Fix Bitmain Predictor::Clone() (#18599)

* update anakin-engine interfaces for content-dnn

test=develop

* support only-gpu mode of Anakin

modify eltwise parse

test=develop

* modification for thread-safe

test=develop

* Integrated template instance

test=develop

* increase template parameters

test=develop

* support MLU predictor

test=develop

* update anakin cmake files

test=develop

* update TargetWrapper::set_device

* update the initialization of anakin subgraph

test=develop

* use the default constructor of base class

test=develop

* load model from buffer with length

test=develop

* modify the access level of class

test=develop

* support anakin for bitmain arch

test=develop

* remove files

* checkout cmakelists

test=develop

* modify interfaces

test=develop

* add cmake dependments

test=develop

* enforce the outputs of net

test=develop

paddle/fluid/inference/api/CMakeLists.txt

@@ -70,9 +70,9 @@ cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_
 if(ANAKIN_FOUND)
   # Do not turn warnings into errors.
   set_source_files_properties(api.cc api_anakin_engine.cc PROPERTIES COMPILE_FLAGS "-Wno-error")
-  cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc)
+  cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS boost xxhash)
   target_link_libraries(inference_anakin_api anakin anakin_saber_common)
-  cc_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc)
+  cc_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS boost xxhash)
   target_link_libraries(inference_anakin_api_shared anakin anakin_saber_common)
   function(anakin_target target_name)
     target_compile_options(${target_name} BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})

paddle/fluid/inference/api/api_anakin_engine.cc

@@ -62,8 +62,9 @@ void PaddleInferenceAnakinPredictor<T, P, R>::InitGraph() {
   } else {
     LOG(FATAL) << "Model load error.";
   }
-  auto inputs = this->graph_p_->get_ins();
+  this->input_names_ = this->graph_p_->get_ins();
-  for (auto &input_str : inputs) {
+  this->output_names_ = this->graph_p_->get_outs();
+  for (auto &input_str : this->input_names_) {
     if (this->config_.init_inputs_shape.find(input_str) ==
         this->config_.init_inputs_shape.end()) {
       LOG(FATAL) << input_str << " should be set in init_inputs_shape.";
@@ -201,7 +202,7 @@ bool PaddleInferenceAnakinPredictor<T, P, R>::RunImpl(
                  << "'s type is not float";
     }
     auto d_tensor_p = this->executor_p_->get_in(input.name);
-    auto net_shape = d_tensor_p->shape();
+    auto net_shape = d_tensor_p->valid_shape();
     if (net_shape.size() != input.shape.size()) {
       LOG(FATAL) << " input  " << input.name
                  << "'s shape size should be equal to that of net";
@@ -250,6 +251,10 @@ bool PaddleInferenceAnakinPredictor<T, P, R>::RunImpl(
     LOG(FATAL) << "At least one output should be set with tensors' names.";
   }
   for (auto &output : *output_data) {
+    if (std::find(this->output_names_.begin(), this->output_names_.end(),
+                  output.name) == this->output_names_.end()) {
+      LOG(FATAL) << output.name << " is not in the outputs of the graph.";
+    }
     auto *d_tensor_p = this->executor_p_->get_out(output.name);
     output.shape = d_tensor_p->valid_shape();
     if (output.data.length() < d_tensor_p->valid_size() * sizeof(float)) {
@@ -264,20 +269,23 @@ bool PaddleInferenceAnakinPredictor<T, P, R>::RunImpl(
   return true;
 }
 template <typename T, Precision P, OpRunType R>
-bool PaddleInferenceAnakinPredictor<T, P, R>::ResetConfig(
+bool PaddleInferenceAnakinPredictor<T, P, R>::Reset(
-    const AnakinConfig &config) {
+    PaddleInferenceAnakinPredictor<T, P, R> *predictor) {
-  this->config_ = config;
+  this->config_ = predictor->GetConfig();
-  return true;
+  this->graph_p_ = predictor->GetGraph();
-}
+  this->input_names_ = predictor->GetInputNames();
-template <typename T, Precision P, OpRunType R>
+  this->output_names_ = predictor->GetOutputNames();
-anakin::Net<T, P, R> &PaddleInferenceAnakinPredictor<T, P, R>::ResetExecuter(
-    std::shared_ptr<anakin::graph::Graph<T, P>> graph_p) {
-  this->graph_p_ = graph_p;
   this->ctx_p_ = std::make_shared<anakin::Context<T>>(
       this->config_.device_id, this->config_.data_stream_id,
       this->config_.compute_stream_id);
   this->InitNet();
-  return *this->executor_p_;
+  return true;
+}
+template <typename T, Precision P, OpRunType R>
+std::unique_ptr<PaddlePredictor>
+PaddleInferenceAnakinPredictor<T, P, R>::New() {
+  return std::unique_ptr<PaddlePredictor>(
+      new PaddleInferenceAnakinPredictor<T, P, R>());
 }
 // the cloned new Predictor of anakin share the same net weights from original
 // Predictor
@@ -285,21 +293,24 @@ template <typename T, Precision P, OpRunType R>
 std::unique_ptr<PaddlePredictor>
 PaddleInferenceAnakinPredictor<T, P, R>::Clone() {
   VLOG(3) << "Anakin Predictor::clone";
-  std::unique_ptr<PaddlePredictor> cls(
+  std::unique_ptr<PaddlePredictor> cls = std::move(this->New());
-      new PaddleInferenceAnakinPredictor<T, P, R>());
-  // construct executer from other graph
   auto anakin_predictor_p =
       dynamic_cast<PaddleInferenceAnakinPredictor<T, P, R> *>(cls.get());
   if (!anakin_predictor_p) {
     LOG(FATAL) << "fail to call Init";
   }
-  anakin_predictor_p->ResetConfig(this->config_);
+  anakin_predictor_p->Reset(this);
-  anakin_predictor_p->ResetExecuter(this->graph_p_);
   return cls;
 }
 
