merge develop

.pre-commit-config.yaml

@@ -42,12 +42,6 @@ repos:
         entry: bash ./tools/codestyle/pylint_pre_commit.hook
         language: system
         files: \.(py)$
--   repo: https://github.com/PaddlePaddle/pre-commit-golang
-    sha: 8337620115c25ff8333f1b1a493bd031049bd7c0
-    hooks:
-    -   id: go-fmt
-        types:
-        - go
 -   repo: local
     hooks:
     -   id: copyright_checker

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
 

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

@@ -39,8 +39,10 @@ IF(WIN32)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5md.lib)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/mklml.dll)
     SET(MKLML_SHARED_IOMP_LIB     ${MKLML_LIB_DIR}/libiomp5md.dll)
-ELSE()  
-    SET(MKLML_VER "mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
+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_LIB                 ${MKLML_LIB_DIR}/libmklml_intel.so)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5.so)

cmake/external/ngraph.cmake

@@ -37,7 +37,7 @@ INCLUDE(GNUInstallDirs)
 INCLUDE(ExternalProject)
 
 SET(NGRAPH_PROJECT         "extern_ngraph")
-SET(NGRAPH_GIT_TAG         "20bd8bbc79ae3a81c57313846a2be7313e5d1dab")
+SET(NGRAPH_GIT_TAG         "a444f7a959b7d87f2c117c9b57a4c387759e481e")
 SET(NGRAPH_SOURCES_DIR     ${THIRD_PARTY_PATH}/ngraph)
 SET(NGRAPH_INSTALL_DIR     ${THIRD_PARTY_PATH}/install/ngraph)
 SET(NGRAPH_INC_DIR         ${NGRAPH_INSTALL_DIR}/include)
@@ -69,7 +69,7 @@ ExternalProject_Add(
     CMAKE_ARGS          -DNGRAPH_DEX_ONLY=TRUE
     CMAKE_ARGS          -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
     CMAKE_ARGS          -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
-    CMAKE_ARGS          -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/lib
+    CMAKE_ARGS          -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}
     CMAKE_ARGS          -DMKLML_LIB_DIR=${MKLML_INSTALL_DIR}/lib
 )
 

cmake/operators.cmake

@@ -153,7 +153,11 @@ function(op_library TARGET)
     # pybind USE_OP_DEVICE_KERNEL for CUDNN
     list(LENGTH cudnn_cu_cc_srcs cudnn_cu_cc_srcs_len)
     if (WITH_GPU AND ${cudnn_cu_cc_srcs_len} GREATER 0)
+      if(${TARGET} STREQUAL "activation")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, CUDNN);\n")
+      else()
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
+      endif()
     endif()
 
     # pybind USE_OP_DEVICE_KERNEL for MIOPEN
@@ -168,6 +172,9 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
       elseif(${MKLDNN_FILE} STREQUAL "conv_mkldnn_op")
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, FP32);\n")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, S8);\n")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, U8);\n")
+        
       else()
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
       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/block_desc.cc

@@ -13,7 +13,11 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/block_desc.h"
+
 #include <queue>
+#include <unordered_set>
+#include <utility>
+
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/program_desc.h"
 
@@ -155,6 +159,16 @@ void BlockDesc::RemoveOp(size_t s, size_t e) {
   ops_.erase(ops_.begin() + s, ops_.begin() + e);
 }
 
+void BlockDesc::RemoveOpInternal(const OpDesc *op_desc) {
+  // TODO(minqiyang): make this faster
+  for (auto it = ops_.begin(); it != ops_.end(); ++it) {
+    if (it->get() == op_desc) {
+      ops_.erase(it);
+      break;
+    }
+  }
+}
+
 std::vector<OpDesc *> BlockDesc::AllOps() const {
   std::vector<OpDesc *> res;
   for (const auto &op : ops_) {
@@ -163,20 +177,6 @@ std::vector<OpDesc *> BlockDesc::AllOps() const {
   return res;
 }
 
-void BlockDesc::Clear() {
-  // clear all ops
-  ops_.clear();
-
-  // clear all vars which are not persistable
-  for (auto it = vars_.begin(); it != vars_.end();) {
-    if (it->second->Persistable()) {
-      ++it;
-    } else {
-      vars_.erase(it++);
-    }
-  }
-}
-
 void BlockDesc::Flush() {
   for (auto &op_desc : ops_) {
     op_desc->Flush();

paddle/fluid/framework/block_desc.h

@@ -93,12 +93,12 @@ class BlockDesc {
    */
   void RemoveOp(size_t s, size_t e);
 
+  void RemoveOpInternal(const OpDesc *op_desc);
+
   void RemoveVar(const std::string &name) { vars_.erase(name); }
 
   std::vector<OpDesc *> AllOps() const;
 
-  void Clear();
-
   size_t OpSize() const { return ops_.size(); }
 
   OpDesc *Op(int idx) const { return ops_.at(idx).get(); }

paddle/fluid/framework/details/build_strategy.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <vector>
 
@@ -76,11 +77,11 @@ struct BuildStrategy {
 
   bool fuse_relu_depthwise_conv_{false};
 
-  bool memory_optimize_{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/fast_threaded_ssa_graph_executor.cc

@@ -12,7 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 #include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <vector>
 #include "paddle/fluid/framework/details/fetch_op_handle.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
@@ -55,7 +57,7 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
   std::vector<FetchOpHandle *> fetch_ops;
 
   for (auto &fetch_var_name : fetch_tensors) {
-    for (auto &var_map : graph_->Get<details::GraphVars>("vars")) {
+    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());

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.
@@ -348,10 +354,6 @@ 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/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/executor.cc

@@ -20,6 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/reader.h"
+#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/distributed/distributed.h"

paddle/fluid/framework/ir/CMakeLists.txt

@@ -105,4 +105,5 @@ if (WITH_MKLDNN)
     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)
     cc_test(test_conv_relu_mkldnn_fuse_pass SRCS mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
     cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
+    cc_test(test_mkldnn_placement_pass SRCS mkldnn/mkldnn_placement_pass_tester.cc DEPS mkldnn_placement_pass)
 endif ()

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_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/mkldnn/mkldnn_placement_pass.cc

@@ -21,7 +21,7 @@ namespace ir {
 
 std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
-  VLOG(3) << "Aplies MKL-DNN placement strategy.";
+  VLOG(3) << "Applies MKL-DNN placement strategy.";
   const auto& op_types_list =
       Get<std::unordered_set<std::string>>("mkldnn_enabled_op_types");
   for (const Node* n : graph->Nodes()) {

paddle/fluid/framework/ir/mkldnn/mkldnn_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/mkldnn/mkldnn_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_mkldnn) {
+  auto* op = prog->MutableBlock(0)->AppendOp();
+
+  op->SetType(type);
+
+  if (!boost::indeterminate(use_mkldnn)) op->SetAttr("use_mkldnn", use_mkldnn);
+
+  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_mkldnn
+// ---------------------------------------
+// (a,b)->concat->c              none
+// (c,weights,bias)->conv->f     none
+// f->relu->g                    false
+// g->pool->h                    false
+// (h,weights2,bias2)->conv->k   true
+// k->relu->l                    true
+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", std::vector<std::string>({"a", "b"}),
+        std::vector<std::string>({"c"}), boost::indeterminate);
+  SetOp(&prog, "conv2d", "conv1",
+        std::vector<std::string>({"c", "weights", "bias"}),
+        std::vector<std::string>({"f"}), boost::indeterminate);
+  SetOp(&prog, "relu", "relu1", std::vector<std::string>({"f"}),
+        std::vector<std::string>({"g"}), false);
+  SetOp(&prog, "pool2d", "pool1", std::vector<std::string>({"g"}),
+        std::vector<std::string>({"h"}), false);
+  SetOp(&prog, "conv2d", "conv2",
+        std::vector<std::string>({"h", "weights2", "bias2"}),
+        std::vector<std::string>({"k"}), true);
+  SetOp(&prog, "relu", "relu2", std::vector<std::string>({"k"}),
+        std::vector<std::string>({"l"}), true);
+
+  return prog;
+}
+
+void MainTest(std::initializer_list<std::string> mkldnn_enabled_op_types,
+              unsigned expected_use_mkldnn_true_count) {
+  auto prog = BuildProgramDesc();
+
+  std::unique_ptr<ir::Graph> graph(new ir::Graph(prog));
+
+  auto pass = PassRegistry::Instance().Get("mkldnn_placement_pass");
+  pass->Set("mkldnn_enabled_op_types",
+            new std::unordered_set<std::string>(mkldnn_enabled_op_types));
+
+  graph = pass->Apply(std::move(graph));
+
+  unsigned use_mkldnn_true_count = 0;
+
+  for (auto* node : graph->Nodes()) {
+    if (node->IsOp()) {
+      auto* op = node->Op();
+      if (op->HasAttr("use_mkldnn") &&
+          boost::get<bool>(op->GetAttr("use_mkldnn"))) {
+        ++use_mkldnn_true_count;
+      }
+    }
+  }
+
+  EXPECT_EQ(use_mkldnn_true_count, expected_use_mkldnn_true_count);
+}
+
+TEST(MKLDNNPlacementPass, enable_conv_relu) {
+  // 1 conv (1 conv is always true) + 2 relu (1 relu is always true) + 0 pool
+  MainTest({"conv2d", "relu"}, 3);
+}
+
+TEST(MKLDNNPlacementPass, enable_relu_pool) {
+  // 1 conv (1 conv is always true) + 2 relu (1 relu is always true) + 1 pool
+  MainTest({"relu", "pool2d"}, 4);
+}
+
+TEST(MKLDNNPlacementPass, enable_all) {
+  // 1 conv (1 conv is always true) + 2 relu (1 relu is always true) + 1 pool
+  MainTest({}, 4);
+}
+
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
+USE_PASS(mkldnn_placement_pass);

paddle/fluid/framework/op_registry.h

@@ -290,7 +290,7 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
       "USE_OP_DEVICE_KERNEL must be in global namespace");                 \
   extern int                                                               \
       TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE##_##customized_name(); \
-  UNUSED static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_##DEFAULT_TYPE##_ = /* NOLINT */ \
+  UNUSED static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_##customized_name##_ = /* NOLINT */ \
       TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE##_##customized_name()
 
 #define USE_OP_DEVICE_KERNEL(op_type, LIBRARY_TYPE) \

paddle/fluid/framework/operator.cc

@@ -467,12 +467,6 @@ const Variable* ExecutionContext::InputVar(const std::string& name) const {
   return it->second.empty() ? nullptr : it->second[0];
 }
 
-const Variable* ExecutionContext::LegacyInputVar(
-    const std::string& name) const {
-  auto ipt = op_.Input(name);
-  return ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
-}
-
 Variable* ExecutionContext::OutputVar(const std::string& name) const {
   auto it = ctx_.outputs.find(name);
   if (it == ctx_.outputs.end()) return nullptr;
@@ -483,22 +477,11 @@ Variable* ExecutionContext::OutputVar(const std::string& name) const {
   return it->second.empty() ? nullptr : it->second[0];
 }
 
-Variable* ExecutionContext::LegacyOutputVar(const std::string& name) const {
-  auto opt = op_.Output(name);
-  return opt == kEmptyVarName ? nullptr : scope_.FindVar(opt);
-}
-
 template <>
 const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
   return Input<LoDTensor>(name);
 }
 
-template <>
-const Tensor* ExecutionContext::LegacyInput<Tensor>(
-    const std::string& name) const {
-  return LegacyInput<LoDTensor>(name);
-}
-
 template <>
 const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
     const std::string& name) const {
@@ -521,35 +504,11 @@ const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
   return res;
 }
 
-template <>
-const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
-    const std::string& name) const {
-  auto names = op().Inputs(name);
-  std::vector<const Tensor*> res;
-  res.reserve(names.size());
-  std::transform(names.begin(), names.end(), std::back_inserter(res),
-                 [&](const std::string& sub_name) -> const Tensor* {
-                   auto var = scope_.FindVar(sub_name);
-                   if (var == nullptr) return nullptr;
-                   PADDLE_ENFORCE(
-                       var->IsType<LoDTensor>(),
-                       "%s should be LoDTensor, but the received type is %s",
-                       sub_name, ToTypeName(var->Type()));
-                   return &(var->Get<LoDTensor>());
-                 });
-  return res;
-}
-
 template <>
 Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
   return Output<LoDTensor>(name);
 }
 
-template <>
-Tensor* ExecutionContext::LegacyOutput<Tensor>(const std::string& name) const {
-  return LegacyOutput<LoDTensor>(name);
-}
-
 template <>
 std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
     const std::string& name) const {
@@ -882,7 +841,8 @@ class RuntimeInferShapeContext : public InferShapeContext {
   const RuntimeContext& ctx_;
 };
 
-static void CheckTensorNANOrInf(const std::string& name,
+static void CheckTensorNANOrInf(const std::string& op_type,
+                                const std::string& name,
                                 const framework::Tensor& tensor) {
   if (tensor.memory_size() == 0) {
     return;
@@ -892,9 +852,9 @@ static void CheckTensorNANOrInf(const std::string& name,
     return;
   }
   PADDLE_ENFORCE(!framework::TensorContainsInf(tensor),
-                 "Tensor %s contains Inf", name);
+                 "Operator %s output Tensor %s contains Inf", op_type, name);
   PADDLE_ENFORCE(!framework::TensorContainsNAN(tensor),
-                 "Tensor %s contains NAN", name);
+                 "Operator %s output Tensor %s contains NAN", op_type, name);
 }
 
 void OperatorWithKernel::RuntimeInferShape(const Scope& scope,
@@ -988,9 +948,10 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
       auto* var = exec_scope.FindVar(vname);
       if (var == nullptr) continue;
       if (var->IsType<framework::LoDTensor>()) {
-        CheckTensorNANOrInf(vname, var->Get<framework::LoDTensor>());
+        CheckTensorNANOrInf(type_, vname, var->Get<framework::LoDTensor>());
       } else if (var->IsType<framework::SelectedRows>()) {
-        CheckTensorNANOrInf(vname, var->Get<framework::SelectedRows>().value());
+        CheckTensorNANOrInf(type_, vname,
+                            var->Get<framework::SelectedRows>().value());
       }
     }
   }

paddle/fluid/framework/operator.h

@@ -16,9 +16,11 @@ limitations under the License. */
 
 #include <algorithm>
 #include <atomic>
+#include <memory>
 #include <string>
 #include <tuple>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 
 #include "glog/logging.h"  // For VLOG
@@ -253,31 +255,6 @@ class ExecutionContext {
     return it->second;
   }
 
-  const std::vector<Variable*> LegacyMultiInputVar(
-      const std::string& name) const {
-    auto names = op_.Inputs(name);
-    std::vector<Variable*> res;
-    res.reserve(names.size());
-    std::transform(names.begin(), names.end(), std::back_inserter(res),
-                   [this](const std::string& name) {
-                     return name == kEmptyVarName ? nullptr
-                                                  : scope_.FindVar(name);
-                   });
-    return res;
-  }
-
-  std::vector<Variable*> LegacyMultiOutputVar(const std::string& name) const {
-    auto names = op_.Outputs(name);
-    std::vector<Variable*> res;
-    res.reserve(names.size());
-    std::transform(names.begin(), names.end(), std::back_inserter(res),
-                   [this](const std::string& name) {
-                     return name == kEmptyVarName ? nullptr
-                                                  : scope_.FindVar(name);
-                   });
-    return res;
-  }
-
   template <typename T>
   const T* Input(const std::string& name) const {
     auto* var = InputVar(name);
@@ -290,22 +267,6 @@ class ExecutionContext {
     return var == nullptr ? nullptr : var->GetMutable<T>();
   }
 
-  template <typename T>
-  const T* LegacyInput(const std::string& name) const {
-    auto* var = LegacyInputVar(name);
-    return var == nullptr ? nullptr : &var->Get<T>();
-  }
-
-  template <typename T>
-  T* LegacyOutput(const std::string& name) const {
-    auto var = LegacyOutputVar(name);
-    return var == nullptr ? nullptr : var->GetMutable<T>();
-  }
-
-  const Variable* LegacyInputVar(const std::string& name) const;
-
-  Variable* LegacyOutputVar(const std::string& name) const;
-
   template <typename T>
   const std::vector<const T*> MultiInput(const std::string& name) const {
     auto it = ctx_.inputs.find(name);
@@ -338,32 +299,6 @@ class ExecutionContext {
     return res;
   }
 
-  template <typename T>
-  const std::vector<const T*> LegacyMultiInput(const std::string& name) const {
-    auto names = op_.Inputs(name);
-    std::vector<const T*> res;
-    res.reserve(names.size());
-    std::transform(names.begin(), names.end(), std::back_inserter(res),
-                   [&](const std::string& sub_name) -> const T* {
-                     auto var = scope_.FindVar(sub_name);
-                     return var == nullptr ? nullptr : &var->Get<T>();
-                   });
-    return res;
-  }
-
-  template <typename T>
-  std::vector<T*> LegacyMultiOutput(const std::string& name) const {
-    auto names = op_.Outputs(name);
-    std::vector<T*> res;
-    res.reserve(names.size());
-    std::transform(names.begin(), names.end(), std::back_inserter(res),
-                   [&](const std::string& sub_name) -> T* {
-                     auto var = scope_.FindVar(sub_name);
-                     return var == nullptr ? nullptr : var->GetMutable<T>();
-                   });
-    return res;
-  }
-
   platform::Place GetPlace() const { return device_context_.GetPlace(); }
 
   template <typename DeviceContextType>
@@ -436,24 +371,13 @@ class ExecutionContext {
 template <>
 const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const;
 
-template <>
-const Tensor* ExecutionContext::LegacyInput<Tensor>(
-    const std::string& name) const;
-
 template <>
 const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
     const std::string& name) const;
 
-template <>
-const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
-    const std::string& name) const;
-
 template <>
 Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
 
-template <>
-Tensor* ExecutionContext::LegacyOutput<Tensor>(const std::string& name) const;
-
 template <>
 std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
     const std::string& name) const;

paddle/fluid/framework/python_headers.h

@@ -24,3 +24,11 @@ limitations under the License. */
 
 #pragma pop_macro("_XOPEN_SOURCE")
 #pragma pop_macro("_POSIX_C_SOURCE")
+
+#if !defined(PYBIND11_HIDDEN)
+#ifdef _WIN32
+#define PYBIND11_HIDDEN __declspec(dllexport)
+#else
+#define PYBIND11_HIDDEN __attribute__((visibility("hidden")))
+#endif
+#endif

paddle/fluid/framework/tensor_util.cc

@@ -14,6 +14,8 @@
 #include "paddle/fluid/framework/tensor_util.h"
 #include <algorithm>
 #include <limits>
+#include <memory>
+#include <utility>
 #include <vector>
 #include "paddle/fluid/framework/data_type.h"
 

paddle/fluid/imperative/layer.cc

@@ -18,6 +18,7 @@
 #include <limits>
 #include <map>
 #include <random>
+#include <unordered_set>
 #include <utility>
 
 #include "paddle/fluid/framework/lod_tensor.h"
@@ -139,6 +140,8 @@ class Autograd {
           }
         }
       }
+
+      ready_op->InvokeBackwardHooks();
     }
   }
 
@@ -156,8 +159,10 @@ class Autograd {
       for (auto it : candidate->pre_ops_) {
         for (OpBase* pre_op : it.second) {
           if (!pre_op) continue;
-          VLOG(5) << "op dep " << candidate->op_desc_->Type() << " <---- "
-                  << it.first << " <---- " << pre_op->op_desc_->Type();
+          VLOG(5) << "op dep " << candidate->op_desc_->Type() << " trace id "
+                  << candidate->trace_id_ << " <---- " << it.first << " <---- "
+                  << pre_op->op_desc_->Type() << " trace id "
+                  << pre_op->trace_id_;
           if (visited.find(pre_op) == visited.end()) {
             visited.insert(pre_op);
             queue.push_back(pre_op);
@@ -211,6 +216,7 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
     return {};
   }
 
+  VLOG(3) << "apply op grad: " << op_desc_->Type();
   std::vector<framework::VariableValueMap> grad_outputs;
   if (backward_id_ > 0) {
     VLOG(3) << "py_layer_grad";
@@ -272,6 +278,22 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
   return input_vars_;
 }
 
