Merge branch 'develop' into docrefine

CMakeLists.txt

@@ -24,6 +24,8 @@ 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}")
+message(STATUS "AR tools: ${CMAKE_AR}")
+
 if(WIN32)
     set(CMAKE_SUPPRESS_REGENERATION ON)
     set(CMAKE_STATIC_LIBRARY_PREFIX lib)

Dockerfile

@@ -75,8 +75,9 @@ RUN curl -s -q https://glide.sh/get | sh
 #    and its size is only one-third of the official one.
 # 2. Manually add ~IPluginFactory() in IPluginFactory class of NvInfer.h, otherwise, it couldn't work in paddle.
 #    See https://github.com/PaddlePaddle/Paddle/issues/10129 for details.
-RUN wget -qO- http://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.0.3.Ubuntu-16.04.4.x86_64-gnu.cuda-8.0.cudnn7.0.tar.gz | \
-    tar -xz -C /usr/local && \
+
+RUN wget -q https://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.1.6-ubuntu14.04.x86_64-gnu.cuda.8.0.cudnn7.0.tar.gz --no-check-certificate && \
+    tar -zxf TensorRT-4.0.1.6-ubuntu14.04.x86_64-gnu.cuda.8.0.cudnn7.0.tar.gz -C /usr/local && \
     cp -rf /usr/local/TensorRT/include /usr && \
     cp -rf /usr/local/TensorRT/lib /usr
 

benchmark/fluid/fluid_benchmark.py

@@ -179,7 +179,6 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
     else:
         build_strategy.reduce_strategy = fluid.BuildStrategy(
         ).ReduceStrategy.AllReduce
-    build_strategy.fuse_broadcast_op = args.fuse_broadcast_op
 
     avg_loss = train_args[0]
 

cmake/external/boost.cmake

@@ -24,7 +24,7 @@ set(BOOST_PROJECT       "extern_boost")
 # So we use 1.41.0 here.
 set(BOOST_VER           "1.41.0")
 set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
-set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
+set(BOOST_URL "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
 
 MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")
 

cmake/external/grpc.cmake

@@ -44,7 +44,7 @@ ExternalProject_Add(
     # 3. keep only zlib, cares, protobuf, boringssl under "third_party",
     #    checkout and clean other dirs under third_party
     # 4. remove .git, and package the directory.
-    URL "http://paddlepaddledeps.cdn.bcebos.com/grpc-v1.10.x.tar.gz"
+    URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.10.x.tar.gz"
     URL_MD5  "1f268a2aff6759839dccd256adcc91cf"
     PREFIX          ${GRPC_SOURCES_DIR}
     UPDATE_COMMAND  ""

cmake/external/mkldnn.cmake

@@ -31,9 +31,17 @@ IF(APPLE)
     return()
 ENDIF()
 
-MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
+# Introduce variables:
+# * CMAKE_INSTALL_LIBDIR
+INCLUDE(GNUInstallDirs)
+SET(LIBDIR "lib")
+if(CMAKE_INSTALL_LIBDIR MATCHES ".*lib64$")
+  SET(LIBDIR "lib64")
+endif()
+
+MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/l${LIBDIR} to runtime path")
 SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
-SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
+SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/${LIBDIR}")
 
 INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
 
@@ -58,7 +66,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/intel/mkl-dnn.git"
-    GIT_TAG             "830a10059a018cd2634d94195140cf2d8790a75a"
+    GIT_TAG             "863ff6e7042cec7d2e29897fe9f0872e0888b0fc"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@@ -79,9 +87,9 @@ ExternalProject_Add(
                         -DMKLROOT:PATH=${MKLML_ROOT}
 )
 if(WIN32)
-    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
+    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/${LIBDIR}/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
 else(WIN32)
-    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
+    SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/${LIBDIR}/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
 endif(WIN32)
 
 ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
@@ -101,7 +109,7 @@ ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
 # copy the real so.0 lib to install dir
 # it can be directly contained in wheel or capi
 if(WIN32)
-    SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/lib/mkldnn.dll)
+    SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/bin/mkldnn.dll)
 else(WIN32)
     SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
     ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}

cmake/external/mklml.cmake

@@ -34,7 +34,7 @@ SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
 SET(TIME_VERSION "2019.0.1.20181227")
 IF(WIN32)
     SET(MKLML_VER "mklml_win_${TIME_VERSION}" CACHE STRING "" FORCE)
-    SET(MKLML_URL "https://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
+    SET(MKLML_URL "https://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
     SET(MKLML_LIB                 ${MKLML_LIB_DIR}/mklml.lib)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5md.lib)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/mklml.dll)
@@ -43,7 +43,7 @@ ELSE()
     #TODO(intel-huying):
     #  Now enable Erf function in mklml library temporarily, it will be updated as offical version later.
     SET(MKLML_VER "Glibc225_vsErf_mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
-    SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
+    SET(MKLML_URL "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
     SET(MKLML_LIB                 ${MKLML_LIB_DIR}/libmklml_intel.so)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5.so)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/libmklml_intel.so)

cmake/operators.cmake

@@ -110,7 +110,7 @@ function(op_library TARGET)
     # Define operators that don't need pybind here.
     foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
 "tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op"
-"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op")
+"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op" "sync_batch_norm_op")
         if ("${TARGET}" STREQUAL "${manual_pybind_op}")
             set(pybind_flag 1)
         endif()

paddle/contrib/float16/README.md

@@ -5,13 +5,13 @@ Kexin Zhao <zhaokexin01@baidu.com>
 ## Introduction
 Deep learning is usually a two-stage work: training and inference. The training stage estimates model parameters (weights) from data.  The inference stage loads the weights and uses them to interpret inputs. Typically, weights are 32-bit float values (float32).  Some new devices, including NVIDIA Volta GPUs, support higher speed computation using 16-bit float values (float16).
 
-This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16.
+This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16.
 
 
 ## What is float16?
 float16 (or FP16) is a half-precision floating-point format that uses 16 bits in memory to represent a value. The advantage over 32-bit single-precision floating-point format (commonly known as float or float32 data type) is that it requires half the storage and bandwidth at the expense of precision and range. Fortunately, DNN inference has a high tolerance for the loss of precision and range when using float16 to represent the weights, and the inference accuracy will only be minimally affected in most cases, which gives us the opportunity to use float16 data type to speed up the inference.
 
-Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type.
+Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type.
 
 ## Why float16?
 The trend in today's deep learning community is to use bigger and deeper model, which translates to larger memory footprint, higher computation demands, and as a result higher energy consumption on computing devices. The advantages of float16 over float32 are correspondingly three-fold:
@@ -24,12 +24,12 @@ The trend in today's deep learning community is to use bigger and deeper model,
 
 ## Fluid implementation of float16 inference
 ### Overview
-Fluid use [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block. 
+Fluid use [Program](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block.
 
 ### Basic requirement
 When an executor runs an operator, it uses a kernel to perform computations on tensors contained in the input variables, and then writes the results to the tensors in the output variables. Each operator has multiple kernels for different combinations of data types, devices, and library types, respectively. The operator will select the appropriate kernel to run based on, among other things, the data type of the input tensors. By default, every Fluid operator has a kernel for float data type that takes float inputs and generates float outputs.
 
-If we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will make the program run in float mode and gives us a final output of float data type. 
+If we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will make the program run in float mode and gives us a final output of float data type.
 
 The same principle applies if we want a program to run in float16 mode. We provide input variable of the float16 data type to the first operator, and every subsequent operator will invoke the float16 kernel until we get the final output in float16. So the preliminary requirements for float16 inference are to add float16 kernels to operators that are needed in a specific kind of neural networks. Our current focus is on Convolutional Neural Networks (CNN) and hence we have added float16 kernels to the following operators: convolution, pooling, GEMM, elementwise addition, batch norm, dropout, various activations including relu and tanh, and softmax.
 
@@ -75,7 +75,7 @@ In this scenario, we already have a float32 inference program and some associate
 
 We can then run various inference experiments in float16 mode and save the float16 program and weights on disk for future deployment. To enhance the code usability, we maintain a consistent API so that user can use the same float32 input data to run inference program in either float32 and float16 mode and obtain output data both of float32 data type. Consequently, we need to add cast operators in the float16 inference program for conversions between the float16 tensor and float32 tensor.
 
-The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference).
+The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference).
 
 ### Experiment results
 Simply running the following commands to reproduce the experiment results presented in this section:
@@ -113,7 +113,7 @@ We repeat the test ten times and get the following results:
 | #10    | 62.53%  | 62.48%   |
 | average| 62.63%  | 62.62%   |
 
-We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over ten tests. 
+We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over ten tests.
 
 #### Performance benchmark
 Currently, Fluid only supports float16 inference on NVIDIA GPUs. There is no motivation to support float16 inference on non-ARM CPUs where float16 is not natively supported, and float16 calculation will only be slower than its float32 counterpart. 
@@ -132,7 +132,7 @@ Average inference time for one mini-batch on Vgg16 model tested on ImageNet data
 |float16|  3.32 | 4.11  |  5.88 |  9.41 | 16.54  | 30.47 |  60.23 |
 |Speedup|  4.22 | 2.36  |  3.91 |  3.00 |  3.26  |  2.77 |   2.97 |
 
-We can see that float16 inference provides **2x ~ 4x** speedup on different batch sizes. 
+We can see that float16 inference provides **2x ~ 4x** speedup on different batch sizes.
 
 Convolution operation is ususally the computational bottleneck of CNN, so we also check the average time spent on the Fluid convolution operators for one mini-batch as follows:
 
@@ -162,7 +162,7 @@ We find that the speedup provided by float16 inference starts relatively small a
 
 We also did the same benchmark on a single NVIDIA GeForce GTX 1080 Ti GPU that does not support Tensor Core. The results show that for Vgg16, float16 inference provides consistent small speedup (around 1.15x) for all mini-batch sizes, while for Resnet50, float16 inference is slower than its float32 counterpart in small batch sizes (mb = 1 and 2) and then delivers around 1.15x speedup for all larger batch sizes. By comparing the benchmarks on 1080 Ti and V100, we find that Tensor Core, which is specialized for float16 computations, is a critical component of high performance float16 inference.
 
-Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/contrib/float16/float16_benchmark.md) for complete benchmark results.
+Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/contrib/float16/float16_benchmark.md) for complete benchmark results.
 
 ### Summary
 1. Fluid is now able to run inference in float16 mode via a float16 transpiler. We currently support CNN programs, including Vgg and Resnet, to run in float16 inference mode.

paddle/fluid/framework/CMakeLists.txt

@@ -38,10 +38,10 @@ if(WITH_GPU)
     nv_library(tensor SRCS tensor.cc .tensor_util.cu DEPS place memory data_type device_context)
     add_dependencies(tensor tensor_util)
   else()
-    nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context )
+    nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context profiler)
   endif(WIN32)
 else()
-  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context )
+  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context profiler)
 endif()
 
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
@@ -174,7 +174,7 @@ else()
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
 
-target_link_libraries(executor garbage_collector)
+target_link_libraries(executor garbage_collector while_op_helper)
 
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
         threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor

paddle/fluid/framework/details/CMakeLists.txt

@@ -9,6 +9,7 @@ cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place
 cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper)
 cc_library(multi_devices_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper)
 cc_library(multi_devices_graph_check_pass SRCS multi_devices_graph_check_pass.cc DEPS multi_devices_helper)
+cc_library(alloc_continuous_space_for_grad_pass SRCS alloc_continuous_space_for_grad_pass.cc DEPS graph graph_helper)
 
 cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
 
@@ -22,6 +23,8 @@ endif()
 if(WITH_GPU)
     nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
             dynload_cuda variable_visitor)
+    nv_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
+            dynload_cuda variable_visitor)
     if(WITH_DISTRIBUTE)
         nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
             ddim dynload_cuda selected_rows_functor sendrecvop_rpc)
@@ -35,6 +38,8 @@ if(WITH_GPU)
 else()
     cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
              variable_visitor)
+    cc_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
+            variable_visitor)
     if(WITH_DISTRIBUTE)
         cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
             ddim selected_rows_functor sendrecvop_rpc)
@@ -46,9 +51,7 @@ else()
     cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
 endif()
 
-cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor)
 cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
-cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
 
 if(WITH_GPU)
 cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info)
@@ -61,14 +64,17 @@ cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_
 cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
 cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle)
 cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper)
-cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass)
+cc_library(while_op_eager_deletion_pass SRCS while_op_eager_deletion_pass.cc DEPS while_op_helper graph_helper pass computation_op_handle)
+cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass while_op_eager_deletion_pass)
 cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle graph graph_helper pass op_graph_view reference_count_pass_helper)
 
 cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass)
 cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
 
 cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
-        scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
+        scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
+
+cc_library(fuse_all_reduce_op_pass SRCS fuse_all_reduce_op_pass.cc DEPS graph graph_helper fused_all_reduce_op_handle)
 
 set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass inplace_op_pass)
 if (WITH_GPU)
@@ -97,5 +103,5 @@ cc_library(build_strategy SRCS build_strategy.cc DEPS
         graph_viz_pass multi_devices_graph_pass
         multi_devices_graph_print_pass multi_devices_graph_check_pass
         fuse_elewise_add_act_pass multi_batch_merge_pass 
-        fuse_relu_depthwise_conv_pass
-        memory_optimize_pass lock_free_optimize_pass)
+        fuse_relu_depthwise_conv_pass 
+        memory_optimize_pass lock_free_optimize_pass alloc_continuous_space_for_grad_pass fuse_all_reduce_op_pass)

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -11,9 +11,8 @@
 // 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 <algorithm>
-
 #include "paddle/fluid/framework/details/all_reduce_op_handle.h"
+#include <algorithm>
 #include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/reduce_and_gather.h"
 #include "paddle/fluid/framework/details/variable_visitor.h"
@@ -56,6 +55,7 @@ void AllReduceOpHandle::RunImpl() {
   platform::RecordEvent record_event(Name());
 
   WaitInputVarGenerated();
+
   auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
   auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
   PADDLE_ENFORCE_EQ(

paddle/fluid/framework/details/alloc_continuous_space_for_grad_pass.cc

+//   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <algorithm>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "paddle/fluid/framework/details/build_strategy.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/framework/op_registry.h"
+DEFINE_uint32(fuse_parameter_memory_size, 0,  // 0 KB
+              "fuse_parameter_memory_size is up limited memory size "
+              "of one group parameters' gradient which is the input "
+              "of communication calling(e.g NCCLAllReduce). "
+              "The default value is 0, it means that "
+              "not set group according to memory_size.");
+DEFINE_int32(
+    fuse_parameter_groups_size, 3,
+    "fuse_parameter_groups_size is the size of one group parameters' gradient. "
+    "The default value is a experimental result. If the "
+    "fuse_parameter_groups_size is 1, it means that the groups size is "
+    "the number of parameters' gradient. If the fuse_parameter_groups_size is "
+    "-1, it means that there are only one group. The default value is 3, it is "
+    "an experimental value.");
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+static const char kUnKnow[] = "@UNKNOW@";
+static framework::proto::VarType::Type kDefaultDtype =
+    framework::proto::VarType::Type::VarType_Type_BOOL;
+
+class AllocContinuousSpaceForGradPass : public ir::Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override {
+    ir::Graph &result = *graph;
+
+    auto &places = Get<const std::vector<platform::Place>>(kPlaces);
+    auto &local_scopes = Get<const std::vector<Scope *>>(kLocalScopes);
+
+    ResetAttribute<ParamsAndGrads>(kParamsAndGrads, &result);
+    ResetAttribute<GroupGradsAndParams>(kGroupGradsAndParams, &result);
+
+    // NOTE: The operator nodes should be in topology order.
+    std::vector<ir::Node *> topo_nodes = ir::TopologySortOperations(result);
+    auto &params_grads = result.Get<ParamsAndGrads>(kParamsAndGrads);
+    for (auto &node : topo_nodes) {
+      RecordParamsAndGrads(node, &params_grads);
+    }
+
+    if (params_grads.size() == 0) {
+      VLOG(10) << "Doesn't find gradients";
+      return std::move(graph);
+    }
+
+    std::unordered_map<std::string, ir::Node *> vars;
+    for (ir::Node *node : result.Nodes()) {
+      if (node->IsVar() && node->Var()) {
+        // Note: The graph may have the same name node. For example, parameter
+        // is the input of operator and it also is the output of optimizer;
+        vars.emplace(node->Var()->Name(), node);
+      }
+    }
+
+    auto &group_grads_params =
+        result.Get<GroupGradsAndParams>(kGroupGradsAndParams);
+
+    // Note: the order of params_grads may be changed by SetGroupGradsAndParams.
+    SetGroupGradsAndParams(vars, params_grads, &group_grads_params);
+
+    params_grads.clear();
+    for (auto &group_p_g : group_grads_params) {
+      params_grads.insert(params_grads.begin(), group_p_g.begin(),
+                          group_p_g.end());
+    }
+    for (auto &p_g : params_grads) {
+      std::swap(p_g.first, p_g.second);
+    }
+
+    // Set Gradients as Persistable to prevent this var becoming reusable.
+    auto dtype = kDefaultDtype;
+    for (auto &p_g : params_grads) {
+      // Get gradient var
+      auto iter = vars.find(p_g.second);
+      PADDLE_ENFORCE(iter != vars.end(), "%s is not found.", p_g.second);
+      iter->second->Var()->SetPersistable(true);
+
+      PADDLE_ENFORCE(IsSupportedVarType(iter->second->Var()->GetType()));
+
+      // Get Dtype
+      auto ele_dtype = iter->second->Var()->GetDataType();
+      if (dtype == kDefaultDtype) {
+        dtype = ele_dtype;
+        PADDLE_ENFORCE_NE(ele_dtype, kDefaultDtype);
+      }
+      PADDLE_ENFORCE_EQ(ele_dtype, dtype);
+    }
+
+    // Create the fused variable name.
+    if (!result.Has(kFusedVars)) {
+      result.Set(kFusedVars, new FusedVars);
+    }
+    const std::string prefix(kFusedVarNamePrefix);
+    // The fused_var_name should be unique.
+    auto fused_var_name = prefix + "GRAD@" + params_grads[0].second;
+    auto &fused_var_set = result.Get<FusedVars>(kFusedVars);
+    PADDLE_ENFORCE_EQ(fused_var_set.count(fused_var_name), 0);
+    fused_var_set.insert(fused_var_name);
+
+    InitFusedVarsAndAllocSpaceForVars(places, local_scopes, vars,
+                                      fused_var_name, params_grads);
+
+    return std::move(graph);
+  }
+
+  template <typename AttrType>
+  void ResetAttribute(const std::string &attr_name, ir::Graph *graph) const {
+    if (graph->Has(attr_name)) {
+      VLOG(10) << attr_name << " is reset.";
+      graph->Erase(attr_name);
+    }
+    graph->Set(attr_name, new AttrType);
+  }
+
+  void SetGroupGradsAndParams(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      const ParamsAndGrads &params_grads,
+      GroupGradsAndParams *group_grads_params) const {
+    SetGroupAccordingToLayers(var_nodes, params_grads, group_grads_params);
+    SetGroupAccordingToMemorySize(var_nodes, group_grads_params);
+    SetGroupAccordingToGroupSize(var_nodes, group_grads_params);
+  }
+
+  void SetGroupAccordingToLayers(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      const ParamsAndGrads &params_grads,
+      GroupGradsAndParams *group_grads_params) const {
+    std::unordered_map<std::string, std::vector<int>> layer_params;
+
+    for (size_t i = 0; i < params_grads.size(); ++i) {
+      auto pos = params_grads[i].first.find_first_of(".");
+      if (pos == std::string::npos) {
+        layer_params[std::string(kUnKnow)].emplace_back(i);
+      } else {
+        layer_params[params_grads[i].first.substr(0, pos)].emplace_back(i);
+      }
+    }
+
+    group_grads_params->reserve(layer_params.size());
+    for (size_t i = 0; i < params_grads.size(); ++i) {
+      auto pos = params_grads[i].first.find_first_of(".");
+      std::string key = kUnKnow;
+      if (pos != std::string::npos) {
+        key = params_grads[i].first.substr(0, pos);
+      }
+      auto iter = layer_params.find(key);
+      if (iter == layer_params.end()) continue;
+
+      group_grads_params->emplace_back();
+      auto &local_group_grads_params = group_grads_params->back();
+      for (auto &idx : iter->second) {
+        local_group_grads_params.emplace_back(
+            std::make_pair(params_grads[idx].second, params_grads[idx].first));
+      }
+      layer_params.erase(iter);
+    }
+
+    VLOG(10) << "SetGroupAccordingToLayers: ";
+    for (size_t i = 0; i < group_grads_params->size(); ++i) {
+      VLOG(10) << "group " << i;
+      std::stringstream out;
+      for (auto &p_g : group_grads_params->at(i)) {
+        out << "(" << p_g.second << ", " << p_g.first << "), ";
+      }
+      VLOG(10) << out.str();
+    }
+  }
+
+  void SetGroupAccordingToMemorySize(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      GroupGradsAndParams *group_grads_params) const {
+    if (FLAGS_fuse_parameter_memory_size == 0) {
+      return;
+    }
+    size_t group_memory_size =
+        static_cast<size_t>(FLAGS_fuse_parameter_memory_size);
+    GroupGradsAndParams local_group_grads_params;
+
+    size_t j = 0;
+    while (j < group_grads_params->size()) {
+      local_group_grads_params.emplace_back();
+      auto &group_p_g = local_group_grads_params.back();
+      size_t local_group_memory_size = 0;
+      while (j < group_grads_params->size()) {
+        std::for_each(
+            group_grads_params->at(j).begin(), group_grads_params->at(j).end(),
+            [&local_group_memory_size,
+             &var_nodes](const std::pair<std::string, std::string> &g_p) {
+              auto iter = var_nodes.find(g_p.second);
+              PADDLE_ENFORCE(iter != var_nodes.end(), "%s is not found.",
+                             g_p.second);
+              auto shape = iter->second->Var()->GetShape();
+              size_t size =
+                  framework::SizeOfType(iter->second->Var()->GetDataType());
+              std::for_each(shape.begin(), shape.end(),
+                            [&size](const int64_t &n) { size *= n; });
+              local_group_memory_size += size;
+            });
+        group_p_g.insert(group_p_g.end(), group_grads_params->at(j).begin(),
+                         group_grads_params->at(j).end());
+        ++j;
+        if (local_group_memory_size >= group_memory_size) {
+          break;
+        }
+      }
+    }
+
+    std::swap(*group_grads_params, local_group_grads_params);
+
+    VLOG(10) << string::Sprintf(
+        "SetGroupAccordingToMemorySize(memory_size: %d):",
+        FLAGS_fuse_parameter_memory_size);
+    for (size_t i = 0; i < group_grads_params->size(); ++i) {
+      VLOG(10) << "group " << i;
+      std::stringstream out;
+      for (auto &g_p : group_grads_params->at(i)) {
+        auto iter = var_nodes.find(g_p.second);
+        PADDLE_ENFORCE(iter != var_nodes.end(), "%s is not found.", g_p.second);
+        auto shape = iter->second->Var()->GetShape();
+        size_t size = framework::SizeOfType(iter->second->Var()->GetDataType());
+        std::for_each(shape.begin(), shape.end(),
+                      [&size](const int64_t &n) { size *= n; });
+        out << string::Sprintf("(%s(%d), %s)", g_p.second, size, g_p.first);
+      }
+      VLOG(10) << out.str();
+    }
+  }
+
+  void SetGroupAccordingToGroupSize(
+      const std::unordered_map<std::string, ir::Node *> &var_nodes,
+      GroupGradsAndParams *group_grads_params) const {
+    if (FLAGS_fuse_parameter_groups_size == 1) {
+      return;
+    }
+    size_t group_size = static_cast<size_t>(FLAGS_fuse_parameter_groups_size);
+    if (FLAGS_fuse_parameter_groups_size == -1) {
+      group_size = group_grads_params->size();
+    }
+    PADDLE_ENFORCE_GT(group_size, 1);
+    size_t groups = (group_grads_params->size() + group_size - 1) / group_size;
+    GroupGradsAndParams local_group_grads_params;
+    local_group_grads_params.reserve(groups);
+
+    size_t j = 0;
+    for (size_t i = 0; i < groups; ++i) {
+      local_group_grads_params.emplace_back();
+      auto &group_p_g = local_group_grads_params.back();
+      group_p_g.reserve(group_size);
+      while (j < group_grads_params->size()) {
+        group_p_g.insert(group_p_g.end(), group_grads_params->at(j).begin(),
+                         group_grads_params->at(j).end());
+        ++j;
+        if (j % group_size == 0) break;
+      }
+    }
+    std::swap(*group_grads_params, local_group_grads_params);
+
+    VLOG(10) << "SetGroupAccordingToGroupSize(group_size: " << group_size
+             << "): ";
+    for (size_t i = 0; i < group_grads_params->size(); ++i) {
+      VLOG(10) << "group " << i;
+      std::stringstream out;
+      for (auto &p_g : group_grads_params->at(i)) {
+        out << "(" << p_g.second << ", " << p_g.first << "), ";
+      }
+      VLOG(10) << out.str();
+    }
+  }
+
+ private:
+  bool IsSupportedVarType(const proto::VarType::Type &type) const {
+    // Current only support LOD_TENSOR.
+    return type == proto::VarType::LOD_TENSOR;
+  }
+
+  void AppendAllocSpaceForVarsOp(const std::vector<std::string> &params_name,
+                                 const std::vector<std::string> &grads_name,
+                                 const std::string &fused_var_name,
+                                 BlockDesc *global_block) const {
+    auto op_desc = global_block->AppendOp();
+    op_desc->SetType("alloc_continuous_space");
+    op_desc->SetInput("Input", params_name);
+    op_desc->SetOutput("Output", grads_name);
+    op_desc->SetOutput("FusedOutput", {fused_var_name});
+  }
+
+  void RecordParamsAndGrads(ir::Node *node,
+                            ParamsAndGrads *params_grads) const {
+    try {
+      bool is_bk_op =
+          static_cast<bool>(boost::get<int>(node->Op()->GetAttr(
+                                OpProtoAndCheckerMaker::OpRoleAttrName())) &
+                            static_cast<int>(OpRole::kBackward));
+      if (!is_bk_op) return;
+
+      // Currently, we assume that once gradient is generated, it can be
+      // broadcast, and each gradient is only broadcast once.
+      auto backward_vars =
+          boost::get<std::vector<std::string>>(node->Op()->GetNullableAttr(
+              OpProtoAndCheckerMaker::OpRoleVarAttrName()));
+      PADDLE_ENFORCE_EQ(backward_vars.size() % 2, static_cast<size_t>(0));
+
+      for (size_t i = 0; i < backward_vars.size(); i += 2) {
+        VLOG(10) << "Trainable parameter: " << backward_vars[i]
+                 << ", gradient: " << backward_vars[i + 1];
+
+        params_grads->emplace_back(std::make_pair(
+            backward_vars[i] /*param*/, backward_vars[i + 1] /*grad*/));
+      }
+    } catch (boost::bad_get e) {
+    }
+  }
+
+  void InitFusedVarsAndAllocSpaceForVars(
+      const std::vector<platform::Place> &places,
+      const std::vector<Scope *> &local_scopes,
+      const std::unordered_map<std::string, ir::Node *> &vars,
+      const std::string &fused_var_name,
+      const ParamsAndGrads &params_grads) const {
+    //  Init Gradients and FusedVars
+    VLOG(10) << "Init FusedVars and Gradients.";
+    for (auto it = local_scopes.rbegin(); it != local_scopes.rend(); ++it) {
+      auto &scope = *it;
+
+      PADDLE_ENFORCE(scope->FindVar(fused_var_name) == nullptr,
+                     "%s has existed in scope.", fused_var_name);
+      scope->Var(fused_var_name)->GetMutable<LoDTensor>();
+
+      for (auto &p_g : params_grads) {
+        auto iter = vars.find(p_g.second);
+        PADDLE_ENFORCE(iter != vars.end());
+        PADDLE_ENFORCE_NOT_NULL(iter->second->Var());
+        PADDLE_ENFORCE_EQ(iter->second->Var()->GetType(),
+                          proto::VarType::LOD_TENSOR);
+        scope->Var(p_g.second)->GetMutable<LoDTensor>();
+      }
+    }
+
+    std::vector<std::string> grads_name;
+    std::vector<std::string> params_name;
+    grads_name.reserve(params_grads.size());
+    params_name.reserve(params_grads.size());
+    for (auto &p_g : params_grads) {
+      params_name.emplace_back(p_g.first);
+      grads_name.emplace_back(p_g.second);
+    }
+    framework::ProgramDesc program_desc;
+    AppendAllocSpaceForVarsOp(params_name, grads_name, fused_var_name,
+                              program_desc.MutableBlock(0));
+
+    // Run Only Once Programs
+    for (size_t i = 0; i < local_scopes.size(); ++i) {
+      for (auto &op_desc : program_desc.Block(0).AllOps()) {
+        auto op = OpRegistry::CreateOp(*op_desc);
+        op->Run(*local_scopes[i], places[i]);
+      }
+    }
+  }
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(alloc_continuous_space_for_grad_pass,
+              paddle::framework::details::AllocContinuousSpaceForGradPass)
+    .RequirePassAttr(paddle::framework::details::kPlaces)
+    .RequirePassAttr(paddle::framework::details::kLocalScopes);