 #ifdef ANAKIN_MLU_PLACE
 template <Precision P, OpRunType R>
+std::unique_ptr<PaddlePredictor>
+PaddleInferenceAnakinMLUPredictor<P, R>::New() {
+  return std::unique_ptr<PaddlePredictor>(
+      new PaddleInferenceAnakinMLUPredictor<P, R>());
+}
+template <Precision P, OpRunType R>
 void PaddleInferenceAnakinMLUPredictor<P, R>::SetContext() {
   this->ctx_p_ = std::make_shared<anakin::Context<anakin::MLU>>(
       this->config_.device_id, this->config_.data_stream_id,
@@ -329,6 +340,11 @@ void PaddleInferenceAnakinMLUPredictor<P, R>::Predict() {
 
 #ifdef ANAKIN_BM_PLACE
 template <Precision P, OpRunType R>
+std::unique_ptr<PaddlePredictor> PaddleInferenceAnakinBMPredictor<P, R>::New() {
+  return std::unique_ptr<PaddlePredictor>(
+      new PaddleInferenceAnakinBMPredictor<P, R>());
+}
+template <Precision P, OpRunType R>
 void PaddleInferenceAnakinBMPredictor<P, R>::OptimizeGraph() {
   if (!this->graph_p_->fusion_optimize()) {
     LOG(FATAL) << "Graph optimization error.";

paddle/fluid/inference/api/api_anakin_engine.h

@@ -20,6 +20,7 @@ limitations under the License. */
 #pragma once
 
 #include <memory>
+#include <string>
 #include <vector>
 
 #include "framework/core/net/net.h"
@@ -51,10 +52,18 @@ class PaddleInferenceAnakinPredictor : public PaddlePredictor {
            int batch_size = -1) override;
 
   std::unique_ptr<PaddlePredictor> Clone() override;
-  virtual bool ResetConfig(const AnakinConfig& config);
+  bool Reset(PaddleInferenceAnakinPredictor<T, P, R>* predictor);
-  virtual anakin::Net<T, P, R>& ResetExecuter(
-      std::shared_ptr<anakin::graph::Graph<T, P>> graph_p);
   void InitPredictor();
+  std::shared_ptr<anakin::graph::Graph<T, P>> GetGraph() {
+    return this->graph_p_;
+  }
+  std::vector<std::string> GetInputNames() override {
+    return this->input_names_;
+  }
+  std::vector<std::string> GetOutputNames() override {
+    return this->output_names_;
+  }
+  const AnakinConfig& GetConfig() const { return this->config_; }
 
   ~PaddleInferenceAnakinPredictor() override;
 
@@ -65,11 +74,14 @@ class PaddleInferenceAnakinPredictor : public PaddlePredictor {
   virtual void InitNet();
   virtual void SetContext();
   virtual void Predict();
+  virtual std::unique_ptr<PaddlePredictor> New();
   static std::mutex mutex_;
   AnakinConfig config_;
   std::shared_ptr<anakin::Context<T>> ctx_p_;
   std::shared_ptr<anakin::graph::Graph<T, P>> graph_p_;
   anakin::Net<T, P, R>* executor_p_{nullptr};
+  std::vector<std::string> input_names_;
+  std::vector<std::string> output_names_;
 
  private:
   bool RunImpl(const std::vector<PaddleTensor>& inputs,
@@ -82,10 +94,12 @@ template <Precision P, OpRunType R>
 class PaddleInferenceAnakinMLUPredictor final
     : public PaddleInferenceAnakinPredictor<anakin::MLU, P, R> {
  public:
+  PaddleInferenceAnakinMLUPredictor() = default;
   explicit PaddleInferenceAnakinMLUPredictor(const AnakinConfig& config) {
-    this->ResetConfig(config);
+    this->config_ = config;
     this->InitPredictor();
   }
+  std::unique_ptr<PaddlePredictor> New() override;
   void SetContext() override;
   void OptimizeGraph() override;
   void InitNet() override;
@@ -98,10 +112,12 @@ template <Precision P, OpRunType R>
 class PaddleInferenceAnakinBMPredictor final
     : public PaddleInferenceAnakinPredictor<anakin::BM, P, R> {
  public:
+  PaddleInferenceAnakinBMPredictor() = default;
   explicit PaddleInferenceAnakinBMPredictor(const AnakinConfig& config) {
-    this->ResetConfig(config);
+    this->config_ = config;
     this->InitPredictor();
   }
+  std::unique_ptr<PaddlePredictor> New() override;
   void OptimizeGraph() override;
   void InitNet() override;
   void Predict() override;