+void OpBase::InvokeBackwardHooks() {
+  VLOG(3) << "call backward hooks, hooks num: " << backward_hooks_.size();
+
+  // call backward hooks
+  for (py::object& callable : backward_hooks_) {
+    callable(this);
+  }
+}
+
+void OpBase::RegisterBackwardHooks(const py::object& callable) {
+  VLOG(3) << "Register backward hooks " << trace_id_;
+
+  // TODO(minqiyang): check the callable format
+  backward_hooks_.push_back(callable);
+}
+
 void VarBase::RunBackward() {
   if (!pre_op_) return;
 

paddle/fluid/imperative/layer.h

@@ -123,22 +123,32 @@ class VarBase {
 
  private:
   VarBase(framework::Variable* var, VarBase* grad, bool stop_gradient)
-      : var_desc_(nullptr),
+      : name_(),
+        var_desc_(nullptr),
         var_(var),
         grads_(grad),
+        block_(nullptr),
+        persistable_(false),
         stop_gradient_(stop_gradient),
         pre_op_(nullptr),
+        pre_op_out_name_(),
         pre_op_out_idx_(-1) {}
 
  public:
   virtual ~VarBase() {
+    // TODO(minqiyang): remove var desc from block desc
     if (var_) {
       delete var_;
+      var_ = nullptr;
     }
 
     if (grads_) {
       delete grads_;
+      grads_ = nullptr;
     }
+
+    pre_op_ = nullptr;
+    pre_op_out_idx_ = -1;
   }
 
   inline OpBase* PreOp() const { return pre_op_; }
@@ -151,6 +161,14 @@ class VarBase {
 
   void RunBackward();
 
+  inline void ResetPreOp(OpBase* op) {
+    if (op == pre_op_) {
+      // clear pre_op info when op equals to var's pre_op
+      pre_op_ = nullptr;
+      pre_op_out_idx_ = -1;
+    }
+  }
+
   void TrackPreOp(OpBase* pre_op, const std::string& pre_op_out_name,
                   int pre_op_out_idx, bool pre_op_stop_gradient) {
     pre_op_ = pre_op;
@@ -184,11 +202,15 @@ class VarBase {
     return string::Sprintf("%s@IGrad", var_desc_->Name());
   }
 
+  std::string name_;
   framework::VarDesc* var_desc_;
 
   framework::Variable* var_;
   VarBase* grads_;
 
+  framework::BlockDesc* block_;
+  bool persistable_;
+
  private:
   bool stop_gradient_;
   OpBase* pre_op_;
@@ -199,15 +221,27 @@ class VarBase {
 /* The wrapper for OpDesc which holds a OpDesc and a OpDesc of its
  * gradient. This object should be managed totally by Python intepreter.
  */
-class OpBase {
+class PYBIND11_HIDDEN OpBase {
  public:
   OpBase()
       : op_desc_(nullptr),
         forward_id_(-1),
         backward_id_(-1),
-        place_(platform::CPUPlace()) {}
+        trace_id_(-1),
+        place_(platform::CPUPlace()),
+        backward_hooks_() {}
 
   virtual ~OpBase() {
+    // TODO(minqiyang): remove op_desc from block_desc in tracer
+    //
+    // reset all output vars' pre op
+    for (auto iter : output_vars_) {
+      for (VarBase* var : iter.second) {
+        var->ResetPreOp(this);
+      }
+    }
+
+    // release resource
     for (framework::OpDesc* desc : grad_op_descs_) {
       delete desc;
     }
@@ -215,6 +249,10 @@ class OpBase {
 
   std::map<std::string, std::vector<VarBase*>> ApplyGrad();
 
+  void RegisterBackwardHooks(const py::object& callable);
+
+  void InvokeBackwardHooks();
+
   // One of `op_desc_` or `forward_id_` is set, not both.
   // For pure python PyLayer, use `forward_id_`, otherwise, use op_desc_.
   framework::OpDesc* op_desc_;
@@ -225,6 +263,7 @@ class OpBase {
   // Note: each fwd op corresponds to a vector of bwd ops.
   std::vector<framework::OpDesc*> grad_op_descs_;
   int backward_id_;
+  int trace_id_;
 
   platform::Place place_;
 
@@ -239,6 +278,8 @@ class OpBase {
   std::vector<framework::VariableValueMap> grad_output_vars_;
 
   framework::BlockDesc* block_;
+
+  std::vector<py::object> backward_hooks_;
 };
 
 class Layer {

paddle/fluid/imperative/tracer.cc

@@ -14,15 +14,32 @@
 
 #include "paddle/fluid/imperative/tracer.h"
 
+#include <memory>
 #include <set>
+#include <unordered_map>
+#include <unordered_set>
 
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
 
+#ifdef WITH_GPERFTOOLS
+#include "gperftools/profiler.h"
+#endif
+
+DEFINE_string(
+    tracer_profile_fname, "",
+    "Profiler filename for imperative tracer, which generated by gperftools."
+    "Only valid when compiled `WITH_PROFILER=ON`. Empty if disable.");
+
 namespace paddle {
 namespace imperative {
 
+static std::once_flag gTracerProfileOnce;
+#ifdef WITH_GPERFTOOLS
+static bool gTracerProfilerStarted = false;
+#endif
+
 void CreateGradOp(const framework::OpDesc& op_desc,
                   const std::unordered_set<std::string>& no_grad_set,
                   const std::vector<framework::BlockDesc*>& grad_sub_block,
@@ -68,15 +85,36 @@ platform::Place GetExpectedPlace(platform::Place place, VarBasePtrMap inputs) {
   return result;
 }
 
+Tracer::Tracer(framework::BlockDesc* root_block) : root_block_(root_block) {
+  if (!FLAGS_tracer_profile_fname.empty()) {
+    std::call_once(gTracerProfileOnce, [] {
+#ifdef WITH_GPERFTOOLS
+      ProfilerStart(FLAGS_tracer_profile_fname.c_str());
+      gTracerProfilerStarted = true;
+#else
+      LOG(WARNING) << "Paddle is not compiled with gperftools. "
+                      "FLAGS_tracer_profile_fname will be ignored";
+#endif
+    });
+  }
+}
+
 std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
                                     const VarBasePtrMap& outputs,
                                     framework::BlockDesc* block,
                                     const platform::Place expected_place,
                                     const bool stop_gradient) {
+#ifdef WITH_GPERFTOOLS
+  if (gTracerProfilerStarted) {
+    ProfilerFlush();
+  }
+#endif
+
   std::map<std::string, VarBase*> vars;
 
   framework::OpDesc* op_desc = op->op_desc_;
-  VLOG(3) << "tracer tracing " << op_desc->Type();
+  VLOG(3) << "tracer tracing " << op_desc->Type() << " trace id "
+          << op->trace_id_;
   op_desc->InferShape(*block);
   op_desc->InferVarType(block);
 
@@ -99,11 +137,13 @@ std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
       if (inp->PreOp() && !inp->IsStopGradient()) {
         op->pre_ops_[it.first].push_back(inp->PreOp());
         op->pre_ops_out_idx_[it.first].push_back(inp->PreOpOutIdx());
+        VLOG(3) << "add pre op " << inp->PreOp()->op_desc_->Type();
       } else {
         op->pre_ops_[it.first].push_back(nullptr);
       }
       VLOG(3) << "input vname " << inp->var_desc_->Name() << " "
-              << inp->var_->IsInitialized();
+              << inp->var_->IsInitialized() << " stop_gradient "
+              << inp->IsStopGradient();
     }
   }
 
@@ -155,6 +195,7 @@ std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
 
     op->grad_input_vars_.resize(op->grad_op_descs_.size());
     op->grad_output_vars_.resize(op->grad_op_descs_.size());
+
     for (size_t i = 0; i < op->grad_op_descs_.size(); ++i) {
       framework::OpDesc* grad_op_desc = op->grad_op_descs_[i];
       for (auto it : grad_op_desc->Inputs()) {
@@ -167,7 +208,6 @@ std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
             PADDLE_ENFORCE(fwd_var_it != vars.end());
             // Forward inputs or outputs.
             grad_in_vars.push_back(fwd_var_it->second->var_);
-            vars_saved_for_backward.insert(it.first);
           } else {
             VarBase* var = vars[var_it->second];
             if (!var->grads_->var_->IsInitialized()) {
@@ -177,6 +217,8 @@ std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
             // Douts.
             grad_in_vars.push_back(var->grads_->var_);
           }
+
+          vars_saved_for_backward.insert(it.first);
         }
       }
 

paddle/fluid/imperative/tracer.h

@@ -40,7 +40,7 @@ platform::Place GetExpectedPlace(platform::Place place, VarBasePtrMap inputs);
 
 class Tracer {
  public:
-  explicit Tracer(framework::BlockDesc* root_block) : root_block_(root_block) {}
+  explicit Tracer(framework::BlockDesc* root_block);
 
   virtual ~Tracer() {}
 

paddle/fluid/inference/CMakeLists.txt

@@ -16,6 +16,7 @@ add_subdirectory(utils)
 if (TENSORRT_FOUND)
   add_subdirectory(tensorrt)
 endif()
+# add_subdirectory(anakin)
 
 get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
 get_property(cuda_modules GLOBAL PROPERTY CUDA_MODULES)

paddle/fluid/inference/anakin/CMakeLists.txt

+cc_library(anakin_engine SRCS engine.cc)
+target_link_libraries(anakin_engine anakin anakin_saber_common)
+cc_test(test_anakin_engine SRCS test_anakin_engine.cc DEPS anakin_engine)
+add_subdirectory(convert)

paddle/fluid/inference/anakin/convert/CMakeLists.txt

+cc_library(anakin_op_converter SRCS fc.cc registrar.cc DEPS anakin_engine framework_proto scope)
+cc_test(test_anakin_fc SRCS test_fc_op.cc DEPS anakin_op_converter mul_op)

paddle/fluid/inference/anakin/convert/fc.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/inference/anakin/convert/fc.h"
+#include <algorithm>
+
+using anakin::graph::GraphGlobalMem;
+using anakin::AK_FLOAT;
+using anakin::Precision;
+using anakin::saber::NV;
+using anakin::saber::X86;
+using anakin::saber::Shape;
+using anakin::PBlock;
+using anakin::PTuple;
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+void FcOpConverter::operator()(const framework::proto::OpDesc &op,
+                               const framework::Scope &scope, bool test_mode) {
+  framework::OpDesc op_desc(op, nullptr);
+  PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
+  PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1);
+  PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);
+
+  auto x_name = op_desc.Input("X").front();
+  auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
+  auto *y_v = scope.FindVar(op_desc.Input("Y").front());
+  PADDLE_ENFORCE_NOT_NULL(y_v);
+  auto *y_t = y_v->GetMutable<framework::LoDTensor>();
+
+  auto input_name = op_desc.Input("X").front();
+  auto output_name = op_desc.Output("Out").front();
+
+  auto weight_shape = framework::vectorize2int(y_t->dims());
+  engine_->AddOp(op_name, "Dense", {input_name}, {output_name});
+  engine_->AddOpAttr(op_name, "bias_term", false);
+  engine_->AddOpAttr(op_name, "axis", 1);
+  int out_dim = weight_shape[1];
+  engine_->AddOpAttr(op_name, "out_dim", out_dim);
+
+  weight_shape.push_back(1);
+  weight_shape.push_back(1);
+  Shape anakin_shape(weight_shape);
+
+  framework::LoDTensor weight_tensor;
+  weight_tensor.Resize(y_t->dims());
+  TensorCopySync((*y_t), platform::CPUPlace(), &weight_tensor);
+
+  auto *weight1 =
+      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(anakin_shape);
+  float *cpu_data = static_cast<float *>(weight1->h_tensor().mutable_data());
+  std::copy_n(weight_tensor.data<float>(), weight_tensor.numel(), cpu_data);
+  weight1->d_tensor().set_shape(anakin_shape);
+  weight1->d_tensor().copy_from(weight1->h_tensor());
+  engine_->AddOpAttr(op_name, "weight_1", *weight1);
+}
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/anakin/convert/fc.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/fluid/inference/anakin/convert/op_converter.h"
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+class FcOpConverter : public AnakinOpConverter {
+ public:
+  FcOpConverter() = default;
+
+  virtual void operator()(const framework::proto::OpDesc &op,
+                          const framework::Scope &scope,
+                          bool test_mode) override;
+  virtual ~FcOpConverter() {}
+
+ private:
+};
+
+static Registrar<FcOpConverter> register_fc_op_converter("fc");
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/anakin/convert/op_converter.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include "framework/core/types.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/inference/anakin/convert/registrar.h"
+#include "paddle/fluid/inference/anakin/engine.h"
+#include "paddle/fluid/inference/utils/singleton.h"
+#include "saber/saber_types.h"
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+using AnakinNvEngine =
+    AnakinEngine<::anakin::saber::NV, ::anakin::Precision::FP32>;
+
+class AnakinOpConverter {
+ public:
+  AnakinOpConverter() = default;
+
+  virtual void operator()(const framework::proto::OpDesc &op,
+                          const framework::Scope &scope, bool test_mode) {}
+  void ConvertOp(const framework::proto::OpDesc &op,
+                 const std::unordered_set<std::string> &parameters,
+                 const framework::Scope &scope, AnakinNvEngine *engine,
+                 bool test_mode = false) {
+    framework::OpDesc op_desc(op, nullptr);
+    std::string op_type = op_desc.Type();
+    std::shared_ptr<AnakinOpConverter> it{nullptr};
+
+    if (op_type == "mul") {
+      PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1UL);
+      std::string Y = op_desc.Input("Y")[0];
+      std::cout << Y << parameters.count(Y) << std::endl;
+      if (parameters.count(Y)) {
+        it = OpRegister::instance()->Get("fc");
+      }
+    }
+
+    if (!it) {
+      it = OpRegister::instance()->Get(op_type);
+    }
+    PADDLE_ENFORCE_NOT_NULL(it, "no OpConverter for optype [%s]", op_type);
+    it->SetEngine(engine);
+    (*it)(op, scope, test_mode);
+  }
+
+  void ConvertBlock(const framework::proto::BlockDesc &block,
+                    const std::unordered_set<std::string> &parameters,
+                    const framework::Scope &scope, AnakinNvEngine *engine) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    for (auto i = 0; i < block.ops_size(); i++) {
+      auto &op = block.ops(i);
+      ConvertOp(op, parameters, scope, engine);
+    }
+  }
+  void SetEngine(AnakinNvEngine *engine) { engine_ = engine; }
+  virtual ~AnakinOpConverter() {}
+
+ protected:
+  bool test_mode_;
+  AnakinNvEngine *engine_{nullptr};
+
+ private:
+  std::unordered_map<std::string, AnakinOpConverter *> converters_;
+  framework::Scope *scope_{nullptr};
+  std::mutex mutex_;
+};
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle
+
+#define REGISTER_ANAKIN_OP_CONVERTER(op_type__, Converter__)                \
+  struct anakin_##op_type__##_converter                                     \
+      : public ::paddle::framework::Registrar {                             \
+    anakin_##op_type__##_converter() {                                      \
+      ::paddle::inference::                                                 \
+          Registry<paddle::inference::anakin::AnakinOpConverter>::Register< \
+              ::paddle::inference::anakin::Converter__>(#op_type__);        \
+    }                                                                       \
+  };                                                                        \
+  anakin_##op_type__##_converter anakin_##op_type__##_converter__;          \
+  int TouchConverterRegister_anakin_##op_type__() {                         \
+    anakin_##op_type__##_converter__.Touch();                               \
+    return 0;                                                               \
+  }
+
+#define USE_ANAKIN_CONVERTER(op_type__)                                    \
+  extern int TouchConverterRegister_anakin_##op_type__();                  \
+  static int use_op_converter_anakin_##op_type__ __attribute__((unused)) = \
+      TouchConverterRegister_anakin_##op_type__();

paddle/fluid/inference/anakin/convert/registrar.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/inference/anakin/convert/registrar.h"
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+std::shared_ptr<AnakinOpConverter> OpRegister::Get(const std::string &name) {
+  auto it = registry_.find(name);
+  if (it == registry_.end()) return nullptr;
+  return it->second();
+}
+
+OpRegister *OpRegister::instance() {
+  static OpRegister factory;
+  return &factory;
+}
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/anakin/convert/registrar.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <functional>
+#include <map>
+#include <memory>
+#include <string>
+#include <utility>
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+class AnakinOpConverter;
+
+class OpRegister {
+ public:
+  OpRegister() = default;
+  std::shared_ptr<AnakinOpConverter> Get(const std::string &name);
+  static OpRegister *instance();
+  void OpRegisterFn(const std::string &name,
+                    std::function<std::shared_ptr<AnakinOpConverter>()> fn) {
+    registry_[name] = fn;
+  }
+
+ private:
+  using RegisterFnType = std::function<std::shared_ptr<AnakinOpConverter>()>;
+  std::map<std::string, std::function<std::shared_ptr<AnakinOpConverter>()>>
+      registry_;
+};
+
+template <typename T, typename... Args>
+class Registrar {
+ public:
+  Registrar(const std::string &name, Args... args) {
+    std::shared_ptr<AnakinOpConverter> converter =
+        std::make_shared<T>(std::move(args)...);
+    OpRegister::instance()->OpRegisterFn(name,
+                                         [converter]() { return converter; });
+  }
+};
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

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

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "paddle/fluid/inference/anakin/convert/fc.h"
+#include "paddle/fluid/inference/anakin/convert/op_converter.h"
+#include "paddle/fluid/inference/anakin/convert/ut_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+TEST(fc_op, test) {
+  auto fc_converter = OpRegister::instance()->Get("fc");
+  ASSERT_TRUE(fc_converter != nullptr);
+  // Registrar<FcOpConverter> register_fc("fc");
+  // auto fc = std::make_shared<FcOpConverter>();
+
+  std::unordered_set<std::string> parameters({"mul_y"});
+  framework::Scope scope;
+  AnakinConvertValidation validator(parameters, scope);
+  validator.DeclInputVar("mul_x", {1, 1, 1, 1});
+  validator.DeclParamVar("mul_y", {1, 2});
+  validator.DeclOutputVar("mul_out", {1, 1, 1, 2});
+
+  // Prepare Op description
+  framework::OpDesc desc;
+  desc.SetType("mul");
+  desc.SetInput("X", {"mul_x"});
+  desc.SetInput("Y", {"mul_y"});
+  desc.SetOutput("Out", {"mul_out"});
+  int num_flatten_dims = 3;
+  desc.SetAttr("x_num_col_dims", num_flatten_dims);
+  validator.SetOp(*desc.Proto());
+
+  validator.Execute(10);
+}
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle
+
+USE_OP(mul);

