
 }
 
+#ifdef PADDLE_WITH_MKLML
+
+#include "paddle/gserver/layers/MKLPackedRecurrentLayer.h"
+
+LayerPtr initMKLPackedLayer(LayerConfig layerConfig,
+                            bool reversed,
+                            int layerSize,
+                            LayerPtr dataLayer,
+                            ParameterPtr para,
+                            ParameterPtr bias = nullptr) {
+  LayerMap layerMap;
+  ParameterMap parameterMap;
+  layerMap[dataLayer->getName()] = dataLayer;
+  parameterMap[para->getName()] = para;
+  if (bias) {
+    parameterMap[bias->getName()] = bias;
+    layerConfig.set_bias_parameter_name("bias_0");
+  }
+
+  layerConfig.set_size(layerSize);
+  layerConfig.set_reversed(reversed);
+  layerConfig.add_inputs();
+  LayerInputConfig& input = *(layerConfig.mutable_inputs(0));
+  input.set_input_layer_name("layer_0");
+  input.set_input_parameter_name("para_0");
+
+  LayerPtr testLayer = Layer::create(layerConfig);
+  layerMap[testLayer->getName()] = testLayer;
+
+  testLayer->init(layerMap, parameterMap);
+  testLayer->setNeedGradient(true);
+
+  return testLayer;
+}
+
+void checkMKLPackedLayer(LayerConfig layerConfig1,
+                         LayerConfig layerConfig2,
+                         bool reversed,
+                         int layerSize,
+                         int batchSize,
+                         bool useBatch1,
+                         bool useBatch2) {
+  LayerPtr dataLayer;
+  ParameterPtr para, bias;
+
+  if (layerConfig1.type() == "recurrent") {
+    dataLayer = creatDataLayer("layer_0", batchSize, layerSize, false);
+    para = creatParameter("para_0", 0, layerSize * layerSize, false);
+    bias = nullptr;
+  } else if (layerConfig1.type() == "gated_recurrent") {
+    dataLayer = creatDataLayer("layer_0", batchSize, layerSize * 3, false);
+    para = creatParameter("para_0", 0, layerSize * layerSize * 3, false);
+    bias = creatParameterBias("bias_0", 1, layerSize * 3, false);
+  }
+
+  LayerPtr testLayer1 = initMKLPackedLayer(
+      layerConfig1, reversed, layerSize, dataLayer, para, bias);
+  LayerPtr testLayer2 = initMKLPackedLayer(
+      layerConfig2, reversed, layerSize, dataLayer, para, bias);
+
+  const VectorPtr& weightGrad =
+      (testLayer1->getParameters()[0])->getBuf(PARAMETER_GRADIENT);
+  const MatrixPtr& inputGrad = testLayer1->getPrev(0)->getOutputGrad();
+  CpuVector wgt_grad1(weightGrad->getSize());
+  CpuVector wgt_grad2(weightGrad->getSize());
+  CpuMatrix input_grad1(inputGrad->getHeight(), inputGrad->getWidth());
+  CpuMatrix input_grad2(inputGrad->getHeight(), inputGrad->getWidth());
+
+  for (int i = 0; i < 2; i++) {
+    FLAGS_rnn_use_batch = useBatch1;
+
+    testLayer1->forward(PASS_GC);
+
+    FLAGS_rnn_use_batch = useBatch2;
+    testLayer2->forward(PASS_GC);
+
+    testLayer1->getOutputGrad()->randomizeUniform();
+    testLayer2->getOutputGrad()->copyFrom(*testLayer1->getOutputGrad());
+
+    weightGrad->zero();
+    inputGrad->zero();
+    FLAGS_rnn_use_batch = useBatch1;
+    testLayer1->backward(nullptr);
+
+    wgt_grad1.copyFrom(*weightGrad);
+    input_grad1.copyFrom(*inputGrad);
+
+    weightGrad->zero();
+    inputGrad->zero();
+    FLAGS_rnn_use_batch = useBatch2;
+    testLayer2->backward(nullptr);
+
+    wgt_grad2.copyFrom(*weightGrad);
+    input_grad2.copyFrom(*inputGrad);
+
+    checkError(*testLayer1->getOutputValue(), *testLayer2->getOutputValue());
+    checkError(wgt_grad1, wgt_grad2);
+    checkError(input_grad1, input_grad2);
+  }
+}
+
+TEST(MKLPackedLayer, RecurrentLayer) {
+  LayerConfig layerConfig1;
+  LayerConfig layerConfig2;
+
+  layerConfig1.set_name("paddle-rnn");
+  layerConfig1.set_type("recurrent");
+  layerConfig1.set_active_type("relu");
+
+  layerConfig2.set_name("mkl-packed-rnn");
+  layerConfig2.set_type("mkl_packed_recurrent");
+  layerConfig2.set_active_type("relu");
+
+  FLAGS_use_gpu = false;
+
+  for (auto layerSize : {32, 64, 128, 256, 512}) {
+    for (auto batchSize : {1, 5, 100, 500}) {
+      for (auto reversed : {true, false}) {
+        for (auto paddle_use_batch : {true, false}) {
+          for (auto MKLPacked_use_batch : {true, false}) {
+            LOG(INFO) << " layerSize=" << layerSize
+                      << " batchSize=" << batchSize << " reversed=" << reversed
+                      << " paddle_use_batch=" << paddle_use_batch
+                      << " MKLPacked_use_batch=" << MKLPacked_use_batch;
+
+            checkMKLPackedLayer(layerConfig1,
+                                layerConfig2,
+                                reversed,
+                                layerSize,
+                                batchSize,
+                                paddle_use_batch,
+                                MKLPacked_use_batch);
+          }
+        }
+      }
+    }
+  }
+}
+#endif
+
 int main(int argc, char** argv) {
-  if (version::isWithGpu()) {
-    testing::InitGoogleTest(&argc, argv);
-    initMain(argc, argv);
-    return RUN_ALL_TESTS();
-  } else {
-    return 0;
+  testing::InitGoogleTest(&argc, argv);
+  initMain(argc, argv);
+  if (!version::isWithGpu()) {
+    testing::GTEST_FLAG(filter) = "-Layer.*";
   }
+  return RUN_ALL_TESTS();
 }

