Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add-fake-reader

test=develop

cmake/cblas.cmake

@@ -40,7 +40,7 @@ set(OPENBLAS_LIB_SEARCH_PATHS
         /usr/local/opt/openblas/lib)
 
 find_path(OPENBLAS_INC_DIR NAMES cblas.h
-  PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})
+  PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS} NO_DEFAULT_PATH)
 find_path(OPENBLAS_LAPACKE_INC_DIR NAMES lapacke.h
   PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})
 find_library(OPENBLAS_LIB NAMES openblas

cmake/external/openblas.cmake

@@ -27,7 +27,7 @@ IF(NOT ${CBLAS_FOUND})
 
     SET(CBLAS_SOURCES_DIR ${THIRD_PARTY_PATH}/openblas)
     SET(CBLAS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/openblas)
-    SET(CBLAS_INCLUDE_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
+    SET(CBLAS_INC_DIR "${CBLAS_INSTALL_DIR}/include" CACHE PATH "openblas include directory." FORCE)
 
     SET(CBLAS_LIBRARIES
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
@@ -96,7 +96,7 @@ IF(NOT ${CBLAS_FOUND})
     ENDIF(NOT WIN32)
     SET(CBLAS_PROVIDER openblas)
     IF(WITH_C_API)
-        INSTALL(DIRECTORY ${CBLAS_INCLUDE_DIR} DESTINATION third_party/openblas)
+        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
         # Because libopenblas.a is a symbolic link of another library, thus need to
         # install the whole directory.
         IF(ANDROID)
@@ -117,8 +117,8 @@ IF(NOT ${CBLAS_FOUND})
 ENDIF(NOT ${CBLAS_FOUND})
 
 MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
-MESSAGE(STATUS "BLAS Include: ${CBLAS_INCLUDE_DIR}")
-INCLUDE_DIRECTORIES(${CBLAS_INCLUDE_DIR})
+MESSAGE(STATUS "BLAS Include: ${CBLAS_INC_DIR}")
+INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
 
 # FIXME(gangliao): generate cblas target to track all high performance
 # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)

paddle/fluid/framework/CMakeLists.txt

+
 # windows treat symbolic file as a real file, which is different with unix
 # We create a hidden file and compile it instead of origin source file.
 function(windows_symbolic TARGET)
@@ -9,11 +10,23 @@ function(windows_symbolic TARGET)
   if (NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc OR NOT EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cu)
       message(FATAL " ${src}.cc and ${src}.cu must exsits, and ${src}.cu must be symbolic file.")
   endif()
-  add_custom_command(OUTPUT .${src}.cu 
+
+  # only copy the xx.cu to .xx.cu when the content are modified
+  set(copy_flag 1)
+  if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu)
+  file(READ ${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc SOURCE_STR)
+  file(READ ${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu TARGET_STR)
+  if (SOURCE_STR STREQUAL TARGET_STR)
+    set(copy_flag 0)
+  endif()
+  endif()
+  if (copy_flag)
+  add_custom_command(OUTPUT .${src}.cu
           COMMAND ${CMAKE_COMMAND} -E remove ${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu
           COMMAND ${CMAKE_COMMAND} -E copy "${CMAKE_CURRENT_SOURCE_DIR}/${src}.cc" "${CMAKE_CURRENT_SOURCE_DIR}/.${src}.cu"
           COMMENT "create hidden file of ${src}.cu")
-  add_custom_target(${TARGET} ALL DEPENDS .${src}.cu)  
+  endif(copy_flag)
+  add_custom_target(${TARGET} ALL DEPENDS .${src}.cu)
   endforeach()
 endfunction()
 
@@ -81,6 +94,8 @@ nv_test(data_device_transform_test SRCS data_device_transform_test.cu
 
 if(WITH_GPU)
   if (WIN32)
+    # windows treat symbolic file as a real file, which is different with unix
+    # We create a hidden file and compile it instead of origin source file.
       windows_symbolic(hidden_file SRCS data_type_transform.cu)
       nv_library(data_type_transform SRCS .data_type_transform.cu DEPS tensor)
       add_dependencies(data_type_transform hidden_file)
@@ -149,7 +164,7 @@ if(WITH_DISTRIBUTE)
   set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 else()
   cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
-  cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass elementwise_add_op)
+  cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
  
 if (NOT WIN32)

paddle/fluid/framework/naive_executor.cc

@@ -146,5 +146,22 @@ void NaiveExecutor::CleanFeedFetchOps() {
   ops_.swap(ops);
 }
 
+void NaiveExecutor::EnableMKLDNN(const ProgramDesc &program) {
+#ifdef PADDLE_WITH_MKLDNN
+  VLOG(3) << "use_mkldnn=True";
+  for (size_t block_id = 0; block_id < program.Size(); ++block_id) {
+    auto *block = const_cast<ProgramDesc &>(program).MutableBlock(block_id);
+    for (auto *op : block->AllOps()) {
+      if (op->HasAttr("use_mkldnn")) {
+        op->SetAttr("use_mkldnn", true);
+      }
+    }
+  }
+#else
+  LOG(WARNING)
+      << "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option";
+#endif
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/naive_executor.h