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

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <map>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/inference/anakin/engine.h"
+#include "paddle/fluid/inference/analysis/helper.h"
+#include "paddle/fluid/inference/utils/singleton.h"
+#include "paddle/fluid/platform/enforce.h"
+
+using anakin::graph::GraphGlobalMem;
+using anakin::AK_FLOAT;
+using anakin::Precision;
+using anakin::saber::NV;
+using anakin::saber::X86;
+using anakin::saber::Shape;
+using anakin::PBlock;
+using anakin::PTuple;
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+/*
+ * Get a random float value between [low, high]
+ */
+float random(float low, float high) {
+  static std::random_device rd;
+  static std::mt19937 mt(rd());
+  std::uniform_real_distribution<double> dist(low, high);
+  return dist(mt);
+}
+
+void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place,
+                     const platform::DeviceContext& ctx) {
+  auto dims = tensor->dims();
+  size_t num_elements = analysis::AccuDims(dims, dims.size());
+  PADDLE_ENFORCE_GT(num_elements, 0);
+
+  platform::CPUPlace cpu_place;
+  framework::LoDTensor temp_tensor;
+  temp_tensor.Resize(dims);
+  auto* temp_data = temp_tensor.mutable_data<float>(cpu_place);
+
+  for (size_t i = 0; i < num_elements; i++) {
+    *(temp_data + i) = random(0., 1.);
+  }
+
+  TensorCopySync(temp_tensor, place, tensor);
+}
+
+/*
+ * Help to validate the correctness between Fluid Op and the corresponding
+ * anakin
+ * layer.
+ */
+class AnakinConvertValidation {
+  using AnakinNvEngineT = AnakinEngine<NV, Precision::FP32>;
+
+ public:
+  AnakinConvertValidation() = delete;
+
+  AnakinConvertValidation(const std::unordered_set<std::string>& parameters,
+                          const framework::Scope& scope)
+      : parameters_(parameters), scope_(scope), place_(0) {
+    PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0);
+    engine_.reset(new AnakinEngine<NV, Precision::FP32>(true));
+  }
+
+  // Declare a Variable as input with random initialization.
+  void DeclInputVar(const std::string& name,
+                    const std::vector<int> tensor_dims) {
+    DeclVar(name, tensor_dims);
+    // should decalre anakin input here.
+  }
+
+  void DeclParamVar(const std::string& name, const std::vector<int> dim_vec) {
+    DeclVar(name, dim_vec);
+  }
+
+  void DeclOutputVar(const std::string& name, const std::vector<int> dim_vec) {
+    DeclVar(name, dim_vec);
+    // should declare anakin output here.
+  }
+
+  void DeclVar(const std::string& name, const std::vector<int> dim_vec) {
+    platform::CUDADeviceContext ctx(place_);
+    auto* x = scope_.Var(name);
+    auto* x_tensor = x->GetMutable<framework::LoDTensor>();
+    x_tensor->Resize(framework::make_ddim(dim_vec));
+    RandomizeTensor(x_tensor, place_, ctx);
+  }
+
+  void SetOp(const framework::proto::OpDesc& desc) {
+    op_ = framework::OpRegistry::CreateOp(desc);
+    op_desc_.reset(new framework::OpDesc(desc, nullptr));
+    // should init anakin engine here.
+
+    Singleton<AnakinOpConverter>::Global().ConvertOp(
+        desc, parameters_, scope_, engine_.get(), true /*test_mode*/);
+    engine_->Freeze();
+    for (const auto& input : op_desc_->InputArgumentNames()) {
+      if (parameters_.count(input)) continue;
+      auto& t = inference::analysis::GetFromScope<framework::LoDTensor>(scope_,
+                                                                        input);
+      auto t_shape = framework::vectorize2int(t.dims());
+      engine_->SetInputShape(input, t_shape);
+    }
+    engine_->Optimize();
+    engine_->InitGraph();
+  }
+
+  // We use the set 'neglected_output' here, because some Ops like batch norm,
+  // the outputs specified in the op des are only used during training,
+  // so we should neglect those output during inference.
+  void Execute(int batch_size,
+               std::unordered_set<std::string> neglected_output = {}) {
+    // Execute Fluid Op
+    platform::CUDADeviceContext ctx(place_);
+    op_->Run(scope_, place_);
+
+    // std::vector<framework::LoDTensor> input_vector;
+    // std::vector<framework::LoDTensor> output_vector;
+    std::map<std::string, framework::LoDTensor*> inputs;
+    for (const auto& input : op_desc_->InputArgumentNames()) {
+      if (parameters_.count(input)) continue;
+      auto* var = scope_.FindVar(input);
+      auto tensor = var->GetMutable<framework::LoDTensor>();
+      inputs.insert({input, tensor});
+    }
+
+    std::map<std::string, framework::LoDTensor*> outputs;
+    std::vector<std::vector<float>> fluid_outputs;
+    for (const auto& output : op_desc_->OutputArgumentNames()) {
+      if (neglected_output.count(output)) continue;
+      std::vector<float> fluid_out;
+      auto* var = scope_.FindVar(output);
+      auto tensor = var->GetMutable<framework::LoDTensor>();
+      framework::TensorToVector(*tensor, ctx, &fluid_out);
+      fluid_outputs.push_back(fluid_out);
+
+      // size_t fluid_out_size = fluid_out.size();
+      /*for (size_t i = 0; i < fluid_out_size; i++) {
+        std::cout << fluid_out[i] << std::endl;
+      }*/
+      outputs.insert({output, tensor});
+    }
+
+    engine_->Execute(inputs, outputs);
+    int i_output = 0;
+    for (const auto& output : op_desc_->OutputArgumentNames()) {
+      if (neglected_output.count(output)) continue;
+      std::vector<float> anakin_out;
+      auto* var = scope_.FindVar(output);
+      auto tensor = var->GetMutable<framework::LoDTensor>();
+      framework::TensorToVector(*tensor, ctx, &anakin_out);
+
+      size_t anakin_out_size = anakin_out.size();
+      auto fluid_out = fluid_outputs[i_output++];
+      for (size_t i = 0; i < anakin_out_size; i++) {
+        LOG(INFO) << "Output[" << i << "]: anakin[" << anakin_out[i] << "], "
+                  << "fluid[" << fluid_out[i] << "]";
+      }
+    }
+  }
+
+  framework::Scope& scope() { return scope_; }
+
+ private:
+  std::unique_ptr<AnakinNvEngineT> engine_{nullptr};
+  cudaStream_t stream_;
+  std::unique_ptr<framework::OperatorBase> op_;
+  std::unique_ptr<framework::OpDesc> op_desc_;
+  const std::unordered_set<std::string>& parameters_;
+  framework::Scope& scope_;
+  platform::CUDAPlace place_;
+};
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/anakin/engine.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/inference/anakin/engine.h"
+#include <algorithm>
+#include <cstring>
+#include <map>
+#include <utility>
+#include "paddle/fluid/framework/ddim.h"
+
+using anakin::Precision;
+using anakin::OpRunType;
+using paddle::framework::LoDTensor;
+template <typename T, Precision P, OpRunType O>
+using AnakinNetT = anakin::Net<T, P, O>;
+
+template <typename T, Precision P>
+using AnakinGraphT = anakin::graph::Graph<T, P>;
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+AnakinEngine<TargetT, PrecisionType, RunType>::AnakinEngine(bool need_summary)
+    : graph_(new AnakinGraphT<TargetT, PrecisionType>()),
+      net_(new AnakinNetT<TargetT, PrecisionType, RunType>(need_summary)) {}
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+AnakinEngine<TargetT, PrecisionType, RunType>::~AnakinEngine() {}
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+void AnakinEngine<TargetT, PrecisionType, RunType>::SetInputShape(
+    const std::string &name, std::vector<int> shape) {
+  graph_->AddOpAttr<::anakin::PTuple<int>>(name, "input_shape",
+                                           std::move(shape));
+}
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+void AnakinEngine<TargetT, PrecisionType, RunType>::InitGraph() {
+  net_->init(*graph_);
+}
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+void AnakinEngine<TargetT, PrecisionType, RunType>::AddOp(
+    const std::string &name, const std::string &type,
+    const std::vector<std::string> &inputs,
+    const std::vector<std::string> &outputs) {
+  PADDLE_ENFORCE(graph_->AddOp(name, type, inputs, outputs), "Add operation.");
+}
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+void AnakinEngine<TargetT, PrecisionType, RunType>::Execute(
+    const std::map<std::string, framework::LoDTensor *> &inputs,
+    const std::map<std::string, framework::LoDTensor *> &outputs) {
+  for (const auto &input : inputs) {
+    auto *tensor = input.second;
+    auto *data = tensor->data<float>();
+    auto shape = framework::vectorize2int(tensor->dims());
+    ::anakin::saber::Shape anakin_shape(shape);
+    auto *anakin_input = net_->get_in(input.first);
+    ::anakin::saber::Tensor<TargetT> tmp_anakin_tensor(data, TargetT(), 0,
+                                                       anakin_shape);
+    anakin_input->share_from(tmp_anakin_tensor);
+  }
+
+  for (const auto &output : outputs) {
+    auto *tensor = output.second;
+    auto *data = tensor->data<float>();
+    auto shape = framework::vectorize2int(tensor->dims());
+    ::anakin::saber::Shape anakin_shape(shape);
+    auto *anakin_output = net_->get_out(output.first);
+    ::anakin::saber::Tensor<TargetT> tmp_anakin_tensor(data, TargetT(), 0,
+                                                       anakin_shape);
+    anakin_output->share_from(tmp_anakin_tensor);
+  }
+  net_->prediction();
+}
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+void AnakinEngine<TargetT, PrecisionType, RunType>::Freeze() {
+  PADDLE_ENFORCE(graph_->Freeze(), "Freeze anakin subgraph.");
+}
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+void AnakinEngine<TargetT, PrecisionType, RunType>::Optimize() {
+  PADDLE_ENFORCE(graph_->Optimize(), "Graph optimization.");
+}
+
+template <typename TargetT, Precision PrecisionType, OpRunType RunType>
+std::unique_ptr<AnakinEngine<TargetT, PrecisionType, RunType>>
+AnakinEngine<TargetT, PrecisionType, RunType>::Clone() {
+  auto *engine = new AnakinEngine();
+  engine->net_ = std::move(net_->Clone());
+  return std::unique_ptr<AnakinEngine>(engine);
+}
+
+template class AnakinEngine<::anakin::saber::NV, ::anakin::Precision::FP32>;
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/anakin/engine.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <algorithm>
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/inference/engine.h"
+#include "paddle/fluid/inference/utils/singleton.h"
+
+#include "framework/core/net/net.h"
+#include "framework/core/types.h"
+#include "framework/graph/graph.h"
+#include "saber/saber_types.h"
+
+namespace anakin {
+
+template <typename, Precision, OpRunType>
+class Net;
+
+namespace graph {
+template <typename, Precision>
+class Graph;
+}  // namespace graph
+}  // namespace anakin
+
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+template <typename TargetT, ::anakin::Precision PrecisionType,
+          ::anakin::OpRunType RunType = ::anakin::OpRunType::ASYNC>
+class AnakinEngine {
+ public:
+  explicit AnakinEngine(bool need_summary = false);
+  ~AnakinEngine();
+  void InitGraph();
+  void SetInputShape(const std::string &name, std::vector<int> shape);
+  void AddOp(const std::string &name, const std::string &type,
+             const std::vector<std::string> &inputs,
+             const std::vector<std::string> &outputs);
+
+  template <typename T>
+  void AddOpAttr(const std::string &op_name, const std::string &attr_name,
+                 const T &attr_value) {
+    PADDLE_ENFORCE(graph_->AddOpAttr(op_name, attr_name, attr_value),
+                   "Add operation's attribution.");
+  }
+
+  std::unique_ptr<AnakinEngine> Clone();
+  void Freeze();
+  void Optimize();
+  void Execute(const std::map<std::string, framework::LoDTensor *> &inputs,
+               const std::map<std::string, framework::LoDTensor *> &outputs);
+
+ private:
+  using NetT = ::anakin::Net<TargetT, PrecisionType, RunType>;
+  using GraphT = ::anakin::graph::Graph<TargetT, PrecisionType>;
+  std::unique_ptr<GraphT> graph_;
+  std::unique_ptr<NetT> net_;
+};
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/anakin/test_anakin_engine.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <glog/logging.h>
+#include <gtest/gtest.h>
+
+#include <map>
+
+#include "framework/core/net/net.h"
+#include "framework/graph/graph.h"
+#include "framework/graph/graph_global_mem.h"
+#include "paddle/fluid/inference/anakin/engine.h"
+
+using anakin::graph::GraphGlobalMem;
+using anakin::AK_FLOAT;
+using anakin::Precision;
+using anakin::saber::NV;
+using anakin::saber::X86;
+using anakin::saber::Shape;
+using anakin::PBlock;
+using anakin::PTuple;
+namespace paddle {
+namespace inference {
+namespace anakin {
+
+class TestAnakinEngine : public ::testing::Test {
+ protected:
+  void SetUp() override;
+  void TearDown() override {}
+
+ protected:
+  using AnakinNvEngineT = AnakinEngine<NV, Precision::FP32>;
+  std::unique_ptr<AnakinNvEngineT> engine_{nullptr};
+};
+
+void TestAnakinEngine::SetUp() {
+  engine_.reset(new AnakinEngine<NV, Precision::FP32>(true));
+}
+
+TEST_F(TestAnakinEngine, Execute) {
+  engine_->AddOp("op1", "Dense", {"x"}, {"y"});
+  engine_->AddOpAttr("op1", "out_dim", 2);
+  engine_->AddOpAttr("op1", "bias_term", false);
+  engine_->AddOpAttr("op1", "axis", 1);
+  std::vector<int> shape = {1, 1, 1, 2};
+  Shape tmp_shape(shape);
+  // PBlock<NV> weight1(tmp_shape);
+  auto *weight1 =
+      GraphGlobalMem<NV>::Global().template new_block<AK_FLOAT>(tmp_shape);
+  // auto *weight1 = new PBlock<NV>(tmp_shape, AK_FLOAT);
+
+  float *cpu_data = static_cast<float *>(weight1->h_tensor().mutable_data());
+  cpu_data[0] = 2.;
+  weight1->d_tensor().set_shape(tmp_shape);
+  weight1->d_tensor().copy_from(weight1->h_tensor());
+  engine_->AddOpAttr("op1", "weight_1", *weight1);
+
+  engine_->Freeze();
+  // PTuple<int> input_shape = {1};
+  // engine_->AddOpAttr("x", "input_shape", input_shape);
+  engine_->SetInputShape("x", {1, 1, 1, 1});
+  engine_->Optimize();
+  engine_->InitGraph();
+  framework::LoDTensor x;
+  framework::LoDTensor y;
+  x.Resize({1, 1, 1, 1});
+  y.Resize({1, 1, 1, 2});
+  auto *x_data = x.mutable_data<float>(platform::CUDAPlace());
+  float x_data_cpu[] = {1.};
+  cudaMemcpy(x_data, x_data_cpu, sizeof(float), cudaMemcpyHostToDevice);
+
+  std::map<std::string, framework::LoDTensor *> inputs = {{"x", &x}};
+  auto *y_data = y.mutable_data<float>(platform::CUDAPlace());
+  std::map<std::string, framework::LoDTensor *> outputs = {{"y", &y}};
+
+  engine_->Execute(inputs, outputs);
+  auto *y_data_gpu = y_data;
+  float y_data_cpu[2];
+  cudaMemcpy(y_data_cpu, y_data_gpu, sizeof(float) * 2, cudaMemcpyDeviceToHost);
+  LOG(INFO) << "output value: " << y_data_cpu[0] << ", " << y_data_cpu[1];
+}
+
+}  // namespace anakin
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/analysis/argument.h

@@ -23,8 +23,12 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
+#include <unordered_map>
+#include <unordered_set>
 #include <vector>
+
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
@@ -133,6 +137,8 @@ struct Argument {
   DECL_ARGUMENT_FIELD(tensorrt_min_subgraph_size, TensorRtMinSubgraphSize, int);
   DECL_ARGUMENT_FIELD(tensorrt_precision_mode, TensorRtPrecisionMode,
                       AnalysisConfig::Precision);
+  DECL_ARGUMENT_FIELD(tensorrt_use_static_engine, TensorRtUseStaticEngine,
+                      bool);
 
   // Memory optimized related.
   DECL_ARGUMENT_FIELD(enable_memory_optim, EnableMemoryOptim, bool);

paddle/fluid/inference/analysis/helper.h

@@ -17,10 +17,12 @@ limitations under the License. */
 #include <sys/stat.h>
 #include <cstdio>
 #include <fstream>
+#include <memory>
 #include <set>
 #include <string>
 #include <typeindex>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 
 #include "paddle/fluid/framework/framework.pb.h"
@@ -217,6 +219,35 @@ static std::string GetTrtCalibTableData(const std::string &model_opt_cache_dir,
   return "";
 }
 
+static std::string GetTrtEngineSerializedPath(const std::string &model_root,
+                                              const std::string &engine_key) {
+  return model_root + "/trt_serialized_" + engine_key;
+}
+
+static std::string GetTrtEngineSerializedData(
+    const std::string &model_opt_cache_dir, const std::string &engine_key) {
+  std::string trt_serialized_path =
+      GetTrtEngineSerializedPath(model_opt_cache_dir, engine_key);
+  if (FileExists(trt_serialized_path)) {
+    VLOG(3) << "Trt serialized file: " << trt_serialized_path
+            << "is found here";
+    std::ifstream infile(trt_serialized_path, std::ios::in);
+    std::stringstream buffer;
+    buffer << infile.rdbuf();
+    std::string trt_engine_serialized_data(buffer.str());
+    return trt_engine_serialized_data;
+  }
+  return "";
+}
+
+static void SaveTrtEngineSerializedDataToFile(
+    const std::string &trt_serialized_path,
+    const std::string &engine_serialized_data) {
+  std::ofstream outfile(trt_serialized_path);
+  outfile << engine_serialized_data;
+  outfile.close();
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -81,6 +81,9 @@ void IRPassManager::CreatePasses(Argument *argument,
       pass->Set(
           "model_opt_cache_dir",
           new std::string(GetOrCreateModelOptCacheDir(model_opt_cache_dir)));
+      pass->Set("gpu_device_id", new int(argument->gpu_device_id()));
+      pass->Set("use_static_engine",
+                new bool(argument->tensorrt_use_static_engine()));
     }
 
     pre_pass = pass_name;

paddle/fluid/inference/analysis/ir_pass_manager.h

@@ -22,7 +22,10 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
+#include <unordered_set>
+#include <utility>
 #include <vector>
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/ir/pass.h"

paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc

@@ -14,13 +14,13 @@
 
 #include <algorithm>
 #include <set>
-#include <string>
-#include <vector>
 
 #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 #include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h"
 #include "paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.h"
+#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
+#include "paddle/fluid/inference/tensorrt/engine.h"
 #include "paddle/fluid/inference/tensorrt/op_teller.h"
 #include "paddle/fluid/string/pretty_log.h"
 
@@ -33,8 +33,15 @@ using framework::ir::Node;
 std::vector<std::string> ExtractParameters(
     const std::unordered_set<Node *> &nodes);
 
-std::unique_ptr<framework::ir::Graph> analysis::TensorRtSubgraphPass::ApplyImpl(
+void RenameAndGetOutputs(
+    const std::vector<framework::ir::Node *> &subgraph_nodes,
+    framework::BlockDesc *block_desc,
+    const std::set<std::string> &input_names_with_id,
+    std::set<std::string> *output_names_with_id,
+    std::set<std::string> *output_names,
+    std::unordered_map<std::string, std::string> *output_name_map);
 
+std::unique_ptr<framework::ir::Graph> analysis::TensorRtSubgraphPass::ApplyImpl(
     std::unique_ptr<framework::ir::Graph> graph) const {
   framework::ir::FusePassBase::Init("tensorrt_subgraph_pass", graph.get());
 
@@ -47,9 +54,16 @@ std::unique_ptr<framework::ir::Graph> analysis::TensorRtSubgraphPass::ApplyImpl(
                       Get<int>("min_subgraph_size") /*min subgraph size*/);
   fuser();
 
+  std::vector<std::string> graph_param_names =
+      ExtractParameters(graph->Nodes());
+  // those parameter already exist in trt, and should not have another copy in
+  // fluid.
+  std::vector<std::string> repetitive_params;
+
   for (auto *node : graph->Nodes()) {
     if (node->IsOp() && !Agent(node).subgraph()->empty()) {
-      CreateTensorRTOp(node, graph.get());
+      CreateTensorRTOp(node, graph.get(), graph_param_names,
+                       &repetitive_params);
 
       std::unordered_set<const Node *> nodes2remove(
           Agent(node).subgraph()->begin(), Agent(node).subgraph()->end());
@@ -64,12 +78,15 @@ std::unique_ptr<framework::ir::Graph> analysis::TensorRtSubgraphPass::ApplyImpl(
     }
   }
   framework::ir::GraphSafeRemoveNodes(graph.get(), nodes2remove);
+  graph->Set(framework::ir::kRepetitiveParamAttr,
+             new std::vector<std::string>(repetitive_params));
 
   return graph;
 }
 
 std::string GenerateEngineKey(const std::set<std::string> &engine_inputs,
-                              const std::set<std::string> &engine_outputs) {
+                              const std::set<std::string> &engine_outputs,
+                              const std::string &predictor_id) {
   std::string engine_hash_key = "";
   for (auto name : engine_inputs) {
     engine_hash_key += name;
@@ -77,12 +94,15 @@ std::string GenerateEngineKey(const std::set<std::string> &engine_inputs,
   for (auto name : engine_outputs) {
     engine_hash_key += name;
   }
+  engine_hash_key += predictor_id;
   auto engine_key = std::to_string(std::hash<std::string>()(engine_hash_key));
   return engine_key;
 }
 
-void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
-                                            Graph *graph) const {
+void TensorRtSubgraphPass::CreateTensorRTOp(
+    framework::ir::Node *node, Graph *graph,
+    const std::vector<std::string> &graph_params,
+    std::vector<std::string> *repetitive_params) const {
   auto *op_desc = node->Op();
   auto &subgraph = *Agent(node).subgraph();
   PADDLE_ENFORCE(!subgraph.empty());
@@ -116,12 +136,16 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
   // is unique.
   std::set<std::string> input_names;
   std::set<std::string> input_names_with_id;
+  std::vector<std::string> params;
+
+  // The node->inputs containes input tensors and parameters.
   for (auto *x : node->inputs) {
     input_names.insert(x->Name());
     input_names_with_id.insert(x->Name() + std::to_string(x->id()));
+    if (std::count(graph_params.begin(), graph_params.end(), x->Name()) > 0) {
+      params.push_back(x->Name());
+    }
   }
-  op_desc->SetInput(
-      "Xs", std::vector<std::string>(input_names.begin(), input_names.end()));
 
   std::set<std::string> output_names;
   std::set<std::string> output_names_with_id;
@@ -130,11 +154,8 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
     output_names_with_id.insert(x->Name() + std::to_string(x->id()));
   }
 