paddle/inference/CMakeLists.txt

+set(FLUID_CORE_MODULES
+    backward proto_desc paddle_memory executor prune init ${GLOB_OP_LIB})
+
+cc_library(paddle_fluid_api
+    SRCS inference.cc
+    DEPS ${FLUID_CORE_MODULES})
+
+# Merge all modules into a simgle static library
+cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES})
+
+# ptools
+# just for testing, we may need to change the storing format for inference_model
+# and move the dependent of pickle.
+# download from http://www.picklingtools.com/
+# build in the C++ sub-directory, using command
+#     make -f Makefile.Linux libptools.so
+set(PTOOLS_LIB)
+set(PTOOLS_ROOT $ENV{PTOOLS_ROOT} CACHE PATH "Folder contains PicklingTools")
+find_path(PTOOLS_INC_DIR chooseser.h PATHS ${PTOOLS_ROOT}/C++)
+find_library(PTOOLS_SHARED_LIB NAMES ptools PATHS ${PTOOLS_ROOT}/C++)
+if(PTOOLS_INC_DIR AND PTOOLS_SHARED_LIB)
+  add_definitions(-DPADDLE_USE_PTOOLS)
+  set(PTOOLS_LIB ptools)
+  message(STATUS "Found PicklingTools: ${PTOOLS_SHARED_LIB}")
+  add_library(${PTOOLS_LIB} SHARED IMPORTED GLOBAL)
+  set_property(TARGET ${PTOOLS_LIB} PROPERTY IMPORTED_LOCATION ${PTOOLS_SHARED_LIB})
+  include_directories(${PTOOLS_ROOT}/C++)
+  include_directories(${PTOOLS_ROOT}/C++/opencontainers_1_8_5/include)
+  add_definitions(-DOC_NEW_STYLE_INCLUDES) # used in ptools
+endif()
+
+add_executable(example example.cc)
+if(APPLE)
+  set(OPTIONAL_LINK_FLAGS)
+  if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang" OR "${CMAKE_CXX_COMPILER_ID}" STREQUAL "AppleClang")
+    set(OPTIONAL_LINK_FLAGS "-undefined dynamic_lookup")
+  endif()
+  target_link_libraries(example
+      -Wl,-force_load paddle_fluid
+      ${OPTIONAL_LINK_FLAGS}
+      ${PTOOLS_LIB})
+else()
+  target_link_libraries(example
+      -Wl,--start-group -Wl,--whole-archive paddle_fluid
+      -Wl,--no-whole-archive -Wl,--end-group
+      ${PTOOLS_LIB})
+endif()

paddle/inference/example.cc

+/* 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. */
+
+#include <time.h>
+#include <iostream>
+#include "gflags/gflags.h"
+#include "paddle/inference/inference.h"
+
+DEFINE_string(dirname, "", "Directory of the inference model.");
+DEFINE_string(feed_var_names, "", "Names of feeding variables");
+DEFINE_string(fetch_var_names, "", "Names of fetching variables");
+
+int main(int argc, char** argv) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  if (FLAGS_dirname.empty() || FLAGS_feed_var_names.empty() ||
+      FLAGS_fetch_var_names.empty()) {
+    // Example:
+    //   ./example --dirname=recognize_digits_mlp.inference.model
+    //             --feed_var_names="x"
+    //             --fetch_var_names="fc_2.tmp_2"
+    std::cout << "Usage: ./example --dirname=path/to/your/model "
+                 "--feed_var_names=x --fetch_var_names=y"
+              << std::endl;
+    exit(1);
+  }
+
+  std::cout << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
+  std::cout << "FLAGS_feed_var_names: " << FLAGS_feed_var_names << std::endl;
+  std::cout << "FLAGS_fetch_var_names: " << FLAGS_fetch_var_names << std::endl;
+
+  std::string dirname = FLAGS_dirname;
+  std::vector<std::string> feed_var_names = {FLAGS_feed_var_names};
+  std::vector<std::string> fetch_var_names = {FLAGS_fetch_var_names};
+
+  paddle::InferenceEngine* engine = new paddle::InferenceEngine();
+  engine->LoadInferenceModel(dirname, feed_var_names, fetch_var_names);
+
+  paddle::framework::LoDTensor input;
+  srand(time(0));
+  float* input_ptr =
+      input.mutable_data<float>({1, 784}, paddle::platform::CPUPlace());
+  for (int i = 0; i < 784; ++i) {
+    input_ptr[i] = rand() / (static_cast<float>(RAND_MAX));
+  }
+
+  std::vector<paddle::framework::LoDTensor> feeds;
+  feeds.push_back(input);
+  std::vector<paddle::framework::LoDTensor> fetchs;
+  engine->Execute(feeds, fetchs);
+
+  for (size_t i = 0; i < fetchs.size(); ++i) {
+    auto dims_i = fetchs[i].dims();
+    std::cout << "dims_i:";
+    for (int j = 0; j < dims_i.size(); ++j) {
+      std::cout << " " << dims_i[j];
+    }
+    std::cout << std::endl;
+    std::cout << "result:";
+    float* output_ptr = fetchs[i].data<float>();
+    for (int j = 0; j < paddle::framework::product(dims_i); ++j) {
+      std::cout << " " << output_ptr[j];
+    }
+    std::cout << std::endl;
+  }
+
+  delete engine;
+  return 0;
+}