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -57,7 +57,7 @@ struct BroadcastOpHandle : public OpHandleBase {
 
   std::string Name() const override;
 
-  bool IsMultiDeviceTransfer() override { return false; };
+  bool IsMultiDeviceTransfer() override { return true; };
 
  protected:
   void RunImpl() override;

paddle/fluid/framework/details/build_strategy.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <glog/logging.h>
 #include <memory>
+#include <utility>
 
 #include "paddle/fluid/framework/details/memory_optimize_helper.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
@@ -45,12 +46,27 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
  public:
   explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy)
       : ir::PassBuilder(), strategy_(strategy) {
+    // Add a graph viz pass to record a graph.
+    if (!strategy_.debug_graphviz_path_.empty()) {
+      auto viz_pass = AppendPass("graph_viz_pass");
+      const std::string graph_path = string::Sprintf(
+          "%s%s", strategy_.debug_graphviz_path_.c_str(), "_original_graph");
+      viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
+    }
+
     if (strategy_.enable_sequential_execution_) {
+      VLOG(10) << "Add sequential_execution_pass";
       AppendPass("sequential_execution_pass");
     }
 
+    // Add op fusion.
+    if (strategy.sync_batch_norm_) {
+      AppendPass("sync_batch_norm_pass");
+    }
+
     // Add op fusion.
     if (strategy.fuse_relu_depthwise_conv_) {
+      VLOG(10) << "Add fuse_relu_depthwise_conv_pass";
       AppendPass("fuse_relu_depthwise_conv_pass");
     }
 
@@ -62,29 +78,30 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
 
     // Add automatically inplace.
     if (strategy_.enable_inplace_) {
+      VLOG(10) << "Add inplace_pass";
       AppendPass("inplace_pass");
     }
 
+    if (strategy.fuse_elewise_add_act_ops_) {
+      VLOG(10) << "Add fuse_elewise_add_act_pass";
+      AppendPass("fuse_elewise_add_act_pass");
+    }
+
+    // for single card training, fuse_all_reduce_ops is unnecessary.
+    // alloc_continuous_space_for_grad_pass should be before of MultiDevPass.
+    if (strategy.fuse_all_reduce_ops_) {
+      VLOG(10) << "Add alloc_continuous_space_for_grad_pass";
+      AppendPass("alloc_continuous_space_for_grad_pass");
+    }
+
     // Add a graph viz pass to record a graph.
-    if (!strategy_.debug_graphviz_path_.empty()) {
+    if (!strategy.debug_graphviz_path_.empty()) {
       auto viz_pass = AppendPass("graph_viz_pass");
       const std::string graph_path = string::Sprintf(
-          "%s%s", strategy_.debug_graphviz_path_.c_str(), "_original_graph");
+          "%s%s", strategy.debug_graphviz_path_.c_str(), "_fused_graph");
       viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
     }
 
-    if (strategy.fuse_elewise_add_act_ops_) {
-      auto fuse_elewise_add_act_pass = AppendPass("fuse_elewise_add_act_pass");
-      // Add a graph viz pass to record a graph.
-      if (!strategy.debug_graphviz_path_.empty()) {
-        auto viz_pass = AppendPass("graph_viz_pass");
-        const std::string graph_path = string::Sprintf(
-            "%s%s", strategy.debug_graphviz_path_.c_str(), "_fused_graph");
-        viz_pass->Set<std::string>("graph_viz_path",
-                                   new std::string(graph_path));
-      }
-    }
-
     CollectiveContext *context = CollectiveContext::GetInstance();
     context->endpoints_ = strategy_.trainers_endpoints_;
     context->trainer_id_ = strategy_.trainer_id_;
@@ -102,11 +119,19 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
     // A side-effect of that, memory optimize cannot forsee the fetched vars
     // , so fetchlist should be set persistable before call the Run interface.
     if (strategy.memory_optimize_) {
-      auto memory_optimize_pass = AppendPass("memory_optimize_pass");
+      VLOG(10) << "Add memory_optimize_pass";
+      AppendPass("memory_optimize_pass");
     }
 
     AppendMultiDevPass(strategy);
 
+    if (strategy.fuse_all_reduce_ops_) {
+      // NOTE: fuse_all_reduce_ops will count the number of all_reduce operator
+      // first, if the number is zero, fuse_all_reduce_ops will do nothing.
+      VLOG(10) << "Add fuse_all_reduce_op_pass";
+      AppendPass("fuse_all_reduce_op_pass");
+    }
+
     // Add a graph print pass to record a graph with device info.
     if (!strategy_.debug_graphviz_path_.empty()) {
       auto multi_devices_print_pass = AppendPass("multi_devices_print_pass");
@@ -122,28 +147,34 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
     // Verify that the graph is correct for multi-device executor.
     AppendPass("multi_devices_check_pass");
 
+    if (VLOG_IS_ON(2)) {
+      AppendPass("all_reduce_deps_pass");
+    }
+
     if (SeqOnlyAllReduceOps(strategy)) {
+      VLOG(10) << "Add all_reduce_deps_pass";
       AppendPass("all_reduce_deps_pass");
     }
 
     if (strategy_.remove_unnecessary_lock_) {
+      VLOG(10) << "Add modify_op_lock_and_record_event_pass";
       AppendPass("modify_op_lock_and_record_event_pass");
     }
   }
 
   // Convert graph to run on multi-devices.
   void AppendMultiDevPass(const BuildStrategy &strategy) {
-    ir::Pass *multi_devices_pass;
+    ir::Pass *multi_devices_pass = nullptr;
     if (strategy_.is_distribution_) {
-      VLOG(3) << "multi device parameter server mode";
+      VLOG(10) << "Add dist_multi_devices_pass";
       multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
     } else {
       if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
-        VLOG(3) << "multi devices collective mode with allreduce";
+        VLOG(10) << "Add all_reduce_mode_multi_devices_pass";
         multi_devices_pass =
-            AppendPass("allreduce_mode_multi_devices_pass").get();
+            AppendPass("all_reduce_mode_multi_devices_pass").get();
       } else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
-        VLOG(3) << "multi deivces collective mode with reduce";
+        VLOG(10) << "Add reduce_mode_multi_devices_pass";
         multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
       } else {
         PADDLE_THROW("Unknown reduce strategy.");
@@ -200,9 +231,26 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
 
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
-      pass->Erase("nccl_ctxs");
-      pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
+      pass->Erase(kNCCLCtxs);
+      pass->SetNotOwned<platform::NCCLContextMap>(kNCCLCtxs, nctx);
+#endif
+    } else if (pass->Type() == "fuse_all_reduce_op_pass") {
+      pass->Erase(kPlaces);
+      pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
+      pass->Erase(kLocalScopes);
+      pass->SetNotOwned<const std::vector<Scope *>>(kLocalScopes,
+                                                    &local_scopes);
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+      platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
+      pass->Erase(kNCCLCtxs);
+      pass->SetNotOwned<platform::NCCLContextMap>(kNCCLCtxs, nctx);
 #endif
+    } else if (pass->Type() == "alloc_continuous_space_for_grad_pass") {
+      pass->Erase(kPlaces);
+      pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
+      pass->Erase(kLocalScopes);
+      pass->SetNotOwned<const std::vector<Scope *>>(kLocalScopes,
+                                                    &local_scopes);
     } else if (pass->Type() == "sequential_execution_pass") {
       LOG(INFO) << "set enable_sequential_execution:"
                 << enable_sequential_execution_;
@@ -227,12 +275,13 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
 }  // namespace framework
 }  // namespace paddle
 
+USE_PASS(sync_batch_norm_pass);
 USE_PASS(fuse_relu_depthwise_conv_pass);
 USE_PASS(fuse_elewise_add_act_pass);
 USE_PASS(graph_viz_pass);
 USE_PASS(multi_batch_merge_pass);
 USE_PASS(reduce_mode_multi_devices_pass);
-USE_PASS(allreduce_mode_multi_devices_pass);
+USE_PASS(all_reduce_mode_multi_devices_pass);
 USE_PASS(dist_multi_devices_pass);
 USE_PASS(multi_devices_check_pass);
 USE_PASS(multi_devices_print_pass);
@@ -242,4 +291,6 @@ USE_PASS(all_reduce_deps_pass);
 USE_PASS(modify_op_lock_and_record_event_pass);
 USE_PASS(inplace_pass);
 USE_PASS(lock_free_optimize_pass);
+USE_PASS(alloc_continuous_space_for_grad_pass);
 USE_PASS(graph_to_program_pass);
+USE_PASS(fuse_all_reduce_op_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -14,7 +14,9 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
+#include <utility>
 #include <vector>
 
 #include "paddle/fluid/framework/ir/pass_builder.h"
@@ -74,13 +76,17 @@ struct BuildStrategy {
 
   bool fuse_elewise_add_act_ops_{false};
 
+  bool fuse_all_reduce_ops_{false};
+
   bool fuse_relu_depthwise_conv_{false};
 
-  bool memory_optimize_{false};
+  bool sync_batch_norm_{false};
+
+  bool memory_optimize_{true};
   // TODO(dzhwinter):
   // make enable_inplace, memory_optimize_
   // memory_early_delete_ true by default
-  bool enable_inplace_{false};
+  bool enable_inplace_{true};
 
   bool enable_sequential_execution_{false};
 

paddle/fluid/framework/details/computation_op_handle.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <vector>
 
@@ -31,6 +32,8 @@ class ComputationOpHandle : public OpHandleBase {
   ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,
                       size_t scope_idx);
 
+  OperatorBase *GetOp() { return op_.get(); }
+
   std::string Name() const override;
 
   const Scope *GetScope() const { return scope_; }

paddle/fluid/framework/details/data_balance_op_handle.cc

-// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#include "paddle/fluid/framework/details/data_balance_op_handle.h"
-#include <algorithm>
-#include "paddle/fluid/framework/details/container_cast.h"
-
-namespace paddle {
-namespace framework {
-namespace details {
-
-#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-DataBalanceOpHandle::DataBalanceOpHandle(
-    ir::Node *node, const std::vector<Scope *> &local_scopes,
-    const std::vector<platform::Place> &places,
-    const platform::NCCLContextMap *ctxs)
-    : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {
-  if (ctxs) {
-    for (auto &p : places_) {
-      this->SetDeviceContext(p, ctxs->DevCtx(p));
-    }
-  }
-}
-#else
-DataBalanceOpHandle::DataBalanceOpHandle(
-    ir::Node *node, const std::vector<Scope *> &local_scopes,
-    const std::vector<platform::Place> &places)
-    : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {}
-#endif
-
-std::string DataBalanceOpHandle::Name() const { return "data balance"; }
-
-std::vector<std::array<int, 3>> DataBalanceOpHandle::GetBalancePlan(
-    const std::vector<int> &device_sizes) {
-  int device_num = device_sizes.size();
-  int total_size = 0;
-  int empty_num = 0;
-  std::vector<std::array<int, 2>> size_device_vec;
-  size_device_vec.reserve(device_num);
-  for (int i = 0; i < device_num; ++i) {
-    if (device_sizes[i] == 0) {
-      ++empty_num;
-    }
-    total_size += device_sizes[i];
-    size_device_vec.push_back({{device_sizes[i], i}});
-  }
-  std::vector<std::array<int, 3>> res;
-  if (empty_num == 0) {
-    // No need to do data balance.
-    return res;
-  }
-  if (total_size < device_num) {
-    // No enough data.
-    PADDLE_THROW_EOF();
-  }
-  std::sort(size_device_vec.begin(), size_device_vec.end(),
-            [](const std::array<int, 2> &a, const std::array<int, 2> &b) {
-              return a[0] > b[0];
-            });
-  int expected_device_size = total_size / device_num;
-  int src_idx = 0;
-  for (int dst_idx = device_num - empty_num; dst_idx < device_num; ++dst_idx) {
-    if (size_device_vec[src_idx][0] <= expected_device_size) {
-      ++src_idx;
-      PADDLE_ENFORCE_LT(
-          src_idx, device_num - empty_num,
-          "In current srategy an empty tensor should not be copy source.");
-    }
-    size_device_vec[src_idx][0] -= expected_device_size;
-    size_device_vec[dst_idx][0] += expected_device_size;
-    res.push_back({{size_device_vec[src_idx][1], size_device_vec[dst_idx][1],
-                    expected_device_size}});
-  }
-  return res;
-}
-
-void DataBalanceOpHandle::RunImpl() {
-  PADDLE_ENFORCE_GT(places_.size(), 1UL,
-                    "Data balance can only be enabled when the number of "
-                    "places to run larger than 1.");
-  auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
-  auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
-  PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0);
-  PADDLE_ENFORCE_EQ(
-      in_var_handles.size(), out_var_handles.size(),
-      "The NoDummyInputSize and NoDummyOutputSize should be equal.");
-  int data_num = in_var_handles.size() / places_.size();
-  WaitInputVarGenerated();
-  std::vector<std::vector<LoDTensor *>> lod_tensors(data_num);
-  std::vector<int> device_sizes;
-  for (int i = 0; i < static_cast<int>(in_var_handles.size()); ++i) {
-    PADDLE_ENFORCE_EQ(in_var_handles[i]->name(), out_var_handles[i]->name(),
-                      "The name of input and output should be equal.");
-    int place_idx = i / data_num;
-    int data_idx = i % data_num;
-    auto *local_scope =
-        local_scopes_[place_idx]->FindVar(kLocalExecScopeName)->Get<Scope *>();
-    auto *tensor_var = local_scope->FindVar(in_var_handles[i]->name());
-    PADDLE_ENFORCE(tensor_var->IsType<LoDTensor>());
-    auto *tensor = tensor_var->GetMutable<LoDTensor>();
-    lod_tensors[data_idx].push_back(tensor);
-    int ins_size =
-        tensor->lod().empty() ? tensor->dims()[0] : tensor->NumElements();
-    if (data_idx == 0) {
-      device_sizes.emplace_back(ins_size);
-    } else {
-      PADDLE_ENFORCE_EQ(
-          ins_size, device_sizes.at(place_idx),
-          "All data on the same device shall have the same batch size.");
-    }
-  }
-  const auto &balance_plan = GetBalancePlan(device_sizes);
-
-  for (const auto &trans : balance_plan) {
-    for (int data_idx = 0; data_idx < data_num; ++data_idx) {
-      LoDTensor *src_tensor = lod_tensors[data_idx][trans[0]];
-      LoDTensor *dst_tensor = lod_tensors[data_idx][trans[1]];
-      int trans_ins_size = trans[2];
-      LoD src_lod = src_tensor->lod();
-      int src_ins_size =
-          src_lod.empty() ? src_tensor->dims()[0] : src_tensor->NumElements();
-      int cut_point = src_ins_size - trans_ins_size;
-      if (!src_lod.empty()) {
-        for (auto &level : src_lod) {
-          cut_point = level[cut_point];
-        }
-      }
-      TensorCopySync(src_tensor->Slice(cut_point, src_tensor->dims()[0]),
-                     dst_tensor->place(), dst_tensor);
-      src_tensor->ShareDataWith(src_tensor->Slice(0, cut_point));
-      if (!src_lod.empty()) {
-        dst_tensor->set_lod(SliceInLevel(
-            src_lod, 0, src_ins_size - trans_ins_size, src_ins_size));
-        src_tensor->set_lod(
-            SliceInLevel(src_lod, 0, 0, src_ins_size - trans_ins_size));
-      }
-    }
-  }
-}
-
-}  // namespace details
-}  // namespace framework
-}  // namespace paddle

paddle/fluid/framework/details/eager_deletion_op_handle.cc

@@ -12,6 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <memory>
+#include <unordered_set>
+#include <utility>
+
 #include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/scope.h"
@@ -45,6 +49,7 @@ EagerDeletionOpHandle::EagerDeletionOpHandle(
     }
   }
 #endif
+  PADDLE_ENFORCE(!var_names_.empty(), "Var names cannot be empty");
 }
 