-  op_desc->SetOutput(
-      "Ys", std::vector<std::string>(output_names.begin(), output_names.end()));
-  op_desc->SetType("tensorrt_engine");
-
   std::unordered_map<std::string, std::string> output_name_map;
+  auto &subgraph_nodes = *Agent(node).subgraph();
 
   // The following procedure is used to rename all the intermediate
   // variables and the output variables of the subgraph.
@@ -148,61 +169,8 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
   // input of a OP, but also the output of a Op, there will be problems.
   // So we have to rename the variable in the subgraph to make sure
   // it is either an OP's input or an OP's output.
-
-  auto &subgraph_nodes = *Agent(node).subgraph();
-  for (size_t index = 0; index < block_desc.OpSize(); ++index) {
-    framework::proto::OpDesc *op = block_desc.Op(index)->Proto();
-    auto correspond_node = subgraph_nodes[index];
-    PADDLE_ENFORCE_EQ(correspond_node->Name(), op->type());
-
-    std::unordered_map<std::string, size_t> var2id;
-    for (auto *in_var : correspond_node->inputs) {
-      var2id[in_var->Name()] = in_var->id();
-    }
-    // rename for the input variables of op inside subgraph
-    for (int i = 0; i < op->inputs_size(); i++) {
-      // one input
-      auto *in_var = op->mutable_inputs(i);
-      std::vector<std::string> replaced_names;
-      for (int k = 0; k < in_var->arguments_size(); k++) {  // all the arguments
-        std::string arg_value = in_var->arguments(k);
-        std::string arg_value_with_id =
-            arg_value + std::to_string(var2id[arg_value]);
-        if (input_names_with_id.count(arg_value_with_id)) {
-          replaced_names.push_back(arg_value);
-        } else {
-          replaced_names.push_back(arg_value_with_id);
-        }
-      }
-      in_var->clear_arguments();
-      for (size_t k = 0; k < replaced_names.size(); k++) {
-        in_var->add_arguments(replaced_names[k]);
-      }
-    }
-    var2id.clear();
-    for (auto out_var : correspond_node->outputs) {
-      var2id[out_var->Name()] = out_var->id();
-    }
-
-    // rename for the output variables of op inside subgraph
-    for (int i = 0; i < op->outputs_size(); i++) {
-      framework::proto::OpDesc_Var *out_var = op->mutable_outputs(i);
-      std::vector<std::string> replaced_names;
-      for (int k = 0; k < out_var->arguments_size(); k++) {
-        std::string arg_value = out_var->arguments(k);
-        std::string arg_value_with_id =
-            arg_value + std::to_string(var2id[arg_value]);
-        if (output_names_with_id.count(arg_value_with_id)) {
-          output_name_map[arg_value] = arg_value_with_id;
-        }
-        replaced_names.push_back(arg_value_with_id);
-      }
-      out_var->clear_arguments();
-      for (size_t k = 0; k < replaced_names.size(); k++) {
-        out_var->add_arguments(replaced_names[k]);
-      }
-    }
-  }
+  RenameAndGetOutputs(subgraph_nodes, &block_desc, input_names_with_id,
+                      &output_names_with_id, &output_names, &output_name_map);
 
   // When tensorrt engine runs at the end of the operation,
   // output_mapping help us copy the data from the renamed ITensor
@@ -212,6 +180,7 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
     PADDLE_ENFORCE(output_name_map.count(name) != 0);
     output_mapping.push_back(output_name_map[name]);
   }
+  PADDLE_ENFORCE(!output_mapping.empty());
 
   auto *vars = block_desc.Proto()->mutable_vars();
   for (framework::ir::Node *node : graph->Nodes()) {
@@ -222,26 +191,83 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
 
   PADDLE_ENFORCE(!block_desc.Proto()->vars().empty(),
                  "the block has no var-desc");
-  PADDLE_ENFORCE(!output_mapping.empty());
+
+  // Set attrs
+  op_desc->SetType("tensorrt_engine");
+  op_desc->SetInput(
+      "Xs", std::vector<std::string>(input_names.begin(), input_names.end()));
+
+  op_desc->SetOutput(
+      "Ys", std::vector<std::string>(output_names.begin(), output_names.end()));
+
   op_desc->SetBlockAttr("sub_block", new_block);
   SetAttr(op_desc->Proto(), "subgraph",
           block_desc.Proto()->SerializeAsString());
-  // Set attrs
   SetAttr(op_desc->Proto(), "max_batch_size", Get<int>("max_batch_size"));
   SetAttr(op_desc->Proto(), "workspace_size", Get<int>("workspace_size"));
-  SetAttr(op_desc->Proto(), "parameters", ExtractParameters(graph->Nodes()));
   SetAttr(op_desc->Proto(), "output_name_mapping", output_mapping);
+  SetAttr(op_desc->Proto(), "parameters", params);
 
   auto enable_int8 = Get<bool>("enable_int8");
-  auto engine_key =
-      GenerateEngineKey(input_names_with_id, output_names_with_id);
+  auto engine_key = GenerateEngineKey(input_names_with_id, output_names_with_id,
+                                      std::to_string(0));
 
+  // Get "" when there is no cached calibration table data.
   std::string calibration_data = GetTrtCalibTableData(
       Get<std::string>("model_opt_cache_dir"), engine_key, enable_int8);
   SetAttr(op_desc->Proto(), "calibration_data", calibration_data);
 
   SetAttr(op_desc->Proto(), "enable_int8", enable_int8);
   SetAttr(op_desc->Proto(), "engine_key", engine_key);
+  SetAttr(op_desc->Proto(), "engine_serialized_data", std::string(""));
+
+  std::unique_ptr<tensorrt::TRTInt8Calibrator> calibrator;
+  if (enable_int8 && calibration_data.size() != 0) {
+    calibrator.reset(new tensorrt::TRTInt8Calibrator(calibration_data));
+  }
+
+  bool use_static_engine = Get<bool>("use_static_engine");
+  // When in int8 mode and calibration_mode, the program just produce the
+  // calibration table data.
+  bool calibration_mode = (enable_int8 && calibration_data.size() == 0);
+  if (!calibration_mode && use_static_engine) {
+    std::copy(params.begin(), params.end(),
+              std::back_inserter(*repetitive_params));
+    std::string trt_engine_serialized_data = GetTrtEngineSerializedData(
+        Get<std::string>("model_opt_cache_dir"), engine_key);
+
+    if (trt_engine_serialized_data.empty()) {
+      LOG(INFO) << "Prepare TRT engine (Optimize model structure, Select OP "
+                   "kernel etc). This process may cost a lot of time.";
+      std::unique_ptr<tensorrt::TensorRTEngine> trt_engine(
+          new tensorrt::TensorRTEngine(
+              Get<int>("max_batch_size"), Get<int>("workspace_size"),
+              enable_int8, calibrator.get(), Get<int>("gpu_device_id")));
+      auto *scope = param_scope();
+      framework::BlockDesc block_desc_temp(nullptr, block_desc.Proto());
+      std::unordered_set<std::string> param_set(params.begin(), params.end());
+      inference::Singleton<inference::tensorrt::OpConverter>::Global()
+          .ConvertBlockToTRTEngine(
+              &block_desc_temp, *scope,
+              std::vector<std::string>(input_names.begin(), input_names.end()),
+              param_set, output_mapping, trt_engine.get());
+      nvinfer1::IHostMemory *serialized_engine_data = trt_engine->Serialize();
+      trt_engine_serialized_data =
+          std::string((const char *)serialized_engine_data->data(),
+                      serialized_engine_data->size());
+      SaveTrtEngineSerializedDataToFile(
+          GetTrtEngineSerializedPath(Get<std::string>("model_opt_cache_dir"),
+                                     engine_key),
+          trt_engine_serialized_data);
+    } else {
+      LOG(INFO) << "Load TRT Optimized Info from "
+                << GetTrtEngineSerializedPath(
+                       Get<std::string>("model_opt_cache_dir"), engine_key);
+    }
+
+    SetAttr(op_desc->Proto(), "engine_serialized_data",
+            trt_engine_serialized_data);
+  }
 }
 
 std::vector<std::string> ExtractParameters(
@@ -253,7 +279,7 @@ std::vector<std::string> ExtractParameters(
   for (const auto &node : nodes) {
     if (!node->IsOp()) continue;
     std::string op_type = node->Op()->Type();
-    if (op_type == "feed") {
+    if (op_type == "feed" || op_type == "fetch") {
       std::vector<std::string> output_names = node->Op()->OutputArgumentNames();
       std::copy(output_names.begin(), output_names.end(),
                 std::back_inserter(feed_outputs));
@@ -272,6 +298,99 @@ std::vector<std::string> ExtractParameters(
   return parameters;
 }
 
+void RenameAndGetOutputs(
+    const std::vector<framework::ir::Node *> &subgraph_nodes,
+    framework::BlockDesc *block_desc,
+    const std::set<std::string> &input_names_with_id,
+    std::set<std::string> *output_names_with_id,
+    std::set<std::string> *output_names,
+    std::unordered_map<std::string, std::string> *output_name_map) {
+  //// In the normal case, the paddle-trt exists bug when runing the googlenet.
+  // When there are more than two convolutions of 1 * 1 with the same input, the
+  // paddle-tensorrt will do the merging optimization, which fuse those conv
+  // into one conv, and then trigger bug. So,  We should use strategy to avoid
+  // this optimization for the time being. This bug will be fixed in the future.
+  std::unordered_map<std::string /*name*/, int /*ITensor_quote_num*/>
+      same_hierarchy_conv2d_num_map;
+
+  for (size_t index = 0; index < block_desc->OpSize(); ++index) {
+    framework::proto::OpDesc *op = block_desc->Op(index)->Proto();
+    framework::OpDesc op_desc(*op, nullptr);
+    auto correspond_node = subgraph_nodes[index];
+    PADDLE_ENFORCE_EQ(correspond_node->Name(), op->type());
+
+    std::unordered_map<std::string, size_t> var2id;
+    std::unordered_map<std::string, framework::ir::Node *> in_vars;
+    for (auto *in_var : correspond_node->inputs) {
+      var2id[in_var->Name()] = in_var->id();
+      in_vars[in_var->Name()] = in_var;
+    }
+    // rename for the input variables of op inside subgraph
+    for (int i = 0; i < op->inputs_size(); i++) {
+      // one input
+      auto *in_var = op->mutable_inputs(i);
+      std::vector<std::string> replaced_names;
+      for (int k = 0; k < in_var->arguments_size(); k++) {  // all the arguments
+        std::string arg_value = in_var->arguments(k);
+        std::string arg_value_with_id =
+            arg_value + std::to_string(var2id[arg_value]);
+        if (input_names_with_id.count(arg_value_with_id)) {
+          replaced_names.push_back(arg_value);
+        } else {
+          replaced_names.push_back(arg_value_with_id);
+        }
+      }
+      in_var->clear_arguments();
+      for (size_t k = 0; k < replaced_names.size(); k++) {
+        in_var->add_arguments(replaced_names[k]);
+      }
+    }
+    var2id.clear();
+    for (auto out_var : correspond_node->outputs) {
+      var2id[out_var->Name()] = out_var->id();
+    }
+
+    if (op_desc.Type() == "conv2d") {
+      auto input_var_name = op_desc.Input("Input").front();
+      auto filter_var_name = op_desc.Input("Filter").front();
+      auto out_var_name = op_desc.Output("Output").front();
+      auto filter_shape = in_vars[filter_var_name]->Var()->GetShape();
+      const std::vector<int> strides =
+          boost::get<std::vector<int>>(op_desc.GetAttr("strides"));
+      const std::vector<int> paddings =
+          boost::get<std::vector<int>>(op_desc.GetAttr("paddings"));
+      if (same_hierarchy_conv2d_num_map[input_var_name] > 0) {
+        (*output_names_with_id)
+            .insert(out_var_name + std::to_string(var2id[out_var_name]));
+        (*output_names).insert(out_var_name);
+      } else if (filter_shape[2] == 1 && filter_shape[3] == 1 &&
+                 strides[0] == 1 && strides[1] == 1 && paddings[0] == 0 &&
+                 paddings[1] == 0) {
+        same_hierarchy_conv2d_num_map[input_var_name] += 1;
+      }
+    }
+
+    // rename for the output variables of op inside subgraph
+    for (int i = 0; i < op->outputs_size(); i++) {
+      framework::proto::OpDesc_Var *out_var = op->mutable_outputs(i);
+      std::vector<std::string> replaced_names;
+      for (int k = 0; k < out_var->arguments_size(); k++) {
+        std::string arg_value = out_var->arguments(k);
+        std::string arg_value_with_id =
+            arg_value + std::to_string(var2id[arg_value]);
+        if (output_names_with_id->count(arg_value_with_id)) {
+          (*output_name_map)[arg_value] = arg_value_with_id;
+        }
+        replaced_names.push_back(arg_value_with_id);
+      }
+      out_var->clear_arguments();
+      for (size_t k = 0; k < replaced_names.size(); k++) {
+        out_var->add_arguments(replaced_names[k]);
+      }
+    }
+  }
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.h

@@ -13,7 +13,12 @@
 // limitations under the License.
 
 #pragma once
-#include <paddle/fluid/framework/ir/fuse_pass_base.h>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/ir/pass.h"
 
 namespace paddle {
@@ -26,8 +31,9 @@ class TensorRtSubgraphPass : public framework::ir::FusePassBase {
       std::unique_ptr<framework::ir::Graph> graph) const override;
 
  private:
-  void CreateTensorRTOp(framework::ir::Node *x,
-                        framework::ir::Graph *graph) const;
+  void CreateTensorRTOp(framework::ir::Node *x, framework::ir::Graph *graph,
+                        const std::vector<std::string> &graph_params,
+                        std::vector<std::string> *repetitive_params) const;
   void CleanIntermediateOutputs(framework::ir::Node *node);
 };
 

paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc

@@ -31,6 +31,13 @@ void IrParamsSyncAmongDevicesPass::RunImpl(Argument *argument) {
   // The parameters are on the cpu, therefore, synchronization is not necessary.
   if (!argument->use_gpu()) return;
 
+  auto &graph = argument->main_graph();
+  std::vector<std::string> repetitive_params;
+
+  if (graph.Has(framework::ir::kRepetitiveParamAttr))
+    repetitive_params = graph.Get<std::vector<std::string>>(
+        framework::ir::kRepetitiveParamAttr);
+
   LOG(INFO) << "Sync params from CPU to GPU";
 
   PADDLE_ENFORCE(argument->gpu_device_id_valid());
@@ -43,6 +50,10 @@ void IrParamsSyncAmongDevicesPass::RunImpl(Argument *argument) {
   // Because there exists the case that new parameter variables are not added to
   // the program in the analysis pass.
   for (auto &var_name : all_vars) {
+    if (std::count(repetitive_params.begin(), repetitive_params.end(),
+                   var_name)) {
+      continue;
+    }
     auto *var = scope->FindLocalVar(var_name);
     PADDLE_ENFORCE(var != nullptr);
     if (var->IsType<framework::LoDTensor>() ||

paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h

@@ -17,6 +17,7 @@
 #include <string>
 #include <vector>
 
+#include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/analysis/analysis_pass.h"
 #include "paddle/fluid/platform/place.h"

paddle/fluid/inference/api/analysis_config.cc

@@ -103,6 +103,7 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
   CP_MEMBER(tensorrt_max_batchsize_);
   CP_MEMBER(tensorrt_min_subgraph_size_);
   CP_MEMBER(tensorrt_precision_mode_);
+  CP_MEMBER(trt_use_static_engine_);
   // MKLDNN related.
   CP_MEMBER(use_mkldnn_);
   CP_MEMBER(mkldnn_enabled_op_types_);
@@ -144,7 +145,7 @@ void AnalysisConfig::EnableMKLDNN() {
 
 void AnalysisConfig::EnableTensorRtEngine(
     int workspace_size, int max_batch_size, int min_subgraph_size,
-    AnalysisConfig::Precision precision_mode) {
+    AnalysisConfig::Precision precision_mode, bool use_static) {
 #ifdef PADDLE_WITH_CUDA
   if (!use_gpu()) {
     LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first";
@@ -156,6 +157,7 @@ void AnalysisConfig::EnableTensorRtEngine(
   tensorrt_max_batchsize_ = max_batch_size;
   tensorrt_min_subgraph_size_ = min_subgraph_size;
   tensorrt_precision_mode_ = precision_mode;
+  trt_use_static_engine_ = use_static;
 
   Update();
 #else

paddle/fluid/inference/api/analysis_predictor.cc

@@ -183,6 +183,9 @@ void AnalysisPredictor::SetMkldnnThreadID(int tid) {
 bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
                             std::vector<PaddleTensor> *output_data,
                             int batch_size) {
+  if (UNLIKELY(config_.cpu_math_library_num_threads() > 1)) {
+    paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
+  }
   VLOG(3) << "Predictor::predict";
   inference::Timer timer;
   timer.tic();
@@ -240,6 +243,8 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
       input_ptr = input.mutable_data<int64_t>(ddim, place_);
     } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
       input_ptr = input.mutable_data<float>(ddim, place_);
+    } else if (inputs[i].dtype == PaddleDType::INT32) {
+      input_ptr = input.mutable_data<int32_t>(ddim, place_);
     } else {
       LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
       return false;
@@ -323,8 +328,11 @@ bool AnalysisPredictor::GetFetch(std::vector<PaddleTensor> *outputs,
     } else if (type == framework::proto::VarType::INT64) {
       GetFetchOne<int64_t>(fetch, output);
       output->dtype = PaddleDType::INT64;
+    } else if (type == framework::proto::VarType::INT32) {
+      GetFetchOne<int32_t>(fetch, output);
+      output->dtype = PaddleDType::INT32;
     } else {
-      LOG(ERROR) << "unknown type, only support float32 and int64 now.";
+      LOG(ERROR) << "unknown type, only support float32, int64 and int32 now.";
     }
   }
   return true;
@@ -362,6 +370,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
     argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_);
     argument_.SetTensorRtMinSubgraphSize(config_.tensorrt_min_subgraph_size_);
     argument_.SetTensorRtPrecisionMode(config_.tensorrt_precision_mode_);
+    argument_.SetTensorRtUseStaticEngine(config_.trt_use_static_engine_);
   }
 
   if (config_.use_mkldnn_) {
@@ -435,12 +444,14 @@ void AnalysisPredictor::PrepareFeedFetch() {
       }
       feeds_[idx] = op;
       feed_names_[op->Output("Out")[0]] = idx;
+      idx2feeds_[idx] = op->Output("Out")[0];
     } else if (op->Type() == "fetch") {
       int idx = boost::get<int>(op->GetAttr("col"));
       if (fetches_.size() <= static_cast<size_t>(idx)) {
         fetches_.resize(idx + 1);
       }
       fetches_[idx] = op;
+      idx2fetches_[idx] = op->Input("X")[0];
     }
   }
 }
@@ -453,6 +464,22 @@ void AnalysisPredictor::CreateFeedFetchVar(framework::Scope *scope) {
   var->GetMutable<framework::FeedFetchList>();
 }
 
+std::vector<std::string> AnalysisPredictor::GetInputNames() {
+  std::vector<std::string> input_names;
+  for (auto &item : idx2feeds_) {
+    input_names.push_back(item.second);
+  }
+  return input_names;
+}
+
+std::vector<std::string> AnalysisPredictor::GetOutputNames() {
+  std::vector<std::string> output_names;
+  for (auto &item : idx2fetches_) {
+    output_names.push_back(item.second);
+  }
+  return output_names;
+}
+
 std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetInputTensor(
     const std::string &name) {
   PADDLE_ENFORCE(executor_->scope()->FindVar(name), "no name called %s", name);
@@ -460,6 +487,13 @@ std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetInputTensor(
       new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
   res->input_or_output_ = true;
   res->SetName(name);
+  if (platform::is_cpu_place(place_)) {
+    res->SetPlace(PaddlePlace::kCPU);
+  } else {
+    auto gpu_place = boost::get<platform::CUDAPlace>(place_);
+    res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
+  }
+
   return res;
 }
 
@@ -470,6 +504,12 @@ std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
       new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
   res->input_or_output_ = false;
   res->SetName(name);
+  if (platform::is_cpu_place(place_)) {
+    res->SetPlace(PaddlePlace::kCPU);
+  } else {
+    auto gpu_place = boost::get<platform::CUDAPlace>(place_);
+    res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
+  }
   return res;
 }
 

paddle/fluid/inference/api/analysis_predictor.h

@@ -15,12 +15,14 @@
 #pragma once
 #include <algorithm>
 #include <map>
+#include <memory>
 #include <string>
 #include <vector>
 #include "paddle/fluid/framework/naive_executor.h"
 #include "paddle/fluid/inference/analysis/analyzer.h"
 #include "paddle/fluid/inference/api/api_impl.h"
 #include "paddle/fluid/inference/api/details/reset_tensor_array.h"
+#include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/string/printf.h"
 #ifdef PADDLE_WITH_TESTING
@@ -53,6 +55,9 @@ class AnalysisPredictor : public PaddlePredictor {
            std::vector<PaddleTensor> *output_data,
            int batch_size = -1) override;
 