paddle/inference/inference.cc

+/* 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. */
+
+#include "inference.h"
+#include <fstream>
+#include "paddle/framework/executor.h"
+#include "paddle/framework/feed_fetch_method.h"
+#include "paddle/framework/init.h"
+#include "paddle/framework/scope.h"
+
+#ifdef PADDLE_USE_PTOOLS
+#include "chooseser.h"
+#endif
+
+namespace paddle {
+
+void InferenceEngine::LoadInferenceModel(
+    const std::string& dirname,
+    const std::vector<std::string>& feed_var_names,
+    const std::vector<std::string>& fetch_var_names) {
+#ifdef PADDLE_USE_PTOOLS
+  std::string model_filename = dirname + "/__model__";
+  LOG(INFO) << "Using PicklingTools, loading model from " << model_filename;
+  Val v;
+  LoadValFromFile(model_filename.c_str(), v, SERIALIZE_P0);
+  std::string program_desc_str = v["program_desc_str"];
+  LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
+// PicklingTools cannot parse the vector of strings correctly.
+#else
+  std::string model_filename = dirname + "/__model__.dat";
+  LOG(INFO) << "loading model from " << model_filename;
+  std::ifstream inputfs(model_filename, std::ios::in | std::ios::binary);
+  std::string program_desc_str;
+  inputfs.seekg(0, std::ios::end);
+  program_desc_str.resize(inputfs.tellg());
+  inputfs.seekg(0, std::ios::beg);
+  LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
+  inputfs.read(&program_desc_str[0], program_desc_str.size());
+  inputfs.close();
+#endif
+  program_ = new framework::ProgramDesc(program_desc_str);
+  GenerateLoadProgram(dirname);
+
+  if (feed_var_names.empty() || fetch_var_names.empty()) {
+    LOG(FATAL) << "Please specify the feed_var_names and fetch_var_names.";
+  }
+  feed_var_names_ = feed_var_names;
+  fetch_var_names_ = fetch_var_names;
+  PrependFeedOp();
+  AppendFetchOp();
+}
+
+bool InferenceEngine::IsParameter(const framework::VarDesc* var) {
+  if (var->Persistable()) {
+    // There are many unreachable variables in the program
+    for (size_t i = 0; i < program_->Size(); ++i) {
+      const framework::BlockDesc& block = program_->Block(i);
+      for (auto* op : block.AllOps()) {
+        for (auto input_argument_name : op->InputArgumentNames()) {
+          if (input_argument_name == var->Name()) {
+            return true;
+          }
+        }
+      }
+    }
+  }
+  return false;
+}
+
+void InferenceEngine::GenerateLoadProgram(const std::string& dirname) {
+  framework::BlockDesc* global_block = program_->MutableBlock(0);
+
+  load_program_ = new framework::ProgramDesc();
+  framework::BlockDesc* load_block = load_program_->MutableBlock(0);
+  for (auto* var : global_block->AllVars()) {
+    if (IsParameter(var)) {
+      LOG(INFO) << "parameter's name: " << var->Name();
+
+      framework::VarDesc* new_var = load_block->Var(var->Name());
+      new_var->SetShape(var->Shape());
+      new_var->SetDataType(var->GetDataType());
+      new_var->SetType(var->GetType());
+      new_var->SetLoDLevel(var->GetLoDLevel());
+      new_var->SetPersistable(true);
+
+      // append_op
+      framework::OpDesc* op = load_block->AppendOp();
+      op->SetType("load");
+      op->SetOutput("Out", {new_var->Name()});
+      op->SetAttr("file_path", {dirname + "/" + new_var->Name()});
+      op->CheckAttrs();
+    }
+  }
+}
+
+void InferenceEngine::PrependFeedOp() {
+  if (!program_) {
+    LOG(FATAL) << "Please initialize the program_ first.";
+  }
+
+  framework::BlockDesc* global_block = program_->MutableBlock(0);
+
+  // create_var
+  framework::VarDesc* feed_var = global_block->Var("feed");
+  feed_var->SetType(framework::proto::VarDesc::FEED_MINIBATCH);
+  feed_var->SetPersistable(true);
+
+  // prepend feed_op
+  for (size_t i = 0; i < feed_var_names_.size(); ++i) {
+    std::string var_name = feed_var_names_[i];
+    LOG(INFO) << "feed var's name: " << var_name;
+
+    // prepend_op
+    framework::OpDesc* op = global_block->PrependOp();
+    op->SetType("feed");
+    op->SetInput("X", {"feed"});
+    op->SetOutput("Out", {var_name});
+    op->SetAttr("col", {static_cast<int>(i)});
+    op->CheckAttrs();
+  }
+}
+
+void InferenceEngine::AppendFetchOp() {
+  if (!program_) {
+    LOG(FATAL) << "Please initialize the program_ first.";
+  }
+
+  framework::BlockDesc* global_block = program_->MutableBlock(0);
+
+  // create_var
+  framework::VarDesc* fetch_var = global_block->Var("fetch");
+  fetch_var->SetType(framework::proto::VarDesc::FETCH_LIST);
+  fetch_var->SetPersistable(true);
+
+  // append fetch_op
+  for (size_t i = 0; i < fetch_var_names_.size(); ++i) {
+    std::string var_name = fetch_var_names_[i];
+    LOG(INFO) << "fetch var's name: " << var_name;
+
+    // append_op
+    framework::OpDesc* op = global_block->AppendOp();
+    op->SetType("fetch");
+    op->SetInput("X", {var_name});
+    op->SetOutput("Out", {"fetch"});
+    op->SetAttr("col", {static_cast<int>(i)});
+    op->CheckAttrs();
+  }
+}
+
+void InferenceEngine::Execute(const std::vector<framework::LoDTensor>& feeds,
+                              std::vector<framework::LoDTensor>& fetchs) {
+  if (!program_ || !load_program_) {
+    LOG(FATAL) << "Please initialize the program_ and load_program_ first.";
+  }
+
+  if (feeds.size() < feed_var_names_.size()) {
+    LOG(FATAL) << "Please feed " << feed_var_names_.size() << " input Tensors.";
+  }
+
+  auto* place = new platform::CPUPlace();
+  framework::InitDevices({"CPU"});
+  framework::Executor* executor = new framework::Executor(*place);
+  framework::Scope* scope = new framework::Scope();
+
+  executor->Run(*load_program_, scope, 0, true, true);
+
+  // set_feed_variable
+  for (size_t i = 0; i < feed_var_names_.size(); ++i) {
+    framework::SetFeedVariable(scope, feeds[i], "feed", i);
+  }
+
+  executor->Run(*program_, scope, 0, true, true);
+
+  // get_fetch_variable
+  fetchs.resize(fetch_var_names_.size());
+  for (size_t i = 0; i < fetch_var_names_.size(); ++i) {
+    fetchs[i] = framework::GetFetchVariable(*scope, "fetch", i);
+  }
+
+  delete place;
+  delete scope;
+  delete executor;
+}
+}  // namespace paddle