 EagerDeletionOpHandle::~EagerDeletionOpHandle() {
@@ -60,15 +65,20 @@ EagerDeletionOpHandle::~EagerDeletionOpHandle() {
 std::string EagerDeletionOpHandle::Name() const { return "eager_deletion"; }
 
 void EagerDeletionOpHandle::RunImpl() {
-  auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
+  Scope *exec_scope = nullptr;
   std::deque<std::shared_ptr<memory::Allocation>> garbages;
   for (auto &name : var_names_) {
     auto it = ref_cnts_->find(name);
-    // Var not found, not reference count has not decreased to 0
+    // Reference count has not decreased to 0
     if (it == ref_cnts_->end() || it->second.fetch_sub(1) != 1) {
       continue;
     }
 
+    if (!exec_scope) {
+      exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
+    }
+
+    // Var not found
     auto *var = exec_scope->FindVar(name);
     if (var == nullptr) {
       continue;

paddle/fluid/framework/details/eager_deletion_pass.cc

@@ -12,20 +12,173 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <algorithm>
+#include <functional>
 #include <queue>
 #include <string>
+#include <tuple>
 #include <vector>
 
 #include "paddle/fluid/framework/details/computation_op_handle.h"
 #include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
-#include "paddle/fluid/framework/details/eager_deletion_pass.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 
+DEFINE_double(memory_fraction_of_eager_deletion, 1.0,
+              "Fraction of eager deletion. If less than 1.0, all variables in "
+              "the program would be sorted according to its memory size, and "
+              "only the FLAGS_memory_fraction_of_eager_deletion of the largest "
+              "variables would be deleted.");
+
 namespace paddle {
 namespace framework {
 namespace details {
 
+// op -> variables which can be deleted after op runs
+using OpToVarNameSetMap =
+    std::unordered_map<ComputationOpHandle *, std::unordered_set<std::string>>;
+
+// Check whether the variable is LoDTensor based on static VarDesc info
+static bool IsLoDTensor(VarDesc *var) {
+  return var->Proto()->type().type() == proto::VarType::LOD_TENSOR;
+}
+
+// Get memory size of LoDTensor
+static int64_t GetMemorySize(
+    const std::unordered_map<std::string, std::vector<VarHandle *>> &vars,
+    const std::string &var_name) {
+  auto *var_desc = TryGetLatestVarDesc(vars.at(var_name));
+  PADDLE_ENFORCE_NOT_NULL(var_desc);
+  PADDLE_ENFORCE(IsLoDTensor(var_desc));
+  auto dims = var_desc->GetShape();
+  return SizeOfType(var_desc->GetDataType()) *
+         std::accumulate(dims.begin(), dims.end(), static_cast<int64_t>(1),
+                         std::multiplies<int64_t>());
+}
+
+// Split all variables in the graph into LoDTensor and Non-LoDTensor (e.g.
+// SelectedRows, LoDTensorArray)
+// Since partial GC is based on static analysis of memory size of each variable
+// So we should skip SelectedRows and LoDTensorArray here
+static void SplitIntoLoDTensorAndNonLoDTensorVars(
+    const OpToVarNameSetMap &m, const GraphVars &vars,
+    OpToVarNameSetMap *lod_tensors, OpToVarNameSetMap *other_vars) {
+  lod_tensors->clear();
+  other_vars->clear();
+
+  for (auto &op_vars_pair : m) {
+    for (auto &var_name : op_vars_pair.second) {
+      auto *var_desc = TryGetLatestVarDesc(
+          vars[op_vars_pair.first->GetScopeIdx()].at(var_name));
+      if (IsLoDTensor(var_desc)) {
+        (*lod_tensors)[op_vars_pair.first].insert(var_name);
+      } else {
+        (*other_vars)[op_vars_pair.first].insert(var_name);
+      }
+    }
+  }
+}
+
+struct GCVarInfo {
+  GCVarInfo(const std::string &name, int64_t memory_size,
+            ComputationOpHandle *op, size_t scope_idx)
+      : name_(name),
+        memory_size_(memory_size),
+        op_(op),
+        scope_idx_(scope_idx) {}
+
+  std::string name_;         // variable name
+  int64_t memory_size_;      // memory size
+  ComputationOpHandle *op_;  // op after which the variable could be deleted
+  size_t scope_idx_;         // scope index where the variable locates
+
+  int64_t AbsMemorySize() const { return std::abs(memory_size_); }
+};
+
+// Delete delete_lod_tensor_only is not used currently
+static OpToVarNameSetMap ShrinkGCVars(
+    const OpToVarNameSetMap &m, const GraphVars &vars,
+    const std::vector<platform::Place> &places, double fraction_of_memory_size,
+    bool delete_lod_tensor_only = false) {
+  // Do not perform gc when fraction_of_memory_size = 0
+  if (fraction_of_memory_size <= 0.0) return {};
+
+  /**
+   * Step 1: Split all variables into LoDTensor and Non-LoDTensor.
+   * We can only calculate memory size of LoDTensors
+   */
+  OpToVarNameSetMap lod_tensors, other_vars;
+  SplitIntoLoDTensorAndNonLoDTensorVars(m, vars, &lod_tensors, &other_vars);
+
+  // Perform complete gc when fraction_of_memory_size >= 1
+  if (fraction_of_memory_size >= 1.0) {
+    return delete_lod_tensor_only ? lod_tensors : m;
+  }
+
+  /**
+   * Step 2: build GCVarInfos, and calculate total memory sizes of each device
+   */
+
+  // place -> variable info (name, memory size, place, scope_idx)
+  std::map<platform::Place, std::vector<GCVarInfo>> place_to_vars;
+
+  // place -> total memory sizes
+  std::map<platform::Place, int64_t> place_to_size;
+  for (auto &op_vars_pair : lod_tensors) {
+    auto *op = op_vars_pair.first;
+    auto &var_names = op_vars_pair.second;
+    auto scope_idx = op->GetScopeIdx();
+    auto &place = places[scope_idx];
+
+    for (auto &var_name : var_names) {
+      auto var_size = GetMemorySize(vars[scope_idx], var_name);
+      GCVarInfo var_info(var_name, var_size, op, scope_idx);
+      place_to_size[place] += var_info.AbsMemorySize();
+      place_to_vars[place].emplace_back(std::move(var_info));
+    }
+  }
+
+  /**
+   * Step 3: sort GCVarInfos, and only delete the largest variables.
+   */
+  OpToVarNameSetMap partial_vars;
+  for (auto &place_to_var_pair : place_to_vars) {
+    auto &place = place_to_var_pair.first;
+    auto &gc_vars = place_to_var_pair.second;
+    std::sort(gc_vars.begin(), gc_vars.end(),
+              [](const GCVarInfo &var1, const GCVarInfo &var2) {
+                return var1.AbsMemorySize() > var2.AbsMemorySize();
+              });
+
+    int64_t accumulated_size = 0;
+    int64_t size_threshold =
+        static_cast<int64_t>(fraction_of_memory_size * place_to_size[place]);
+    for (size_t i = 0; i < gc_vars.size() && accumulated_size < size_threshold;
+         ++i) {
+      partial_vars[gc_vars[i].op_].insert(gc_vars[i].name_);
+      accumulated_size += gc_vars[i].AbsMemorySize();
+    }
+  }
+
+  /**
+   * Step 4: Combine other vars (SelectedRows, LoDTensorArray)
+   */
+  if (!delete_lod_tensor_only) {
+    for (auto &op_vars_pair : other_vars) {
+      partial_vars[op_vars_pair.first].insert(op_vars_pair.second.begin(),
+                                              op_vars_pair.second.end());
+    }
+  }
+
+  return partial_vars;
+}
+
+class EagerDeletionPass : public ir::Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+};
+
 std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
   auto &ref_cnts =
@@ -43,9 +196,7 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
 
   // a reverse map of last_live_ops
   //   i.e., last op --> variable names which can be deleted.
-  std::unordered_map<ComputationOpHandle *, std::unordered_set<std::string>>
-      op_vars_map;
-
+  OpToVarNameSetMap op_vars_map;
   for (auto &var_ops_map : last_live_ops) {
     for (auto &var_ops_pair : var_ops_map) {
       const std::string &var_name = var_ops_pair.first;
@@ -55,6 +206,9 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     }
   }
 
+  op_vars_map = ShrinkGCVars(op_vars_map, vars, places,
+                             FLAGS_memory_fraction_of_eager_deletion);
+
   for (auto &pair : op_vars_map) {
     auto *op = pair.first;
     auto &var_names = pair.second;
@@ -85,8 +239,13 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     eager_deletion_op->AddOutput(dummy_leaf);
   }
 
+  VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = "
+           << FLAGS_memory_fraction_of_eager_deletion;
   VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
-  return graph;
+
+  auto while_op_eager_deletion_pass =
+      ir::PassRegistry::Instance().Get("while_op_eager_deletion_pass");
+  return while_op_eager_deletion_pass->Apply(std::move(graph));
 }
 
 }  // namespace details
@@ -99,3 +258,5 @@ REGISTER_PASS(eager_deletion_pass,
     .RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars)
     .RequirePassAttr(paddle::framework::details::kAllPlaces)
     .RequirePassAttr(paddle::framework::details::kGarbageCollector);
+
+USE_PASS(while_op_eager_deletion_pass);

paddle/fluid/framework/details/fetch_op_handle.cc

@@ -82,6 +82,8 @@ void FetchOpHandle::WaitInputVarGenerated(const platform::Place &place) {
   }
 }
 
+bool FetchOpHandle::IsMultiDeviceTransfer() { return true; }
+
 std::string FetchOpHandle::Name() const { return "Fetch"; }
 
 }  // namespace details

paddle/fluid/framework/details/fetch_op_handle.h

@@ -39,6 +39,8 @@ struct FetchOpHandle : public OpHandleBase {
 
   std::string Name() const override;
 
+  bool IsMultiDeviceTransfer() override;
+
  protected:
   void RunImpl() override;
 

paddle/fluid/framework/details/fuse_all_reduce_op_pass.cc

+//   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <algorithm>
+#include <string>
+#include <vector>
+
+#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
+#include "paddle/fluid/framework/details/container_cast.h"
+#include "paddle/fluid/framework/details/fused_all_reduce_op_handle.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+class FuseAllReduceOpPass : public ir::Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override {
+    ir::Graph &result = *graph;
+
+    auto &places = Get<const std::vector<platform::Place>>(kPlaces);
+    auto &local_scopes = Get<const std::vector<Scope *>>(kLocalScopes);
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+    auto *nccl_ctxs = &Get<platform::NCCLContextMap>(kNCCLCtxs);
+#endif
+
+    std::unordered_set<std::string> grads;
+    auto &params_grads = result.Get<ParamsAndGrads>(kParamsAndGrads);
+    size_t num_of_all_reduce = params_grads.size();
+    grads.reserve(num_of_all_reduce);
+    for (auto p_g : params_grads) {
+      grads.insert(p_g.second);
+    }
+
+    size_t num_place = places.size();
+    std::unordered_map<std::string, ir::Node *> all_reduce_ops;
+    all_reduce_ops.reserve(grads.size());
+    for (auto &node : result.Nodes()) {
+      if (node->IsOp()) {
+        PADDLE_ENFORCE(node->IsWrappedBy<OpHandleBase>());
+        auto *all_reduce_op_handle =
+            dynamic_cast<AllReduceOpHandle *>(&node->Wrapper<OpHandleBase>());
+        if (all_reduce_op_handle) {
+          auto inputs = DynamicCast<VarHandle>(all_reduce_op_handle->Inputs());
+          PADDLE_ENFORCE_EQ(inputs.size(), num_place);
+          // The inputs' name should be the same.
+          auto &grad_name = inputs[0]->name();
+          for (size_t i = 1; i < inputs.size(); ++i) {
+            PADDLE_ENFORCE_EQ(inputs[i]->name(), grad_name,
+                              "The input name should be the same.");
+          }
+          PADDLE_ENFORCE_NE(grads.count(grad_name), static_cast<size_t>(0));
+          all_reduce_ops.emplace(grad_name, node);
+        }
+      }
+    }
+
+    VLOG(10) << "Find all_reduce_ops: " << all_reduce_ops.size();
+    if (all_reduce_ops.size() == 0) {
+      return std::move(graph);
+    }
+
+    PADDLE_ENFORCE_EQ(all_reduce_ops.size(), grads.size(),
+                      "The number of all_reduce OpHandle is not equal to the "
+                      "number of grads. Maybe some gradients are sparse type, "
+                      "it is not supported currently.");
+    VLOG(10) << "Insert fused_all_reduce";
+
+    auto &group_grads_params =
+        graph->Get<GroupGradsAndParams>(kGroupGradsAndParams);
+
+    for (auto &group_g_p : group_grads_params) {
+      size_t group_size = group_g_p.size();
+      PADDLE_ENFORCE_GT(group_size, static_cast<size_t>(0));
+      std::vector<ir::Node *> group_all_reduce_ops;
+      group_all_reduce_ops.reserve(group_size);
+      for (auto &g_p : group_g_p) {
+        group_all_reduce_ops.emplace_back(all_reduce_ops.at(g_p.first));
+      }
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+      InsertFusedAllReduce(places, local_scopes, group_size,
+                           group_all_reduce_ops, nccl_ctxs, &result);
+#else
+      InsertFusedAllReduce(places, local_scopes, group_size,
+                           group_all_reduce_ops, &result);
+#endif
+    }
+    return std::move(graph);
+  }
+
+  void InsertFusedAllReduce(const std::vector<platform::Place> &places,
+                            const std::vector<Scope *> &local_scopes,
+                            const size_t num_of_all_reduce,
+                            const std::vector<ir::Node *> &all_reduce_ops,
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+                            const platform::NCCLContextMap *nccl_ctxs,
+#endif
+                            ir::Graph *result) const {
+    std::vector<VarHandleBase *> inputs;
+    std::vector<VarHandleBase *> outputs;
+    for (auto &op : all_reduce_ops) {
+      auto &op_handle = op->Wrapper<OpHandleBase>();
+      inputs.insert(inputs.end(), op_handle.Inputs().begin(),
+                    op_handle.Inputs().end());
+      // Remove output
+      for_each(op_handle.Inputs().begin(), op_handle.Inputs().end(),
+               [&op_handle](VarHandleBase *var_handle) {
+                 var_handle->RemoveOutput(&op_handle, op_handle.Node());
+               });
+
+      outputs.insert(outputs.end(), op_handle.Outputs().begin(),
+                     op_handle.Outputs().end());
+      // Remove Input
+      for_each(
+          op_handle.Outputs().begin(), op_handle.Outputs().end(),
+          [](VarHandleBase *var_handle) { var_handle->ClearGeneratedOp(); });
+
+      result->RemoveNode(op_handle.Node());
+    }
+
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+    CreateFusedAllReduceOp(inputs, outputs, num_of_all_reduce, places,
+                           local_scopes, nccl_ctxs, result);
+#else
+    CreateFusedAllReduceOp(inputs, outputs, num_of_all_reduce, places,
+                           local_scopes, result);
+#endif
+  }
+
+ private:
+  void CreateFusedAllReduceOp(const std::vector<VarHandleBase *> &inputs,
+                              const std::vector<VarHandleBase *> &outputs,
+                              const size_t num_of_all_reduce,
+                              const std::vector<platform::Place> &places,
+                              const std::vector<Scope *> &local_scopes,
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+                              const platform::NCCLContextMap *nccl_ctxs,
+#endif
+                              ir::Graph *result) const {
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+    auto *op_handle = new FusedAllReduceOpHandle(
+        result->CreateEmptyNode("fused_all_reduce", ir::Node::Type::kOperation),
+        local_scopes, places, num_of_all_reduce, nccl_ctxs);
+#else
+    auto *op_handle = new FusedAllReduceOpHandle(
+        result->CreateEmptyNode("fused_all_reduce", ir::Node::Type::kOperation),
+        local_scopes, places, num_of_all_reduce);
+#endif
+
+    for (auto in : inputs) {
+      op_handle->AddInput(in);
+    }
+
+    for (auto out : outputs) {
+      op_handle->AddOutput(out);
+    }
+
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+    if (!nccl_ctxs) {
+      SetCommunicationContext(places, op_handle);
+    }
+#else
+    SetCommunicationContext(places, op_handle);
+#endif
+  }
+
+  void SetCommunicationContext(const std::vector<platform::Place> &places,
+                               FusedAllReduceOpHandle *op_handle) const {
+    for (size_t i = 0; i < places.size(); ++i) {
+      op_handle->SetDeviceContext(
+          places[i], platform::DeviceContextPool::Instance().Get(places[i]));
+    }
+  }
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(fuse_all_reduce_op_pass,
+              paddle::framework::details::FuseAllReduceOpPass);

paddle/fluid/framework/details/fused_all_reduce_op_handle.cc

+//   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+#include "paddle/fluid/framework/details/fused_all_reduce_op_handle.h"
+#include <algorithm>
+#include <utility>
+#include "paddle/fluid/framework/details/container_cast.h"
+#include "paddle/fluid/framework/details/reduce_and_gather.h"
+#include "paddle/fluid/framework/details/variable_visitor.h"
+#include "paddle/fluid/platform/profiler.h"
+
+DEFINE_bool(skip_fused_all_reduce_check, false, "");
+namespace paddle {
+namespace framework {
+namespace details {
+
+typedef std::vector<std::vector<std::pair<std::string, const LoDTensor *>>>
+    GradientAndLoDTensor;
+
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+FusedAllReduceOpHandle::FusedAllReduceOpHandle(
+    ir::Node *node, const std::vector<Scope *> &local_scopes,
+    const std::vector<platform::Place> &places, const size_t num_of_all_reduce,
+    const platform::NCCLContextMap *ctxs)
+    : OpHandleBase(node),
+      local_scopes_(local_scopes),
+      places_(places),
+      num_of_all_reduce_(num_of_all_reduce),
+      nccl_ctxs_(ctxs) {
+  if (nccl_ctxs_) {
+    for (auto &p : places_) {
+      this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
+    }
+  }
+  PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
+}
+#else
+
+FusedAllReduceOpHandle::FusedAllReduceOpHandle(
+    ir::Node *node, const std::vector<Scope *> &local_scopes,
+    const std::vector<platform::Place> &places, const size_t num_of_all_reduce)
+    : OpHandleBase(node),
+      local_scopes_(local_scopes),
+      places_(places),
+      num_of_all_reduce_(num_of_all_reduce) {
+  PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
+}
+
+#endif
+
+void FusedAllReduceOpHandle::RunImpl() {
+  platform::RecordEvent record_event(Name());
+
+  VLOG(4) << this->DebugString();
+
+  WaitInputVarGenerated();
+  // The input: grad0(dev0), grad0(dev1), grad1(dev0), grad1(dev1)...
+  // The output: grad0(dev0), grad0(dev1), grad1(dev0), grad1(dev1)...
+  auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
+  auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
+
+  size_t place_num = places_.size();
+  PADDLE_ENFORCE_EQ(
+      in_var_handles.size(), place_num * num_of_all_reduce_,
+      "The NoDummyInputSize should be equal to the number of places.");
+  PADDLE_ENFORCE_EQ(
+      in_var_handles.size(), out_var_handles.size(),
+      "The NoDummyInputSize and NoDummyOutputSize should be equal.");
+
+  GradientAndLoDTensor grads_tensor;
+  grads_tensor.resize(place_num);
+
+  int64_t numel = -1;
+  auto dtype = static_cast<framework::proto::VarType::Type>(0);
+  for (size_t scope_idx = 0; scope_idx < local_scopes_.size(); ++scope_idx) {
+    auto &g_tensor = grads_tensor.at(scope_idx);
+    g_tensor.reserve(num_of_all_reduce_);
+
+    GetGradLoDTensor(scope_idx, in_var_handles, out_var_handles, &g_tensor);
+
+    int64_t element_num = 0;
+    framework::proto::VarType::Type ele_dtype =
+        static_cast<framework::proto::VarType::Type>(0);
+    GetDTypeAndNumel(g_tensor, &ele_dtype, &element_num);
+
+    if (numel == -1) {
+      numel = element_num;
+    }
+    if (dtype == static_cast<framework::proto::VarType::Type>(0)) {
+      dtype = ele_dtype;
+      PADDLE_ENFORCE_NE(ele_dtype,
+                        static_cast<framework::proto::VarType::Type>(0));
+    }
+    PADDLE_ENFORCE_EQ(ele_dtype, dtype);
+
+    // Check whether the address space is contiguous.
+    std::sort(
+        g_tensor.begin(), g_tensor.end(),
+        [](const std::pair<std::string, const LoDTensor *> &grad1,
+           const std::pair<std::string, const LoDTensor *> &grad2) -> bool {
+          return grad1.second->data<void>() < grad2.second->data<void>();
+        });
+
+    for (size_t k = 1; k < g_tensor.size(); ++k) {
+      const void *cur_address = g_tensor.at(k - 1).second->data<void>();
+      int64_t len = g_tensor.at(k - 1).second->numel();
+      auto offset = len * framework::SizeOfType(dtype);
+      void *infer_next_address = reinterpret_cast<void *>(
+          reinterpret_cast<uintptr_t>(cur_address) + offset);
+      const void *next_address = g_tensor.at(k).second->data<void>();
+
+      VLOG(10) << string::Sprintf(
+          "Input[%d](%s) address: 0X%02x, Input[%d](%s) address: 0X%02x, Infer "
+          "input[%d] address: 0X%02x. The offset: %d",
+          k - 1, g_tensor.at(k - 1).first, cur_address, g_tensor.at(k).first, k,
+          next_address, k, infer_next_address, offset);
+      PADDLE_ENFORCE_EQ(infer_next_address, next_address,
+                        "The address is not consistent.");
+    }
+  }
+
+  if (!FLAGS_skip_fused_all_reduce_check) {
+    for (size_t scope_idx = 0; scope_idx < place_num; ++scope_idx) {
+      for (size_t j = 1; j < num_of_all_reduce_; ++j) {
+        PADDLE_ENFORCE_EQ(grads_tensor.at(0).at(j).first,
+                          grads_tensor.at(scope_idx).at(j).first);
+      }
+    }
+  }
+
+  std::vector<const void *> lod_tensor_data;
+  for (size_t scope_idx = 0; scope_idx < place_num; ++scope_idx) {
+    auto data = grads_tensor.at(scope_idx).at(0).second->data<void>();
+    lod_tensor_data.emplace_back(data);
+  }
+
+  if (platform::is_gpu_place(places_[0])) {
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+    PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
+    int nccl_dtype = platform::ToNCCLDataType(dtype);
+    std::vector<std::function<void()>> all_reduce_calls;
+    for (size_t i = 0; i < local_scopes_.size(); ++i) {
+      auto &p = places_[i];
+      void *buffer = const_cast<void *>(lod_tensor_data.at(i));
+
+      int dev_id = boost::get<platform::CUDAPlace>(p).device;
+      auto &nccl_ctx = nccl_ctxs_->at(dev_id);
+      auto stream = nccl_ctx.stream();
+      auto comm = nccl_ctx.comm_;
+      all_reduce_calls.emplace_back([=] {
+        PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
+            buffer, buffer, numel, static_cast<ncclDataType_t>(nccl_dtype),
+            ncclSum, comm, stream));
+      });
+    }
+
+    this->RunAndRecordEvent([&] {
+      if (all_reduce_calls.size() == 1UL) {
+        // Do not use NCCLGroup when manage NCCL by per thread per device
+        all_reduce_calls[0]();
+      } else {
+        platform::NCCLGroupGuard guard;
+        for (auto &call : all_reduce_calls) {
+          call();
+        }
+      }
+    });
+#else
+    PADDLE_THROW("Not compiled with CUDA");
+#endif
+  } else {
+    // Special handle CPU only Operator's gradient. Like CRF
+    auto grad_name = grads_tensor.at(0).at(0).first;
+    auto &trg = *this->local_scopes_[0]
+                     ->FindVar(kLocalExecScopeName)
+                     ->Get<Scope *>()
+                     ->FindVar(grad_name)
+                     ->GetMutable<framework::LoDTensor>();
+
+    // Reduce All data to trg in CPU
+    ReduceBufferData func(lod_tensor_data, trg.data<void>(), numel);
+    VisitDataType(trg.type(), func);
+
+    for (size_t i = 1; i < local_scopes_.size(); ++i) {
+      auto &scope =
+          *local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
+      auto &p = places_[i];
+      auto *var = scope.FindVar(grad_name);
+      auto *dev_ctx = dev_ctxes_.at(p);
+      size_t size = numel * SizeOfType(trg.type());
+      RunAndRecordEvent(p, [&trg, var, dev_ctx, p, size] {
+        auto dst_ptr = var->GetMutable<framework::LoDTensor>()->data<void>();
+        platform::CPUPlace cpu_place;
+        memory::Copy(cpu_place, dst_ptr, cpu_place, trg.data<void>(), size);
+      });
+    }
+  }
+}
+
+void FusedAllReduceOpHandle::GetGradLoDTensor(
+    const size_t &scope_idx, const std::vector<VarHandle *> &in_var_handles,
+    const std::vector<VarHandle *> &out_var_handles,
+    std::vector<std::pair<std::string, const LoDTensor *>> *grad_tensor) const {
+  auto *local_scope =
+      local_scopes_.at(scope_idx)->FindVar(kLocalExecScopeName)->Get<Scope *>();
+  size_t place_num = places_.size();
+
+  for (size_t j = 0; j < in_var_handles.size(); j += place_num) {
+    auto var_name = in_var_handles[j]->name();
+    PADDLE_ENFORCE_EQ(var_name, out_var_handles[j]->name());
+    auto &lod_tensor = local_scope->FindVar(var_name)->Get<LoDTensor>();
+    PADDLE_ENFORCE_EQ(lod_tensor.place(), places_.at(scope_idx));
+    grad_tensor->emplace_back(std::make_pair(var_name, &lod_tensor));
+  }
+}
+
+void FusedAllReduceOpHandle::GetDTypeAndNumel(
+    const std::vector<std::pair<std::string, const LoDTensor *>> &grad_tensor,
+    proto::VarType::Type *dtype, int64_t *numel) const {
+  *numel = 0;
+  for (size_t i = 0; i < grad_tensor.size(); ++i) {
+    // Get element number
+    int64_t len = grad_tensor.at(i).second->numel();
+    PADDLE_ENFORCE_GT(len, 0);
+    *numel += len;
+
+    // Get dtype
+    auto ele_type = grad_tensor.at(i).second->type();
+    if (i == 0) {
+      *dtype = ele_type;
+    }
+    PADDLE_ENFORCE_EQ(ele_type, *dtype);
+  }
+}
+
+std::string FusedAllReduceOpHandle::Name() const { return "fused_all_reduce"; }
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/data_balance_op_handle.h → paddle/fluid/framework/details/fused_all_reduce_op_handle.h

-// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -15,6 +15,7 @@
 #pragma once
 
 #include <string>
+#include <utility>
 #include <vector>
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
@@ -27,31 +28,47 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-struct DataBalanceOpHandle : public OpHandleBase {
- public:
+struct FusedAllReduceOpHandle : public OpHandleBase {
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-  DataBalanceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
-                      const std::vector<platform::Place> &places,
-                      const platform::NCCLContextMap *ctxs);
+  FusedAllReduceOpHandle(ir::Node *node,
+                         const std::vector<Scope *> &local_scopes,
+                         const std::vector<platform::Place> &places,
+                         const size_t num_of_all_reduce,
+                         const platform::NCCLContextMap *ctxs);
 #else
-  DataBalanceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
-                      const std::vector<platform::Place> &places);
+  FusedAllReduceOpHandle(ir::Node *node,
+                         const std::vector<Scope *> &local_scopes,
+                         const std::vector<platform::Place> &places,
+                         const size_t num_of_all_reduce);
 #endif
-
   std::string Name() const override;
 
-  bool IsMultiDeviceTransfer() override { return false; };
+  // Delay and buffer nccl_all_reduce together can significantly increase
+  // performance. Disable this feature by returning false.
+  bool IsMultiDeviceTransfer() override { return true; };
 
  protected:
   void RunImpl() override;
 
  private:
-  // std::vector<(src_dev_id, dst_dev_id, trans_size)>
-  std::vector<std::array<int, 3>> GetBalancePlan(
-      const std::vector<int> &batch_size_per_device);
+  std::vector<Scope *> local_scopes_;
+  std::vector<platform::Place> places_;
+  size_t num_of_all_reduce_;
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
+  const platform::NCCLContextMap *nccl_ctxs_;
+#endif
+
+  // Check the dtype of the input
+  void GetDTypeAndNumel(
+      const std::vector<std::pair<std::string, const LoDTensor *>> &g_tensor,
+      proto::VarType::Type *dtype, int64_t *total_num) const;
 
-  const std::vector<Scope *> local_scopes_;
-  const std::vector<platform::Place> places_;
+  // Get gradient's name and LoDTensor
+  void GetGradLoDTensor(const size_t &scope_idx,
+                        const std::vector<VarHandle *> &in_var_handles,
+                        const std::vector<VarHandle *> &out_var_handles,
+                        std::vector<std::pair<std::string, const LoDTensor *>>
+                            *grad_tensor) const;
 };
 
 }  // namespace details

paddle/fluid/framework/details/graph_test_base.h

@@ -68,11 +68,11 @@ class SplitOpMaker : public OpProtoAndCheckerMaker {
 
 class DummyVarTypeInference : public VarTypeInference {
  public:
-  void operator()(const OpDesc& op_desc, BlockDesc* block) const override {
-    auto& inputs = op_desc.Input("X");
-    auto type = block->Var(inputs.front())->GetType();
-    auto out_var_name = op_desc.Output("Out").front();
-    block->Var(out_var_name)->SetType(type);
+  void operator()(framework::InferVarTypeContext* ctx) const override {
+    auto& inputs = ctx->Input("X");
+    auto type = ctx->GetType(inputs.front());
+    auto out_var_name = ctx->Output("Out").front();
+    ctx->SetType(out_var_name, type);
   }
 };
 

paddle/fluid/framework/details/inplace_op_pass.cc

@@ -16,6 +16,7 @@
 #include <algorithm>
 #include <deque>
 #include <iterator>
+#include <memory>
 #include <stack>
 #include <string>
 #include <unordered_map>
@@ -263,6 +264,10 @@ void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
 void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
                                           ir::Graph* graph) const {
   VLOG(4) << "Try to inplace op " << op->Name();
+  // FIXME(liuwei1031): Graph is not aware of the existence of BlockDescs and
+  // ProgramDescs.
+  // The operations related to BlockDesc or ProgramDesc should perform on Graph
+  // or Node directly!
   PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
                  "op_desc is nullptr");
   // some pre-requirments need to meet if the op want to inplaced.

paddle/fluid/framework/details/memory_optimize_helper.cc

@@ -20,6 +20,9 @@
 #include <numeric>
 #include <sstream>
 #include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/var_desc.h"
 #include "paddle/fluid/platform/cpu_info.h"
 
@@ -302,7 +305,10 @@ std::string OrderedSet::ToString() const {
 
 bool NodeCanReused(ir::Node* node) {
   // valid the node is a var node
-  if (node == nullptr || !node->IsVar() || node->IsCtrlVar()) return false;
+  // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad
+  if (node == nullptr || !node->IsVar() || node->IsCtrlVar() ||
+      node->Name() == kEmptyVarName)
+    return false;
 
   bool flag = true;
   // op output force generated in cpu, can not be reused.
@@ -331,7 +337,6 @@ bool NodeCanReused(const VarDesc& node) {
   auto type = node.GetType();
   // only these types holds bulk of gpu memory
   if (!(type == proto::VarType::LOD_TENSOR ||
-        type == proto::VarType::SELECTED_ROWS ||
         type == proto::VarType::LOD_TENSOR_ARRAY)) {
     return false;
   }
@@ -348,10 +353,6 @@ bool NodeCanReused(const VarDesc& node) {
   if (shape.empty() || size < MinChunkSize()) {
     return false;
   }
-  // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad
-  std::string name = node.Name();
-  if (!name.empty() && name[0] == '@' && name[name.size() - 1] == '@')
-    return false;
   return true;
 }
 

paddle/fluid/framework/details/memory_optimize_pass.cc

@@ -24,6 +24,7 @@
 #include <sstream>
 #include <string>
 #include <type_traits>
+#include <unordered_set>
 #include <vector>
 #include "gflags/gflags.h"
 #include "paddle/fluid/framework/data_type.h"
@@ -191,6 +192,10 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
           // immediately to make the subblock variable reuse strategy take
           // effect. Because it is a single op in graph. No need to
           // update the ir nodes.
+          // FIXME(liuwei1031): Graph is not aware of the existence of
+          // BlockDescs and ProgramDescs.
+          // The operations related to BlockDesc or ProgramDesc should perform
+          // on Graph or Node directly!
           sub_op_desc->Rename(var->Name(), cache->Name());
           if (sub_op_desc->Block() != nullptr &&
               sub_op_desc->Block()->HasVar(var->Name())) {

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -11,18 +11,19 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
+#include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
 #include <algorithm>
 #include <fstream>
+#include <memory>
 #include <string>
+#include <unordered_map>
+#include <unordered_set>
 #include <utility>
 #include <vector>
-
 #include "paddle/fluid/framework/details/all_reduce_op_handle.h"
 #include "paddle/fluid/framework/details/broadcast_op_handle.h"
 #include "paddle/fluid/framework/details/computation_op_handle.h"
-#include "paddle/fluid/framework/details/data_balance_op_handle.h"
 #include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
-#include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
 #include "paddle/fluid/framework/details/reduce_op_handle.h"
 #include "paddle/fluid/framework/details/rpc_op_handle.h"
 #include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
@@ -134,21 +135,26 @@ void AddOutputToLeafOps(ir::Graph *graph) {
 }
 }  // namespace
 
+void MultiDevSSAGraphBuilderBase::CheckGraph(const ir::Graph &graph) const {}
+
 void MultiDevSSAGraphBuilderBase::Init() const {
   all_vars_.clear();
 
   loss_var_name_ = Get<const std::string>(kLossVarName);
+  VLOG(10) << "Init MultiDevSSAGraphBuilder, loss name: " << loss_var_name_;
   places_ = Get<const std::vector<platform::Place>>(kPlaces);
   local_scopes_ = Get<const std::vector<Scope *>>(kLocalScopes);
   strategy_ = Get<const BuildStrategy>(kStrategy);
 #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
-  nccl_ctxs_ = &Get<platform::NCCLContextMap>("nccl_ctxs");
+  nccl_ctxs_ = &Get<platform::NCCLContextMap>(kNCCLCtxs);
 #endif
+  PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
 }
 
 std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
   Init();
+  CheckGraph(*graph);
   std::vector<ir::Node *> sorted_ops = SortOperations(*graph);
 
   auto nodes = graph->ReleaseNodes();
@@ -166,7 +172,6 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
   result.Set(kGraphOps, new GraphOps);
 
   bool is_forwarding = true;
-  bool insert_collection_ops = NeedCollectiveOps();
 
   for (ir::Node *node : sorted_ops) {
     if (DealWithSpecialOp(&result, node)) {
@@ -185,8 +190,8 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
         CreateComputationalOps(&result, node, places_.size());
       }
 
-      // Insert collection ops
-      if (!is_forwarding && insert_collection_ops) {
+      // Insert collective ops if nranks > 1
+      if (!is_forwarding && Get<size_t>(kNRanks) > 1) {
         try {
           bool is_bk_op =
               static_cast<bool>(boost::get<int>(node->Op()->GetAttr(
@@ -200,13 +205,13 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
               boost::get<std::vector<std::string>>(node->Op()->GetNullableAttr(
                   OpProtoAndCheckerMaker::OpRoleVarAttrName()));
           PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
-
           for (size_t i = 0; i < backward_vars.size(); i += 2) {
             auto &p_name = backward_vars[i];
             auto &g_name = backward_vars[i + 1];
             VLOG(10) << "Bcast " << g_name << " for parameter " << p_name;
-
-            InsertCollectiveOp(&result, p_name, g_name);
+            if (NeedCollectiveForGrad(g_name, sorted_ops)) {
+              InsertCollectiveOp(&result, p_name, g_name);
+            }
           }
         } catch (boost::bad_get e) {
         }
@@ -226,6 +231,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilderBase::ApplyImpl(
    * Only variables should be the leaves of graph.
    */
   AddOutputToLeafOps(&result);
+
   result.Erase(kGraphOps);
   return graph;
 }
@@ -258,6 +264,11 @@ void MultiDevSSAGraphBuilderBase::InsertScaleLossGradOp(
   }
 }
 
+bool MultiDevSSAGraphBuilderBase::DealWithSpecialOp(ir::Graph *result,
+                                                    ir::Node *node) const {
+  return false;
+}
+
 std::vector<ir::Node *> MultiDevSSAGraphBuilderBase::SortOperations(
     const ir::Graph &graph) const {
   return ir::TopologySortOperations(graph);
@@ -271,8 +282,20 @@ bool MultiDevSSAGraphBuilderBase::UseGPU() const {
   return use_gpu;
 }
 
-bool MultiDevSSAGraphBuilderBase::NeedCollectiveOps() const {
-  return Get<size_t>(kNRanks) > 1;
+bool MultiDevSSAGraphBuilderBase::NeedCollectiveForGrad(
+    const std::string &grad_name, std::vector<ir::Node *> ops) const {
+  // if we have allreduce_op for current gradient variable in the graph,
+  // then we don't need to add allreduce_op_handle for this gradient
+  // NOTE: This is for the case that all gradients should add collective ops
+  for (auto *node : ops) {
+    if (node->Op()->Type() != "allreduce") continue;
+    for (auto in_name : node->Op()->InputArgumentNames()) {
+      if (in_name == grad_name) {
+        return false;
+      }
+    }
+  }
+  return true;
 }
 
 void MultiDevSSAGraphBuilderBase::CreateOpHandleIOs(ir::Graph *result,
@@ -496,20 +519,17 @@ VarHandle *MultiDevSSAGraphBuilderBase::CreateReduceOp(ir::Graph *result,
 }
 
 bool MultiDevSSAGraphBuilderBase::IsScaleLossOp(ir::Node *node) const {
-  return boost::get<int>(
+  return !loss_var_name_.empty() && node->Op() &&
+         boost::get<int>(
              node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) ==
              (static_cast<int>(OpRole::kBackward) |
-              static_cast<int>(OpRole::kLoss)) &&
-         !loss_var_name_.empty();  // If loss_var is empty. This is test mode
+              static_cast<int>(OpRole::kLoss));
 }
 
 bool MultiDevSSAGraphBuilderBase::IsSparseGradient(
     const std::string &og) const {
   PADDLE_ENFORCE(all_vars_.count(og) != 0);
-  if (all_vars_.at(og)->GetType() == proto::VarType::SELECTED_ROWS) {
-    return true;
-  }
-  return false;
+  return all_vars_.at(og)->GetType() == proto::VarType::SELECTED_ROWS;
 }
 
 void AllReduceSSAGraphBuilder::InsertCollectiveOp(
@@ -995,7 +1015,7 @@ static int MultiDevSSAGraphBuilderRegister(const std::string &builder_mode) {
 REGISTER_MULTI_DEVICES_PASS(reduce_mode_multi_devices_pass,
                             paddle::framework::details::ReduceSSAGraphBuilder);
 REGISTER_MULTI_DEVICES_PASS(
-    allreduce_mode_multi_devices_pass,
+    all_reduce_mode_multi_devices_pass,
     paddle::framework::details::AllReduceSSAGraphBuilder);
 REGISTER_MULTI_DEVICES_PASS(dist_multi_devices_pass,
                             paddle::framework::details::DistSSAGraphBuilder);

paddle/fluid/framework/details/multi_devices_graph_pass.h

@@ -14,7 +14,10 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
+#include <unordered_map>
+#include <unordered_set>
 #include <utility>
 #include <vector>
 
@@ -31,12 +34,6 @@ namespace framework {
 class Scope;
 namespace details {
 
-constexpr char kLossVarName[] = "loss_var_name";
-constexpr char kPlaces[] = "places";
-constexpr char kLocalScopes[] = "local_scopes";
-constexpr char kStrategy[] = "strategy";
-constexpr char kNRanks[] = "nranks";
-
 class MultiDevSSAGraphBuilderBase : public ir::Pass {
  protected:
   std::unique_ptr<ir::Graph> ApplyImpl(
@@ -44,18 +41,21 @@ class MultiDevSSAGraphBuilderBase : public ir::Pass {
 
   virtual void Init() const;
 
+  virtual void CheckGraph(const ir::Graph &graph) const;
+
   virtual std::vector<ir::Node *> SortOperations(const ir::Graph &graph) const;
 
   virtual void InsertCollectiveOp(ir::Graph *result, const std::string &p_name,
                                   const std::string &g_name) const = 0;
 
-  virtual bool DealWithSpecialOp(ir::Graph *result, ir::Node *node) const = 0;
+  virtual bool DealWithSpecialOp(ir::Graph *result, ir::Node *node) const;
 
   virtual void InsertPostprocessOps(ir::Graph *result) const = 0;
 
   bool UseGPU() const;
 
-  bool NeedCollectiveOps() const;
+  bool NeedCollectiveForGrad(const std::string &grad_name,
+                             std::vector<ir::Node *> ops) const;
 
   bool IsScaleLossOp(ir::Node *node) const;
 
@@ -109,10 +109,6 @@ class AllReduceSSAGraphBuilder : public MultiDevSSAGraphBuilderBase {
   virtual void InsertCollectiveOp(ir::Graph *result, const std::string &p_name,
                                   const std::string &g_name) const;
 
-  virtual bool DealWithSpecialOp(ir::Graph *result, ir::Node *node) const {
-    return false;
-  }
-
   virtual void InsertPostprocessOps(ir::Graph *result) const {}
 };
 

paddle/fluid/framework/details/multi_devices_helper.h

@@ -16,6 +16,9 @@
 
 #include <memory>
 #include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
 #include <vector>
 
 #include "paddle/fluid/framework/details/op_handle_base.h"
@@ -44,6 +47,26 @@ const char kGraphVars[] = "vars";
 typedef std::unordered_set<VarHandleBase *> GraphDepVars;
 const char kGraphDepVars[] = "dep_vars";
 
+constexpr char kNCCLCtxs[] = "nccl_ctxs";
+
+constexpr char kLossVarName[] = "loss_var_name";
+constexpr char kPlaces[] = "places";
+constexpr char kLocalScopes[] = "local_scopes";
+constexpr char kStrategy[] = "strategy";
+constexpr char kNRanks[] = "nranks";
+
+typedef std::unordered_set<std::string> FusedVars;
+constexpr char kFusedVars[] = "fused_vars";
+
+typedef std::vector<std::pair<std::string, std::string>> ParamsAndGrads;
+constexpr char kParamsAndGrads[] = "params_grads";
+
+typedef std::vector<std::vector<std::pair<std::string, std::string>>>
+    GroupGradsAndParams;
+constexpr char kGroupGradsAndParams[] = "group_grads_params";
+
+constexpr char kFusedVarNamePrefix[] = "@FUSEDVAR@";
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/op_handle_base.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include <map>
+#include <unordered_set>
 
 namespace paddle {
 namespace framework {
@@ -41,15 +42,42 @@ OpHandleBase::~OpHandleBase() {
 
 void OpHandleBase::Run(bool use_cuda) {
 #ifdef PADDLE_WITH_CUDA
-  if (events_.empty() && use_cuda) {
+  if (events_.empty() && use_cuda && dev_ctxes_.size() > 0) {
     for (auto &p : dev_ctxes_) {
       int dev_id = boost::get<platform::CUDAPlace>(p.first).device;
       PADDLE_ENFORCE(cudaSetDevice(dev_id));
       PADDLE_ENFORCE(
           cudaEventCreateWithFlags(&events_[dev_id], cudaEventDisableTiming));
     }
+    if (IsMultiDeviceTransfer() && dev_ctxes_.size() > 0) {
+      for (auto &out_var : outputs_) {
+        auto *out_var_handle = dynamic_cast<VarHandle *>(out_var);
+        if (out_var_handle) {
+          int dev_id =
+              boost::get<platform::CUDAPlace>(out_var_handle->place()).device;
+          out_var_handle->SetGenerateEvent(events_[dev_id]);
+        }
+      }
+    } else {
+      PADDLE_ENFORCE_EQ(dev_ctxes_.size(), 1UL,
+                        "%s should have only one dev_ctx.", Name());
+      auto &place = dev_ctxes_.begin()->first;
+      int dev_id = boost::get<platform::CUDAPlace>(place).device;
+      for (auto &out_var : outputs_) {
+        auto *out_var_handle = dynamic_cast<VarHandle *>(out_var);
+        if (out_var_handle) {
+          PADDLE_ENFORCE(
+              platform::is_same_place(place, out_var_handle->place()),
+              "The place of input(%s) is not consistent with the "
+              "place of current op(%s).",
+              out_var_handle->Name(), Name());
+          out_var_handle->SetGenerateEvent(events_[dev_id]);
+        }
+      }
+    }
   }
 #else
+
   PADDLE_ENFORCE(!use_cuda);
 #endif
 
