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

test=develop

Author: velconia

cmake/operators.cmake

@@ -109,7 +109,8 @@ function(op_library TARGET)
 
     # Define operators that don't need pybind here.
     foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
-"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op")
+"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op"
+"fusion_transpose_flatten_concat_op")
         if ("${TARGET}" STREQUAL "${manual_pybind_op}")
             set(pybind_flag 1)
         endif()

paddle/fluid/framework/CMakeLists.txt

@@ -117,7 +117,7 @@ cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 
 if (NOT WIN32)
-cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto)
+cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto device_context)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
     shape_inference data_transform lod_tensor profiler transfer_scope_cache)
 else()

paddle/fluid/framework/executor.cc

@@ -392,8 +392,8 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
 
   int64_t max_memory_size = GetEagerDeletionThreshold();
   std::unique_ptr<GarbageCollector<Tensor>> gc;
-  // WhileOp would set keep_kids to false
-  // WhileGradOp would need the scopes created in WhileOp
+  // WhileOp would set keep_kids to true,
+  // because WhileGradOp needs the scopes created in WhileOp.
   // Perhaps, we should not perform eager deletion in WhileOp
   // The scopes and variables created by WhileOp would be deleted
   // in WhileGradOp.

paddle/fluid/framework/transfer_scope_cache.cc

@@ -17,16 +17,28 @@
 namespace paddle {
 namespace framework {
 
+// Holds all the transfer scope across the process.
 std::unordered_map<size_t, Scope*>& global_transfer_data_cache() {
-  thread_local auto* x = new std::unordered_map<size_t, Scope*>;
+  typedef std::unordered_map<size_t, Scope*> map_t;
+  thread_local std::unique_ptr<map_t> x(new map_t);
   return *x;
 }
 
+// Holds all the transfer scope for this thread.
 std::unordered_set<Scope*>& global_transfer_scope_cache() {
-  thread_local auto* x = new std::unordered_set<Scope*>;
+  typedef std::unordered_set<Scope*> set_t;
+  thread_local std::unique_ptr<set_t> x(new set_t);
   return *x;
 }
 
+// Try to create a transfer scope. If one cached scope has match the
+// requirement, just return that one.
+// Inputs:
+// @type0: the source kernel type.
+// @type1: the target kernel type.
+// @scope: the execution scope of this op.
+// Returns: A scope used to hold the transfer data across the different kernel
+// type.
 Scope* TryCreateTransferScope(OpKernelType type0, OpKernelType type1,
                               const Scope* scope) {
   Scope* new_scope{nullptr};
@@ -46,27 +58,5 @@ Scope* TryCreateTransferScope(OpKernelType type0, OpKernelType type1,
   return new_scope;
 }
 
-void RemoveKidsFromTransferScopeCache(Scope* scope) {
-  auto it = global_transfer_scope_cache().find(scope);
-  if (it != global_transfer_scope_cache().end()) {
-    global_transfer_scope_cache().erase(it);
-  }
-  for (auto* s : scope->kids()) {
-    auto it = global_transfer_scope_cache().find(s);
-    if (it != global_transfer_scope_cache().end()) {
-      global_transfer_scope_cache().erase(it);
-    }
-  }
-
-  // remove global transfer data cache
-  auto& cache = global_transfer_data_cache();
-  for (auto it = cache.begin(); it != cache.end();) {
-    if (it->second == scope)
-      it = cache.erase(it);
-    else
-      it++;
-  }
-}
-
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/inference/CMakeLists.txt

@@ -4,6 +4,7 @@ endif()
 # analysis and tensorrt must be added before creating static library,
 # otherwise, there would be undefined reference to them in static library.
 add_subdirectory(analysis)
+add_subdirectory(utils)
 if (TENSORRT_FOUND)
   add_subdirectory(tensorrt)
 endif()

paddle/fluid/inference/api/CMakeLists.txt

@@ -30,7 +30,9 @@ cc_library(paddle_pass_builder SRCS paddle_pass_builder.cc)
 cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor reset_tensor_array analysis_config paddle_pass_builder ir_pass_manager)
 cc_library(zero_copy_tensor SRCS details/zero_copy_tensor.cc DEPS scope lod_tensor enforce)
 cc_library(zero_copy_tensor_dummy SRCS details/zero_copy_tensor_dummy.cc)
-cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS lod_tensor scope paddle_pass_builder reset_tensor_array analysis_config analysis_config paddle_pass_builder DEPS zero_copy_tensor)
+cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS
+           lod_tensor scope paddle_pass_builder reset_tensor_array analysis_config
+           analysis_config paddle_pass_builder zero_copy_tensor reset_tensor_array)
 