paddle/inference/inference.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
+
+#include "paddle/framework/block_desc.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/framework/program_desc.h"
+
+namespace paddle {
+
+class InferenceEngine {
+public:
+  InferenceEngine() : program_(nullptr), load_program_(nullptr) {}
+  ~InferenceEngine() {
+    delete program_;
+    delete load_program_;
+  }
+
+  void LoadInferenceModel(const std::string& dirname,
+                          const std::vector<std::string>& feed_var_names,
+                          const std::vector<std::string>& fetch_var_names);
+  void Execute(const std::vector<framework::LoDTensor>& feeds,
+               std::vector<framework::LoDTensor>& fetchs);
+
+private:
+  bool IsParameter(const framework::VarDesc* var);
+  void GenerateLoadProgram(const std::string& dirname);
+  void PrependFeedOp();
+  void AppendFetchOp();
+
+private:
+  framework::ProgramDesc* program_;
+  framework::ProgramDesc* load_program_;
+  std::vector<std::string> feed_var_names_;
+  std::vector<std::string> fetch_var_names_;
+};
+
+}  // namespace paddle

paddle/operators/CMakeLists.txt

@@ -61,106 +61,28 @@ function(op_library TARGET)
                 ${op_common_deps})
     endif()
 
-    # net_op doesn't need pybind
-    if ("${TARGET}" STREQUAL "net_op")
-        set(pybind_flag 1)
-    endif()
-
-    if ("${TARGET}" STREQUAL "compare_op")
-        set(pybind_flag 1)
-        file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(equal);\n")
-    endif()
-
-    # conv_op contains several operators
-    if ("${TARGET}" STREQUAL "conv_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(conv2d);\n")
-    endif()
-
-    # conv_cudnn_op contains several operators
-    if ("${TARGET}" STREQUAL "conv_cudnn_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(conv2d_cudnn);\n")
-    endif()
-
-    # pool_op contains several operators
-    if ("${TARGET}" STREQUAL "pool_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(pool2d);\n")
-    endif()
-
-    # pool_cudnn_op contains several operators
-    if ("${TARGET}" STREQUAL "pool_cudnn_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(pool2d_cudnn);\n")
-    endif()
-
-    if ("${TARGET}" STREQUAL "logical_op")
-        set(pybind_flag 1)
-        file(APPEND ${pybind_file} "USE_OP(logical_and);\n")
-    endif()
-
-    # pool_with_index_op contains several operators
-    if ("${TARGET}" STREQUAL "pool_with_index_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(max_pool2d_with_index);\n")
-    endif()
-
-    # conv_transpose_op contains several operators
-    if ("${TARGET}" STREQUAL "conv_transpose_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(conv2d_transpose);\n")
-    endif()
-
-    # conv_transpose_cudnn_op contains two operators
-    if ("${TARGET}" STREQUAL "conv_transpose_cudnn_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(conv2d_transpose_cudnn);\n")
-    endif()
-
-    # save_restore_op contains several operators
-    if ("${TARGET}" STREQUAL "save_restore_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(save);\n")
-    endif()
-
-    # activation_op contains several operators
-    if ("${TARGET}" STREQUAL "activation_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(sigmoid);\n")
-    endif()
-
-    # nccl_op contains several operators
-    if ("${TARGET}" STREQUAL "nccl_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n")
-    endif()
-
-    # reduce_op contains several operators
-    if ("${TARGET}" STREQUAL "reduce_op")
-        set(pybind_flag 1)
-        # It's enough to just adding one operator to pybind
-        file(APPEND ${pybind_file} "USE_OP(reduce_sum);\n")
-    endif()
+    # Define operators that don't need pybind here.
+    foreach(manual_pybind_op "net_op" "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op")
+        if ("${TARGET}" STREQUAL "${manual_pybind_op}")
+            set(pybind_flag 1)
+        endif()
+    endforeach()
 