+  std::vector<std::string> GetInputNames();
+  std::vector<std::string> GetOutputNames();
+
   std::unique_ptr<ZeroCopyTensor> GetInputTensor(
       const std::string &name) override;
   std::unique_ptr<ZeroCopyTensor> GetOutputTensor(
@@ -131,7 +136,11 @@ class AnalysisPredictor : public PaddlePredictor {
   std::shared_ptr<framework::ProgramDesc> inference_program_;
   std::vector<framework::OpDesc *> feeds_;
   std::map<std::string, size_t> feed_names_;
+  // Sorted according to the idx.
+  std::map<size_t, std::string> idx2feeds_;
   std::vector<framework::OpDesc *> fetches_;
+  std::map<size_t, std::string> idx2fetches_;
+
   // Memory buffer for feed inputs. The temporary LoDTensor will cause serious
   // concurrency problems, wrong results and memory leak, so cache them.
   std::vector<framework::LoDTensor> feed_tensors_;

paddle/fluid/inference/api/api.cc

@@ -28,6 +28,8 @@ int PaddleDtypeSize(PaddleDType dtype) {
       return sizeof(float);
     case PaddleDType::INT64:
       return sizeof(int64_t);
+    case PaddleDType::INT32:
+      return sizeof(int32_t);
     default:
       assert(false);
       return -1;

paddle/fluid/inference/api/api_impl.cc

@@ -131,6 +131,9 @@ NativePaddlePredictor::~NativePaddlePredictor() {
 bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
                                 std::vector<PaddleTensor> *output_data,
                                 int batch_size) {
+  if (UNLIKELY(config_.cpu_math_library_num_threads() > 1)) {
+    paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
+  }
   VLOG(3) << "Predictor::predict";
   Timer timer;
   timer.tic();
@@ -200,6 +203,8 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
       input_ptr = input.mutable_data<int64_t>(ddim, place_);
     } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
       input_ptr = input.mutable_data<float>(ddim, place_);
+    } else if (inputs[i].dtype == PaddleDType::INT32) {
+      input_ptr = input.mutable_data<int32_t>(ddim, place_);
     } else {
       LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
       return false;
@@ -278,8 +283,11 @@ bool NativePaddlePredictor::GetFetch(std::vector<PaddleTensor> *outputs,
     } else if (type == framework::DataTypeTrait<int64_t>::DataType) {
       GetFetchOne<int64_t>(fetch, output);
       output->dtype = PaddleDType::INT64;
+    } else if (type == framework::DataTypeTrait<int32_t>::DataType) {
+      GetFetchOne<int32_t>(fetch, output);
+      output->dtype = PaddleDType::INT32;
     } else {
-      LOG(ERROR) << "unknown type, only support float32 and int64 now.";
+      LOG(ERROR) << "unknown type, only support float32, int64 and int32 now.";
     }
   }
   return true;

paddle/fluid/inference/api/api_impl_tester.cc

@@ -42,6 +42,9 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
   } else if (t->type() == framework::proto::VarType::FP32) {
     pt.data.Reset(t->data<void>(), t->numel() * sizeof(float));
     pt.dtype = PaddleDType::FLOAT32;
+  } else if (t->type() == framework::proto::VarType::INT32) {
+    pt.data.Reset(t->data<void>(), t->numel() * sizeof(int32_t));
+    pt.dtype = PaddleDType::INT32;
   } else {
     LOG(FATAL) << "unsupported type.";
   }

paddle/fluid/inference/api/demo_ci/utils.h

@@ -88,13 +88,20 @@ void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
       }
       break;
     }
-    case PaddleDType::FLOAT32:
+    case PaddleDType::FLOAT32: {
       for (size_t i = 0; i < numel; ++i) {
         CHECK_LT(
             fabs(static_cast<float*>(output.data.data())[i] - refer.data[i]),
             1e-5);
       }
       break;
+    }
+    case PaddleDType::INT32: {
+      for (size_t i = 0; i < numel; ++i) {
+        CHECK_EQ(static_cast<int32_t*>(output.data.data())[i], refer.data[i]);
+      }
+      break;
+    }
   }
 }
 
@@ -113,11 +120,18 @@ static std::string SummaryTensor(const PaddleTensor& tensor) {
       }
       break;
     }
-    case PaddleDType::FLOAT32:
+    case PaddleDType::FLOAT32: {
       for (int i = 0; i < std::min(num_elems, 10); i++) {
         ss << static_cast<float*>(tensor.data.data())[i] << " ";
       }
       break;
+    }
+    case PaddleDType::INT32: {
+      for (int i = 0; i < std::min(num_elems, 10); i++) {
+        ss << static_cast<int32_t*>(tensor.data.data())[i] << " ";
+      }
+      break;
+    }
   }
   return ss.str();
 }

paddle/fluid/inference/api/details/zero_copy_tensor.cc

@@ -15,6 +15,7 @@
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
+#include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {
@@ -73,6 +74,61 @@ T *ZeroCopyTensor::data(PaddlePlace *place, int *size) const {
   return res;
 }
 
+template <typename T>
+void ZeroCopyTensor::copy_from_cpu(const T *data) {
+  EAGER_GET_TENSOR;
+  PADDLE_ENFORCE_GE(
+      tensor->numel(), 0,
+      "You should call ZeroCopyTensor::Reshape(const std::vector<int> &shape)"
+      "function before copy data from cpu.");
+  size_t ele_size = tensor->numel() * sizeof(T);
+
+  if (place_ == PaddlePlace::kCPU) {
+    auto *t_data = tensor->mutable_data<T>(platform::CPUPlace());
+    std::memcpy(static_cast<void *>(t_data), data, ele_size);
+  } else {
+#ifdef PADDLE_WITH_CUDA
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    platform::CUDAPlace gpu_place(device_);
+    auto *t_data = tensor->mutable_data<T>(gpu_place);
+    auto *dev_ctx =
+        static_cast<const platform::CUDADeviceContext *>(pool.Get(gpu_place));
+
+    memory::Copy(gpu_place, static_cast<void *>(t_data), platform::CPUPlace(),
+                 data, ele_size, dev_ctx->stream());
+#else
+    PADDLE_THROW("Not compile with CUDA, should not reach here.");
+#endif
+  }
+}
+
+template <typename T>
+void ZeroCopyTensor::copy_to_cpu(T *data) {
+  EAGER_GET_TENSOR;
+  auto ele_num = tensor->numel();
+  auto *t_data = tensor->data<T>();
+  auto t_place = tensor->place();
+
+  if (platform::is_cpu_place(t_place)) {
+    std::memcpy(static_cast<void *>(data), t_data, ele_num * sizeof(T));
+  } else {
+#ifdef PADDLE_WITH_CUDA
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    auto gpu_place = boost::get<platform::CUDAPlace>(t_place);
+    auto *dev_ctx =
+        static_cast<const platform::CUDADeviceContext *>(pool.Get(gpu_place));
+    memory::Copy(platform::CPUPlace(), static_cast<void *>(data), gpu_place,
+                 t_data, ele_num * sizeof(T), dev_ctx->stream());
+#else
+    PADDLE_THROW("Not compile with CUDA, should not reach here.");
+#endif
+  }
+}
+template void ZeroCopyTensor::copy_from_cpu<float>(const float *data);
+template void ZeroCopyTensor::copy_from_cpu<int64_t>(const int64_t *data);
+template void ZeroCopyTensor::copy_to_cpu<float>(float *data);
+template void ZeroCopyTensor::copy_to_cpu<int64_t>(int64_t *data);
+
 template float *ZeroCopyTensor::data<float>(PaddlePlace *place,
                                             int *size) const;
 template int64_t *ZeroCopyTensor::data<int64_t>(PaddlePlace *place,
@@ -92,10 +148,10 @@ void *ZeroCopyTensor::FindTensor() const {
   return tensor;
 }
 
-std::vector<int64_t> ZeroCopyTensor::shape() const {
+std::vector<int> ZeroCopyTensor::shape() const {
   EAGER_GET_TENSOR;
   PADDLE_ENFORCE(tensor_, "not found tensor called %s in the scope", name_);
-  return framework::vectorize(tensor->dims());
+  return framework::vectorize2int(tensor->dims());
 }
 
 void ZeroCopyTensor::SetLoD(const std::vector<std::vector<size_t>> &x) {

paddle/fluid/inference/api/details/zero_copy_tensor_dummy.cc

@@ -37,7 +37,7 @@ template int64_t *ZeroCopyTensor::mutable_data(PaddlePlace place);
 
 void *ZeroCopyTensor::FindTensor() const { return nullptr; }
 
-std::vector<int64_t> ZeroCopyTensor::shape() const { return {}; }
+std::vector<int> ZeroCopyTensor::shape() const { return {}; }
 
 void ZeroCopyTensor::SetLoD(const std::vector<std::vector<size_t>> &x) {}
 

paddle/fluid/inference/api/helper.h

@@ -50,6 +50,11 @@ class Timer {
   }
 };
 
+static int GetUniqueId() {
+  static int id = 0;
+  return id++;
+}
+
 static void split(const std::string &str, char sep,
                   std::vector<std::string> *pieces) {
   pieces->clear();
@@ -81,6 +86,13 @@ static void split_to_int64(const std::string &str, char sep,
   std::transform(pieces.begin(), pieces.end(), std::back_inserter(*is),
                  [](const std::string &v) { return std::stoi(v); });
 }
+static void split_to_int(const std::string &str, char sep,
+                         std::vector<int> *is) {
+  std::vector<std::string> pieces;
+  split(str, sep, &pieces);
+  std::transform(pieces.begin(), pieces.end(), std::back_inserter(*is),
+                 [](const std::string &v) { return std::stoi(v); });
+}
 template <typename T>
 std::string to_string(const std::vector<T> &vec) {
   std::stringstream ss;
@@ -197,6 +209,9 @@ static std::string DescribeTensor(const PaddleTensor &tensor,
     case PaddleDType::INT64:
       os << "int64";
       break;
+    case PaddleDType::INT32:
+      os << "int32";
+      break;
     default:
       os << "unset";
   }

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -135,7 +135,8 @@ struct AnalysisConfig {
    */
   void EnableTensorRtEngine(int workspace_size = 1 << 20,
                             int max_batch_size = 1, int min_subgraph_size = 3,
-                            Precision precision = Precision::kFloat32);
+                            Precision precision = Precision::kFloat32,
+                            bool use_static = true);
   /** A boolean state telling whether the TensorRT engine is used.
    */
   bool tensorrt_engine_enabled() const { return use_tensorrt_; }
@@ -233,6 +234,7 @@ struct AnalysisConfig {
   //  subgraph, 3 as default value.
   int tensorrt_min_subgraph_size_{3};
   Precision tensorrt_precision_mode_;
+  bool trt_use_static_engine_;
 
   // memory reuse related.
   bool enable_memory_optim_{false};

paddle/fluid/inference/api/paddle_api.h

@@ -36,6 +36,7 @@ namespace paddle {
 enum PaddleDType {
   FLOAT32,
   INT64,
+  INT32,
   // TODO(Superjomn) support more data types if needed.
 };
 
@@ -160,11 +161,21 @@ class ZeroCopyTensor {
   template <typename T>
   T* data(PaddlePlace* place, int* size) const;
 
-  std::vector<int64_t> shape() const;
+  template <typename T>
+  void copy_from_cpu(const T* data);
+
+  template <typename T>
+  void copy_to_cpu(T* data);
+
+  std::vector<int> shape() const;
 
   void SetLoD(const std::vector<std::vector<size_t>>& x);
   std::vector<std::vector<size_t>> lod() const;
   const std::string& name() const { return name_; }
+  void SetPlace(PaddlePlace place, int device = -1) {
+    place_ = place;
+    device_ = device;
+  }
 
  protected:
   explicit ZeroCopyTensor(void* scope) : scope_{scope} {}
@@ -179,6 +190,8 @@ class ZeroCopyTensor {
   // The corresponding tensor pointer inside Paddle workspace is cached for
   // performance.
   mutable void* tensor_{nullptr};
+  PaddlePlace place_;
+  int device_;
 };
 
 /** A simple Inference API for Paddle.
@@ -200,6 +213,14 @@ class PaddlePredictor {
                    std::vector<PaddleTensor>* output_data,
                    int batch_size = -1) = 0;
 
+  /** \brief Get input names of the model
+   */
+  virtual std::vector<std::string> GetInputNames() { return {}; }
+
+  /** \brief Get output names of the model
+   */
+  virtual std::vector<std::string> GetOutputNames() { return {}; }
+
   /** \brief Get a mutable tensor directly.
    *
    * NOTE Only works in AnalysisPredictor.

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -13,7 +13,9 @@
 // limitations under the License.
 
 #include "paddle/fluid/inference/api/paddle_pass_builder.h"
-
+#ifdef PADDLE_WITH_CUDA
+#include <cudnn.h>
+#endif
 #include <glog/logging.h>
 
 namespace paddle {

paddle/fluid/inference/engine.h

@@ -49,11 +49,6 @@ class EngineBase {
   // Execute the engine, that will run the inference network.
   virtual void Execute(int batch_size) = 0;
 
-  // Return the IO buffer that allocated in engine. One can read/write directly
-  // on the buffer. If the buffer's buffer is nullptr, one can also allocate
-  // memory and maintain it outside the engine.
-  virtual Buffer& buffer(const std::string& name) = 0;
-
   virtual ~EngineBase() {}
 };  // class EngineBase
 

paddle/fluid/inference/tensorrt/convert/conv2d_op.cc

@@ -18,21 +18,6 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-bool to_skip_merging_optimize(TensorRTEngine* engine,
-                              const std::vector<int>& filters,
-                              const std::vector<int>& strides,
-                              const std::vector<int>& paddings,
-                              std::string input_name) {
-  if (engine->itensor_quote_num[input_name] > 0) {
-    return true;
-  }
-  if (filters[0] == 1 && filters[1] == 1 && strides[0] == 1 &&
-      strides[1] == 1 && paddings[0] == 0 && paddings[1] == 0)
-    engine->itensor_quote_num[input_name] += 1;
-
-  return false;
-}
-
 template <typename RegistFunc, typename SetDilationFunc>
 void ConvertConv2d(TensorRTEngine* engine, const framework::proto::OpDesc& op,
                    const framework::Scope& scope, bool test_mode,
@@ -59,7 +44,7 @@ void ConvertConv2d(TensorRTEngine* engine, const framework::proto::OpDesc& op,
   weight_tensor->Resize(Y_t->dims());
   TensorCopySync((*Y_t), cpu_place, weight_tensor.get());
 
-  auto* weight_data = weight_tensor->mutable_data<float>(platform::CPUPlace());
+  auto* weight_data = weight_tensor->mutable_data<float>(cpu_place);
 
   PADDLE_ENFORCE_EQ(weight_tensor->dims().size(), 4UL);
   const int n_output = weight_tensor->dims()[0];
@@ -100,9 +85,7 @@ void ConvertConv2d(TensorRTEngine* engine, const framework::proto::OpDesc& op,
   layer->getOutput(0)->setName(output_name.c_str());
   engine->SetITensor(output_name, layer->getOutput(0));
 
-  if (test_mode ||
-      to_skip_merging_optimize(engine, {filter_h, filter_w}, strides, paddings,
-                               op_desc.Input("Input").front())) {
+  if (test_mode) {
     engine->DeclareOutput(output_name);
   }
 }

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

@@ -153,7 +153,6 @@ class ElementwiseTensorOpConverter : public OpConverter {
     if (CheckDims(dims_x, dims_y)) {
       // The two input tensor should have the same dims
       VLOG(3) << "Convert a fluid elementwise op to TensorRT IElementWiseLayer";
-
       nvinfer1::IElementWiseLayer* layer = TRT_ENGINE_ADD_LAYER(
           engine_, ElementWise, *const_cast<nvinfer1::ITensor*>(X),
           *const_cast<nvinfer1::ITensor*>(Y), op_pair->second);
@@ -166,7 +165,7 @@ class ElementwiseTensorOpConverter : public OpConverter {
                  "ElementWisePluginLayer";
 
       plugin::ElementWisePlugin* plugin =
-          new plugin::ElementWisePlugin(op_pair->second, dims_x, dims_y, axis);
+          new plugin::ElementWisePlugin(op_type_, dims_x, dims_y, axis);
       plugin->AddInput(X);
       plugin->AddInput(Y);
       nvinfer1::IPluginLayer* layer = engine_->AddPlugin(

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

@@ -85,10 +85,10 @@ class FcOpConverter : public OpConverter {
            Y_t->dims()[0] * Y_t->dims()[1] * sizeof(float));
     TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT,
                                   static_cast<void*>(weight_data),
-                                  Y_t->memory_size() / sizeof(float)};
+                                  static_cast<size_t>(Y_t->numel())};
     TensorRTEngine::Weight tmp_weight(nvinfer1::DataType::kFLOAT,
                                       static_cast<void*>(tmp->data<float>()),
-                                      Y_t->memory_size() / sizeof(float));
+                                      static_cast<size_t>(Y_t->numel()));
     weight.dims.assign({Y_t->dims()[0], Y_t->dims()[1]});
     tmp_weight.dims = weight.dims;
 

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

@@ -16,9 +16,12 @@ limitations under the License. */
 
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
+#include <vector>
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
 #include "paddle/fluid/inference/utils/singleton.h"
 
@@ -26,6 +29,37 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 
+using FluidDT = framework::proto::VarType_Type;
+using TRT_DT = nvinfer1::DataType;
+
+namespace {  // NOLINT
+
+TRT_DT FluidDataType2TRT(FluidDT type) {
+  switch (type) {
+    case FluidDT::VarType_Type_FP32:
+      return TRT_DT::kFLOAT;
+    case FluidDT::VarType_Type_INT32:
+      return TRT_DT::kINT32;
+    default:
+      return TRT_DT::kINT32;
+  }
+  PADDLE_THROW("unkown type");
+  return TRT_DT::kINT32;
+}
+
+nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t>& shape) {
+  PADDLE_ENFORCE_GT(shape.size(), 1UL,
+                    "TensorRT' tensor input requires at least 2 dimensions");
+  PADDLE_ENFORCE_LE(shape.size(), 4UL,
+                    "TensorRT' tensor input requires at most 4 dimensions");
+  PADDLE_ENFORCE(shape.size() == 4UL || shape.size() == 2UL);
+  if (shape.size() == 4UL)
+    return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]);
+  return nvinfer1::DimsCHW(shape[1], 1, 1);
+}
+
+}  // namespace // NOLINT
+
 /*
  * Convert Op from Fluid to TensorRT Engine.
  */
@@ -110,6 +144,34 @@ class OpConverter {
     }
   }
 
+  // The scope  here should be inited with the parameter vars.
+  void ConvertBlockToTRTEngine(
+      framework::BlockDesc* block_desc, const framework::Scope& scope,
+      const std::vector<std::string>& inputs,
+      const std::unordered_set<std::string>& parameters,
+      const std::vector<std::string>& outputs, TensorRTEngine* engine) {
+    engine->InitNetwork();
+    for (auto& input : inputs) {
+      if (parameters.count(input)) continue;
+      auto* var = block_desc->FindVar(input);
+      PADDLE_ENFORCE(var, "no variable called %s", input);
+      PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR,
+                        "TensorRT engine only takes LoDTensor as input");
+      auto var_shape = var->GetShape();
+
+      engine->DeclareInput(
+          input, FluidDataType2TRT(
+                     var->Proto()->type().lod_tensor().tensor().data_type()),
+          Vec2TRT_Dims(var_shape));
+    }
+    framework::proto::BlockDesc* block_proto = block_desc->Proto();
+    ConvertBlock(*block_proto, parameters, scope, engine);
+    for (auto& output : outputs) {
+      engine->DeclareOutput(output);
+    }
+    engine->FreezeNetwork();
+  }
+
   void SetEngine(TensorRTEngine* engine) { engine_ = engine; }
 
   virtual ~OpConverter() {}

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

@@ -43,23 +43,20 @@ class PReluOpConverter : public OpConverter {
     PADDLE_ENFORCE_NOT_NULL(alpha_var);
     auto* alpha_tensor = alpha_var->GetMutable<framework::LoDTensor>();
 