@@ -14,6 +14,8 @@
 
 #pragma once
 
+#include <string>
+#include <vector>
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
@@ -46,6 +48,8 @@ class NaiveExecutor {
 
   void CleanFeedFetchOps();
 
+  void EnableMKLDNN(const ProgramDesc& program);
+
  protected:
   void CreateVariables(const ProgramDesc& desc, Scope* scope, int block_id);
 

paddle/fluid/framework/rw_lock.h

@@ -46,6 +46,7 @@ struct RWLock {
  private:
   pthread_rwlock_t lock_;
 };
+// TODO(paddle-dev): Support RWLock for WIN32 for correctness.
 #else
 // https://stackoverflow.com/questions/7125250/making-pthread-rwlock-wrlock-recursive
 // In windows, rw_lock seems like a hack. Use empty object and do nothing.

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -20,8 +20,6 @@ cc_test(test_node SRCS node_tester.cc DEPS analysis)
 cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
 cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid)
 
-set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
-
 function (inference_analysis_test TARGET)
     if(WITH_TESTING)
         set(options "")

paddle/fluid/inference/api/CMakeLists.txt

@@ -31,7 +31,6 @@ function(inference_api_test TARGET_NAME)
         set(multiValueArgs ARGS)
         cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
 
-        set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
         cc_test(${TARGET_NAME}
                 SRCS ${inference_test_SRC}
                 DEPS "${inference_deps}"

paddle/fluid/inference/api/analysis_predictor.cc

@@ -71,6 +71,11 @@ bool AnalysisPredictor::Init(
   } else {
     inference_program_ = program;
   }
+
+  if (config_._use_mkldnn) {
+    executor_->EnableMKLDNN(*inference_program_);
+  }
+
   executor_->Prepare(scope_.get(), *inference_program_, 0,
                      config_.use_feed_fetch_ops);
 
@@ -92,6 +97,7 @@ bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
     LOG(ERROR) << "fail to set feed";
     return false;
   }
+
   // Run the inference program
   // if share variables, we need not create variables
   executor_->Run();

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

@@ -70,6 +70,14 @@ if (NOT EXISTS ${OCR_INSTALL_DIR})
 endif()
 inference_analysis_api_test(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc)
 
+# resnet50
+set(RESNET50_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/resnet50")
+if (NOT EXISTS ${RESNET50_INSTALL_DIR})
+    inference_download_and_uncompress(${RESNET50_INSTALL_DIR} ${INFERENCE_URL} "resnet50_model.tar.gz")
+endif()
+inference_analysis_test(test_analyzer_resnet50 SRCS analyzer_resnet50_tester.cc
+    EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} ARGS --infer_model=${RESNET50_INSTALL_DIR}/model)
+
 # anakin
 if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
    # anakin rnn1

paddle/fluid/inference/tests/api/analyzer_resnet50_tester.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 <fstream>
+#include <iostream>
+#include "paddle/fluid/inference/tests/api/tester_helper.h"
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+void SetConfig(AnalysisConfig *cfg) {
+  cfg->param_file = FLAGS_infer_model + "/params";
+  cfg->prog_file = FLAGS_infer_model + "/model";
+  cfg->use_gpu = false;
+  cfg->device = 0;
+  cfg->enable_ir_optim = true;
+  cfg->specify_input_name = true;
+}
+
+void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
+  PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data.");
+
+  PaddleTensor input;
+  // channel=3, height/width=318
+  std::vector<int> shape({FLAGS_batch_size, 3, 318, 318});
+  input.shape = shape;
+  input.dtype = PaddleDType::FLOAT32;
+
+  // fill input data, for profile easily, do not use random data here.
+  size_t size = FLAGS_batch_size * 3 * 318 * 318;
+  input.data.Resize(size * sizeof(float));
+  float *input_data = static_cast<float *>(input.data.data());
+  for (size_t i = 0; i < size; i++) {
+    *(input_data + i) = static_cast<float>(i) / size;
+  }
+
+  std::vector<PaddleTensor> input_slots;
+  input_slots.assign({input});
+  (*inputs).emplace_back(input_slots);
+}
+
+// Easy for profiling independently.
+TEST(Analyzer_resnet50, profile) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  std::vector<PaddleTensor> outputs;
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+
+  if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
+    PADDLE_ENFORCE_EQ(outputs.size(), 1UL);
+    size_t size = GetSize(outputs[0]);
+    // output is a 512-dimension feature
+    EXPECT_EQ(size, 512 * FLAGS_batch_size);
+  }
+}
+
+// Check the fuse status
+TEST(Analyzer_resnet50, fuse_statis) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  int num_ops;
+  auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
+  auto fuse_statis = GetFuseStatis(
+      static_cast<AnalysisPredictor *>(predictor.get()), &num_ops);
+  ASSERT_TRUE(fuse_statis.count("fc_fuse"));
+  EXPECT_EQ(fuse_statis.at("fc_fuse"), 1);
+}
+
+// Compare result of NativeConfig and AnalysisConfig
+TEST(Analyzer_resnet50, compare) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareNativeAndAnalysis(cfg, input_slots_all);
+}
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle

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

@@ -270,10 +270,11 @@ TEST(Analyzer_rnn1, multi_thread) {
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
   SetInput(&input_slots_all);
-  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
+  TestPrediction(cfg, input_slots_all, &outputs, 4 /* multi_thread */);
 }
 
-bool CompareTensors(framework::Scope &a_scope, framework::Scope &b_scope,
+bool CompareTensors(const framework::Scope &a_scope,
+                    const framework::Scope &b_scope,
                     const std::vector<std::string> &tensors) {
   for (auto &x : tensors) {
     auto *a_var = a_scope.FindVar(x);

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

@@ -61,8 +61,6 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->ir_passes.push_back("fc_gru_fuse_pass");
 #ifdef PADDLE_WITH_MKLDNN
   cfg->_use_mkldnn = true;
-  // disable mkldnn fuse since it should have some bugs
-  cfg->ir_passes.push_back("conv_relu_mkldnn_fuse_pass");
 #endif
 }
 

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

@@ -4,7 +4,6 @@ function(inference_test TARGET_NAME)
   set(multiValueArgs ARGS)
   cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
 
-  set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
   set(arg_list "")
   if(inference_test_ARGS)
     foreach(arg ${inference_test_ARGS})

paddle/fluid/operators/while_op.cc

@@ -224,10 +224,12 @@ class WhileGradOp : public framework::OperatorBase {
         if (cur_scope_iter == step_scopes->rbegin()) {
           auto *var = (*cur_scope_iter)->FindVar(inside_grad_name);
           PADDLE_ENFORCE_NOT_NULL(var, "Can not find var %s", inside_grad_name);
-          PADDLE_ENFORCE(var->IsType<framework::LoDTensorArray>() ||
-                             var->IsType<LoDTensor>(),
-                         "Currently the type of var only can be LoDTensorArray "
-                         "or LoDTensor.");
+          PADDLE_ENFORCE(
+              var->IsType<framework::LoDTensorArray>() ||
+                  var->IsType<LoDTensor>(),
+              "Currently the type of var only can be LoDTensorArray, "
+              "or LoDTensor, but the received var[%s] is %s.",
+              inside_grad_name, var->Type().name());
 
           if (var->IsType<LoDTensor>()) {
             auto &inside_tensor = var->Get<framework::LoDTensor>();

paddle/fluid/platform/gpu_info.cc

@@ -20,8 +20,11 @@ limitations under the License. */
 #include "paddle/fluid/platform/enforce.h"
 
 DEFINE_double(fraction_of_gpu_memory_to_use, 0.92,
-              "Default use 92% of GPU memory for PaddlePaddle,"
-              "reserve the rest for page tables, etc");
+              "Allocate a trunk of gpu memory that is this fraction of the "
+              "total gpu memory size. Future memory usage will be allocated "
+              "from the trunk. If the trunk doesn't have enough gpu memory, "
+              "additional trunks of the same size will be requested from gpu "
+              "until the gpu has no memory left for another trunk.");
 
 namespace paddle {
 namespace platform {

paddle/fluid/train/CMakeLists.txt

@@ -4,7 +4,6 @@ function(train_test TARGET_NAME)
     set(multiValueArgs ARGS)
     cmake_parse_arguments(train_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
 
-    set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
     set(arg_list "")
     if(train_test_ARGS)
         foreach(arg ${train_test_ARGS})

python/paddle/fluid/layers/control_flow.py

@@ -1570,6 +1570,10 @@ class DynamicRNN(object):
 
     The dynamic RNN can mark multiple variables as its output. Use `drnn()` to
     get the output sequence.
+    
+    NOTES:
+        Currently it is not supported that setting is_sparse to True of any 
+        layers within DynamicRNN.
     """
     BEFORE_RNN = 0
     IN_RNN = 1

python/paddle/fluid/tests/CMakeLists.txt

+set(PYTHON_TESTS_DIR ${CMAKE_CURRENT_BINARY_DIR} CACHE PATH "python tests directory")
 file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
 string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
 

python/paddle/fluid/transpiler/inference_transpiler.py

@@ -124,7 +124,7 @@ class InferenceTranspiler(object):
                 next_op = self.block.ops[i + 1]
                 if next_op.type == 'relu':
                     # modify bnorm OP to include relu
-                    current_op.set_attr("fuse_relu", True)
+                    current_op._set_attr("fuse_relu", True)
                     # remove relu OP
                     self.block._remove_op(i + 1)
             i = i + 1
@@ -454,7 +454,7 @@ class InferenceTranspiler(object):
         :type eltwise_op: Operator
         '''
 
-        conv_op.set_attr("fuse_eltwise", True)
+        conv_op._set_attr("fuse_eltwise", True)
         self.input_map[conv_op.output("Output")[0]] = eltwise_op.input("Y")[0]
         self.input_map[eltwise_op.output("Out")[0]] = eltwise_op.input("Y")[0]