-    if ("${TARGET}" STREQUAL "tensor_array_read_write_op")
-        set(pybind_flag 1)
-        file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(read_from_array);\nUSE_NO_KERNEL_OP(write_to_array);\n")
+    # The registration of USE_OP, please refer to paddle/framework/op_registry.h.
+    # Note that it's enough to just adding one operator to pybind in a *_op.cc file.
+    # And for detail pybind information, please see generated paddle/pybind/pybind.h.
+    file(READ ${TARGET}.cc TARGET_CONTENT)
+    string(REGEX MATCH "REGISTER_OP\\(.*REGISTER_OP\\(" multi_register "${TARGET_CONTENT}")
+    string(REGEX MATCH "REGISTER_OP\\([a-z0-9_]*," one_register "${multi_register}")
+    if (one_register STREQUAL "")
+        string(REPLACE "_op" "" TARGET "${TARGET}")
+    else ()
+        string(REPLACE "REGISTER_OP(" "" TARGET "${one_register}")
+        string(REPLACE "," "" TARGET "${TARGET}")
     endif()
 
     # pybind USE_NO_KERNEL_OP
     # HACK: if REGISTER_OP_CPU_KERNEL presents the operator must have kernel
-    file(READ ${TARGET}.cc TARGET_CONTENT)
     string(REGEX MATCH "REGISTER_OP_CPU_KERNEL" regex_result "${TARGET_CONTENT}")
     string(REPLACE "_op" "" TARGET "${TARGET}")
     if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
@@ -171,7 +93,6 @@ function(op_library TARGET)
     # pybind USE_CPU_ONLY_OP
     list(LENGTH cu_srcs cu_srcs_len)
     list(LENGTH cu_cc_srcs cu_cc_srcs_len)
-
     if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0)
         file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
         set(pybind_flag 1)
@@ -188,6 +109,7 @@ add_subdirectory(nccl)
 
 if(WITH_GPU)
     op_library(nccl_op DEPS nccl_common)
+    file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n")
 else()
     set(DEPS_OPS ${DEPS_OPS} nccl_op)
 endif()
@@ -230,6 +152,7 @@ op_library(conv_transpose_op DEPS vol2col)
 op_library(gru_op DEPS sequence2batch gru_compute)
 op_library(recurrent_op DEPS executor)
 op_library(cos_sim_op DEPS cos_sim_functor)
+op_library(parallel_do_op DEPS executor)
 # FIXME(typhoonzero): save/load depends lodtensor serialization functions
 op_library(save_op DEPS lod_tensor)
 op_library(load_op DEPS lod_tensor)
@@ -238,6 +161,8 @@ list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})
     op_library(${src})
 endforeach()
+file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")
+
 
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
 

paddle/operators/accuracy_op.cc

@@ -53,7 +53,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Out")->type()),

paddle/operators/activation_op.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/detail/safe_ref.h"
 
 namespace paddle {
 namespace operators {
@@ -26,12 +27,16 @@ class ActivationKernel
   using T = typename Functor::ELEMENT_TYPE;
 
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* X = context.Input<framework::Tensor>("X");
-    auto* Out = context.Output<framework::Tensor>("Out");
-    Out->mutable_data<T>(context.GetPlace());
-
-    auto x = framework::EigenVector<T>::Flatten(*X);
-    auto out = framework::EigenVector<T>::Flatten(*Out);
+    auto& X = detail::Ref(context.Input<framework::Tensor>("X"),
+                          "Cannot get input tensor X, variable name = %s",
+                          context.op().Input("X"));
+
+    auto& Out = detail::Ref(context.Output<framework::Tensor>("Out"),
+                            "Cannot get output tensor Out, variable name = %s",
+                            context.op().Output("Out"));
+    Out.mutable_data<T>(context.GetPlace());
+    auto x = framework::EigenVector<T>::Flatten(X);
+    auto out = framework::EigenVector<T>::Flatten(Out);
     auto* place =
         context.template device_context<DeviceContext>().eigen_device();
     Functor functor;

paddle/operators/adagrad_op.h

@@ -47,8 +47,7 @@ class AdagradOpKernel : public framework::OpKernel<T> {
           *ctx.Input<framework::Tensor>("Grad"));
       auto moment = framework::EigenVector<T>::Flatten(
           *ctx.Input<framework::Tensor>("Moment"));
-      auto lr = framework::EigenVector<T>::Flatten(
-          *ctx.Input<framework::Tensor>("LearningRate"));
+      auto* learning_rate = ctx.Input<framework::Tensor>("LearningRate");
 
       auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
       auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
@@ -56,8 +55,16 @@ class AdagradOpKernel : public framework::OpKernel<T> {
 
       moment_out.device(*place) = moment + grad * grad;
       Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
-      param_out.device(*place) =
-          param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
+      if (platform::is_cpu_place(ctx.GetPlace())) {
+        auto* lr = learning_rate->data<T>();
+        param_out.device(*place) =
+            param - lr[0] * grad / (moment_out.sqrt() + epsilon);
+      } else {
+        auto lr = framework::EigenVector<T>::Flatten(*learning_rate);
+        param_out.device(*place) =
+            param -
+            lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
+      }
     } else if (grad_var->IsType<framework::SelectedRows>()) {
       auto* param_tensor = ctx.Input<framework::Tensor>("Param");
       PADDLE_ENFORCE_EQ(param_tensor, param_out_tensor);

paddle/operators/auc_op.cc

@@ -39,7 +39,7 @@ class AucOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Out")->type()),

paddle/operators/batch_norm_op.cc

@@ -64,7 +64,7 @@ class BatchNormOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(x_dims.size() >= 2 && x_dims.size() <= 5,
                    "Input X must have 2 to 5 dimensions.");
 