-    platform::CUDAPlace place;
-    std::unique_ptr<framework::LoDTensor> alpha_tensor_device(
+    platform::CPUPlace cpu_place;
+    std::unique_ptr<framework::LoDTensor> alpha_tensor_temp(
         new framework::LoDTensor());
-    alpha_tensor_device->Resize(alpha_tensor->dims());
-    TensorCopySync(*alpha_tensor, place, alpha_tensor_device.get());
-    float* alpha_data = alpha_tensor_device->mutable_data<float>(place);
+    alpha_tensor_temp->Resize(alpha_tensor->dims());
+    TensorCopySync(*alpha_tensor, cpu_place, alpha_tensor_temp.get());
+    float* alpha_data = alpha_tensor_temp->mutable_data<float>(cpu_place);
 
-    // Transform alpha to TensorRTEngine::Weight
-    TensorRTEngine::Weight alpha_rt(nvinfer1::DataType::kFLOAT,
-                                    static_cast<void*>(alpha_data),
-                                    alpha_tensor_device->numel());
-    plugin::PReluPlugin* plugin = new plugin::PReluPlugin(alpha_rt, mode);
+    plugin::PReluPlugin* plugin =
+        new plugin::PReluPlugin(alpha_data, alpha_tensor_temp->numel(), mode);
     nvinfer1::IPluginLayer* layer =
         engine_->AddPlugin(&input, input_num, plugin);
     // keep alpha tensor to avoid release it's memory
     engine_->weight_map[op_desc.Input("Alpha")[0]] =
-        std::move(alpha_tensor_device);
+        std::move(alpha_tensor_temp);
 
     std::string layer_name = "prelu (Output: ";
     auto output_name = op_desc.Output("Out")[0];

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

@@ -19,7 +19,9 @@ limitations under the License. */
 
 #pragma once
 
+#include <memory>
 #include <string>
+#include <unordered_set>
 #include <vector>
 
 #include "paddle/fluid/framework/lod_tensor.h"
@@ -79,7 +81,8 @@ class TRTConvertValidation {
         if_add_batch_(if_add_batch),
         max_batch_size_(max_batch_size) {
     PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0);
-    engine_.reset(new TensorRTEngine(max_batch_size, workspace_size, stream_));
+    engine_.reset(
+        new TensorRTEngine(max_batch_size, workspace_size, false, nullptr, 0));
     engine_->InitNetwork();
   }
 
@@ -114,13 +117,12 @@ class TRTConvertValidation {
   }
 
   void DeclVar(const std::string& name, const std::vector<int> dim_vec) {
-    platform::CUDAPlace place;
-    platform::CUDADeviceContext ctx(place);
+    platform::CUDADeviceContext ctx(place_);
 
     auto* x = scope_.Var(name);
     auto* x_tensor = x->GetMutable<framework::LoDTensor>();
     x_tensor->Resize(framework::make_ddim(dim_vec));
-    RandomizeTensor(x_tensor, place, ctx);
+    RandomizeTensor(x_tensor, place_, ctx);
   }
   // Declare a variable in a fluid Scope.
   void DeclVar(const std::string& name, const nvinfer1::Dims& dims,
@@ -146,19 +148,6 @@ class TRTConvertValidation {
 
     // Declare outputs.
     op_desc_.reset(new framework::OpDesc(desc, nullptr));
-
-    // Set Inputs.
-    for (const auto& input : op_desc_->InputArgumentNames()) {
-      if (parameters_.count(input)) continue;
-      auto* var = scope_.FindVar(input);
-      PADDLE_ENFORCE(var);
-      auto tensor = var->GetMutable<framework::LoDTensor>();
-
-      engine_->SetInputFromGPU(
-          input, static_cast<void*>(tensor->data<void>()),
-          sizeof(float) *
-              analysis::AccuDims(tensor->dims(), tensor->dims().size()));
-    }
   }
 
   // We use the set 'neglected_output' here, because some Ops like batch norm,
@@ -168,43 +157,71 @@ class TRTConvertValidation {
                std::unordered_set<std::string> neglected_output = {}) {
     // Execute Fluid Op
     PADDLE_ENFORCE_LE(batch_size, max_batch_size_);
-    platform::CUDAPlace place;
-    platform::CUDADeviceContext ctx(place);
-    op_->Run(scope_, place);
-    // Execute TRT.
-    engine_->Execute(batch_size);
-    cudaStreamSynchronize(engine_->stream());
+    platform::CUDADeviceContext ctx(place_);
+    op_->Run(scope_, place_);
 
-    ASSERT_FALSE(op_desc_->OutputArgumentNames().empty());
-    const size_t output_space_size = 3000;
+    std::vector<std::string> input_output_names;
+
+    // Note: we need filter the parameter
+    for (const auto& input : op_desc_->InputArgumentNames()) {
+      if (parameters_.count(input)) continue;
+      input_output_names.push_back(input);
+    }
+
+    // Collect the fluid outputs.
+    std::vector<std::vector<float>> fluid_outs;
     for (const auto& output : op_desc_->OutputArgumentNames()) {
       if (neglected_output.count(output)) continue;
+      input_output_names.push_back(output);
       std::vector<float> fluid_out;
-      std::vector<float> trt_out(output_space_size);
-      engine_->GetOutputInCPU(output, &trt_out[0], output_space_size);
-      cudaStreamSynchronize(engine_->stream());
-
       auto* var = scope_.FindVar(output);
-      auto tensor = var->GetMutable<framework::LoDTensor>();
+      auto* tensor = var->GetMutable<framework::LoDTensor>();
       framework::TensorToVector(*tensor, ctx, &fluid_out);
+      fluid_outs.push_back(fluid_out);
+    }
+
+    // Bind input and output for TRT.
+    const int num_bindings = input_output_names.size();
+    std::vector<void*> buffers(num_bindings);
+
+    for (const std::string& name : input_output_names) {
+      auto* var = scope_.FindVar(name);
+      auto* tensor = var->GetMutable<framework::LoDTensor>();
+      const int bind_index = engine_->engine()->getBindingIndex(name.c_str());
+      buffers[bind_index] =
+          static_cast<void*>(tensor->mutable_data<float>(place_));
+    }
+
+    // Execute TRT.
+    engine_->Execute(batch_size, &buffers, stream_);
 
-      size_t fluid_out_size = fluid_out.size();
+    ASSERT_FALSE(op_desc_->OutputArgumentNames().empty());
+    int index = 0;
+    for (const auto& output : op_desc_->OutputArgumentNames()) {
+      if (neglected_output.count(output)) continue;
+      std::vector<float> trt_out;
+      auto* var = scope_.FindVar(output);
+      auto* tensor = var->GetMutable<framework::LoDTensor>();
+      framework::TensorToVector(*tensor, ctx, &trt_out);
+
+      size_t fluid_out_size = fluid_outs[index].size();
       if (if_add_batch_ == true) {
         fluid_out_size =
             batch_size * (framework::product(tensor->dims()) / max_batch_size_);
       }
-      // Compare two output
-      ASSERT_FALSE(fluid_out.empty());
+
       for (size_t i = 0; i < fluid_out_size; i++) {
         // Loose the threshold for CI in different machine model.
-        EXPECT_LT(std::abs(fluid_out[i] - trt_out[i]), 2e-5);
+        EXPECT_LT(std::abs(fluid_outs[index][i] - trt_out[i]), 2e-5);
       }
+      index += 1;
     }
   }
 
   framework::Scope& scope() { return scope_; }
 
  private:
+  platform::CUDAPlace place_;
   std::unique_ptr<TensorRTEngine> engine_;
   cudaStream_t stream_;
   std::unique_ptr<framework::OperatorBase> op_;

paddle/fluid/inference/tensorrt/engine.cc

@@ -32,36 +32,18 @@ void TensorRTEngine::Build(const DescType &paddle_model) {
   PADDLE_ENFORCE(false, "not implemented");
 }
 
-void TensorRTEngine::Execute(int batch_size) {
+void TensorRTEngine::Execute(int batch_size, std::vector<void *> *buffers,
+                             cudaStream_t stream) {
   freshDeviceId();
   batch_size_ = batch_size;
-  std::vector<void *> buffers;
-  for (auto &buf : buffers_) {
-    PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated");
-    PADDLE_ENFORCE_GT(buf.max_size, 0);
-    PADDLE_ENFORCE(buf.device == DeviceType::GPU);
-    buffers.push_back(buf.buffer);
-  }
-  infer_context_->enqueue(batch_size, buffers.data(), stream_, nullptr);
-  cudaStreamSynchronize(stream_);
+  infer_context_->enqueue(batch_size, buffers->data(), stream, nullptr);
+  cudaStreamSynchronize(stream);
   SetRuntimeBatch(batch_size);
 }
 
-TensorRTEngine::~TensorRTEngine() {
-  cudaStreamSynchronize(stream_);
-  // clean buffer
-  for (auto &buf : buffers_) {
-    if (buf.device == DeviceType::GPU && buf.buffer != nullptr) {
-      PADDLE_ENFORCE_EQ(0, cudaFree(buf.buffer));
-      buf.buffer = nullptr;
-      buf.max_size = 0;
-    }
-  }
-}
-
 void TensorRTEngine::FreezeNetwork() {
-  VLOG(3) << "TRT to freeze network";
   freshDeviceId();
+  VLOG(3) << "TRT to freeze network";
   PADDLE_ENFORCE(infer_builder_ != nullptr,
                  "Call InitNetwork first to initialize network.");
   PADDLE_ENFORCE(infer_network_ != nullptr,
@@ -81,30 +63,6 @@ void TensorRTEngine::FreezeNetwork() {
   PADDLE_ENFORCE(infer_engine_ != nullptr, "build cuda engine failed!");
 
   infer_context_.reset(infer_engine_->createExecutionContext());
-
-  // allocate GPU buffers.
-  buffers_.resize(buffer_sizes_.size());
-  for (auto &item : buffer_sizes_) {
-    // The output buffers are not set in the network building phrase, need to
-    // infer from the TesorRT network.
-    if (item.second == 0) {
-      auto slot_offset = infer_engine_->getBindingIndex(item.first.c_str());
-      auto dims = infer_engine_->getBindingDimensions(slot_offset);
-      item.second = kDataTypeSize[static_cast<int>(
-                        infer_engine_->getBindingDataType(slot_offset))] *
-                    analysis::AccuDims(dims.d, dims.nbDims) * max_batch_;
-      PADDLE_ENFORCE_GT(item.second, 0);
-    }
-
-    auto &buf = buffer(item.first);
-    buf.max_size = item.second * max_batch_;
-    CHECK(buf.buffer == nullptr);  // buffer should be allocated only once.
-
-    PADDLE_ENFORCE_EQ(0, cudaMalloc(&buf.buffer, item.second * max_batch_));
-    buf.size = 0;
-    PADDLE_ENFORCE_LE(buf.max_size, 1 << 30);  // 10G
-    buf.device = DeviceType::GPU;
-  }
 }
 
 nvinfer1::ITensor *TensorRTEngine::DeclareInput(const std::string &name,
@@ -158,83 +116,6 @@ void TensorRTEngine::DeclareOutput(const std::string &name) {
   buffer_sizes_[name] = 0;
 }
 
-void *TensorRTEngine::GetOutputInGPU(const std::string &name) {
-  return buffer(name).buffer;
-}
-
-void TensorRTEngine::GetOutputInGPU(const std::string &name, void *dst,
-                                    size_t max_size) {
-  // determine data size
-  auto *output = TensorRTEngine::GetITensor(name);
-  nvinfer1::Dims dims = output->getDimensions();
-  auto dim_size = analysis::AccuDims(dims.d, dims.nbDims);
-  size_t dst_size = dim_size * runtime_batch_ *
-                    kDataTypeSize[static_cast<int>(output->getType())];
-
-  auto it = buffer_sizes_.find(name);
-  PADDLE_ENFORCE(it != buffer_sizes_.end());
-  PADDLE_ENFORCE_GT(it->second, 0);
-  PADDLE_ENFORCE_LE(dst_size, it->second);
-  PADDLE_ENFORCE_GE(max_size, dst_size);
-  auto &buf = buffer(name);
-  PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before");
-  PADDLE_ENFORCE_EQ(cudaMemcpyAsync(dst, buf.buffer, dst_size,
-                                    cudaMemcpyDeviceToDevice, stream_),
-                    0);
-}
-
-void TensorRTEngine::GetOutputInCPU(const std::string &name, void *dst,
-                                    size_t max_size) {
-  // determine data size
-
-  auto *output = TensorRTEngine::GetITensor(name);
-  nvinfer1::Dims dims = output->getDimensions();
-  auto dim_size = analysis::AccuDims(dims.d, dims.nbDims);
-  size_t dst_size = dim_size * runtime_batch_ *
-                    kDataTypeSize[static_cast<int>(output->getType())];
-  auto it = buffer_sizes_.find(name);
-  PADDLE_ENFORCE(it != buffer_sizes_.end());
-  PADDLE_ENFORCE_GT(it->second, 0);
-  PADDLE_ENFORCE_LE(dst_size, it->second);
-  PADDLE_ENFORCE_GE(max_size, dst_size);
-  auto &buf = buffer(name);
-  PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before");
-  PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buf.buffer, dst_size,
-                                       cudaMemcpyDeviceToHost, stream_));
-}
-
-Buffer &TensorRTEngine::buffer(const std::string &name) {
-  PADDLE_ENFORCE(infer_engine_ != nullptr, "call FreezeNetwork first.");
-  auto it = buffer_sizes_.find(name);
-  PADDLE_ENFORCE(it != buffer_sizes_.end(), "tried to access buffer named %s",
-                 name);
-  auto slot_offset = infer_engine_->getBindingIndex(name.c_str());
-  return buffers_[slot_offset];
-}
-
-void TensorRTEngine::SetInputFromCPU(const std::string &name, const void *data,
-                                     size_t size) {
-  auto &buf = buffer(name);
-  PADDLE_ENFORCE_NOT_NULL(buf.buffer);
-  PADDLE_ENFORCE_NOT_NULL(data);
-  PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small");
-  PADDLE_ENFORCE(buf.device == DeviceType::GPU);
-  buf.size = size;
-  PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(buf.buffer, data, size,
-                                       cudaMemcpyHostToDevice, stream_));
-}
-
-void TensorRTEngine::SetInputFromGPU(const std::string &name, const void *data,
-                                     size_t size) {
-  auto &buf = buffer(name);
-  buf.size = size;
-  PADDLE_ENFORCE_NOT_NULL(buf.buffer);
-  PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small");
-  PADDLE_ENFORCE(buf.device == DeviceType::GPU);
-  PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(buf.buffer, data, size,
-                                       cudaMemcpyDeviceToDevice, stream_));
-}
-
 void TensorRTEngine::SetITensor(const std::string &name,
                                 nvinfer1::ITensor *tensor) {
   PADDLE_ENFORCE(tensor != nullptr);
@@ -254,13 +135,6 @@ void TensorRTEngine::SetRuntimeBatch(size_t batch_size) {
 
 int TensorRTEngine::GetRuntimeBatch() { return runtime_batch_; }
 
-void TensorRTEngine::freshDeviceId() {
-  int count;
-  cudaGetDeviceCount(&count);
-  PADDLE_ENFORCE_LT(device_, count);
-  cudaSetDevice(device_);
-}
-
 nvinfer1::IPluginLayer *TensorRTEngine::AddPlugin(
     nvinfer1::ITensor *const *inputs, int num_inputs,
     plugin::PluginTensorRT *plugin) {
@@ -268,6 +142,13 @@ nvinfer1::IPluginLayer *TensorRTEngine::AddPlugin(
   return infer_network_.get()->addPluginExt(inputs, num_inputs, *plugin);
 }
 
+void TensorRTEngine::freshDeviceId() {
+  int count;
+  cudaGetDeviceCount(&count);
+  PADDLE_ENFORCE_LT(device_id_, count);
+  cudaSetDevice(device_id_);
+}
+
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tensorrt/engine.h

@@ -23,6 +23,7 @@ limitations under the License. */
 #include "paddle/fluid/inference/engine.h"
 #include "paddle/fluid/inference/tensorrt/helper.h"
 #include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin_factory.h"
 #include "paddle/fluid/inference/tensorrt/trt_int8_calibrator.h"
 #include "paddle/fluid/inference/utils/singleton.h"
 
@@ -37,7 +38,9 @@ class TRTInt8Calibrator;
  * There are two alternative ways to use it, one is  to build from a paddle
  * protobuf model, another way is to manully construct the network.
  */
-class TensorRTEngine : public EngineBase {
+class TensorRTEngine {
+  using DescType = ::paddle::framework::proto::BlockDesc;
+
  public:
   // Weight is model parameter.
   class Weight {
@@ -56,28 +59,28 @@ class TensorRTEngine : public EngineBase {
     nvinfer1::Weights w_;
   };
 
-  TensorRTEngine(int max_batch, int max_workspace, cudaStream_t stream,
-                 int device = 0, bool enable_int8 = false,
-                 TRTInt8Calibrator* calibrator = nullptr,
+  TensorRTEngine(int max_batch, int max_workspace, bool enable_int8 = false,
+                 TRTInt8Calibrator* calibrator = nullptr, int device_id = 0,
                  nvinfer1::ILogger& logger = NaiveLogger::Global())
       : max_batch_(max_batch),
         max_workspace_(max_workspace),
-        stream_(stream),
-        device_(device),
         enable_int8_(enable_int8),
         calibrator_(calibrator),
+        device_id_(device_id),
         logger_(logger) {}
 
-  virtual ~TensorRTEngine();
+  ~TensorRTEngine() {}
 
   // TODO(Superjomn) implement it later when graph segmentation is supported.
-  void Build(const DescType& paddle_model) override;
+  void Build(const DescType& paddle_model);
 
-  void Execute(int batch_size) override;
+  void Execute(int batch_size, std::vector<void*>* buffers,
+               cudaStream_t stream);
 
   // Initialize the inference network, so that TensorRT layers can add to this
   // network.
   void InitNetwork() {
+    freshDeviceId();
     infer_builder_.reset(createInferBuilder(&logger_));
     infer_network_.reset(infer_builder_->createNetwork());
   }
@@ -98,37 +101,34 @@ class TensorRTEngine : public EngineBase {
   // Check if the ITensor has been declared
   bool HasDeclared(const std::string& name);
 
-  // GPU memory address for an ITensor with specific name. One can operate on
-  // these memory directly for acceleration, for example, output the converted
-  // data directly to the buffer to save data copy overhead.
-  // NOTE this should be used after calling `FreezeNetwork`.
-  Buffer& buffer(const std::string& name) override;
-
-  cudaStream_t stream() { return stream_; }
-
-  // Fill an input from CPU memory with name and size.
-  void SetInputFromCPU(const std::string& name, const void* data, size_t size);
-  // TODO(Superjomn) is this method necessary given that buffer(xxx) can be
-  // accessed directly. Fill an input from GPU memory with name and size.
-  void SetInputFromGPU(const std::string& name, const void* data, size_t size);
-  // Get an output called name, the output of tensorrt is in GPU, so this method
-  // Return the output's GPU memory address without copy.
-  void* GetOutputInGPU(const std::string& name);
-  // Copy data into dst inside the GPU device.
-  void GetOutputInGPU(const std::string& name, void* dst, size_t max_size);
-  // LOW EFFICENCY! Get output to CPU, this will trigger a memory copy from GPU
-  // to CPU.
-  void GetOutputInCPU(const std::string& name, void* dst, size_t max_size);
-  // Fill an ITensor into map itensor_map_.
   void SetITensor(const std::string& name, nvinfer1::ITensor* tensor);
   // Get an ITensor called name.
   nvinfer1::ITensor* GetITensor(const std::string& name);
 
   nvinfer1::ICudaEngine* engine() { return infer_engine_.get(); }
   nvinfer1::INetworkDefinition* network() { return infer_network_.get(); }
+
+  nvinfer1::IHostMemory* Serialize() {
+    PADDLE_ENFORCE(infer_engine_ != nullptr,
+                   "You should build engine first and then serialize");
+    ihost_memory_.reset(infer_engine_->serialize());
+    return ihost_memory_.get();
+  }
+
+  void Deserialize(const std::string& engine_serialized_data) {
+    freshDeviceId();
+    infer_ptr<nvinfer1::IRuntime> runtime(createInferRuntime(&logger_));
+    infer_engine_.reset(runtime->deserializeCudaEngine(
+        engine_serialized_data.c_str(), engine_serialized_data.size(),
+        &inference::Singleton<plugin::PluginFactoryTensorRT>::Global()));
+    PADDLE_ENFORCE(infer_engine_ != nullptr,
+                   "build cuda engine failed when deserialize engine info.!");
+    infer_context_.reset(infer_engine_->createExecutionContext());
+  }
+
   void SetRuntimeBatch(size_t batch_size);
   int GetRuntimeBatch();
-  int GetDevice() { return device_; }
+  int GetDeviceId() { return device_id_; }
   nvinfer1::IPluginLayer* AddPlugin(nvinfer1::ITensor* const* inputs,
                                     int num_inputs, plugin::PluginTensorRT*);
 