 cc_test(test_paddle_inference_api
         SRCS api_tester.cc

paddle/fluid/inference/api/analysis_predictor.cc

@@ -31,6 +31,7 @@
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #endif
 #include "paddle/fluid/inference/utils/singleton.h"
+#include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -174,7 +175,6 @@ bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
   inference::Timer timer;
   timer.tic();
   // set feed variable
-  std::vector<framework::LoDTensor> feeds;
   framework::Scope *scope = sub_scope_ ? sub_scope_ : scope_.get();
   if (!SetFeed(inputs, scope)) {
     LOG(ERROR) << "fail to set feed";
@@ -215,17 +215,29 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
     framework::DDim ddim = framework::make_ddim(inputs[i].shape);
     void *input_ptr;
     if (inputs[i].dtype == PaddleDType::INT64) {
-      input_ptr = input.mutable_data<int64_t>(ddim, platform::CPUPlace());
+      input_ptr = input.mutable_data<int64_t>(ddim, place_);
     } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
-      input_ptr = input.mutable_data<float>(ddim, platform::CPUPlace());
+      input_ptr = input.mutable_data<float>(ddim, place_);
     } else {
       LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
       return false;
     }
 
-    // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
-    std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
-                inputs[i].data.length());
+    if (platform::is_cpu_place(place_)) {
+      // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
+      std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
+                  inputs[i].data.length());
+    } else {
+#ifdef PADDLE_WITH_CUDA
+      auto dst_gpu_place = boost::get<platform::CUDAPlace>(place_);
+      memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
+                   platform::CPUPlace(), inputs[i].data.data(),
+                   inputs[i].data.length(),
+                   0);  // stream 0 for sync copy
+#else
+      PADDLE_THROW("Not compile with CUDA, should not reach here.");
+#endif
+    }
     // TODO(Superjomn) Low performance, need optimization for heavy LoD copy.
     framework::LoD lod;
     for (auto &level : inputs[i].lod) {

paddle/fluid/inference/api/api_impl.cc

@@ -24,6 +24,7 @@ limitations under the License. */
 #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/memory/memcpy.h"
 #include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 
@@ -138,7 +139,6 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
   Timer timer;
   timer.tic();
   // set feed variable
-  std::vector<framework::LoDTensor> feeds;
   framework::Scope *scope = sub_scope_ != nullptr ? sub_scope_ : scope_.get();
   if (!SetFeed(inputs, scope)) {
     LOG(ERROR) << "fail to set feed";
@@ -194,17 +194,30 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
     framework::DDim ddim = framework::make_ddim(inputs[i].shape);
     void *input_ptr;
     if (inputs[i].dtype == PaddleDType::INT64) {
-      input_ptr = input.mutable_data<int64_t>(ddim, platform::CPUPlace());
+      input_ptr = input.mutable_data<int64_t>(ddim, place_);
     } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
-      input_ptr = input.mutable_data<float>(ddim, platform::CPUPlace());
+      input_ptr = input.mutable_data<float>(ddim, place_);
     } else {
       LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
       return false;
     }
 
-    // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
-    std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
-                inputs[i].data.length());
+    if (platform::is_cpu_place(place_)) {
+      // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
+      std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
+                  inputs[i].data.length());
+    } else {
+#ifdef PADDLE_WITH_CUDA
+      auto dst_gpu_place = boost::get<platform::CUDAPlace>(place_);
+      memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
+                   platform::CPUPlace(), inputs[i].data.data(),
+                   inputs[i].data.length(),
+                   0);  // stream 0 for sync copy
+#else
+      PADDLE_THROW("Not compile with CUDA, should not reach here.");
+#endif
+    }
+
     // TODO(Superjomn) Low performance, need optimization for heavy LoD copy.
     framework::LoD lod;
     for (auto &level : inputs[i].lod) {