-    const int C =
+    const int64_t C =
         (data_layout == DataLayout::kNCHW ? x_dims[1]
                                           : x_dims[x_dims.size() - 1]);
 
@@ -78,6 +78,7 @@ class BatchNormOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("VarianceOut", {C});
     ctx->SetOutputDim("SavedMean", {C});
     ctx->SetOutputDim("SavedVariance", {C});
+    ctx->ShareLoD("X", "Y");
   }
 };
 
@@ -305,7 +306,7 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
     const auto *var = ctx.InputVar(framework::GradVarName("Y"));
     if (var == nullptr) {

paddle/operators/chunk_eval_op.cc

@@ -55,10 +55,10 @@ class ChunkEvalOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(framework::proto::DataType::FP32,
-                                   ctx.device_context());
+                                   platform::CPUPlace());
   }
 };
 

paddle/operators/chunk_eval_op.h

@@ -145,6 +145,7 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
         context.Attr<std::vector<int>>("excluded_chunk_types").end());
 
     auto* inference = context.Input<LoDTensor>("Inference");
+    auto place = inference->place();
     auto* label = context.Input<LoDTensor>("Label");
     auto* precision = context.Output<Tensor>("Precision");
     auto* recall = context.Output<Tensor>("Recall");
@@ -155,15 +156,15 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
 
     const int64_t* inference_data = inference->data<int64_t>();
     const int64_t* label_data = label->data<int64_t>();
-    T* precision_data = precision->mutable_data<T>(context.GetPlace());
-    T* racall_data = recall->mutable_data<T>(context.GetPlace());
-    T* f1_data = f1->mutable_data<T>(context.GetPlace());
+    T* precision_data = precision->mutable_data<T>(place);
+    T* racall_data = recall->mutable_data<T>(place);
+    T* f1_data = f1->mutable_data<T>(place);
     int64_t* num_infer_chunks_data =
-        num_infer_chunks->mutable_data<int64_t>(context.GetPlace());
+        num_infer_chunks->mutable_data<int64_t>(place);
     int64_t* num_label_chunks_data =
-        num_label_chunks->mutable_data<int64_t>(context.GetPlace());
+        num_label_chunks->mutable_data<int64_t>(place);
     int64_t* num_correct_chunks_data =
-        num_correct_chunks->mutable_data<int64_t>(context.GetPlace());
+        num_correct_chunks->mutable_data<int64_t>(place);
     *num_infer_chunks_data = 0;
     *num_label_chunks_data = 0;
     *num_correct_chunks_data = 0;

paddle/operators/compare_op.cc

@@ -66,9 +66,9 @@ class CompareOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
-    framework::OpKernelType kt = OperatorWithKernel::GetActualKernelType(ctx);
+    framework::OpKernelType kt = OperatorWithKernel::GetExpectedKernelType(ctx);
     // CompareOp kernel's device type is decided by input tensor place
     kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
     return kt;

paddle/operators/conv_cudnn_op.cu.cc

@@ -315,10 +315,7 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_KERNEL(conv2d, CUDNN, paddle::platform::CUDAPlace,
-                   paddle::operators::CudnnConvOpKernel<float>,
-                   paddle::operators::CudnnConvOpKernel<double>);
-
+// TODO(dzhwinter) : below register should be removed
 REGISTER_OP_CUDA_KERNEL(conv2d_cudnn,
                         paddle::operators::CudnnConvOpKernel<float>,
                         paddle::operators::CudnnConvOpKernel<double>);

paddle/operators/conv_op.cc

@@ -44,14 +44,12 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
       paddings.size(), strides.size(),
       "Conv paddings dimension and Conv strides dimension should be the same.");
 
-  int input_channels = in_dims[1];
-  PADDLE_ENFORCE_EQ(input_channels, filter_dims[1] * groups,
+  PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[1] * groups,
                     "The number of input channels should be equal to filter "
                     "channels * groups.");
 
-  int output_channels = filter_dims[0];
   PADDLE_ENFORCE_EQ(
-      output_channels % groups, 0,
+      filter_dims[0] % groups, 0,
       "The number of output channels should be divided by groups.");
 
   std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
@@ -66,6 +64,7 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
                                       dilations[i], paddings[i], strides[i]));
   }
   ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
+  ctx->ShareLoD("Input", "Output");
 }
 
 Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker)

paddle/operators/crf_decoding_op.cc

@@ -120,17 +120,11 @@ class CRFDecodingOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type()),
-        ctx.device_context());
-  }
-
-  framework::OpKernelType GetExpectedKernelType(
-      const framework::OpKernelType& actual_kernel_type) const override {
-    return framework::OpKernelType(actual_kernel_type.data_type_,
-                                   platform::CPUPlace());
+        platform::CPUPlace());
   }
 };
 }  // namespace operators

paddle/operators/crf_decoding_op.h

@@ -28,9 +28,6 @@ template <typename DeviceContext, typename T>
 class CRFDecodingOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
-                   "The crf_decoding operator can only run on CPU.");
-
     auto* emission_weights = ctx.Input<LoDTensor>("Emission");
     auto* transition_weights = ctx.Input<Tensor>("Transition");
     auto* label = ctx.Input<LoDTensor>("Label");

paddle/operators/cross_entropy_op.cc

@@ -51,7 +51,7 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of computation kernel of cross_entropy
   // is determined by its input "X".
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
@@ -101,7 +101,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of computation kernel of cross_entropy
   // is determined by its input "X".
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),

paddle/operators/detail/recv_impl.cc

@@ -21,14 +21,9 @@ namespace detail {
 Status SendRecvServerImpl::SendVariable(ServerContext *context,
                                         const VariableMessage *in_var,
                                         VoidMessage *out_var) {
-  // TODO(typhoonzero): support different variable types.
-  std::istringstream iss(in_var->serialized());
-  framework::LoDTensor t;
-  framework::DeserializeFromStream(iss, &t);
-  TensorWithName tensor_with_name =
-      std::make_pair(in_var->varname(), std::move(t));
-
-  var_recv_queue_.Push(std::move(tensor_with_name));
+  MessageWithName msg_with_name =
+      std::make_pair(in_var->varname(), std::move(*in_var));
+  var_recv_queue_.Push(std::move(msg_with_name));
   return Status::OK;
 }
 