@@ -140,17 +140,12 @@ class TensorRTEngine : public EngineBase {
   std::unordered_map<std::string /*name*/, std::unique_ptr<framework::Tensor>>
       weight_map;
 
-  // TODO(NHZLX)
-  // In the normal case, the paddle-trt exists bug when runing the googlenet.
-  // When there are more than two convolutions of 1 * 1 with the same input, the
-  // paddle-tensorrt will do the merging optimization, which fuse those conv
-  // into one conv, and then trigger bug. So,  We should use strategy to avoid
-  // this
-  // optimization for the time being. This bug will be fixed in the future.
-  std::unordered_map<std::string /*name*/, int /*ITensor_quote_num*/>
-      itensor_quote_num;
-
  private:
+  // Each ICudaEngine object is bound to a specific GPU when it is instantiated,
+  // ensure that the thread is associated with the correct device by calling
+  // freshDeviceId().
+  void freshDeviceId();
+
   // the max batch size
   int max_batch_;
   // the runtime batch size
@@ -158,18 +153,14 @@ class TensorRTEngine : public EngineBase {
   // the max memory size the engine uses
   int max_workspace_;
 
-  cudaStream_t stream_;
-  // The specific GPU id that the TensorRTEngine bounded to.
-  int device_;
-
   bool enable_int8_;
   TRTInt8Calibrator* calibrator_;
   // batch size of the current data, will be updated each Executation.
   int batch_size_{-1};
 
+  int device_id_;
   nvinfer1::ILogger& logger_;
 
-  std::vector<Buffer> buffers_;
   // max data size for the buffers.
   std::unordered_map<std::string /*name*/, size_t /*max size*/> buffer_sizes_;
   std::unordered_map<std::string /*name*/, nvinfer1::ITensor* /*ITensor*/>
@@ -192,15 +183,11 @@ class TensorRTEngine : public EngineBase {
   infer_ptr<nvinfer1::INetworkDefinition> infer_network_;
   infer_ptr<nvinfer1::ICudaEngine> infer_engine_;
   infer_ptr<nvinfer1::IExecutionContext> infer_context_;
-  // Each ICudaEngine object is bound to a specific GPU when it is instantiated,
-  // ensure that the thread is associated with the correct device by calling
-  // freshDeviceId().
-  void freshDeviceId();
+  infer_ptr<nvinfer1::IHostMemory> ihost_memory_;
 };  // class TensorRTEngine
 
 // Add an layer__ into engine__ with args ARGS.
 // For example:
-//   TRT_ENGINE_ADD_LAYER(xxx, FullyConnected, input, dim, weights, bias)
 //
 // Reference
 // https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#charRNN_define_network

paddle/fluid/inference/tensorrt/helper.h

@@ -17,6 +17,9 @@
 #include <NvInfer.h>
 #include <cuda.h>
 #include <glog/logging.h>
+#include <string>
+#include <utility>
+#include <vector>
 #include "paddle/fluid/platform/dynload/tensorrt.h"
 #include "paddle/fluid/platform/enforce.h"
 
@@ -74,6 +77,32 @@ class NaiveLogger : public nvinfer1::ILogger {
   ~NaiveLogger() override {}
 };
 
+class NaiveProfiler : public nvinfer1::IProfiler {
+ public:
+  typedef std::pair<std::string, float> Record;
+  std::vector<Record> mProfile;
+
+  virtual void reportLayerTime(const char* layerName, float ms) {
+    auto record =
+        std::find_if(mProfile.begin(), mProfile.end(),
+                     [&](const Record& r) { return r.first == layerName; });
+    if (record == mProfile.end())
+      mProfile.push_back(std::make_pair(layerName, ms));
+    else
+      record->second += ms;
+  }
+
+  void printLayerTimes() {
+    float totalTime = 0;
+    for (size_t i = 0; i < mProfile.size(); i++) {
+      printf("%-40.40s %4.3fms\n", mProfile[i].first.c_str(),
+             mProfile[i].second);
+      totalTime += mProfile[i].second;
+    }
+    printf("Time over all layers: %4.3f\n", totalTime);
+  }
+};
+
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/tensorrt/plugin/CMakeLists.txt

 nv_library(tensorrt_plugin
-           SRCS trt_plugin.cc split_op_plugin.cu elementwise_op_plugin.cu prelu_op_plugin.cu
+           SRCS trt_plugin.cc split_op_plugin.cu elementwise_op_plugin.cu
+           prelu_op_plugin.cu  trt_plugin_factory.cc
            avg_pool_op_plugin.cu
            DEPS enforce tensorrt_engine prelu)

paddle/fluid/inference/tensorrt/plugin/avg_pool_op_plugin.cu

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/inference/tensorrt/plugin/avg_pool_op_plugin.h"
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin_factory.h"
 #include "paddle/fluid/operators/math/pooling.h"
 
 namespace paddle {
@@ -20,6 +21,12 @@ namespace inference {
 namespace tensorrt {
 namespace plugin {
 
+AvgPoolPlugin* CreateAvgPoolPluginDeserialize(const void* buffer,
+                                              size_t length) {
+  return new AvgPoolPlugin(buffer, length);
+}
+REGISTER_TRT_PLUGIN("avg_pool_plugin", CreateAvgPoolPluginDeserialize);
+
 nvinfer1::Dims AvgPoolPlugin::getOutputDimensions(
     int index, const nvinfer1::Dims* inputDims, int nbInputs) {
   assert(nbInputs == 1);

paddle/fluid/inference/tensorrt/plugin/avg_pool_op_plugin.h

@@ -33,24 +33,27 @@ class AvgPoolPlugin : public PluginTensorRT {
 
  protected:
   size_t getSerializationSize() override {
-    return SerializedSize(ceil_mode_) + SerializedSize(ksize_) +
-           SerializedSize(strides_) + SerializedSize(paddings_) +
-           SerializedSize(input_shape_) + getBaseSerializationSize();
+    return SerializedSize(getPluginType()) + SerializedSize(ceil_mode_) +
+           SerializedSize(ksize_) + SerializedSize(strides_) +
+           SerializedSize(paddings_) + SerializedSize(input_shape_) +
+           SerializedSize(output_shape_) + getBaseSerializationSize();
   }
 
   // TRT will call this func when we need to serialize the configuration of
   // tensorrt.
-  // It should not be called by users.
   void serialize(void *buffer) override {
+    SerializeValue(&buffer, getPluginType());
     serializeBase(buffer);
     SerializeValue(&buffer, ceil_mode_);
     SerializeValue(&buffer, ksize_);
     SerializeValue(&buffer, strides_);
     SerializeValue(&buffer, paddings_);
     SerializeValue(&buffer, input_shape_);
+    SerializeValue(&buffer, output_shape_);
   }
 
  public:
+  AvgPoolPlugin() {}
   AvgPoolPlugin(bool ceil_mode, std::vector<int> ksize,
                 std::vector<int> strides, std::vector<int> paddings,
                 std::vector<int> input_shape)
@@ -89,6 +92,7 @@ class AvgPoolPlugin : public PluginTensorRT {
     DeserializeValue(&serialData, &serialLength, &strides_);
     DeserializeValue(&serialData, &serialLength, &paddings_);
     DeserializeValue(&serialData, &serialLength, &input_shape_);
+    DeserializeValue(&serialData, &serialLength, &output_shape_);
   }
 
   AvgPoolPlugin *clone() const override {
@@ -96,7 +100,7 @@ class AvgPoolPlugin : public PluginTensorRT {
                              input_shape_);
   }
 
-  const char *getPluginType() const override { return "avg_pool"; }
+  const char *getPluginType() const override { return "avg_pool_plugin"; }
   int getNbOutputs() const override { return 1; }
   nvinfer1::Dims getOutputDimensions(int index, const nvinfer1::Dims *inputs,
                                      int nbInputDims) override;

paddle/fluid/inference/tensorrt/plugin/elementwise_op_plugin.cu

@@ -14,12 +14,19 @@ limitations under the License. */
 
 #include <glog/logging.h>
 #include "paddle/fluid/inference/tensorrt/plugin/elementwise_op_plugin.h"
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin_factory.h"
 
 namespace paddle {
 namespace inference {
 namespace tensorrt {
 namespace plugin {
 
+ElementWisePlugin* CreateElementWisePluginDeserialize(const void* buffer,
+                                                      size_t length) {
+  return new ElementWisePlugin(buffer, length);
+}
+REGISTER_TRT_PLUGIN("elementwise_plugin", CreateElementWisePluginDeserialize);
+
 namespace details {
 
 template <typename T>
@@ -119,10 +126,10 @@ int ElementWisePlugin::enqueue(int batch_size, const void* const* inputs,
   const float* y = reinterpret_cast<const float*>(inputs[1]);
   float* out = reinterpret_cast<float*>(outputs[0]);
 
-  if (type_ == nvinfer1::ElementWiseOperation::kSUM) {
+  if (type_ == "add") {
     details::ElementWise(details::Add<float>(), x, y, out, batch_size,
                          prev_size_, midd_size_, post_size_, stream);
-  } else if (type_ == nvinfer1::ElementWiseOperation::kPROD) {
+  } else if (type_ == "mul") {
     details::ElementWise(details::Mul<float>(), x, y, out, batch_size,
                          prev_size_, midd_size_, post_size_, stream);
   } else {

paddle/fluid/inference/tensorrt/plugin/elementwise_op_plugin.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
 #include <vector>
 #include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
 
@@ -24,9 +25,8 @@ namespace plugin {
 
 class ElementWisePlugin : public PluginTensorRT {
  public:
-  ElementWisePlugin(nvinfer1::ElementWiseOperation type,
-                    nvinfer1::Dims const &dims_x, nvinfer1::Dims const &dims_y,
-                    int axis)
+  ElementWisePlugin(std::string type, nvinfer1::Dims const &dims_x,
+                    nvinfer1::Dims const &dims_y, int axis)
       : type_(type),
         dims_x_(dims_x),
         dims_y_(dims_y),
@@ -37,6 +37,9 @@ class ElementWisePlugin : public PluginTensorRT {
 
   ElementWisePlugin(void const *serial_data, size_t serial_length) {
     deserializeBase(serial_data, serial_length);
+    const char *elementwise_type;
+    DeserializeValue(&serial_data, &serial_length, &elementwise_type);
+    type_ = std::string(elementwise_type);
     DeserializeValue(&serial_data, &serial_length, &axis_);
     DeserializeValue(&serial_data, &serial_length, &dims_x_);
     DeserializeValue(&serial_data, &serial_length, &dims_y_);
@@ -47,7 +50,7 @@ class ElementWisePlugin : public PluginTensorRT {
     return nullptr;
   }
 
-  const char *getPluginType() const override { return "elementwise"; }
+  const char *getPluginType() const override { return "elementwise_plugin"; }
 
   nvinfer1::Dims getOutputDimensions(int index,
                                      const nvinfer1::Dims *input_dims,
@@ -61,18 +64,21 @@ class ElementWisePlugin : public PluginTensorRT {
 
  protected:
   size_t getSerializationSize() override {
-    return SerializedSize(axis_) + SerializedSize(dims_x_) +
-           SerializedSize(dims_y_) + getBaseSerializationSize();
+    return SerializedSize(getPluginType()) + SerializedSize(axis_) +
+           SerializedSize(dims_x_) + SerializedSize(dims_y_) +
+           getBaseSerializationSize();
   }
 
   void serialize(void *buffer) override {
+    SerializeValue(&buffer, getPluginType());
     serializeBase(buffer);
+    SerializeValue(&buffer, type_.c_str());
     SerializeValue(&buffer, axis_);
     SerializeValue(&buffer, dims_x_);
     SerializeValue(&buffer, dims_y_);
   }
 
-  nvinfer1::ElementWiseOperation type_;
+  std::string type_;
   nvinfer1::Dims dims_x_;
   nvinfer1::Dims dims_y_;
   int axis_;

paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.cu

@@ -17,6 +17,7 @@
 #include <vector>
 #include "glog/logging.h"
 #include "paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.h"
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin_factory.h"
 #include "paddle/fluid/operators/math/prelu.h"
 
 namespace paddle {
@@ -24,6 +25,17 @@ namespace inference {
 namespace tensorrt {
 namespace plugin {
 
+PReluPlugin *CreatePreluPluginDeserialize(const void *buffer, size_t length) {
+  return new PReluPlugin(buffer, length);
+}
+REGISTER_TRT_PLUGIN("prelu_plugin", CreatePreluPluginDeserialize);
+
+int PReluPlugin::initialize() {
+  cudaMalloc(&p_gpu_weight_, sizeof(float) * weight_.size());
+  cudaMemcpy(p_gpu_weight_, weight_.data(), weight_.size() * sizeof(float),
+             cudaMemcpyHostToDevice);
+}
+
 nvinfer1::Dims PReluPlugin::getOutputDimensions(int index,
                                                 const nvinfer1::Dims *inputDims,
                                                 int nbInputs) {
@@ -39,7 +51,8 @@ int PReluPlugin::enqueue(int batch_size, const void *const *inputs,
   // input dims is CHW.
   const auto &input_dims = this->getInputDims(0);
   const float *input = reinterpret_cast<const float *>(inputs[0]);
-  const float *alpha = reinterpret_cast<const float *>(alpha_.get().values);
+  // const float *alpha = reinterpret_cast<const float *>(alpha_.get().values);
+  const float *alpha = p_gpu_weight_;
   float *output = reinterpret_cast<float **>(outputs)[0];
 
   std::vector<int> input_shape;

paddle/fluid/inference/tensorrt/plugin/prelu_op_plugin.h

@@ -14,7 +14,12 @@
 
 #pragma once
 
+#include <algorithm>
 #include <string>
+#include <vector>
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/framework/tensor_util.h"
+
 #include "paddle/fluid/inference/tensorrt/engine.h"
 #include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
 
@@ -24,39 +29,51 @@ namespace tensorrt {
 namespace plugin {
 
 class PReluPlugin : public PluginTensorRT {
-  TensorRTEngine::Weight alpha_;
+  std::vector<float> weight_;
+  float *p_gpu_weight_;
   std::string mode_;
 
  protected:
   size_t getSerializationSize() override {
-    // return getBaseSerializationSize(alpha_) + SerializedSize(mode_);
-    return 0;
+    return getBaseSerializationSize() + SerializedSize(mode_.c_str()) +
+           SerializedSize(weight_) + SerializedSize(getPluginType());
   }
 
   // TRT will call this func when we need to serialize the configuration of
   // tensorrt.
   // It should not be called by users.
   void serialize(void *buffer) override {
-    // serializeBase(buffer);
-    // SerializeValue(&buffer, alpha_);
-    // SerializeValue(&buffer, mode_);
+    SerializeValue(&buffer, getPluginType());
+    serializeBase(buffer);
+    SerializeValue(&buffer, weight_);
+    SerializeValue(&buffer, mode_.c_str());
   }
 
  public:
-  PReluPlugin(TensorRTEngine::Weight const &alpha, std::string const &mode)
-      : alpha_(alpha), mode_(mode) {}
+  PReluPlugin(const float *weight, const int weight_num,
+              std::string const &mode)
+      : mode_(mode) {
+    weight_.resize(weight_num);
+    std::copy(weight, weight + weight_num, weight_.data());
+  }
 
   // It was used for tensorrt deserialization.
   // It should not be called by users.
   PReluPlugin(void const *serialData, size_t serialLength) {
-    // deserializeBase(serialData, serialLength);
-    // DeserializeValue(&serialData, &serialLength, &alpha_);
-    // DeserializeValue(&serialData, &serialLength, &mode_);
+    deserializeBase(serialData, serialLength);
+    DeserializeValue(&serialData, &serialLength, &weight_);
+    const char *prelu_mode;
+    DeserializeValue(&serialData, &serialLength, &prelu_mode);
+    mode_ = std::string(prelu_mode);
   }
+  ~PReluPlugin() { cudaFree(p_gpu_weight_); }
+  int initialize() override;
 
-  PReluPlugin *clone() const override { return new PReluPlugin(alpha_, mode_); }
+  PReluPlugin *clone() const override {
+    return new PReluPlugin(weight_.data(), weight_.size(), mode_);
+  }
 
-  const char *getPluginType() const override { return "prelu"; }
+  const char *getPluginType() const override { return "prelu_plugin"; }
   int getNbOutputs() const override { return 1; }
   nvinfer1::Dims getOutputDimensions(int index, const nvinfer1::Dims *inputs,
                                      int nbInputDims) override;

paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu

@@ -15,12 +15,18 @@
 #include <cuda_fp16.h>
 #include <algorithm>
 #include "paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h"
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin_factory.h"
 
 namespace paddle {
 namespace inference {
 namespace tensorrt {
 namespace plugin {
 
+SplitPlugin* CreateSplitPluginDeserialize(const void* buffer, size_t length) {
+  return new SplitPlugin(buffer, length);
+}
+REGISTER_TRT_PLUGIN("split_plugin", CreateSplitPluginDeserialize);
+
 // copied from operators::math::SplitFunctor
 template <typename T>
 __global__ void SplitKernel(const T* input_data, const int in_row,

paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h

@@ -15,6 +15,7 @@
 #pragma once
 
 #include <thrust/device_vector.h>
+#include <utility>
 #include <vector>
 #include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
 
@@ -25,6 +26,7 @@ namespace plugin {
 
 class SplitPlugin : public PluginTensorRT {
  public:
+  SplitPlugin() {}
   SplitPlugin(int axis, std::vector<int> const &output_lengths)
       : axis_(axis), same_shape_(true), output_length_(output_lengths) {}
 
@@ -38,7 +40,7 @@ class SplitPlugin : public PluginTensorRT {
     return new SplitPlugin(axis_, output_length_);
   }
 
-  const char *getPluginType() const override { return "split"; }
+  const char *getPluginType() const override { return "split_plugin"; }
   int getNbOutputs() const override { return output_length_.size(); }
   nvinfer1::Dims getOutputDimensions(int index,
                                      const nvinfer1::Dims *input_dims,
@@ -50,11 +52,12 @@ class SplitPlugin : public PluginTensorRT {
 
  protected:
   size_t getSerializationSize() override {
-    return SerializedSize(axis_) + SerializedSize(output_length_) +
-           getBaseSerializationSize();
+    return SerializedSize(getPluginType()) + SerializedSize(axis_) +
+           SerializedSize(output_length_) + getBaseSerializationSize();
   }
 
   void serialize(void *buffer) override {
+    SerializeValue(&buffer, getPluginType());
     serializeBase(buffer);
     SerializeValue(&buffer, axis_);
     SerializeValue(&buffer, output_length_);

paddle/fluid/inference/tensorrt/plugin/trt_plugin.h

@@ -17,9 +17,10 @@
 #include <NvInfer.h>
 #include <cstring>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 
-#include "paddle/fluid/inference/tensorrt/plugin/serialize.h"
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin_utils.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -30,6 +31,13 @@ namespace inference {
 namespace tensorrt {
 namespace plugin {
 
+class PluginTensorRT;
+
+typedef std::function<PluginTensorRT*(const void*, size_t)>
+    PluginDeserializeFunc;
+
+typedef std::function<PluginTensorRT*(void)> PluginConstructFunc;
+
 class PluginTensorRT : public nvinfer1::IPluginExt {
  public:
   PluginTensorRT() {}

paddle/fluid/inference/tensorrt/plugin/trt_plugin_factory.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin_factory.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+namespace plugin {
+
+PluginTensorRT* PluginFactoryTensorRT::createPlugin(const char* layer_name,
+                                                    const void* serial_data,
+                                                    size_t serial_length) {
+  const char* plugin_type;
+  DeserializeValue(&serial_data, &serial_length, &plugin_type);
+
+  PADDLE_ENFORCE(Has(plugin_type),
+                 "trt plugin type %s does not exists, check it.", plugin_type);
+  auto plugin = plugin_registry_[plugin_type](serial_data, serial_length);
+  owned_plugins_.emplace_back(plugin);
+
+  return plugin;
+}
+
+bool PluginFactoryTensorRT::RegisterPlugin(
+    const std::string& op_name, PluginDeserializeFunc deserialize_func) {
+  if (Has(op_name)) return false;
+  auto ret = plugin_registry_.emplace(op_name, deserialize_func);
+  return ret.second;
+}
+
+void PluginFactoryTensorRT::DestroyPlugins() { owned_plugins_.clear(); }
+
+}  // namespace plugin
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tensorrt/plugin/trt_plugin_factory.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <NvInfer.h>
+#include <cstring>
+#include <list>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
+#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin_utils.h"
+#include "paddle/fluid/inference/utils/singleton.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+namespace plugin {
+
+class PluginFactoryTensorRT : public nvinfer1::IPluginFactory,
+                              public DeleteHelper {
+ public:
+  // Deserialization method
+  PluginTensorRT* createPlugin(const char* layer_name, const void* serial_data,
+                               size_t serial_length) override;
+
+  bool RegisterPlugin(const std::string& op_name,
+                      PluginDeserializeFunc deserialize_func);
+
+  bool Has(const std::string& op_name) {
+    return plugin_registry_.find(op_name) != plugin_registry_.end();
+  }
+
+  void DestroyPlugins();
+
+ protected:
+  std::unordered_map<std::string, PluginDeserializeFunc> plugin_registry_;
+
+  std::list<std::unique_ptr<PluginTensorRT>> owned_plugins_;
+};
+
+class TrtPluginRegistrar {
+ public:
+  TrtPluginRegistrar(const std::string& name,
+                     PluginDeserializeFunc deserialize_func) {
+    inference::Singleton<PluginFactoryTensorRT>::Global().RegisterPlugin(
+        name, deserialize_func);
+  }
+};
+
+#define REGISTER_TRT_PLUGIN(name, deserialize_func) \
+  REGISTER_TRT_PLUGIN_UNIQ(__COUNTER__, name, deserialize_func)
+
+#define REGISTER_TRT_PLUGIN_UNIQ(ctr, name, deserialize_func)      \
+  static paddle::inference::tensorrt::plugin::TrtPluginRegistrar   \
+      trt_plugin_registrar##ctr __attribute__((unused)) =          \
+          paddle::inference::tensorrt::plugin::TrtPluginRegistrar( \
+              name, deserialize_func)
+
+}  // namespace plugin
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tensorrt/plugin/serialize.h → paddle/fluid/inference/tensorrt/plugin/trt_plugin_utils.h

@@ -13,8 +13,8 @@
 // limitations under the License.
 