@@ -93,17 +121,48 @@ void OpHandleBase::AddOutput(VarHandleBase *out) {
 void OpHandleBase::WaitInputVarGenerated() {
   for (auto in_var : inputs_) {
     if (NeedWait(in_var)) {
-      for (auto &pair : dev_ctxes_) {
-        in_var->GeneratedOp()->RecordWaitEventOnCtx(pair.second);
+      // Dummy Variable is used to represent dependencies between operators, so
+      // there doesn't add event for it.
+      auto *in_var_handle = dynamic_cast<VarHandle *>(in_var);
+      if (in_var_handle) {
+        auto &place = in_var_handle->place();
+        if (platform::is_gpu_place(place)) {
+#ifdef PADDLE_WITH_CUDA
+          auto stream =
+              static_cast<platform::CUDADeviceContext *>(dev_ctxes_.at(place))
+                  ->stream();
+          PADDLE_ENFORCE(
+              cudaStreamWaitEvent(stream, in_var_handle->GetEvent(), 0));
+#else
+          PADDLE_THROW("Doesn't compile the GPU.");
+#endif
+        }
+        // There are nothing to do when the place is CPUPlace.
       }
     }
   }
 }
 
 void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) {
-  for (auto *in : inputs_) {
-    if (NeedWait(in)) {
-      in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(place));
+  for (auto in_var : inputs_) {
+    if (NeedWait(in_var)) {
+      // Dummy Variable is used to represent dependencies between operators, so
+      // there doesn't add event for it.
+      auto *in_var_handle = dynamic_cast<VarHandle *>(in_var);
+      if (in_var_handle) {
+        if (platform::is_gpu_place(in_var_handle->place())) {
+#ifdef PADDLE_WITH_CUDA
+          auto stream = static_cast<platform::CUDADeviceContext *>(
+                            dev_ctxes_.at(in_var_handle->place()))
+                            ->stream();
+          PADDLE_ENFORCE(
+              cudaStreamWaitEvent(stream, in_var_handle->GetEvent(), 0));
+#else
+          PADDLE_THROW("Doesn't compile the GPU.");
+#endif
+        }
+        // There are nothing to do when the place is CPUPlace.
+      }
     }
   }
 }

paddle/fluid/framework/details/op_registry.h

@@ -16,6 +16,8 @@ limitations under the License. */
 
 #include <string>
 #include <tuple>
+#include <unordered_map>
+#include <unordered_set>
 #include <vector>
 #include "paddle/fluid/framework/grad_op_desc_maker.h"
 #include "paddle/fluid/framework/inplace_op_inference.h"
@@ -127,9 +129,9 @@ struct OpInfoFiller<T, kGradOpDescMaker> {
 template <typename T>
 struct OpInfoFiller<T, kVarTypeInference> {
   void operator()(const char* op_type, OpInfo* info) const {
-    info->infer_var_type_ = [](const OpDesc& fwd_op, BlockDesc* block) {
+    info->infer_var_type_ = [](InferVarTypeContext* context) {
       T inference;
-      inference(fwd_op, block);
+      inference(context);
     };
   }
 };

paddle/fluid/framework/details/parallel_ssa_graph_executor.cc

@@ -13,6 +13,8 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
+#include <memory>
+#include <utility>
 #include "paddle/fluid/framework/ir/graph_helper.h"
 
 namespace paddle {
@@ -29,6 +31,11 @@ ParallelSSAGraphExecutor::SeparateMultiDevicesGraph(ir::Graph *graph) {
     auto &g = graphs.back();
     g->Set(kGraphVars, new GraphVars(1UL));
     g->Set(kGraphDepVars, new GraphDepVars);
+    auto &stale_ops =
+        graph->Get<const std::vector<OpDesc *>>(details::kStaleProgramOpDescs);
+    g->Erase(details::kStaleProgramOpDescs);
+    g->Set<const std::vector<OpDesc *>>(details::kStaleProgramOpDescs,
+                                        new std::vector<OpDesc *>(stale_ops));
   }
   auto op_handles = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
 

paddle/fluid/framework/details/reduce_and_gather.h

@@ -53,6 +53,31 @@ struct ReduceLoDTensor {
   }
 };
 
+struct ReduceBufferData {
+  const std::vector<const void *> &src_data_;
+  void *dst_data_;
+  int64_t numel_;
+
+  ReduceBufferData(const std::vector<const void *> &src, void *dst,
+                   int64_t numel)
+      : src_data_(src), dst_data_(dst), numel_(numel) {}
+
+  template <typename T>
+  void apply() const {
+    T *dst_data = reinterpret_cast<T *>(dst_data_);
+    for (size_t i = 0; i < src_data_.size(); ++i) {
+      auto srd_data = reinterpret_cast<const T *>(src_data_[i]);
+      VLOG(10) << "dst: " << dst_data_ << ", " << srd_data;
+      if (srd_data == dst_data_) {
+        continue;
+      }
+
+      std::transform(srd_data, srd_data + numel_, dst_data, dst_data,
+                     [](T a, T b) -> T { return a + b; });
+    }
+  }
+};
+
 inline void GatherLocalSelectedRows(
     const std::vector<const SelectedRows *> &src_selecte_rows_,
     const std::vector<platform::Place> &in_places,

paddle/fluid/framework/details/reference_count_pass.cc

@@ -12,9 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <memory>
 #include <queue>
 #include <string>
 #include <type_traits>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
 #include <vector>
 
 #include "paddle/fluid/framework/details/computation_op_handle.h"
@@ -189,15 +193,6 @@ ExtractComputationOpFromLastLivedVar(VarHandle *var, size_t scope_idx,
   return shrink_func(computation_op);
 }
 
-static VarDesc *TryGetLatestVarDesc(const std::vector<VarHandle *> &vars) {
-  VarDesc *var_desc = nullptr;
-  std::find_if(vars.rbegin(), vars.rend(), [&](VarHandle *var_handle) -> bool {
-    var_desc = var_handle->Node()->Var();
-    return var_desc != nullptr;
-  });
-  return var_desc;
-}
-
 std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
   auto &ref_cnts = Get<std::vector<ReferenceCountMap>>(kGlobalReferenceCount);

paddle/fluid/framework/details/reference_count_pass_helper.cc

@@ -13,9 +13,22 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/reference_count_pass_helper.h"
+#include "paddle/fluid/framework/details/var_handle.h"
+#include "paddle/fluid/framework/var_desc.h"
 
 namespace paddle {
 namespace framework {
-namespace details {}  // namespace details
+namespace details {
+
+VarDesc *TryGetLatestVarDesc(const std::vector<VarHandle *> &vars) {
+  VarDesc *var_desc = nullptr;
+  std::find_if(vars.rbegin(), vars.rend(), [&](VarHandle *var_handle) -> bool {
+    var_desc = var_handle->Node()->Var();
+    return var_desc != nullptr;
+  });
+  return var_desc;
+}
+
+}  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/reference_count_pass_helper.h

@@ -16,6 +16,7 @@
 
 #include <atomic>
 #include <map>
+#include <memory>
 #include <string>
 #include <unordered_map>
 #include <unordered_set>
@@ -25,6 +26,10 @@
 
 namespace paddle {
 namespace framework {
+
+class VarDesc;
+class VarHandle;
+
 namespace details {
 
 class ComputationOpHandle;
@@ -43,9 +48,11 @@ const char kGarbageCollector[] = "garbage_collector";
 const char kAllPlaces[] = "all_places";
 
 using LastLiveOpsOfVars =
-    std::unordered_map<std::string, std::unordered_set<ComputationOpHandle*>>;
+    std::unordered_map<std::string, std::unordered_set<ComputationOpHandle *>>;
 const char kLastLiveOpsOfVars[] = "last_live_ops_of_var";
 
+VarDesc *TryGetLatestVarDesc(const std::vector<VarHandle *> &vars);
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -14,7 +14,6 @@
 
 #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
 
-#include "paddle/fluid/framework/details/multi_devices_helper.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -27,62 +26,49 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
     : graph_(graph),
       pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_)
                                        : nullptr),
+      prepare_pool_(1),
       local_scopes_(local_scopes),
       places_(places),
       fetch_ctxs_(places),
-      running_ops_(0),
-      strategy_(strategy) {}
+      strategy_(strategy) {
+  PrepareOpDeps();
+  CopyOpDeps();
+}
 
 FeedFetchList ThreadedSSAGraphExecutor::Run(
     const std::vector<std::string> &fetch_tensors) {
   std::unique_ptr<platform::RecordEvent> event(
       new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare"));
-  std::unordered_map<OpHandleBase *, size_t> pending_ops;
-  std::unordered_set<VarHandleBase *> pending_vars;
-  auto ready_vars = std::make_shared<BlockingQueue<VarHandleBase *>>();
-  std::unordered_set<OpHandleBase *> ready_ops;
+  std::unique_ptr<OpDependentData> op_deps = op_deps_futures_.get();
+  CopyOpDeps();
+  VLOG(10) << "ThreadedSSAGraphExecutor::Run";
+  std::shared_ptr<BlockingQueue<VarHandleBase *>> ready_vars(
+      new BlockingQueue<VarHandleBase *>);
+  auto &pending_ops = op_deps->pending_ops_;
+  auto &pending_vars = op_deps->pending_vars_;
+  auto &ready_ops = op_deps->ready_ops_;
+
   // For ops (e.g. nccl_all_reduce) that need to coordinate multiple
   // streams from multiple GPUs, it's faster to buffer them and schedule
   // together since we currently cannot overlap computation and memcpy streams.
   // Should revisit it if overlapping is available.
   std::unordered_set<OpHandleBase *> delayed_ops;
 
-  // Transform SSAGraph to pending_ops & pending_vars
-  for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
-    for (auto &name_pair : var_map) {
-      for (auto &version_pair : name_pair.second) {
-        InsertPendingVar(&pending_vars, ready_vars.get(), version_pair);
-      }
-    }
-  }
-  for (auto &var : graph_->Get<details::GraphDepVars>(details::kGraphDepVars)) {
-    InsertPendingVar(&pending_vars, ready_vars.get(), var);
-  }
-
-  for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(*graph_)) {
-    if (op->Inputs().empty()) {  // Special case, Op has no input.
-      ready_ops.insert(op);
-    } else {
-      InsertPendingOp(&pending_ops, op);
-    }
-  }
-
   // Step 2. Insert FetchOps
   std::vector<FetchOpHandle *> fetch_ops;
   std::unordered_set<VarHandleBase *> fetch_dependencies;
   FeedFetchList fetch_data(fetch_tensors.size());
 
-  InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops,
-                 &pending_vars, ready_vars.get(), &fetch_data);
+  InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &ready_ops,
+                 &pending_ops, &pending_vars, &fetch_data);
 
   auto run_all_ops = [&](std::unordered_set<OpHandleBase *> &set) {
     for (auto *op : set) {
-      running_ops_++;
       RunOp(ready_vars, op);
     }
     set.clear();
   };
-
+  auto run_all_op = [&](OpHandleBase *op) { RunOp(ready_vars, op); };
   // Clean run context
   run_op_futures_.clear();
   exception_holder_.Clear();
@@ -91,19 +77,11 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
   while (!pending_vars.empty()) {
     // 1. Run All Ready ops
     // Keep loop until all vars are ready.
-    //
-    // NOTE: DelayedOps have a lower priority. It will be scheduled after all
-    // ready_ops have been performed.
-    if (ready_ops.empty() && strategy_.allow_op_delay_ && running_ops_ == 0) {
-      run_all_ops(delayed_ops);
-    } else {
-      run_all_ops(ready_ops);
-    }
+    run_all_ops(ready_ops);
 
     // 2. Find ready variable
     bool timeout;
     auto cur_ready_vars = ready_vars->PopAll(1, &timeout);
-
     if (timeout) {
       if (exception_holder_.IsCaught()) {
         for (auto &run_op_future : run_op_futures_) {
@@ -115,6 +93,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
         continue;
       }
     }
+
     // 3. Remove the dependency of ready_var.
     // Find the ready_ops after the ready_var.
     for (auto ready_var : cur_ready_vars) {
@@ -123,11 +102,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
         auto &deps = pending_ops[op];
         --deps;
         if (deps == 0) {
-          if (op->IsMultiDeviceTransfer() && strategy_.allow_op_delay_) {
-            delayed_ops.insert(op);
-          } else {
-            ready_ops.insert(op);
-          }
+          run_all_op(op);
         }
       }
     }
@@ -143,16 +118,17 @@ void ThreadedSSAGraphExecutor::InsertFetchOps(
     const std::vector<std::string> &fetch_tensors,
     std::vector<FetchOpHandle *> *fetch_ops,
     std::unordered_set<VarHandleBase *> *fetch_dependencies,
+    std::unordered_set<OpHandleBase *> *ready_ops,
     std::unordered_map<OpHandleBase *, size_t> *pending_ops,
     std::unordered_set<VarHandleBase *> *pending_vars,
-    BlockingQueue<VarHandleBase *> *ready_vars, FeedFetchList *fetch_data) {
+    FeedFetchList *fetch_data) {
   std::unordered_map<std::string, std::vector<VarHandleBase *>> fetched_vars;
-
+  std::unordered_set<VarHandleBase *> local_ready_vars;
   for (auto &fetch_var_name : fetch_tensors) {
     for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
       auto it = var_map.find(fetch_var_name);
       if (it != var_map.end()) {
-        fetched_vars[fetch_var_name].push_back(*it->second.rbegin());
+        fetched_vars[fetch_var_name].emplace_back(*it->second.rbegin());
       }
     }
   }
@@ -161,8 +137,9 @@ void ThreadedSSAGraphExecutor::InsertFetchOps(
     auto &var_name = fetch_tensors[i];
     auto fetched_var_it = fetched_vars.find(var_name);
     PADDLE_ENFORCE(fetched_var_it != fetched_vars.end(),
-                   "Cannot find fetched variable.(Perhaps the main_program "
-                   "is not set to ParallelExecutor)");
+                   "Cannot find fetched variable(%s).(Perhaps the main_program "
+                   "is not set to ParallelExecutor)",
+                   var_name);
 
     auto &vars = fetched_var_it->second;
 
@@ -184,9 +161,23 @@ void ThreadedSSAGraphExecutor::InsertFetchOps(
     auto *fetch_dummy = new DummyVarHandle(fetch_var);
     op->AddOutput(fetch_dummy);
     fetch_dependencies->emplace(fetch_dummy);
-    this->InsertPendingVar(pending_vars, ready_vars, fetch_dummy);
-    this->InsertPendingOp(pending_ops, op);
+
+    this->InsertPendingVar(pending_vars, &local_ready_vars, fetch_dummy);
+
+    size_t wait_input_num = 0;
+    std::unordered_set<VarHandleBase *> input_set(vars.begin(), vars.end());
+    for (auto *var : input_set) {
+      if (pending_vars->count(var)) {
+        ++wait_input_num;
+      }
+    }
+    if (wait_input_num) {
+      pending_ops->insert({op, wait_input_num});
+    } else {
+      ready_ops->insert(static_cast<OpHandleBase *>(op));
+    }
   }
+  PADDLE_ENFORCE_EQ(local_ready_vars.size(), 0);
 }
 
 void ThreadedSSAGraphExecutor::InsertPendingOp(
@@ -197,11 +188,63 @@ void ThreadedSSAGraphExecutor::InsertPendingOp(
 
 void ThreadedSSAGraphExecutor::InsertPendingVar(
     std::unordered_set<VarHandleBase *> *pending_vars,
-    BlockingQueue<VarHandleBase *> *ready_vars, VarHandleBase *var) const {
+    std::unordered_set<VarHandleBase *> *ready_vars, VarHandleBase *var) const {
   pending_vars->insert(var);
   if (var->GeneratedOp() == nullptr) {
-    ready_vars->Push(var);
+    ready_vars->insert(var);
+  }
+}
+
+void ThreadedSSAGraphExecutor::PrepareOpDeps() {
+  op_deps_.reset(new OpDependentData());
+  std::unordered_map<OpHandleBase *, size_t> &pending_ops =
+      op_deps_->pending_ops_;
+  std::unordered_set<VarHandleBase *> &pending_vars = op_deps_->pending_vars_;
+  std::unordered_set<OpHandleBase *> &ready_ops = op_deps_->ready_ops_;
+  std::unordered_set<VarHandleBase *> ready_vars;
+
+  // Transform SSAGraph to pending_ops & pending_vars
+  for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
+    for (auto &name_pair : var_map) {
+      for (auto &version_pair : name_pair.second) {
+        InsertPendingVar(&pending_vars, &ready_vars, version_pair);
+      }
+    }
+  }
+  for (auto &var : graph_->Get<details::GraphDepVars>(details::kGraphDepVars)) {
+    InsertPendingVar(&pending_vars, &ready_vars, var);
+  }
+
+  for (auto &op : ir::FilterByNodeWrapper<OpHandleBase>(*graph_)) {
+    if (op->Inputs().empty()) {  // Special case, Op has no input.
+      ready_ops.insert(op);
+    } else {
+      InsertPendingOp(&pending_ops, op);
+    }
   }
+  for (auto ready_var : ready_vars) {
+    pending_vars.erase(ready_var);
+    for (auto *op : ready_var->PendingOps()) {
+      auto &deps = pending_ops[op];
+      --deps;
+      if (deps == 0) {
+        ready_ops.insert(op);
+      }
+    }
+  }
+}
+
+void ThreadedSSAGraphExecutor::CopyOpDeps() {
+  op_deps_futures_ = prepare_pool_.enqueue([&] {
+    auto *op_deps = new OpDependentData();
+    op_deps->pending_ops_.insert(op_deps_->pending_ops_.begin(),
+                                 op_deps_->pending_ops_.end());
+    op_deps->pending_vars_.insert(op_deps_->pending_vars_.begin(),
+                                  op_deps_->pending_vars_.end());
+    op_deps->ready_ops_.insert(op_deps_->ready_ops_.begin(),
+                               op_deps_->ready_ops_.end());
+    return std::unique_ptr<OpDependentData>(op_deps);
+  });
 }
 
 void ThreadedSSAGraphExecutor::RunOp(
@@ -216,7 +259,6 @@ void ThreadedSSAGraphExecutor::RunOp(
         op->Run(strategy_.use_cuda_);
       }
       VLOG(10) << op << " " << op->Name() << " Done ";
-      running_ops_--;
       ready_var_q->Extend(op->Outputs());
       VLOG(10) << op << " " << op->Name() << " Signal posted";
     } catch (...) {

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -15,18 +15,20 @@
 #pragma once
 
 #include <deque>
+#include <functional>
 #include <list>
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <unordered_set>
 #include <utility>
 #include <vector>
-
-#include <functional>
 #include "ThreadPool.h"  // ThreadPool in thrird party
 #include "paddle/fluid/framework/blocking_queue.h"
 #include "paddle/fluid/framework/details/exception_holder.h"
 #include "paddle/fluid/framework/details/execution_strategy.h"
 #include "paddle/fluid/framework/details/fetch_op_handle.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
 #include "paddle/fluid/framework/details/ssa_graph_executor.h"
 #include "paddle/fluid/framework/ir/graph.h"
 
@@ -36,6 +38,12 @@ class Scope;
 
 namespace details {
 
+struct OpDependentData {
+  std::unordered_map<OpHandleBase *, size_t> pending_ops_;
+  std::unordered_set<VarHandleBase *> pending_vars_;
+  std::unordered_set<OpHandleBase *> ready_ops_;
+};
+
 class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
  public:
   ThreadedSSAGraphExecutor(const ExecutionStrategy &strategy,
@@ -57,29 +65,35 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
  private:
   ir::Graph *graph_;
   std::unique_ptr<::ThreadPool> pool_;
+  ::ThreadPool prepare_pool_;
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
   platform::DeviceContextPool fetch_ctxs_;
   ExceptionHolder exception_holder_;
-  std::atomic<int> running_ops_;
 
   void InsertPendingOp(std::unordered_map<OpHandleBase *, size_t> *pending_ops,
                        OpHandleBase *op_instance) const;
 
   void InsertPendingVar(std::unordered_set<VarHandleBase *> *pending_vars,
-                        BlockingQueue<VarHandleBase *> *ready_vars,
+                        std::unordered_set<VarHandleBase *> *ready_vars,
                         VarHandleBase *var) const;
 
   void InsertFetchOps(const std::vector<std::string> &fetch_tensors,
                       std::vector<FetchOpHandle *> *fetch_ops,
                       std::unordered_set<VarHandleBase *> *fetch_dependencies,
+                      std::unordered_set<OpHandleBase *> *ready_ops,
                       std::unordered_map<OpHandleBase *, size_t> *pending_ops,
                       std::unordered_set<VarHandleBase *> *pending_vars,
-                      BlockingQueue<VarHandleBase *> *ready_vars,
                       FeedFetchList *fetch_data);
 
+  void PrepareOpDeps();
+  void CopyOpDeps();
+
  private:
+  std::future<std::unique_ptr<OpDependentData>> op_deps_futures_;
+
   ExecutionStrategy strategy_;
+  std::unique_ptr<OpDependentData> op_deps_;
   // use std::list because clear(), push_back, and for_each are O(1)
   std::list<std::future<void>> run_op_futures_;
 };

paddle/fluid/framework/details/var_handle.h

@@ -43,6 +43,7 @@ struct VarHandleBase {
   virtual ~VarHandleBase();
 
   virtual std::string DebugString() const = 0;
+  virtual const std::string& Name() const = 0;
 
   void AddInput(OpHandleBase* in, ir::Node* node) {
     node_->inputs.clear();
@@ -95,8 +96,6 @@ struct VarHandleBase {
 //
 // NOTE: runtime variables have place.
 struct VarHandle : public VarHandleBase {
-  explicit VarHandle(ir::Node* node) : VarHandleBase(node) {}
-
   virtual ~VarHandle();
 
   std::string DebugString() const override;
@@ -109,6 +108,20 @@ struct VarHandle : public VarHandleBase {
         name_(std::move(name)),
         place_(std::move(place)) {}
 
+#ifdef PADDLE_WITH_CUDA
+  bool HasEvent() { return has_event_; }
+
+  const cudaEvent_t& GetEvent() {
+    PADDLE_ENFORCE(HasEvent(), "The event is not set.");
+    return event_;
+  }
+
+  void SetGenerateEvent(const cudaEvent_t& event) {
+    has_event_ = true;
+    event_ = event;
+  }
+#endif
+
   // version field currently is not used, however, just store the version to
   // debug easily.
  private:
@@ -116,6 +129,11 @@ struct VarHandle : public VarHandleBase {
   size_t scope_idx_;
   std::string name_;
   platform::Place place_;
+#ifdef PADDLE_WITH_CUDA
+  // Only when this event is triggered, var is generated.
+  cudaEvent_t event_;
+  bool has_event_{false};
+#endif
 
  public:
   bool IsTheSameVar(const VarHandle& o) const {
@@ -125,6 +143,7 @@ struct VarHandle : public VarHandleBase {
 
   size_t version() const { return version_; }
   size_t scope_idx() const { return scope_idx_; }
+  const std::string& Name() const override { return name_; }
   const std::string& name() const { return name_; }
   const platform::Place& place() const { return place_; }
 };
@@ -136,6 +155,10 @@ struct DummyVarHandle : public VarHandleBase {
   virtual ~DummyVarHandle();
 
   std::string DebugString() const override;
+
+ public:
+  const std::string& Name() const override { return name_; }
+  std::string name_{"DummyVar"};
 };
 