@@ -37,14 +32,8 @@ Status SendRecvServerImpl::GetVariable(ServerContext *context,
                                        VariableMessage *out_var) {
   std::string get_var_name = in_var->varname();
   auto *var = scope_->FindVar(get_var_name);
-  auto tensor = var->Get<framework::LoDTensor>();
-  std::ostringstream oss;
-  framework::SerializeToStream(oss, tensor, platform::CPUDeviceContext());
 
-  std::string *varname = out_var->mutable_varname();
-  *varname = get_var_name;
-  std::string *serialized = out_var->mutable_serialized();
-  *serialized = oss.str();
+  SerializeToMessage(get_var_name, var, platform::CPUDeviceContext(), out_var);
   return Status::OK;
 }
 

paddle/operators/detail/send_impl.cc

@@ -27,14 +27,8 @@ bool RPCClient::SendVariable(const framework::Scope& scope,
   auto ctx = platform::CPUDeviceContext();
   auto* var = scope.FindVar(inname);
   PADDLE_ENFORCE(var);
-  // TODO(typhoonzero): support SelectedRows
-  PADDLE_ENFORCE(var->IsType<framework::LoDTensor>(),
-                 "Only support LoDTensor, %s has wrong type", inname);
-  const framework::LoDTensor& tensor = var->Get<framework::LoDTensor>();
-  std::ostringstream oss;
-  framework::SerializeToStream(oss, tensor, ctx);
-  msg.set_varname(inname);
-  msg.set_serialized(oss.str());
+  SerializeToMessage(inname, var, ctx, &msg);
+
   Status status = stub_->SendVariable(&context, msg, &out_msg);
   if (!status.ok()) {
     LOG(ERROR) << "gRPC error: " << status.error_message();
@@ -50,19 +44,15 @@ bool RPCClient::GetVariable(const framework::Scope& scope,
   call_msg.set_varname(outname);
   auto ctx = platform::CPUDeviceContext();
   Status status = stub_->GetVariable(&context, call_msg, &ret_msg);
+  auto* outvar = scope.FindVar(outname);
   if (!status.ok()) {
     LOG(ERROR) << "gRPC error: " << status.error_message();
     return false;
   }
 
   std::istringstream iss(ret_msg.serialized());
+  DeserializeFromMessage(ret_msg, ctx, outvar);
 
-  framework::LoDTensor ret_tensor;
-  framework::DeserializeFromStream(iss, &ret_tensor);
-  auto* outvar = scope.FindVar(outname);
-  framework::LoDTensor* out_tensor = outvar->GetMutable<framework::LoDTensor>();
-  // FIXME(typhoonzero): do not copy.
-  framework::CopyFrom(ret_tensor, ctx.GetPlace(), ctx, out_tensor);
   return true;
 }
 

paddle/operators/detail/send_recv.proto

-/* 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.
+/* 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
@@ -13,7 +12,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 syntax = "proto3";
-
 package sendrecv;
 
 service SendRecvService {
@@ -29,12 +27,18 @@ service SendRecvService {
 
 // VariableMessage is serialized paddle variable message.
 // It can be:
-// Tensor
 // LoDTensor
 // SelectedRows
+enum VarType {
+  LOD_TENSOR = 0;
+  SELECTED_ROWS = 1;
+}
+
 message VariableMessage {
   string varname = 1;
-  bytes serialized = 2;
+  // TODO(Yancey1989): reference framework::proto::VarDesc::VarType
+  VarType type = 2;
+  bytes serialized = 3;
 }
 
 message VoidMessage {}

paddle/operators/detail/send_recv_impl.h

@@ -14,10 +14,10 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/framework/data_type.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/selected_rows.h"
+#include "paddle/framework/var_type.h"
 #include "paddle/operators/detail/simple_block_queue.h"
 
 #include "paddle/operators/detail/send_recv.grpc.pb.h"
@@ -44,7 +44,7 @@ namespace paddle {
 namespace operators {
 namespace detail {
 
-typedef std::pair<std::string, framework::LoDTensor> TensorWithName;
+typedef std::pair<std::string, sendrecv::VariableMessage> MessageWithName;
 
 class SendRecvServerImpl final : public SendRecvService::Service {
  public:
@@ -60,13 +60,13 @@ class SendRecvServerImpl final : public SendRecvService::Service {
   void Done();
   void SetScope(framework::Scope *scope) { scope_ = scope; };
 
-  const TensorWithName Get() { return this->var_recv_queue_.Pop(); }
+  const MessageWithName Get() { return this->var_recv_queue_.Pop(); }
 
-  void Push(const TensorWithName &msg) { this->var_recv_queue_.Push(msg); }
+  void Push(const MessageWithName &msg) { this->var_recv_queue_.Push(msg); }
 
  private:
   // received variable from RPC, operators fetch variable from this queue.
-  SimpleBlockQueue<TensorWithName> var_recv_queue_;
+  SimpleBlockQueue<MessageWithName> var_recv_queue_;
   framework::Scope *scope_;
   // condition of the sub program
   std::mutex mutex_;
@@ -89,6 +89,53 @@ class RPCClient {
   std::unique_ptr<SendRecvService::Stub> stub_;
 };
 