 #pragma once
-
 #include <cstring>
+#include <string>
 #include <type_traits>
 #include <vector>
 #include "paddle/fluid/platform/enforce.h"
@@ -24,6 +24,13 @@ namespace inference {
 namespace tensorrt {
 namespace plugin {
 
+// Some trt base classes lack of the destructor.
+// We use a assisted class to fix this.
+struct DeleteHelper {
+ protected:
+  virtual ~DeleteHelper() {}
+};
+
 template <typename T>
 inline void SerializeValue(void** buffer, T const& value);
 

paddle/fluid/inference/tensorrt/test_engine.cc

@@ -17,6 +17,8 @@ limitations under the License. */
 #include <glog/logging.h>
 #include <gtest/gtest.h>
 
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/inference/tensorrt/engine.h"
 #include "paddle/fluid/platform/enforce.h"
 
@@ -27,19 +29,34 @@ namespace tensorrt {
 class TensorRTEngineTest : public ::testing::Test {
  protected:
   void SetUp() override {
-    ASSERT_EQ(0, cudaStreamCreate(&stream_));
-    engine_ = new TensorRTEngine(10, 1 << 10, stream_);
+    ctx_ = new platform::CUDADeviceContext(platform::CUDAPlace(0));
+
+    engine_ = new TensorRTEngine(10, 1 << 10);
     engine_->InitNetwork();
   }
 
   void TearDown() override {
-    delete engine_;
-    cudaStreamDestroy(stream_);
+    if (engine_) {
+      delete engine_;
+      engine_ = nullptr;
+    }
+  }
+
+  void PrepareInputOutput(const std::vector<float> &input,
+                          std::vector<int> output_shape) {
+    TensorFromVector(input, *ctx_, &input_);
+    output_.Resize(framework::make_ddim(output_shape));
+  }
+
+  void GetOutput(std::vector<float> *output) {
+    TensorToVector(output_, *ctx_, output);
   }
 
  protected:
-  TensorRTEngine* engine_;
-  cudaStream_t stream_;
+  framework::Tensor input_;
+  framework::Tensor output_;
+  TensorRTEngine *engine_;
+  platform::CUDADeviceContext *ctx_;
 };
 
 TEST_F(TensorRTEngineTest, add_layer) {
@@ -48,12 +65,14 @@ TEST_F(TensorRTEngineTest, add_layer) {
   float raw_weight[size] = {2.};  // Weight in CPU memory.
   float raw_bias[size] = {3.};
 
+  std::vector<void *> buffers(2);  // TRT binded inputs
+
   LOG(INFO) << "create weights";
   TensorRTEngine::Weight weight(nvinfer1::DataType::kFLOAT, raw_weight, size);
   TensorRTEngine::Weight bias(nvinfer1::DataType::kFLOAT, raw_bias, size);
-  auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
+  auto *x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
                                   nvinfer1::DimsCHW{1, 1, 1});
-  auto* fc_layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, *x, size,
+  auto *fc_layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, *x, size,
                                         weight.get(), bias.get());
   PADDLE_ENFORCE(fc_layer != nullptr);
 
@@ -63,18 +82,24 @@ TEST_F(TensorRTEngineTest, add_layer) {
   ASSERT_EQ(engine_->engine()->getNbBindings(), 2);
 
   // fill in real data
-  float x_v = 1234;
-  engine_->SetInputFromCPU("x", reinterpret_cast<void*>(&x_v),
-                           1 * sizeof(float));
+  std::vector<float> x_v = {1234};
+  std::vector<float> y_cpu;
+  PrepareInputOutput(x_v, {1});
+
+  auto *x_v_gpu_data = input_.mutable_data<float>(ctx_->GetPlace());
+  auto *y_gpu_data = output_.mutable_data<float>(ctx_->GetPlace());
+
+  buffers[0] = reinterpret_cast<void *>(x_v_gpu_data);
+  buffers[1] = reinterpret_cast<void *>(y_gpu_data);
+
   LOG(INFO) << "to execute";
-  engine_->Execute(1);
+  engine_->Execute(1, &buffers, ctx_->stream());
 
   LOG(INFO) << "to get output";
-  float y_cpu;
-  engine_->GetOutputInCPU("y", &y_cpu, 1 * sizeof(float));
+  GetOutput(&y_cpu);
 
   LOG(INFO) << "to checkout output";
-  ASSERT_EQ(y_cpu, x_v * 2 + 3);
+  ASSERT_EQ(y_cpu[0], x_v[0] * 2 + 3);
 }
 
 TEST_F(TensorRTEngineTest, add_layer_multi_dim) {
@@ -83,12 +108,13 @@ TEST_F(TensorRTEngineTest, add_layer_multi_dim) {
   // instead of row-major, which is [[1.0, 1.1], [3.3, 4.4]]
   float raw_weight[4] = {1.0, 1.1, 3.3, 4.4};
   float raw_bias[2] = {1.3, 2.4};
+  std::vector<void *> buffers(2);  // TRT binded inputs
 
   TensorRTEngine::Weight weight(nvinfer1::DataType::kFLOAT, raw_weight, 4);
   TensorRTEngine::Weight bias(nvinfer1::DataType::kFLOAT, raw_bias, 2);
-  auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
+  auto *x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
                                   nvinfer1::DimsCHW{1, 2, 1});
-  auto* fc_layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, *x, 2,
+  auto *fc_layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected, *x, 2,
                                         weight.get(), bias.get());
   PADDLE_ENFORCE(fc_layer != nullptr);
 
@@ -96,19 +122,27 @@ TEST_F(TensorRTEngineTest, add_layer_multi_dim) {
   engine_->FreezeNetwork();
   ASSERT_EQ(engine_->engine()->getNbBindings(), 2);
 
-  float x_v[2] = {1.0, 2.0};
-  engine_->SetInputFromCPU("x", reinterpret_cast<void*>(&x_v),
-                           2 * sizeof(float));
-  engine_->Execute(1);
+  // fill in real data
+  std::vector<float> x_v = {1.0, 2.0};
+  std::vector<float> y_cpu;
+  PrepareInputOutput(x_v, {2});
+
+  auto *x_v_gpu_data = input_.mutable_data<float>(ctx_->GetPlace());
+  auto *y_gpu_data = output_.mutable_data<float>(ctx_->GetPlace());
+
+  buffers[0] = reinterpret_cast<void *>(x_v_gpu_data);
+  buffers[1] = reinterpret_cast<void *>(y_gpu_data);
+
+  engine_->Execute(1, &buffers, ctx_->stream());
 
   LOG(INFO) << "to get output";
-  float y_cpu[2] = {-1., -1.};
+  GetOutput(&y_cpu);
 
   auto dims = engine_->GetITensor("y")->getDimensions();
   ASSERT_EQ(dims.nbDims, 3);
   ASSERT_EQ(dims.d[0], 2);
   ASSERT_EQ(dims.d[1], 1);
-  engine_->GetOutputInCPU("y", &y_cpu[0], 2 * sizeof(float));
+
   ASSERT_EQ(y_cpu[0], 4.5);
   ASSERT_EQ(y_cpu[1], 14.5);
 }
@@ -117,12 +151,13 @@ TEST_F(TensorRTEngineTest, test_conv2d) {
   // Weight in CPU memory.
   float raw_weight[9] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
   float raw_bias[1] = {0};
+  std::vector<void *> buffers(2);  // TRT binded inputs
 
   TensorRTEngine::Weight weight(nvinfer1::DataType::kFLOAT, raw_weight, 9);
   TensorRTEngine::Weight bias(nvinfer1::DataType::kFLOAT, raw_bias, 1);
-  auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
+  auto *x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
                                   nvinfer1::Dims3{1, 3, 3});
-  auto* conv_layer =
+  auto *conv_layer =
       TRT_ENGINE_ADD_LAYER(engine_, Convolution, *x, 1, nvinfer1::DimsHW{3, 3},
                            weight.get(), bias.get());
   PADDLE_ENFORCE(conv_layer != nullptr);
@@ -133,28 +168,36 @@ TEST_F(TensorRTEngineTest, test_conv2d) {
   engine_->FreezeNetwork();
   ASSERT_EQ(engine_->engine()->getNbBindings(), 2);
 
-  float x_v[18] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0,
-                   1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
-  engine_->SetInputFromCPU("x", reinterpret_cast<void*>(&x_v),
-                           18 * sizeof(float));
-  engine_->Execute(2);
+  // fill in real data
+  std::vector<float> x_v = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0,
+                            1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
+  std::vector<float> y_cpu;
+  PrepareInputOutput(x_v, {18});
+
+  auto *x_v_gpu_data = input_.mutable_data<float>(ctx_->GetPlace());
+  auto *y_gpu_data = output_.mutable_data<float>(ctx_->GetPlace());
+
+  buffers[0] = reinterpret_cast<void *>(x_v_gpu_data);
+  buffers[1] = reinterpret_cast<void *>(y_gpu_data);
+
+  engine_->Execute(2, &buffers, ctx_->stream());
 
   LOG(INFO) << "to get output";
-  float* y_cpu = new float[18];
-  engine_->GetOutputInCPU("y", &y_cpu[0], 18 * sizeof(float));
+  GetOutput(&y_cpu);
+
   ASSERT_EQ(y_cpu[0], 4.0);
   ASSERT_EQ(y_cpu[1], 6.0);
 }
 
 TEST_F(TensorRTEngineTest, test_pool2d) {
   // Weight in CPU memory.
-  auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
+  auto *x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
                                   nvinfer1::Dims3{1, 2, 2});
 
+  std::vector<void *> buffers(2);  // TRT binded inputs
   nvinfer1::PoolingType pool_t = nvinfer1::PoolingType::kAVERAGE;
-  auto* pool_layer =
-      TRT_ENGINE_ADD_LAYER(engine_, Pooling, *const_cast<nvinfer1::ITensor*>(x),
-                           pool_t, nvinfer1::DimsHW{2, 2});
+  auto *pool_layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, *x, pool_t,
+                                          nvinfer1::DimsHW{2, 2});
 
   PADDLE_ENFORCE(pool_layer != nullptr);
   pool_layer->setStride(nvinfer1::DimsHW{1, 1});
@@ -164,14 +207,21 @@ TEST_F(TensorRTEngineTest, test_pool2d) {
   engine_->FreezeNetwork();
   ASSERT_EQ(engine_->engine()->getNbBindings(), 2);
 
-  float x_v[8] = {1.0, 2.0, 5.0, 0.0, 2.0, 3.0, 5.0, 10.0};
-  engine_->SetInputFromCPU("x", reinterpret_cast<void*>(&x_v),
-                           8 * sizeof(float));
-  engine_->Execute(2);
+  // fill in real data
+  std::vector<float> x_v = {1.0, 2.0, 5.0, 0.0, 2.0, 3.0, 5.0, 10.0};
+  std::vector<float> y_cpu;
+  PrepareInputOutput(x_v, {2});
+
+  auto *x_v_gpu_data = input_.mutable_data<float>(ctx_->GetPlace());
+  auto *y_gpu_data = output_.mutable_data<float>(ctx_->GetPlace());
+
+  buffers[0] = reinterpret_cast<void *>(x_v_gpu_data);
+  buffers[1] = reinterpret_cast<void *>(y_gpu_data);
+
+  engine_->Execute(2, &buffers, ctx_->stream());
 
   LOG(INFO) << "to get output";
-  float* y_cpu = new float[2];
-  engine_->GetOutputInCPU("y", &y_cpu[0], 2 * sizeof(float));
+  GetOutput(&y_cpu);
 
   ASSERT_EQ(y_cpu[0], 2.0);
   ASSERT_EQ(y_cpu[1], 5.0);

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

@@ -105,6 +105,13 @@ set(SEQ_CONV1_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/seq_conv1")
 download_model_and_data(${SEQ_CONV1_INSTALL_DIR} "seq_conv1_model.tar.gz" "seq_conv1_data.txt.tar.gz")
 inference_analysis_api_test(test_analyzer_seq_conv1 ${SEQ_CONV1_INSTALL_DIR} analyzer_seq_conv1_tester.cc)
 
+# transformer, the dataset only works on batch_size=8 now
+set(TRANSFORMER_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/transformer")
+download_model_and_data(${TRANSFORMER_INSTALL_DIR} "temp%2Ftransformer_model.tar.gz" "temp%2Ftransformer_data.txt.tar.gz")
+inference_analysis_test(test_analyzer_transformer SRCS analyzer_transformer_tester.cc 
+  EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
+  ARGS --infer_model=${TRANSFORMER_INSTALL_DIR}/model --infer_data=${TRANSFORMER_INSTALL_DIR}/data.txt --batch_size=8)
+
 # ocr
 set(OCR_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/ocr")
 if (NOT EXISTS ${OCR_INSTALL_DIR})

paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc

@@ -366,15 +366,17 @@ TEST(Analyzer_rnn1, ZeroCopyMultiThread) {
 #define NEW_TENSOR(name__) \
   auto name__##_tensor = predictor->GetInputTensor(#name__);
 
-  auto base_predictor = CreatePaddlePredictor<AnalysisConfig>(config);
+  std::vector<std::unique_ptr<PaddlePredictor>> predictors;
+  predictors.emplace_back(CreatePaddlePredictor<AnalysisConfig>(config));
+  for (int tid = 1; tid < FLAGS_num_threads; tid++) {
+    predictors.emplace_back(predictors.front()->Clone());
+  }
   double total_time_of_threads{0};
   std::vector<std::thread> threads;
 
   for (int tid = 0; tid < FLAGS_num_threads; tid++) {
     threads.emplace_back([&, tid] {
-      // To ensure the thread binding correctly,
-      // please clone inside the threadpool.
-      auto predictor = base_predictor->Clone();
+      auto &predictor = predictors[tid];
       NEW_TENSOR(data_lod_attention);
       NEW_TENSOR(cell_init);
       NEW_TENSOR(data);

paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc

@@ -266,15 +266,17 @@ TEST(Analyzer_seq_pool1, zerocopy_profile_threads) {
   SetConfig(&config);
   config.SwitchUseFeedFetchOps(false);
 
-  auto base_predictor = CreatePaddlePredictor<AnalysisConfig>(config);
+  std::vector<std::unique_ptr<PaddlePredictor>> predictors;
+  predictors.emplace_back(CreatePaddlePredictor<AnalysisConfig>(config));
+  for (int tid = 1; tid < FLAGS_num_threads; tid++) {
+    predictors.emplace_back(predictors.front()->Clone());
+  }
   double total_time_of_threads{0};
   std::vector<std::thread> threads;
 
   for (int tid = 0; tid < FLAGS_num_threads; tid++) {
     threads.emplace_back([&, tid] {
-      // To ensure the thread binding correctly,
-      // please clone inside the threadpool.
-      auto predictor = base_predictor->Clone();
+      auto &predictor = predictors[tid];
       std::vector<std::unique_ptr<ZeroCopyTensor>> inputs;
       PrepareZeroCopyInputs(predictor, &inputs);
       auto output_tensor = predictor->GetOutputTensor(out_var_name);

paddle/fluid/inference/tests/api/tester_helper.h

@@ -17,13 +17,14 @@
 #include <gtest/gtest.h>
 
 #include <algorithm>
+#include <memory>
 #include <string>
 #include <thread>  // NOLINT
+#include <unordered_map>
 #include <vector>
 #ifdef WITH_GPERFTOOLS
 #include <gperftools/profiler.h>
 #endif
-
 #include "paddle/fluid/framework/ir/fuse_pass_base.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/analysis/analyzer.h"
@@ -95,6 +96,14 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
         }
         break;
       }
+      case PaddleDType::INT32: {
+        int32_t *pdata = static_cast<int32_t *>(out.data.data());
+        int32_t *pdata_ref = static_cast<int32_t *>(ref_out.data.data());
+        for (size_t j = 0; j < size; ++j) {
+          EXPECT_EQ(pdata_ref[j], pdata[j]);
+        }
+        break;
+      }
     }
   }
 }
@@ -252,7 +261,11 @@ void TestMultiThreadPrediction(
   int batch_size = FLAGS_batch_size;
   int num_times = FLAGS_repeat;
   std::vector<std::thread> threads;
-  auto main_predictor = CreateTestPredictor(config, use_analysis);
+  std::vector<std::unique_ptr<PaddlePredictor>> predictors;
+  predictors.emplace_back(CreateTestPredictor(config, use_analysis));
+  for (int tid = 1; tid < num_threads; tid++) {
+    predictors.emplace_back(predictors.front()->Clone());
+  }
 
   size_t total_time{0};
   for (int tid = 0; tid < num_threads; ++tid) {
@@ -260,9 +273,7 @@ void TestMultiThreadPrediction(
       // Each thread should have local inputs and outputs.
       // The inputs of each thread are all the same.
       std::vector<PaddleTensor> outputs_tid;
-      // To ensure the thread binding correctly,
-      // please clone inside the threadpool.
-      auto predictor = main_predictor->Clone();
+      auto &predictor = predictors[tid];
 #ifdef PADDLE_WITH_MKLDNN
       if (use_analysis) {
         static_cast<AnalysisPredictor *>(predictor.get())

paddle/fluid/inference/tests/api/trt_models_tester.cc

@@ -54,7 +54,8 @@ void SetConfig<AnalysisConfig>(AnalysisConfig* config, std::string model_dir,
   if (use_gpu) {
     config->EnableUseGpu(100, 0);
     if (use_tensorrt) {
-      config->EnableTensorRtEngine(1 << 10, batch_size);
+      config->EnableTensorRtEngine(1 << 10, batch_size, 3,
+                                   AnalysisConfig::Precision::kFloat32, false);
       config->pass_builder()->DeletePass("conv_bn_fuse_pass");
       config->pass_builder()->DeletePass("fc_fuse_pass");
       config->pass_builder()->TurnOnDebug();

paddle/fluid/operators/beam_search_op.cc

@@ -51,9 +51,9 @@ class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("selected_scores",
               "A LoDTensor containing the accumulated scores corresponding to "
               "Output(selected_ids).");
-    AddOutput(
-        "parent_idx",
-        "A Tensor preserving the selected_ids' parent indice in pre_ids.");
+    AddOutput("parent_idx",
+              "A Tensor preserving the selected_ids' parent indice in pre_ids.")
+        .AsDispensable();
 
     // Attributes stored in AttributeMap
     AddAttr<int>("level", "the level of LoDTensor");

paddle/fluid/operators/beam_search_op.h

@@ -44,7 +44,6 @@ class BeamSearchOpKernel : public framework::OpKernel<T> {
     auto* parent_idx = context.Output<framework::Tensor>("parent_idx");
     PADDLE_ENFORCE_NOT_NULL(selected_ids);
     PADDLE_ENFORCE_NOT_NULL(selected_scores);
-    PADDLE_ENFORCE_NOT_NULL(parent_idx);
 
     math::BeamSearchFunctor<DeviceContext, T> alg;
     alg(context.template device_context<DeviceContext>(), pre_ids, pre_scores,