 }  // namespace details

paddle/fluid/framework/details/while_op_eager_deletion_pass.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/details/computation_op_handle.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/operators/controlflow/while_op_helper.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+class WhileOpEagerDeletionPass : public ir::Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override {
+    auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
+
+    // Find all while_op and while_grad_op
+    std::unordered_map<size_t, std::pair<std::vector<OperatorBase *>,
+                                         std::vector<OperatorBase *>>>
+        target_ops;
+    for (auto *op : all_ops) {
+      auto compute_op = dynamic_cast<ComputationOpHandle *>(op);
+      if (compute_op == nullptr) continue;
+
+      if (compute_op->Name() == "while") {
+        target_ops[compute_op->GetScopeIdx()].first.emplace_back(
+            compute_op->GetOp());
+      } else if (compute_op->Name() == "while_grad") {
+        target_ops[compute_op->GetScopeIdx()].second.emplace_back(
+            compute_op->GetOp());
+      }
+    }
+
+    for (auto &ops_pair : target_ops) {
+      auto &while_ops = ops_pair.second.first;
+      auto &while_grad_ops = ops_pair.second.second;
+      operators::PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp(
+          while_ops, while_grad_ops);
+    }
+    return graph;
+  }
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(while_op_eager_deletion_pass,
+              paddle::framework::details::WhileOpEagerDeletionPass);

paddle/fluid/framework/executor.cc

@@ -14,6 +14,10 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/executor.h"
 #include <deque>
+#include <memory>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
 
 #include "paddle/fluid/framework/feed_fetch_method.h"
 #include "paddle/fluid/framework/lod_rank_table.h"
@@ -23,17 +27,18 @@ limitations under the License. */
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/framework/transfer_scope_cache.h"
 #include "paddle/fluid/framework/variable_helper.h"
+#include "paddle/fluid/operators/controlflow/while_op_helper.h"
 #include "paddle/fluid/operators/distributed/distributed.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
 
 #ifdef PADDLE_WITH_NGRAPH
 #include "paddle/fluid/operators/ngraph/ngraph_engine.h"
+DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
 #endif
 
 DECLARE_bool(benchmark);
 DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
-DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
 
 namespace paddle {
 namespace framework {
@@ -75,11 +80,11 @@ static std::unordered_map<std::string, size_t> GetNonPersistableReferenceCounts(
 
 ExecutorPrepareContext::ExecutorPrepareContext(
     const framework::ProgramDesc& prog, size_t block_id,
-    const std::vector<std::string>& skip_ref_cnt_vars)
-    : prog_(prog), block_id_(block_id) {
-  if (GetEagerDeletionThreshold() >= 0) {
-    global_ref_cnts_ = GetNonPersistableReferenceCounts(prog.Block(block_id),
-                                                        skip_ref_cnt_vars);
+    const std::vector<std::string>& keep_vars, bool force_disable_gc)
+    : prog_(prog), block_id_(block_id), force_disable_gc_(force_disable_gc) {
+  if (GetEagerDeletionThreshold() >= 0 && !force_disable_gc_) {
+    global_ref_cnts_ =
+        GetNonPersistableReferenceCounts(prog.Block(block_id), keep_vars);
   }
 }
 
@@ -184,13 +189,12 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
 }
 
 void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
-                   bool create_local_scope, bool create_vars) {
+                   bool create_local_scope, bool create_vars,
+                   const std::vector<std::string>& skip_ref_cnt_vars,
+                   bool force_disable_gc) {
   platform::RecordBlock b(block_id);
   if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc);
-#ifdef PADDLE_WITH_NGRAPH
-  if (FLAGS_use_ngraph) operators::NgraphEngine::EnableNgraph(pdesc);
-#endif
-  auto ctx = Prepare(pdesc, block_id);
+  auto ctx = Prepare(pdesc, block_id, skip_ref_cnt_vars, force_disable_gc);
   RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars);
 }
 
@@ -357,20 +361,27 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
 
 std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
     const ProgramDesc& program, int block_id,
-    const std::vector<std::string>& skip_ref_cnt_vars) {
-  std::unique_ptr<ExecutorPrepareContext> ctx(
-      new ExecutorPrepareContext(program, block_id, skip_ref_cnt_vars));
+    const std::vector<std::string>& skip_ref_cnt_vars, bool force_disable_gc) {
+  std::unique_ptr<ExecutorPrepareContext> ctx(new ExecutorPrepareContext(
+      program, block_id, skip_ref_cnt_vars, force_disable_gc));
   PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
   auto& block = program.Block(block_id);
   for (auto& op_desc : block.AllOps()) {
     ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
   }
+#ifdef PADDLE_WITH_NGRAPH
+  if (FLAGS_use_ngraph) {
+    paddle::operators::NgraphEngine::FuseNgraphOps(
+        ctx->prog_.Block(ctx->block_id_), &ctx->ops_);
+  }
+#endif
   return ctx;
 }
 
 std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
     const ProgramDesc& program, const std::vector<int>& block_ids,
-    const std::vector<std::vector<std::string>>& skip_ref_cnt_vars) {
+    const std::vector<std::vector<std::string>>& skip_ref_cnt_vars,
+    bool force_disable_gc) {
   PADDLE_ENFORCE(
       skip_ref_cnt_vars.empty() || skip_ref_cnt_vars.size() == block_ids.size(),
       "skip_ref_cnt_vars should be either empty or equals to block number %d",
@@ -380,9 +391,11 @@ std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
   for (auto& bid : block_ids) {
     ExecutorPrepareContext* ctx;
     if (skip_ref_cnt_vars.empty()) {
-      ctx = new ExecutorPrepareContext(program, bid);
+      ctx = new ExecutorPrepareContext(program, bid, std::vector<std::string>(),
+                                       force_disable_gc);
     } else {
-      ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx]);
+      ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx],
+                                       force_disable_gc);
     }
     PADDLE_ENFORCE_LT(static_cast<size_t>(bid), program.Size());
     auto& block = program.Block(bid);
@@ -409,8 +422,9 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
 
   int64_t max_memory_size = GetEagerDeletionThreshold();
   std::unique_ptr<GarbageCollector> gc;
-  // skip while_op and while_grad_op temporarily
-  if (max_memory_size >= 0 && !keep_kids) {
+  // FIXME(zjl): recurrent_op is rather complex, we would
+  // disable gc forcely in recurrent_op
+  if (!ctx->force_disable_gc_ && max_memory_size >= 0) {
     ctx->ResetReferenceCount();
 #ifdef PADDLE_WITH_CUDA
     if (platform::is_gpu_place(place_)) {
@@ -428,6 +442,11 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
 #ifdef PADDLE_WITH_CUDA
     }
 #endif
+    // If gc is enabled and block size > 1
+    if (gc && ctx->prog_.Size() > 1) {
+      operators::PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp(ctx->block_id_,
+                                                                 ctx->ops_);
+    }
   }
 
   for (auto& op : ctx->ops_) {

paddle/fluid/framework/executor.h

@@ -15,7 +15,9 @@ limitations under the License. */
 #pragma once
 
 #include <map>
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <vector>
 #include "paddle/fluid/framework/garbage_collector.h"
 #include "paddle/fluid/framework/op_info.h"
@@ -30,7 +32,8 @@ namespace framework {
 struct ExecutorPrepareContext {
   ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id,
                          const std::vector<std::string>& skip_ref_cnt_vars =
-                             std::vector<std::string>());
+                             std::vector<std::string>(),
+                         bool force_disable_gc = false);
 
   ~ExecutorPrepareContext();
 
@@ -38,6 +41,7 @@ struct ExecutorPrepareContext {
 
   const framework::ProgramDesc& prog_;
   size_t block_id_;
+  bool force_disable_gc_;
   std::vector<std::unique_ptr<OperatorBase>> ops_;
 
   std::unordered_map<std::string, size_t> global_ref_cnts_;
@@ -66,7 +70,10 @@ class Executor {
    *  Scope
    */
   void Run(const ProgramDesc& prog, Scope* scope, int block_id,
-           bool create_local_scope = true, bool create_vars = true);
+           bool create_local_scope = true, bool create_vars = true,
+           const std::vector<std::string>& skip_ref_cnt_vars =
+               std::vector<std::string>(),
+           bool force_disable_gc = false);
 
   // This API is very slow.
   void Run(const ProgramDesc& program, Scope* scope,
@@ -79,12 +86,14 @@ class Executor {
   static std::unique_ptr<ExecutorPrepareContext> Prepare(
       const ProgramDesc& program, int block_id,
       const std::vector<std::string>& skip_ref_cnt_vars =
-          std::vector<std::string>());
+          std::vector<std::string>(),
+      bool force_disable_gc = false);
 
   static std::vector<std::shared_ptr<ExecutorPrepareContext>> Prepare(
       const ProgramDesc& program, const std::vector<int>& block_ids,
       const std::vector<std::vector<std::string>>& skip_ref_cnt_vars =
-          std::vector<std::vector<std::string>>());
+          std::vector<std::vector<std::string>>(),
+      bool force_disable_gc = false);
 
   void CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id);
 

paddle/fluid/framework/grad_op_desc_maker.h

@@ -14,7 +14,9 @@ limitations under the License. */
 
 #pragma once
 #include <algorithm>
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 #include "paddle/fluid/framework/op_desc.h"
@@ -55,11 +57,11 @@ class GradOpDescMakerBase {
                    std::back_inserter(ret_val),
                    [this](const std::string& fwd_var_name) -> std::string {
                      auto g_name = GradVarName(fwd_var_name);
-                     if (no_grad_set_.count(g_name)) {
-                       return kEmptyVarName;
-                     } else {
+                     if (no_grad_set_.empty() || !no_grad_set_.count(g_name)) {
                        (*this->grad_to_var_)[g_name] = fwd_var_name;
                        return g_name;
+                     } else {
+                       return kEmptyVarName;
                      }
                    });
     if (!drop_empty_grad) {

paddle/fluid/framework/ir/CMakeLists.txt

@@ -46,6 +46,9 @@ cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS pass)
 pass_library(graph_to_program_pass base)
 pass_library(graph_viz_pass base)
 pass_library(lock_free_optimize_pass base)
+pass_library(cpu_quantize_placement_pass base)
+pass_library(cpu_quantize_pass inference)
+pass_library(cpu_quantize_squash_pass inference)
 pass_library(fc_fuse_pass inference)
 pass_library(attention_lstm_fuse_pass inference)
 pass_library(infer_clean_graph_pass inference)
@@ -66,6 +69,8 @@ pass_library(conv_elementwise_add_fuse_pass inference)
 pass_library(conv_affine_channel_fuse_pass inference)
 pass_library(transpose_flatten_concat_fuse_pass inference)
 pass_library(identity_scale_op_clean_pass base)
+pass_library(sync_batch_norm_pass base)
+pass_library(runtime_context_cache_pass base)
 
 # There may be many transpose-flatten structures in a model, and the output of
 # these structures will be used as inputs to the concat Op. This pattern will
@@ -100,6 +105,12 @@ cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS g
 cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
 cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto)
 cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
+cc_test(test_cpu_quantize_placement_pass SRCS cpu_quantize_placement_pass_tester.cc DEPS cpu_quantize_placement_pass)
+cc_test(test_cpu_quantize_pass SRCS cpu_quantize_pass_tester.cc DEPS cpu_quantize_pass naive_executor)
+cc_test(test_cpu_quantize_squash_pass SRCS cpu_quantize_squash_pass_tester.cc DEPS cpu_quantize_squash_pass naive_executor)
+if(NOT WIN32)
+    cc_test(test_sync_batch_norm_pass SRCS sync_batch_norm_pass_tester.cc DEPS sync_batch_norm_pass)
+endif()
 if (WITH_MKLDNN)
     cc_test(test_depthwise_conv_mkldnn_pass SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
     cc_test(test_conv_bias_mkldnn_fuse_pass SRCS mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc DEPS conv_bias_mkldnn_fuse_pass naive_executor)

paddle/fluid/framework/ir/cpu_quantize_pass.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/ir/cpu_quantize_pass.h"
+#include <utility>
+#include <vector>
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/string/pretty_log.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+namespace {
+
+void UnlinkNodes(ir::Node* a, ir::Node* b) {
+  a->outputs.erase(std::remove(a->outputs.begin(), a->outputs.end(), b),
+                   a->outputs.end());
+  b->inputs.erase(std::remove(b->inputs.begin(), b->inputs.end(), a),
+                  b->inputs.end());
+}
+
+}  // namespace
+
+enum { U8_MAX = 255, S8_MAX = 127 };
+
+using EigenVectorArrayMap = Eigen::Map<Eigen::Array<double, Eigen::Dynamic, 1>>;
+using string::PrettyLogDetail;
+
+void CPUQuantizePass::QuantizeInput(Graph* g, Node* op, Node* input,
+                                    std::string input_name, double scale_to_one,
+                                    bool is_unsigned,
+                                    std::string scale_attr_name) const {
+  unsigned max = is_unsigned ? U8_MAX : S8_MAX;
+  float scale = scale_to_one * max;
+
+  // Create quantize output variable
+  VarDesc quantize_out_desc(patterns::PDNodeName("quantize", "out"));
+  auto* quantize_out_node = g->CreateVarNode(&quantize_out_desc);
+
+  // create a quantize op node
+  OpDesc q_desc;
+  q_desc.SetType("quantize");
+  q_desc.SetInput("Input", std::vector<std::string>({input->Name()}));
+  q_desc.SetOutput("Output",
+                   std::vector<std::string>({quantize_out_node->Name()}));
+  q_desc.SetAttr("Scale", scale);
+  q_desc.SetAttr("is_negative_input", !is_unsigned);
+  auto quantize_op = g->CreateOpNode(&q_desc);  // OpDesc will be copied.
+
+  // update op's input
+  op->Op()->SetInput(input_name,
+                     std::vector<std::string>({quantize_out_node->Name()}));
+
+  // link quantize op
+  UnlinkNodes(input, op);
+  IR_NODE_LINK_TO(input, quantize_op);
+  IR_NODE_LINK_TO(quantize_op, quantize_out_node);
+  IR_NODE_LINK_TO(quantize_out_node, op);
+
+  if (!scale_attr_name.empty()) op->Op()->SetAttr(scale_attr_name, scale);
+}
+
+void CPUQuantizePass::DequantizeOutput(Graph* g, Node* op, Node* output,
+                                       std::string output_name,
+                                       double scale_to_one, bool is_unsigned,
+                                       std::string scale_attr_name) const {
+  unsigned max = is_unsigned ? U8_MAX : S8_MAX;
+  float scale = scale_to_one * max;
+
+  // Create dequantize input variable
+  VarDesc dequantize_in_desc(patterns::PDNodeName("dequantize", "in"));
+  auto* dequantize_in_node = g->CreateVarNode(&dequantize_in_desc);
+
+  // create a dequantize op node for output.
+  OpDesc deq_desc;
+  deq_desc.SetType("dequantize");
+  deq_desc.SetInput("Input",
+                    std::vector<std::string>({dequantize_in_node->Name()}));
+  deq_desc.SetOutput("Output", std::vector<std::string>({output->Name()}));
+  deq_desc.SetAttr("Scale", scale);
+  auto dequantize_op = g->CreateOpNode(&deq_desc);  // OpDesc will be copied.
+
+  // update op's output
+  op->Op()->SetOutput(output_name,
+                      std::vector<std::string>({dequantize_in_node->Name()}));
+
+  // link dequantize op
+  UnlinkNodes(op, output);
+  IR_NODE_LINK_TO(op, dequantize_in_node);
+  IR_NODE_LINK_TO(dequantize_in_node, dequantize_op);
+  IR_NODE_LINK_TO(dequantize_op, output);
+
+  if (!scale_attr_name.empty()) op->Op()->SetAttr(scale_attr_name, scale);
+}
+
+void CPUQuantizePass::QuantizeConv(Graph* graph,
+                                   bool with_residual_data) const {
+  GraphPatternDetector gpd;
+  auto pattern = gpd.mutable_pattern();
+  patterns::ConvResidual conv_pattern{pattern, name_scope_};
+  conv_pattern(with_residual_data);
+
+  int quantize_conv_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "Quantize conv2d op";
+    GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern);
+    auto* conv_op_desc = conv_op->Op();
+
+    // skip if should not be quantized
+    if (!conv_op_desc->HasAttr("use_quantizer") ||
+        !boost::get<bool>(conv_op_desc->GetAttr("use_quantizer")))
+      return;
+
+    GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern);
+
+    // get scales calculated after warmup, they scale variables to MAX=1.0
+    auto scales = Get<VarQuantScale>("quant_var_scales");
+
+    auto input_scale = scales[conv_input->Name()].second.data<double>()[0];
+    bool is_input_unsigned = scales[conv_input->Name()].first;
+    QuantizeInput(g, conv_op, conv_input, "Input", input_scale,
+                  is_input_unsigned, "Scale_in");
+
+    auto filter_scale_tensor = scales[conv_filter->Name()].second;
+    EigenVectorArrayMap eigen_tensor{filter_scale_tensor.data<double>(),
+                                     filter_scale_tensor.numel(), 1};
+    eigen_tensor *= static_cast<double>(S8_MAX);
+    std::vector<float> filter_scale{
+        filter_scale_tensor.data<double>(),
+        filter_scale_tensor.data<double>() + filter_scale_tensor.numel()};
+
+    conv_op->Op()->SetAttr("Scale_weights", filter_scale);
+
+    if (with_residual_data) {
+      GET_IR_NODE_FROM_SUBGRAPH(conv_residual_data, conv_residual_data,
+                                conv_pattern);
+      auto residual_scale =
+          scales[conv_residual_data->Name()].second.data<double>()[0];
+      bool is_residual_unsigned = scales[conv_residual_data->Name()].first;
+
+      QuantizeInput(g, conv_op, conv_residual_data, "ResidualData",
+                    residual_scale, is_residual_unsigned, "Scale_in_eltwise");
+    }
+
+    auto output_scale = scales[conv_output->Name()].second.data<double>()[0];
+    bool is_output_unsigned = scales[conv_output->Name()].first;
+    DequantizeOutput(g, conv_op, conv_output, "Output", output_scale,
+                     is_output_unsigned, "Scale_out");
+
+    ++quantize_conv_count;
+  };
+
+  gpd(graph, handler);
+  AddStatis(quantize_conv_count);
+
+  std::stringstream msg_ss;
+  msg_ss << "---    quantized " << quantize_conv_count << " conv2d ops";
+  if (with_residual_data) msg_ss << " with residual connection";
+  PrettyLogDetail(msg_ss.str().c_str());
+}
+
+void CPUQuantizePass::QuantizePool(Graph* graph) const {
+  GraphPatternDetector gpd;
+  auto pattern = gpd.mutable_pattern();
+  patterns::Pool pool_pattern{pattern, name_scope_};
+  pool_pattern();
+
+  int quantize_pool_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "Quantize pool2d op";
+    GET_IR_NODE_FROM_SUBGRAPH(pool_op, pool_op, pool_pattern);
+    auto* pool_op_desc = pool_op->Op();
+
+    // skip if should not be quantized
+    if (!pool_op_desc->HasAttr("use_quantizer") ||
+        !boost::get<bool>(pool_op_desc->GetAttr("use_quantizer")))
+      return;
+
+    GET_IR_NODE_FROM_SUBGRAPH(pool_input, pool_input, pool_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(pool_output, pool_output, pool_pattern);
+
+    // get scales calculated after warmup, they scale variables to MAX=1.0
+    auto scales = Get<VarQuantScale>("quant_var_scales");
+
+    auto input_scale = scales[pool_input->Name()].second.data<double>()[0];
+    bool is_input_unsigned = scales[pool_input->Name()].first;
+    QuantizeInput(g, pool_op, pool_input, "X", input_scale, is_input_unsigned);
+
+    auto output_scale = scales[pool_output->Name()].second.data<double>()[0];
+    bool is_output_unsigned = scales[pool_output->Name()].first;
+    DequantizeOutput(g, pool_op, pool_output, "Out", output_scale,
+                     is_output_unsigned);
+
+    ++quantize_pool_count;
+  };
+
+  gpd(graph, handler);
+  AddStatis(quantize_pool_count);
+
+  PrettyLogDetail("---    quantized %d pool2d ops", quantize_pool_count);
+}
+
+std::unique_ptr<ir::Graph> CPUQuantizePass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  VLOG(3) << "Quantizing the graph.";
+  PADDLE_ENFORCE(graph.get());
+  FusePassBase::Init(name_scope_, graph.get());
+
+  PADDLE_ENFORCE(param_scope());
+
+  QuantizeConv(graph.get(), false /* with_residual_data */);
+  QuantizeConv(graph.get(), true /* with_residual_data */);
+  QuantizePool(graph.get());
+
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(cpu_quantize_pass, paddle::framework::ir::CPUQuantizePass)
+    .RequirePassAttr("quant_var_scales");