+inline void SerializeToMessage(const std::string &name,
+                               const framework::Variable *var,
+                               const platform::DeviceContext &ctx,
+                               VariableMessage *msg) {
+  msg->set_varname(name);
+  std::ostringstream oss;
+  switch (framework::ToVarType(var->Type())) {
+    case framework::proto::VarDesc_VarType_LOD_TENSOR:
+      msg->set_type(sendrecv::VarType::LOD_TENSOR);
+      framework::SerializeToStream(oss, var->Get<framework::LoDTensor>(), ctx);
+      break;
+    case framework::proto::VarDesc_VarType_SELECTED_ROWS:
+      msg->set_type(sendrecv::VarType::SELECTED_ROWS);
+      framework::SerializeToStream(oss, var->Get<framework::SelectedRows>(),
+                                   ctx);
+      break;
+    default: {
+      PADDLE_THROW("Serialize does not support type: %s",
+                   typeid(var->Type()).name());
+      break;
+    }
+  }
+  msg->set_serialized(oss.str());
+}
+
+inline void DeserializeFromMessage(const VariableMessage &msg,
+                                   const platform::DeviceContext &ctx,
+                                   framework::Variable *var) {
+  using namespace paddle::framework::proto;
+  std::istringstream iss(msg.serialized());
+  switch (msg.type()) {
+    case sendrecv::VarType::LOD_TENSOR:
+      DeserializeFromStream(iss, var->GetMutable<framework::LoDTensor>(), ctx);
+      break;
+    case sendrecv::VarType::SELECTED_ROWS: {
+      DeserializeFromStream(iss, var->GetMutable<framework::SelectedRows>(),
+                            ctx);
+      break;
+    }
+    default: {
+      PADDLE_THROW("Deserialize does not support type: %s",
+                   typeid(var->Type()).name());
+      break;
+    }
+  }
+}
+
 }  // namespace detail
 }  // namespace operators
 }  // namespace paddle

paddle/operators/fill_constant_batch_size_like_op.cc

@@ -49,7 +49,7 @@ class FillConstantBatchSizeLikeOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::proto::DataType>(ctx.Attr<int>("dtype")),

paddle/operators/gather_op.cc

@@ -40,7 +40,7 @@ class GatherOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
@@ -57,7 +57,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),

paddle/operators/gaussian_random_op.cc

@@ -60,7 +60,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::proto::DataType>(ctx.Attr<int>("dtype")),

paddle/operators/linear_chain_crf_op.cc

@@ -183,7 +183,7 @@ class LinearChainCRFOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of computation kernel of linear_chain_crf
   // is determined by its input "Emission".
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type()),
@@ -242,7 +242,7 @@ class LinearChainCRFGradOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of output of the linear_chain_crf_grad
   // operator is determined by its input: gradients of LogLikelihood.
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(

paddle/operators/load_op.cc

@@ -38,10 +38,10 @@ class LoadOp : public framework::OperatorBase {
                    out_var_name);
 
     auto *tensor = out_var->GetMutable<framework::LoDTensor>();
-    DeserializeFromStream(fin, tensor);
 
     platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
     auto &dev_ctx = *pool.Get(place);
+    DeserializeFromStream(fin, tensor, dev_ctx);
 
     if (platform::is_gpu_place(place)) {
       // copy CPU to GPU

paddle/operators/lod_reset_op.cc

@@ -38,7 +38,7 @@ class LoDResetOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
@@ -97,7 +97,7 @@ class LoDResetGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),

paddle/operators/logical_op.cc

@@ -99,9 +99,9 @@ class LogicalOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext &ctx) const override {
-    framework::OpKernelType kt = OperatorWithKernel::GetActualKernelType(ctx);
+    framework::OpKernelType kt = OperatorWithKernel::GetExpectedKernelType(ctx);
     // LogicalOp kernel's device type is decided by input tensor place
     kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
     return kt;

paddle/operators/lookup_table_op.cc

@@ -41,7 +41,7 @@ class LookupTableOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("W")->type()),
@@ -98,7 +98,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("W")->type()),

paddle/operators/lstm_op.cc

@@ -92,7 +92,7 @@ class LSTMOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("Input")->type()),
@@ -260,7 +260,7 @@ class LSTMGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("Input")->type()),

paddle/operators/multiplex_op.cc

@@ -51,7 +51,7 @@ class MultiplexOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.MultiInput<Tensor>("X")[0]->type()),
@@ -102,7 +102,7 @@ class MultiplexGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.MultiInput<Tensor>("X")[0]->type()),

paddle/operators/nce_op.cc

@@ -63,7 +63,7 @@ class NCEOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
@@ -166,7 +166,7 @@ class NCEOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetActualKernelType(
+  framework::OpKernelType GetExpectedKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Input")->type()),

paddle/operators/net_op.cc

@@ -56,11 +56,11 @@ void NetOp::CompleteAddOp(bool calc) {
   std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs));
 }
 
-std::string NetOp::DebugString() const {
+std::string NetOp::DebugStringEx(const framework::Scope* scope) const {
   std::ostringstream os;
-  os << OperatorBase::DebugString() << std::endl;
+  os << OperatorBase::DebugStringEx(scope) << std::endl;
   for (auto& op : ops_) {
-    std::istringstream is(op->DebugString());
+    std::istringstream is(op->DebugStringEx(scope));
     for (std::string line; std::getline(is, line);) {
       os << "    " << line << std::endl;
     }