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

@@ -46,8 +46,6 @@ if(WITH_GPU)
     endif()
   endif(NOT WIN32)
 endif()
-
-include_directories("D:/Paddle/")
 include_directories("${PADDLE_LIB}")
 include_directories("${PADDLE_LIB}/third_party/install/protobuf/include")
 include_directories("${PADDLE_LIB}/third_party/install/glog/include")

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

@@ -74,7 +74,7 @@ inference_analysis_api_test(test_analyzer_seq_conv1 ${SEQ_CONV1_INSTALL_DIR} ana
 # ocr
 set(OCR_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/ocr")
 if (NOT EXISTS ${OCR_INSTALL_DIR})
-  inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz")
+    inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz")
 endif()
 inference_analysis_api_test(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc)
 
@@ -88,31 +88,31 @@ inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet
 
 # anakin
 if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
-   # anakin rnn1
-   set(ANAKIN_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/anakin")
-   set(ANAKIN_RNN1_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/rnn1")
-   inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn.anakin2.model.bin")
-   inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn_data.txt")
-   cc_test(test_anakin_rnn1 SRCS anakin_rnn1_tester.cc
-           ARGS --model=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn.anakin2.model.bin
-                --datapath=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn_data.txt
-           DEPS inference_anakin_api_shared SERIAL)
-   # anakin mobilenet
-   if(WITH_GPU)
-       set(ANAKIN_MOBILENET_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/mobilenet")
-       inference_download(${ANAKIN_MOBILENET_INSTALL_DIR} ${INFERENCE_URL} "mobilenet_v2.anakin.bin")
-       cc_test(test_anakin_mobilenet SRCS anakin_mobilenet_tester.cc
-               ARGS --model=${ANAKIN_MOBILENET_INSTALL_DIR}/mobilenet_v2.anakin.bin
-               DEPS inference_anakin_api_shared dynload_cuda SERIAL)
-   endif()
+    # anakin rnn1
+    set(ANAKIN_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/anakin")
+    set(ANAKIN_RNN1_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/rnn1")
+    inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn.anakin2.model.bin")
+    inference_download(${ANAKIN_RNN1_INSTALL_DIR} ${INFERENCE_URL} "anakin_test%2Fditu_rnn_data.txt")
+    cc_test(test_anakin_rnn1 SRCS anakin_rnn1_tester.cc
+            ARGS --model=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn.anakin2.model.bin
+                 --datapath=${ANAKIN_RNN1_INSTALL_DIR}/anakin_test%2Fditu_rnn_data.txt
+            DEPS inference_anakin_api_shared SERIAL)
+    # anakin mobilenet
+    if(WITH_GPU)
+        set(ANAKIN_MOBILENET_INSTALL_DIR "${ANAKIN_INSTALL_DIR}/mobilenet")
+        inference_download(${ANAKIN_MOBILENET_INSTALL_DIR} ${INFERENCE_URL} "mobilenet_v2.anakin.bin")
+        cc_test(test_anakin_mobilenet SRCS anakin_mobilenet_tester.cc
+                ARGS --model=${ANAKIN_MOBILENET_INSTALL_DIR}/mobilenet_v2.anakin.bin
+                DEPS inference_anakin_api_shared dynload_cuda SERIAL)
+    endif()
 endif()
 
 if(WITH_GPU AND TENSORRT_FOUND)
-   set(TRT_MODEL_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/trt")
-   if (NOT EXISTS ${TRT_MODEL_INSTALL_DIR})
-       inference_download_and_uncompress(${TRT_MODEL_INSTALL_DIR} ${INFERENCE_URL}/tensorrt_test "trt_test_models.tar.gz")
-   endif()
-   inference_analysis_test(test_trt_models SRCS trt_models_tester.cc
-      EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
-        ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_test_models SERIAL)
+    set(TRT_MODEL_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/trt")
+    if (NOT EXISTS ${TRT_MODEL_INSTALL_DIR})
+        inference_download_and_uncompress(${TRT_MODEL_INSTALL_DIR} ${INFERENCE_URL}/tensorrt_test "trt_test_models.tar.gz")
+    endif()
+    inference_analysis_test(test_trt_models SRCS trt_models_tester.cc
+            EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
+            ARGS --infer_model=${TRT_MODEL_INSTALL_DIR}/trt_test_models SERIAL)
 endif()