paddle/fluid/framework/ir/cpu_quantize_pass.h

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+/*
+ * Map variable name to tensor of scaling factors scaling it to MAX=1.0.
+ * bool denotes whether quantization of the variable should be done to unsigned
+ * type.
+ */
+using VarQuantScale =
+    std::unordered_map<std::string, std::pair<bool, LoDTensor>>;
+
+/*
+ * Quantize all supported operators.
+ */
+class CPUQuantizePass : public FusePassBase {
+ public:
+  virtual ~CPUQuantizePass() {}
+
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+
+  void QuantizeConv(Graph* graph, bool with_residual_data = false) const;
+
+  void QuantizePool(Graph* graph) const;
+
+  void QuantizeInput(Graph* g, Node* op, Node* input, std::string input_name,
+                     double scale_to_one, bool is_unsigned,
+                     std::string scale_attr_name = "") const;
+
+  void DequantizeOutput(Graph* g, Node* op, Node* output,
+                        std::string output_name, double scale_to_one,
+                        bool is_unsigned,
+                        std::string scale_attr_name = "") const;
+
+  const std::string name_scope_{"quantize"};
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/cpu_quantize_pass_tester.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/ir/cpu_quantize_pass.h"
+#include <gtest/gtest.h>
+#include "paddle/fluid/framework/naive_executor.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
+           const std::vector<std::string>& inputs,
+           const std::vector<std::string>& outputs, bool use_mkldnn,
+           bool use_quantizer = false) {
+  auto* op = prog->MutableBlock(0)->AppendOp();
+  op->SetType(type);
+  op->SetAttr("use_mkldnn", use_mkldnn);
+  op->SetAttr("name", name);
+  if (type == "conv2d") {
+    op->SetInput("Input", {inputs[0]});
+    op->SetInput("Filter", {inputs[1]});
+    if (inputs.size() > 2)
+      op->SetInput("Bias", {inputs[2]});
+    else
+      op->SetInput("Bias", {});
+    if (inputs.size() > 3) {
+      op->SetInput("ResidualData", {inputs[3]});
+      op->SetAttr("fuse_residual_connection", true);
+    } else {
+      op->SetInput("ResidualData", {});
+      op->SetAttr("fuse_residual_connection", false);
+    }
+    op->SetOutput("Output", {outputs[0]});
+    op->SetAttr("use_quantizer", use_quantizer);
+    op->SetAttr("Scale_in", 1.0f);
+    op->SetAttr("Scale_out", 1.0f);
+    op->SetAttr("Scale_weights", std::vector<float>{1.0f});
+  } else if (type == "pool2d") {
+    op->SetInput("X", {inputs[0]});
+    op->SetOutput("Out", {outputs[0]});
+    op->SetAttr("use_quantizer", use_quantizer);
+  } else if (type == "dropout") {
+    op->SetInput("X", {inputs[0]});
+    op->SetOutput("Out", {outputs[0]});
+  } else if (type == "fc") {
+    op->SetInput("Input", {inputs[0]});
+    if (inputs.size() > 1) op->SetInput("W", {inputs[1]});
+    if (inputs.size() > 2) op->SetInput("Bias", {inputs[2]});
+    op->SetOutput("Out", {outputs[0]});
+  }
+}
+
+static const std::initializer_list<std::string> variable_names{
+    "a", "w1", "c",  "d", "w2", "e",  "f", "g",
+    "h", "w3", "b1", "i", "j",  "w4", "b2"};
+// (a,w1)->Conv1->c and c->Pool1->d
+//
+// (d,w2)->Conv2->e and e->Pool2->f
+//
+// d->Dropout1->g and g->Fc1->h and (h,w3,b1,i)->Conv3->j
+//
+// (d,w4, b2)->Conv4->i
+ProgramDesc BuildProgramDesc(bool use_mkldnn, bool use_quantizer) {
+  ProgramDesc prog;
+  for (auto& v : variable_names) {
+    auto* var = prog.MutableBlock(0)->Var(v);
+    if (v.find("w") == 0 || v.find("b") == 0) {
+      var->SetPersistable(true);
+    }
+  }
+
+  SetOp(&prog, "conv2d", "Conv1", {"a", "w1"}, {"c"}, use_mkldnn,
+        use_quantizer);
+  SetOp(&prog, "pool2d", "Pool1", {"c"}, {"d"}, use_mkldnn, use_quantizer);
+
+  SetOp(&prog, "conv2d", "Conv2", {"d", "w2"}, {"e"}, use_mkldnn,
+        use_quantizer);
+  SetOp(&prog, "pool2d", "Pool2", {"e"}, {"f"}, use_mkldnn, use_quantizer);
+
+  SetOp(&prog, "dropout", "Dropout1", {"d"}, {"g"}, use_mkldnn);
+  SetOp(&prog, "fc", "Fc1", {"g"}, {"h"}, use_mkldnn);
+  SetOp(&prog, "conv2d", "Conv3", {"h", "w3", "b1", "i"}, {"j"}, use_mkldnn,
+        use_quantizer);
+
+  SetOp(&prog, "conv2d", "Conv4", {"c", "w4", "b2"}, {"i"}, use_mkldnn,
+        use_quantizer);
+
+  return prog;
+}
+
+void InitTensorHolder(Scope* scope, const paddle::platform::Place& place,
+                      const char* var_name) {
+  auto x = scope->Var(var_name);
+  auto tensor = x->GetMutable<LoDTensor>();
+  tensor->mutable_data(place, proto::VarType::FP32,
+                       ::paddle::memory::Allocator::kDefault, 1);
+}
+
+void MainTest(const ProgramDesc& prog, int conv_count, int pool_count,
+              int quant_count, int dequant_count, int added_nodes_count,
+              float scale) {
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  // Init scope, as it is used in pass
+  auto place = paddle::platform::CPUPlace();
+  NaiveExecutor exe{place};
+  Scope scope;
+  exe.CreateVariables(prog, 0, true, &scope);
+
+  auto* scales = new VarQuantScale();
+
+  for (auto& v : variable_names) {
+    InitTensorHolder(&scope, place, v.c_str());
+    LoDTensor tensor;
+    tensor.Resize({1});
+    auto* ptr = tensor.mutable_data<double>(place);
+    ptr[0] = 2.0;
+
+    (*scales)[v] = std::make_pair(false, std::move(tensor));
+  }
+
+  graph->Set(kParamScopeAttr, new framework::Scope*(&scope));
+
+  auto pass = PassRegistry::Instance().Get("cpu_quantize_pass");
+  pass->Set("quant_var_scales", scales);
+
+  int original_nodes_num = graph->Nodes().size();
+
+  graph = pass->Apply(std::move(graph));
+
+  int current_nodes_num = graph->Nodes().size();
+
+  int quantize_nodes_count = 0;
+  int dequantize_nodes_count = 0;
+  int conv2d_nodes_count = 0;
+  int pool2d_nodes_count = 0;
+  for (auto* node : graph->Nodes()) {
+    if (node->IsOp()) {
+      auto* op = node->Op();
+      if (op->Type() == "conv2d") {
+        conv2d_nodes_count++;
+        auto op_name = boost::get<std::string>(op->GetAttr("name"));
+        EXPECT_EQ(boost::get<float>(op->GetAttr("Scale_in")), scale)
+            << "Scale_in for node '" + op_name + "'.";
+        EXPECT_EQ(boost::get<float>(op->GetAttr("Scale_out")), scale)
+            << "Scale_out for node '" + op_name + "'.";
+        EXPECT_EQ(
+            boost::get<std::vector<float>>(op->GetAttr("Scale_weights"))[0],
+            scale)
+            << "Scale_weights for node '" + op_name + "'.";
+      } else if (op->Type() == "pool2d") {
+        pool2d_nodes_count++;
+      } else if (op->Type() == "quantize") {
+        quantize_nodes_count++;
+      } else if (op->Type() == "dequantize") {
+        dequantize_nodes_count++;
+      }
+    }
+  }
+  EXPECT_EQ(conv2d_nodes_count, conv_count);
+  EXPECT_EQ(pool2d_nodes_count, pool_count);
+  EXPECT_EQ(quantize_nodes_count, quant_count);
+  EXPECT_EQ(dequantize_nodes_count, dequant_count);
+  EXPECT_EQ(original_nodes_num + added_nodes_count, current_nodes_num);
+}
+
+TEST(CpuQuantizePass, quantize) {
+  bool use_mkldnn = true;
+  bool use_quantizer = true;
+  // (a->QUANT1->IN1,w1)->Conv1->OUT1->DEQUANT1->c and
+  // c->QUANT2->IN2->Pool1->OUT2->DEQUANT2->d
+  //
+  // (d->QUANT3->IN3,w2)->Conv2->OUT3->DEQUANT3->e and
+  // e->QUANT4->IN4->Pool2->OUT4->DEQUANT4->f
+  //
+  // d->Dropout1->g and g->Fc1->h and
+  // (h->QUANT5->IN5,w3,b1,i->QUANT6->IN6)->Conv3->OUT5->DEQUANT5->j
+  //
+  // (d->QUANT7->IN7,w4, b2)->Conv4->DEQUANT6->OUT6->i
+  // Insert nodes: 7 Quant + 7 IN + 6 OUT + 6 DEQUANT
+  int added_nodes = 7 + 7 + 6 + 6;
+  MainTest(BuildProgramDesc(use_mkldnn, use_quantizer), 4, 2, 7, 6, added_nodes,
+           2.0f * 127);
+}
+
+TEST(CpuQuantizePass, do_not_quantize) {
+  bool use_mkldnn = true;
+  bool use_quantizer = false;
+  int added_nodes = 0;
+  MainTest(BuildProgramDesc(use_mkldnn, use_quantizer), 4, 2, 0, 0, added_nodes,
+           1.0f);
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(cpu_quantize_pass);

paddle/fluid/framework/ir/cpu_quantize_placement_pass.cc

+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/ir/cpu_quantize_placement_pass.h"
+#include <string>
+#include <unordered_set>
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+std::unique_ptr<ir::Graph> CPUQuantizePlacementPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  VLOG(3) << "Marks operators which are to be quantized.";
+  const auto& excluded_ids_list =
+      Get<std::unordered_set<int>>("quantize_excluded_op_ids");
+  const auto& op_types_list =
+      Get<std::unordered_set<std::string>>("quantize_enabled_op_types");
+  for (const Node* n : graph->Nodes()) {
+    if (n->IsOp()) {
+      if (std::find(excluded_ids_list.begin(), excluded_ids_list.end(),
+                    n->id()) != excluded_ids_list.end())
+        continue;
+      auto* op = n->Op();
+      if (op->HasAttr("use_quantizer") || op->HasProtoAttr("use_quantizer")) {
+        if (op_types_list.empty()) {
+          op->SetAttr("use_quantizer", true);
+        } else if (std::find(op_types_list.begin(), op_types_list.end(),
+                             n->Name()) != op_types_list.end()) {
+          op->SetAttr("use_quantizer", true);
+        }
+      }
+    }
+  }
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(cpu_quantize_placement_pass,
+              paddle::framework::ir::CPUQuantizePlacementPass)
+    // a vector of operator type names to be quantized ("conv2d" etc.)
+    .RequirePassAttr("quantize_enabled_op_types")
+    // a vector of operator ids that are to be excluded from quantization
+    .RequirePassAttr("quantize_excluded_op_ids");

paddle/fluid/framework/ir/cpu_quantize_placement_pass.h

+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <memory>
+#include "paddle/fluid/framework/ir/pass.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+/*
+ * Specifies which operators should be quantized.
+ */
+class CPUQuantizePlacementPass : public Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+};
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/cpu_quantize_placement_pass_tester.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/ir/cpu_quantize_placement_pass.h"
+
+#include <gtest/gtest.h>
+#include <boost/logic/tribool.hpp>
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
+           const std::vector<std::string>& inputs,
+           const std::vector<std::string>& outputs,
+           boost::tribool use_quantizer) {
+  auto* op = prog->MutableBlock(0)->AppendOp();
+
+  op->SetType(type);
+
+  if (!boost::indeterminate(use_quantizer))
+    op->SetAttr("use_quantizer", use_quantizer);
+
+  if (type == "conv2d") {
+    op->SetAttr("name", name);
+    op->SetInput("Input", {inputs[0]});
+    op->SetInput("Filter", {inputs[1]});
+    op->SetInput("Bias", {inputs[2]});
+  } else if (type == "relu") {
+    op->SetInput("X", inputs);
+  } else if (type == "concat") {
+    op->SetAttr("axis", 1);
+    op->SetInput("X", {inputs[0], inputs[1]});
+  } else if (type == "pool2d") {
+    op->SetInput("X", {inputs[0]});
+  } else {
+    FAIL() << "Unexpected operator type.";
+  }
+  op->SetOutput("Out", {outputs[0]});
+}
+
+// operator                      use_quantizer
+// ---------------------------------------
+// (a,b)->concat->c              none
+// (c,weights,bias)->conv->f     false
+// f->relu->g                    none
+// g->pool->h                    false
+// (h,weights2,bias2)->conv->k   false
+// k->pool->l                    false
+ProgramDesc BuildProgramDesc() {
+  ProgramDesc prog;
+
+  for (auto& v :
+       std::vector<std::string>({"a", "b", "c", "weights", "bias", "f", "g",
+                                 "h", "weights2", "bias2", "k", "l"})) {
+    auto* var = prog.MutableBlock(0)->Var(v);
+    var->SetType(proto::VarType::SELECTED_ROWS);
+    if (v == "weights" || v == "bias") {
+      var->SetPersistable(true);
+    }
+  }
+
+  SetOp(&prog, "concat", "concat1", {"a", "b"}, {"c"}, boost::indeterminate);
+  SetOp(&prog, "conv2d", "conv1", {"c", "weights", "bias"}, {"f"}, false);
+  SetOp(&prog, "relu", "relu1", {"f"}, {"g"}, boost::indeterminate);
+  SetOp(&prog, "pool2d", "pool1", {"g"}, {"h"}, false);
+  SetOp(&prog, "conv2d", "conv2", {"h", "weights2", "bias2"}, {"k"}, false);
+  SetOp(&prog, "pool2d", "pool2", {"k"}, {"l"}, false);
+
+  return prog;
+}
+
+void MainTest(std::initializer_list<std::string> quantize_enabled_op_types,
+              std::initializer_list<int> quantize_excluded_op_ids,
+              unsigned expected_use_quantizer_true_count) {
+  auto prog = BuildProgramDesc();
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  auto pass = PassRegistry::Instance().Get("cpu_quantize_placement_pass");
+  pass->Set("quantize_enabled_op_types",
+            new std::unordered_set<std::string>(quantize_enabled_op_types));
+  pass->Set("quantize_excluded_op_ids",
+            new std::unordered_set<int>(quantize_excluded_op_ids));
+
+  graph = pass->Apply(std::move(graph));
+
+  unsigned use_quantizer_true_count = 0;
+
+  for (auto* node : graph->Nodes()) {
+    if (node->IsOp()) {
+      auto* op = node->Op();
+      if (op->HasAttr("use_quantizer") &&
+          boost::get<bool>(op->GetAttr("use_quantizer"))) {
+        ++use_quantizer_true_count;
+      }
+    }
+  }
+
+  EXPECT_EQ(use_quantizer_true_count, expected_use_quantizer_true_count);
+}
+
+TEST(QuantizerPlacementPass, enabled_pool) { MainTest({"pool2d"}, {}, 2); }
+
+TEST(QuantizerPlacementPass, enabled_conv_excluded_one) {
+  MainTest({"conv2d"}, {4}, 1);
+}
+
+TEST(QuantizerPlacementPass, excluded_none) {
+  // 2 conv + 2 pool
+  MainTest({}, {}, 4);
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(cpu_quantize_placement_pass);

paddle/fluid/framework/ir/cpu_quantize_squash_pass.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file eint8_outcept 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 eint8_outpress or
+// implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/ir/cpu_quantize_squash_pass.h"
+#include <string>
+#include <vector>
+#include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/string/pretty_log.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+using string::PrettyLogDetail;
+
+void CPUQuantizeSquashPass::FindNodesToKeep(
+    Graph* graph,
+    std::unordered_map<const Node*, int>* nodes_keep_counter) const {
+  GraphPatternDetector gpd;
+  patterns::DequantAny deq_any_pattern{gpd.mutable_pattern(), "deqant_any"};
+  deq_any_pattern();
+
+  int found_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    GET_IR_NODE_FROM_SUBGRAPH(dequant_out, dequant_out, deq_any_pattern);
+
+    if (nodes_keep_counter->find(dequant_out) == nodes_keep_counter->end())
+      (*nodes_keep_counter)[dequant_out] = 1;
+    else
+      (*nodes_keep_counter)[dequant_out] += 1;
+
+    found_count++;
+  };
+  gpd(graph, handler);
+  AddStatis(found_count);
+}
+
+void CPUQuantizeSquashPass::Squash(
+    Graph* graph,
+    std::unordered_map<const Node*, int>* nodes_keep_counter) const {
+  GraphPatternDetector gpd;
+  patterns::DequantQuantAny squash_pattern{gpd.mutable_pattern(), "squash"};
+  squash_pattern();
+
+  int found_squash_count = 0;
+  auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
+                     Graph* g) {
+    VLOG(4) << "squash requantize-quantize ops pair";
+
+    GET_IR_NODE_FROM_SUBGRAPH(dequant_in, dequant_in, squash_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(dequant_op, dequant_op, squash_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(dequant_out, dequant_out, squash_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(quant_op, quant_op, squash_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(quant_out, quant_out, squash_pattern);
+    GET_IR_NODE_FROM_SUBGRAPH(next_op, next_op, squash_pattern);
+
+    auto* next_op_desc = next_op->Op();
+    float dequant_scale = boost::get<float>(dequant_op->Op()->GetAttr("Scale"));
+    float quant_scale = boost::get<float>(quant_op->Op()->GetAttr("Scale"));
+    PADDLE_ENFORCE(nodes_keep_counter->find(dequant_out) !=
+                   nodes_keep_counter->end());
+
+    // check if dequantize op should be kept or removed, decrease the counter
+    bool keep_dequant = (*nodes_keep_counter)[dequant_out]-- > 1;
+
+    if (dequant_scale == quant_scale) {
+      // squash dequantize-quantize to nothing
+      auto quant_out_var_name = quant_out->Name();
+      auto next_op_inputs = next_op_desc->InputNames();
+      for (const auto& name : next_op_inputs) {
+        auto var_name = next_op_desc->Input(name)[0];
+        if (var_name.compare(quant_out_var_name) == 0) {
+          next_op_desc->SetInput(
+              name, std::vector<std::string>({dequant_in->Name()}));
+          break;
+        }
+      }
+
+      if (keep_dequant)
+        GraphSafeRemoveNodes(graph, {quant_op, quant_out});
+      else
+        GraphSafeRemoveNodes(graph,
+                             {dequant_op, quant_op, dequant_out, quant_out});
+
+      IR_NODE_LINK_TO(dequant_in, next_op);
+
+      found_squash_count++;
+    } else {
+      // squash dequantize-quantize to requantize op
+      OpDesc desc;
+      desc.SetType("requantize");
+      desc.SetInput("Input", std::vector<std::string>({dequant_in->Name()}));
+      desc.SetOutput("Output", std::vector<std::string>({quant_out->Name()}));
+      desc.SetAttr("Scale_in", dequant_scale);
+      desc.SetAttr("Scale_out", quant_scale);
+
+      auto requant_op = g->CreateOpNode(&desc);
+
+      if (keep_dequant)
+        GraphSafeRemoveNodes(graph, {quant_op});
+      else
+        GraphSafeRemoveNodes(graph, {dequant_op, quant_op, dequant_out});
+
+      IR_NODE_LINK_TO(dequant_in, requant_op);
+      IR_NODE_LINK_TO(requant_op, quant_out);
+
+      found_squash_count++;
+    }
+  };
+  gpd(graph, handler);
+  AddStatis(found_squash_count);
+  PrettyLogDetail("---    squashed %d dequantize-quantize pairs",
+                  found_squash_count);
+}
+
+std::unique_ptr<ir::Graph> CPUQuantizeSquashPass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  PADDLE_ENFORCE(graph.get());
+  FusePassBase::Init("cpu_quantize_squash_pass", graph.get());
+
+  std::unordered_map<const Node*, int> nodes_keep_counter;
+  FindNodesToKeep(graph.get(), &nodes_keep_counter);
+  Squash(graph.get(), &nodes_keep_counter);
+
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(cpu_quantize_squash_pass,
+              paddle::framework::ir::CPUQuantizeSquashPass);

paddle/fluid/framework/details/fuse_vars_op_handle.h → paddle/fluid/framework/ir/cpu_quantize_squash_pass.h

-//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -14,52 +14,45 @@
 
 #pragma once
 
-#include <map>
+#include <memory>
 #include <string>
-#include <vector>
-
-#include "paddle/fluid/framework/details/container_cast.h"
-#include "paddle/fluid/framework/details/op_handle_base.h"
-#include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/platform/device_context.h"
+#include <unordered_map>
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
+#include "paddle/fluid/framework/ir/pass.h"
 
 namespace paddle {
 namespace framework {
-namespace details {
+namespace ir {
 
-struct FuseVarsOpHandle : public OpHandleBase {
+/*
+ * Squash dequantize->quantize pair pattern into requantize op
+ */
+class CPUQuantizeSquashPass : public FusePassBase {
  public:
-  FuseVarsOpHandle(ir::Node *node, Scope *local_scope,
-                   const platform::Place &place,
-                   const std::unordered_map<std::string, int64_t> &inputs_numel,
-                   const proto::VarType::Type var_type)
-      : OpHandleBase(node),
-        local_scope_(local_scope),
-        place_(place),
-        inputs_numel_(inputs_numel),
-        type_(var_type) {
-    total_numel_ = 0;
-    for (auto in_numel : inputs_numel) {
-      PADDLE_ENFORCE_GT(in_numel.second, 0);
-      total_numel_ += in_numel.second;
-    }
-  }
-
-  std::string Name() const override;
-
-  bool IsMultiDeviceTransfer() override { return false; };
+  virtual ~CPUQuantizeSquashPass() {}
 
  protected:
-  void RunImpl() override;
-
- private:
-  Scope *local_scope_;
-  const platform::Place place_;
-  const std::unordered_map<std::string, int64_t> inputs_numel_;
-  const proto::VarType::Type type_;
-  int64_t total_numel_;
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+
+  /*
+   * For each dequantize's output find the number of operators it is an input to
+   */
+  void FindNodesToKeep(
+      Graph* graph,
+      std::unordered_map<const Node*, int>* nodes_keep_counter) const;
+
+  /*
+   * Squash dequantize-quantize ops pairs into requantize or nothing
+   */
+  void Squash(Graph* graph,
+              std::unordered_map<const Node*, int>* nodes_keep_counter) const;
+
+  const std::string name_scope_{"squash"};
 };
-}  // namespace details
+
+}  // namespace ir
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/cpu_quantize_squash_pass_tester.cc

+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/framework/ir/cpu_quantize_squash_pass.h"
+#include <gtest/gtest.h>
+#include "paddle/fluid/framework/naive_executor.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name,
+           const std::vector<std::string>& inputs,
+           const std::vector<std::string>& outputs, bool use_mkldnn,
+           float scale = 0) {
+  auto* op = prog->MutableBlock(0)->AppendOp();
+  op->SetType(type);
+  op->SetAttr("use_mkldnn", use_mkldnn);
+  op->SetAttr("name", name);
+  if (type == "conv2d") {
+    op->SetInput("Input", {inputs[0]});
+    if (inputs.size() > 1) op->SetInput("Filter", {inputs[1]});
+    if (inputs.size() > 2) op->SetInput("Bias", {inputs[2]});
+    op->SetOutput("Output", {outputs[0]});
+  } else if (type == "quantize") {
+    op->SetInput("Input", {inputs[0]});
+    op->SetOutput("Output", {outputs[0]});
+    op->SetAttr("Scale", scale);
+  } else if (type == "dequantize") {
+    op->SetInput("Input", {inputs[0]});
+    op->SetOutput("Output", {outputs[0]});
+    op->SetAttr("Scale", scale);
+  }
+}
+
+// (a,w1,b1)->Conv1->d
+// d->Dequant->e
+// e->Quant->f
+// (f,w2,b2)->Conv2->i
+ProgramDesc BuildProgramDesc(bool use_mkldnn, float scale1, float scale2) {
+  ProgramDesc prog;
+  for (auto& v : std::initializer_list<std::string>(
+           {"a", "w1", "b1", "d", "e", "f", "w2", "b2", "i"})) {
+    auto* var = prog.MutableBlock(0)->Var(v);
+    if (v.find("w") == 0 || v.find("b") == 0) {
+      var->SetPersistable(true);
+    }
+  }
+
+  SetOp(&prog, "conv2d", "Conv1", {"a", "w1", "b1"}, {"d"}, use_mkldnn);
+  SetOp(&prog, "dequantize", "Dequant", {"d"}, {"e"}, use_mkldnn, scale1);
+  SetOp(&prog, "quantize", "Quant", {"e"}, {"f"}, use_mkldnn, scale2);
+  SetOp(&prog, "conv2d", "Conv2", {"f", "w2", "b2"}, {"i"}, use_mkldnn);
+  return prog;
+}
+
+static const std::initializer_list<std::string> variable_names{
+    "a", "b", "c", "d", "e", "f", "g", "h"};
+// a->Conv1->b
+// b->Dequant->c
+//
+// c->Quant1->d and d->Conv2->e
+//
+// c->Conv3->f
+//
+// c->Quant2->g and g->Conv4->h
+//
+ProgramDesc BuildProgramDesc2(bool use_mkldnn, float scale1, float scale2,
+                              float scale3) {
+  ProgramDesc prog;
+  for (auto& v : variable_names) {
+    prog.MutableBlock(0)->Var(v);
+  }
+
+  SetOp(&prog, "conv2d", "Conv1", {"a"}, {"b"}, use_mkldnn);
+  SetOp(&prog, "dequantize", "Dequant", {"b"}, {"c"}, use_mkldnn, scale1);
+
+  SetOp(&prog, "quantize", "Quant1", {"c"}, {"d"}, use_mkldnn, scale2);
+  SetOp(&prog, "conv2d", "Conv2", {"d"}, {"e"}, use_mkldnn);
+
+  SetOp(&prog, "conv2d", "Conv3", {"c"}, {"f"}, use_mkldnn);
+
+  SetOp(&prog, "quantize", "Quant2", {"c"}, {"g"}, use_mkldnn, scale3);
+  SetOp(&prog, "conv2d", "Conv4", {"g"}, {"h"}, use_mkldnn);
+
+  return prog;
+}
+
+void InitTensorHolder(Scope* scope, const paddle::platform::Place& place,
+                      const char* var_name) {
+  auto x = scope->Var(var_name);
+  auto tensor = x->GetMutable<LoDTensor>();
+  tensor->mutable_data(place, proto::VarType::FP32,
+                       ::paddle::memory::Allocator::kDefault, 1);
+}
+
+void MainTest(const ProgramDesc& prog, int removed_nodes_num) {
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  // Init scope, as it is used in pass
+  auto place = paddle::platform::CPUPlace();
+  NaiveExecutor exe{place};
+  Scope scope;
+  exe.CreateVariables(prog, 0, true, &scope);
+
+  for (auto& v : variable_names) {
+    InitTensorHolder(&scope, place, v.c_str());
+  }
+
+  graph->Set(kParamScopeAttr, new framework::Scope*(&scope));
+
+  auto pass = PassRegistry::Instance().Get("cpu_quantize_squash_pass");
+
+  int original_nodes_num = graph->Nodes().size();
+
+  graph = pass->Apply(std::move(graph));
+
+  int current_nodes_num = graph->Nodes().size();
+
+  EXPECT_EQ(original_nodes_num - removed_nodes_num, current_nodes_num);
+}
+
+TEST(CpuQuantizeSquashPass, equal_scales) {
+  auto scale = 1.2345f;
+  auto use_mkldnn = true;
+  // Remove 4 nodes: Dequant, Quant, e, f
+  auto remove_nodes = 4;
+  MainTest(BuildProgramDesc(use_mkldnn, scale, scale), remove_nodes);
+
+  use_mkldnn = !use_mkldnn;
+  MainTest(BuildProgramDesc(use_mkldnn, scale, scale), remove_nodes);
+}
+
+TEST(CpuQuantizeSquashPass, inequal_scales) {
+  auto scale1 = 1.2345f;
+  auto scale2 = 21.0f;
+  auto use_mkldnn = true;
+  // Remove 3 nodes: Dequant, Quant, e
+  // Insert 1 node: requantize
+  auto remove_nodes = 2;
+  MainTest(BuildProgramDesc(use_mkldnn, scale1, scale2), remove_nodes);
+
+  use_mkldnn = !use_mkldnn;
+  MainTest(BuildProgramDesc(use_mkldnn, scale1, scale2), remove_nodes);
+}
+
+TEST(CpuQuantizeSquashPass, branch_to_equal_inequal_and_fp32) {
+  // Delete both quantize ops,
+  // bypass dequantize in both branches,
+  // insert requantize on one branch
+  auto scale = 1.2345f;
+  auto scale2 = 21.0f;
+  auto use_mkldnn = true;
+  // Remove 3 nodes: Quant1, Quant2, g
+  // Insert 1 node: requantize
+  auto remove_nodes = 2;
+  MainTest(BuildProgramDesc2(use_mkldnn, scale, scale, scale2), remove_nodes);
+
+  use_mkldnn = !use_mkldnn;
+  MainTest(BuildProgramDesc2(use_mkldnn, scale, scale, scale2), remove_nodes);
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(cpu_quantize_squash_pass);

paddle/fluid/framework/ir/fuse_pass_base.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <string>
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/pass.h"
 #include "paddle/fluid/framework/scope.h"
@@ -24,6 +25,10 @@ namespace ir {
 
 static const char kParamScopeAttr[] = "__param_scope__";
 static const char kFuseStatisAttr[] = "__fuse_statis__";
+// When we use trt or other third_party lib, the parameters are managed by
+// the lib, but not the fluid. So we need to record them to avoid duplicate
+// allocation.
+static const char kRepetitiveParamAttr[] = "__repetitive_param__";
 
 enum FuseOptions {
   DO_NOT_FUSE,  // fusing will not be done

paddle/fluid/framework/ir/graph.cc

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <algorithm>
-#include <unordered_set>
+#include <unordered_map>
 
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/op_proto_maker.h"
@@ -152,6 +152,39 @@ void Graph::ResolveHazard(
   }
 }
 
+std::shared_ptr<Graph> Graph::Clone() {
+  auto cloned_graph = std::make_shared<Graph>(this->program_);
+  cloned_graph->ReleaseNodes();
+  cloned_graph->num_node_created_ = 0;
+  std::unordered_map<ir::Node *, ir::Node *> origin_to_cloned;
+  for (auto *n : this->node_set_) {
+    ir::Node *cloned_node = nullptr;
+    if (n->IsCtrlVar()) {
+      cloned_node = cloned_graph->CreateControlDepVar();
+    } else if (!n->var_desc_ && !n->op_desc_) {  // empty node
+      cloned_node = cloned_graph->CreateEmptyNode(n->Name(), n->NodeType());
+    } else if (n->IsVar()) {
+      cloned_node = cloned_graph->CreateVarNode(n->Var());
+    } else if (n->IsOp()) {
+      cloned_node = cloned_graph->CreateOpNode(n->Op());
+    }
+    if (cloned_node) {
+      origin_to_cloned[n] = cloned_node;
+    } else {
+      PADDLE_THROW("The cloned node's type is not supported!");
+    }
+  }
+  for (auto *n : this->node_set_) {
+    for (auto it = n->inputs.begin(); it != n->inputs.end(); it++) {
+      origin_to_cloned[n]->inputs.push_back(origin_to_cloned[*it]);
+    }
+    for (auto it = n->outputs.begin(); it != n->outputs.end(); it++) {
+      origin_to_cloned[n]->outputs.push_back(origin_to_cloned[*it]);
+    }
+  }
+  return cloned_graph;
+}
+
 bool IsControlDepVar(const ir::Node &var) {
   return var.Name().find(ir::Node::kControlDepVarName) != std::string::npos;
 }

paddle/fluid/framework/ir/graph.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <map>
 #include <memory>
 #include <string>
+#include <unordered_set>
 #include <vector>
 
 #include "paddle/fluid/framework/ir/node.h"
@@ -199,7 +200,12 @@ class Graph {
   // WARN: After a series of passes, the current graph can be quite
   // different from OriginProgram. Caller shouldn't assume much from
   // the returned OriginProgram.
-  const ProgramDesc &OriginProgram() const { return program_; }
+  const ProgramDesc &OriginProgram() const {
+    LOG(WARNING) << "WARN: After a series of passes, the current graph can be "
+                    "quite different from OriginProgram. So, please avoid "
+                    "using the `OriginProgram()` method!";
+    return program_;
+  }
 
   // This method takes ownership of `node`.
   ir::Node *AddNode(ir::Node *node) {
@@ -212,6 +218,10 @@ class Graph {
   void ResolveHazard(
       const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
 
+  // Create a new and duplicated graph.
+  // WARN: The method only clones the graph structure, not its attributes.
+  std::shared_ptr<Graph> Clone();
+
  private:
   std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
       const ProgramDesc &program);

paddle/fluid/framework/ir/graph_helper.cc

@@ -130,15 +130,21 @@ std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
     if (adj_list.find(n) == adj_list.end()) {
       adj_list[n] = std::unordered_set<ir::Node *>();
     }
+    std::vector<ir::Node *> nodes;
     for (auto &var : n->inputs) {
       for (auto &adj_n : var->inputs) {
         PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
         VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
                 << " -> " << n->Name() << reinterpret_cast<void *>(n)
                 << "  via " << var->Name() << reinterpret_cast<void *>(var);
-        adj_list[n].insert(adj_n);
+        nodes.push_back(adj_n);
       }
     }
+    std::sort(nodes.begin(), nodes.end(), [](ir::Node *node1, ir::Node *node2) {
+      return node1->id() > node2->id();
+    });
+    adj_list[n].insert(std::make_move_iterator(nodes.begin()),
+                       std::make_move_iterator(nodes.end()));
   }
   return adj_list;
 }

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -90,7 +90,8 @@ void GraphPatternDetector::operator()(Graph *graph,
   ValidateByNodeRole(&subgraphs);
 
   if (subgraphs.empty()) return;
-  PrettyLogEndl(Style::detail(), "---  detect %d subgraphs", subgraphs.size());
+  PrettyLogEndl(Style::detail(), "---  detected %d subgraphs",
+                subgraphs.size());
   int id = 0;
   for (auto &g : subgraphs) {
     VLOG(3) << "optimizing #" << id++ << " subgraph";
@@ -598,10 +599,19 @@ bool VarLinksToOp(Node *node, const std::string &op_type) {
 bool IsNthInput(Node *var, Node *op, const std::string &argument, size_t nth) {
   PADDLE_ENFORCE(var->IsVar());
   PADDLE_ENFORCE(op->IsOp());
-  if (op->Op()->Input(argument).size() <= nth) return false;
+  if (!HasInput(op, argument) || op->Op()->Input(argument).size() <= nth)
+    return false;
   return var->Name() == op->Op()->Input(argument)[nth];
 }
 
+bool HasInput(Node *op, const std::string &argument) {
+  PADDLE_ENFORCE(op->IsOp());
+  auto const &names = op->Op()->InputNames();
+  if (std::find(names.begin(), names.end(), argument) == names.end())
+    return false;
+  return true;
+}
+
 bool IsNthOutput(Node *var, Node *op, const std::string &argument, size_t nth) {
   PADDLE_ENFORCE(var->IsVar());
   PADDLE_ENFORCE(op->IsOp());
@@ -1074,9 +1084,60 @@ PDNode *patterns::Conv::operator()() {
                         ->AsOutput()
                         ->assert_is_op_output("conv2d", "Output");
 
-  conv_op->LinksFrom({input_var, filter_var});
-  conv_op->LinksTo({output_var});
+  conv_op->LinksFrom({input_var, filter_var}).LinksTo({output_var});
+  return output_var;
+}
+
+PDNode *patterns::ConvResidual::operator()(bool with_residual_data) {
+  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
+
+  if (!with_residual_data) {
+    conv_op->assert_more([&](Node *x) {
+      auto node_names = x->Op()->InputNames();
+      if (!HasInput(x, "ResidualData") ||
+          x->Op()->Input("ResidualData").size() == 0)
+        return true;
+      return false;
+    });
+  }
+
+  auto input_var = pattern->NewNode(conv_input_repr())
+                       ->AsInput()
+                       ->assert_is_op_input("conv2d", "Input");
+
+  auto filter_var = pattern->NewNode(conv_filter_repr())
+                        ->AsInput()
+                        ->assert_is_op_input("conv2d", "Filter");
+
+  auto output_var = pattern->NewNode(conv_output_repr())
+                        ->AsOutput()
+                        ->assert_is_op_output("conv2d", "Output");
+
+  std::vector<PDNode *> links_from{input_var, filter_var};
 
+  if (with_residual_data) {
+    auto res_conn_var = pattern->NewNode(conv_residual_data_repr())
+                            ->AsInput()
+                            ->assert_is_op_input("conv2d", "ResidualData");
+    links_from.push_back(res_conn_var);
+  }
+
+  conv_op->LinksFrom(links_from).LinksTo({output_var});
+  return output_var;
+}
+
+PDNode *patterns::Pool::operator()() {
+  auto pool_op = pattern->NewNode(pool_op_repr())->assert_is_op("pool2d");
+
+  auto input_var = pattern->NewNode(pool_input_repr())
+                       ->AsInput()
+                       ->assert_is_op_input("pool2d", "X");
+
+  auto output_var = pattern->NewNode(pool_output_repr())
+                        ->AsOutput()
+                        ->assert_is_op_output("pool2d", "Out");
+
+  pool_op->LinksFrom({input_var}).LinksTo({output_var});
   return output_var;
 }
 
@@ -1301,6 +1362,51 @@ PDNode *patterns::ConvAffineChannel::operator()(
   return ac_out_var;
 }
 
+PDNode *patterns::DequantQuantAny::operator()() {
+  auto *dequant_in = pattern->NewNode(dequant_in_repr())
+                         ->AsInput()
+                         ->assert_is_op_input("dequantize", "Input");
+
+  auto *dequant_op =
+      pattern->NewNode(dequant_op_repr())->assert_is_op("dequantize");
+
+  auto *dequant_out = pattern->NewNode(dequant_out_repr())
+                          ->AsOutput()
+                          ->assert_is_op_output("dequantize", "Output");
+
+  auto *quant_op = pattern->NewNode(quant_op_repr())
+                       ->assert_is_op("quantize")
+                       ->AsIntermediate();
+
+  auto *quant_out = pattern->NewNode(quant_out_repr())
+                        ->AsOutput()
+                        ->assert_is_op_output("quantize");
+
+  auto *next_op = pattern->NewNode(next_op_repr())->assert_is_op();
+
+  dequant_op->LinksFrom({dequant_in}).LinksTo({dequant_out});
+  quant_op->LinksFrom({dequant_out}).LinksTo({quant_out});
+  next_op->LinksFrom({quant_out});
+
+  return quant_out;
+}
+
+PDNode *patterns::DequantAny::operator()() {
+  auto *dequant_op =
+      pattern->NewNode(dequant_op_repr())->assert_is_op("dequantize");
+
+  auto *dequant_out = pattern->NewNode(dequant_out_repr())
+                          ->AsOutput()
+                          ->assert_is_op_output("dequantize", "Output");
+
+  auto *next_op = pattern->NewNode(next_op_repr())->assert_is_op();
+
+  dequant_op->LinksTo({dequant_out});
+  next_op->LinksFrom({dequant_out});
+
+  return dequant_out;
+}
+
 // a -> transpose_op(1) -> transpose_out_a -> flatten_op(1) -> flatten_out_a
 // b -> transpose_op(2) -> transpose_out_b -> flatten_op(2) -> flatten_out_b
 // ...

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -18,8 +18,11 @@
 #include <gtest/gtest_prod.h>
 #endif
 
+#include <memory>
 #include <numeric>
 #include <string>
+#include <unordered_map>
+#include <unordered_set>
 #include <utility>
 #include <vector>
 #include "paddle/fluid/framework/ir/graph.h"
@@ -302,6 +305,9 @@ bool VarLinksFromOp(Node* node, const std::string& op_type);
 // Check whether a var node is a op node's nth input.
 bool IsNthInput(Node* var, Node* op, const std::string& argument, size_t nth);
 
+// Check whether the op node has input of given name.
+bool HasInput(Node* op, const std::string& argument);
+
 // Tell whether a var node is a op node's nth output.
 bool IsNthOutput(Node* var, Node* op, const std::string& argument, size_t nth);
 
@@ -656,6 +662,35 @@ struct Conv : public PatternBase {
   PATTERN_DECL_NODE(conv_output);
 };
 
+// Convolution op with residual data
+struct ConvResidual : public PatternBase {
+  ConvResidual(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "conv_residual") {}
+
+  PDNode* operator()(bool with_residual_data);
+
+  PATTERN_DECL_NODE(conv_op);
+  PATTERN_DECL_NODE(conv_input);
+  PATTERN_DECL_NODE(conv_filter);
+  PATTERN_DECL_NODE(conv_residual_data);
+  PATTERN_DECL_NODE(conv_output);
+};
+
+// Pool op
+// Forward pass for pooling.
+// pool_input is the input.
+// pool_output is a result of the operator.
+struct Pool : public PatternBase {
+  Pool(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "pooling") {}
+
+  PDNode* operator()();
+
+  PATTERN_DECL_NODE(pool_op);
+  PATTERN_DECL_NODE(pool_input);
+  PATTERN_DECL_NODE(pool_output);
+};
+
 // ElementwiseAdd used in residual connections.
 // y_var is used and convolution output.
 // The operator is removed, when residual
@@ -766,6 +801,34 @@ struct ConvAffineChannel : public PatternBase {
   PATTERN_DECL_NODE(ac_out);  // Out
 };
 
+// Dequantize + Quantize + anyOP
+// This pattern is used for squashing the dequantize-quantize pairs.
+struct DequantQuantAny : public PatternBase {
+  DequantQuantAny(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "dequant_quant_any") {}
+  PDNode* operator()();
+
+  PATTERN_DECL_NODE(dequant_in);
+  PATTERN_DECL_NODE(dequant_op);
+  PATTERN_DECL_NODE(dequant_out);
+  PATTERN_DECL_NODE(quant_op);
+  PATTERN_DECL_NODE(quant_out);
+  PATTERN_DECL_NODE(next_op);
+};
+
+// Dequantize + anyOP
+// This quantize is used for getting number of ops the Dequantize's
+// output is an input to.
+struct DequantAny : public PatternBase {
+  DequantAny(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "dequant_any") {}
+  PDNode* operator()();
+
+  PATTERN_DECL_NODE(dequant_op);
+  PATTERN_DECL_NODE(dequant_out);
+  PATTERN_DECL_NODE(next_op);
+};
+
 struct TransposeFlattenConcat : public PatternBase {
   TransposeFlattenConcat(PDPattern* pattern, const std::string& name_scope)
       : PatternBase(pattern, name_scope, "transpose_flatten_concat") {}

paddle/fluid/framework/ir/graph_test.cc

@@ -43,20 +43,20 @@ class SumOpMaker : public OpProtoAndCheckerMaker {
 
 class SumOpVarTypeInference : public VarTypeInference {
  public:
-  void operator()(const OpDesc &op_desc, BlockDesc *block) const override {
-    auto &inputs = op_desc.Input("X");
+  void operator()(InferVarTypeContext *ctx) const override {
+    auto &inputs = ctx->Input("X");
     auto default_var_type = proto::VarType::SELECTED_ROWS;
 
     bool any_input_is_lod_tensor = std::any_of(
-        inputs.begin(), inputs.end(), [block](const std::string &name) {
-          return block->Var(name)->GetType() == proto::VarType::LOD_TENSOR;
+        inputs.begin(), inputs.end(), [&ctx](const std::string &name) {
+          return ctx->GetType(name) == proto::VarType::LOD_TENSOR;
         });
     if (any_input_is_lod_tensor) {
       default_var_type = proto::VarType::LOD_TENSOR;
     }
 
-    auto out_var_name = op_desc.Output("Out").front();
-    block->Var(out_var_name)->SetType(default_var_type);
+    auto out_var_name = ctx->Output("Out").front();
+    ctx->SetType(out_var_name, default_var_type);
   }
 };
 
@@ -71,7 +71,7 @@ class DummyOpMaker : public OpProtoAndCheckerMaker {
 
 class DummyOpVarTypeInference : public VarTypeInference {
  public:
-  void operator()(const OpDesc &op_desc, BlockDesc *block) const override {}
+  void operator()(framework::InferVarTypeContext *ctx) const override {}
 };
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.h

@@ -14,12 +14,16 @@ limitations under the License. */
 
 #pragma once
 
+#include <memory>
 #include "paddle/fluid/framework/ir/pass.h"
 
 namespace paddle {
 namespace framework {
 namespace ir {
 
+/*
+ * Specifies which operators should use MKLDNN.
+ */
 class MKLDNNPlacementPass : public Pass {
  protected:
   std::unique_ptr<ir::Graph> ApplyImpl(

paddle/fluid/framework/ir/node.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <typeindex>
 #include <typeinfo>

paddle/fluid/framework/ir/runtime_context_cache_pass.cc

+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/framework/ir/runtime_context_cache_pass.h"
+#include <memory>
+#include "paddle/fluid/framework/operator.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+
+std::unique_ptr<ir::Graph> RuntimeContextCachePass::ApplyImpl(
+    std::unique_ptr<ir::Graph> graph) const {
+  VLOG(3) << "Applies Runtime Context Cache strategy.";
+  for (const Node* n : graph->Nodes()) {
+    if (n->IsOp()) {
+      n->Op()->SetAttr(kEnableCacheRuntimeContext, true);
+    }
+  }
+  return graph;
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(runtime_context_cache_pass,
+              paddle::framework::ir::RuntimeContextCachePass);

paddle/fluid/framework/scope.cc

@@ -34,7 +34,7 @@ DEFINE_double(
     "Memory size threshold (GB) when the garbage collector clear tensors."
     "Disabled when this value is less than 0");
 
-DEFINE_bool(fast_eager_deletion_mode, false,
+DEFINE_bool(fast_eager_deletion_mode, true,
             "Fast eager deletion mode. If enabled, memory would release "
             "immediately without waiting GPU kernel ends.");