support anakin for bitmain arch

test=develop

Author: Shixiaowei02

paddle/fluid/inference/api/CMakeLists.txt

@@ -60,9 +60,14 @@ cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_
         ARGS --dirname=${WORD2VEC_MODEL_DIR})
 
 if(ANAKIN_FOUND)
+  # Do not turn warnings into errors.
+  set_source_files_properties(api.cc api_anakin_engine.cc PROPERTIES COMPILE_FLAGS "-Wno-error")
   if (ANAKIN_MLU AND NOT WITH_GPU AND NOT ANAKIN_X86)
     message(STATUS "Compile with anakin mlu place.")
     add_definitions(-DANAKIN_MLU_PLACE)
+  elseif(ANAKIN_BM AND NOT WITH_GPU AND NOT ANAKIN_X86)
+    message(STATUS "Compile with anakin bm place.")
+    add_definitions(-DANAKIN_BM_PLACE)
   elseif(ANAKIN_X86)
     message(STATUS "Compile with anakin x86 place.")
     add_definitions(-DANAKIN_X86_PLACE)

paddle/fluid/inference/api/api_anakin_engine.cc

@@ -34,10 +34,10 @@ extern std::once_flag PaddleInferenceAnakinPredictor<T, P, R>::init_anakin_;
 
 template <typename T, Precision P, OpRunType R>
 void PaddleInferenceAnakinPredictor<T, P, R>::InitEnv() {
-  anakin::TargetWrapper<T>::set_device(this->config_.device_id);
   std::call_once(this->init_anakin_, [this]() {
     anakin::Env<T>::env_init(this->config_.max_stream);
   });
+  anakin::TargetWrapper<T>::set_device(this->config_.device_id);
 }
 template <typename T, Precision P, OpRunType R>
 void PaddleInferenceAnakinPredictor<T, P, R>::InitNet() {
@@ -194,6 +194,7 @@ template <typename T, Precision P, OpRunType R>
 bool PaddleInferenceAnakinPredictor<T, P, R>::RunImpl(
     const std::vector<PaddleTensor> &inputs,
     std::vector<PaddleTensor> *output_data) {
+  anakin::TargetWrapper<T>::set_device(this->config_.device_id);
   for (const auto &input : inputs) {
     if (input.dtype != PaddleDType::FLOAT32) {
       LOG(FATAL) << "Only support float type inputs. " << input.name
@@ -326,6 +327,27 @@ void PaddleInferenceAnakinMLUPredictor<P, R>::Predict() {
 }
 #endif
 
+#ifdef ANAKIN_BM_PLACE
+template <Precision P, OpRunType R>
+void PaddleInferenceAnakinBMPredictor<P, R>::OptimizeGraph() {
+  if (!this->graph_p_->fusion_optimize()) {
+    LOG(FATAL) << "Graph optimization error.";
+  }
+}
+template <Precision P, OpRunType R>
+void PaddleInferenceAnakinBMPredictor<P, R>::InitNet() {
+  std::unique_lock<std::mutex> lock(this->mutex_);
+  this->executor_p_ = new anakin::Net<anakin::BM, P, R>();
+  this->executor_p_->fusion_init(*this->graph_p_, this->ctx_p_, true);
+}
+template <Precision P, OpRunType R>
+void PaddleInferenceAnakinBMPredictor<P, R>::Predict() {
+  anakin::TargetWrapper<anakin::BM>::device_sync();
+  this->executor_p_->fusion_prediction();
+  anakin::TargetWrapper<anakin::BM>::device_sync();
+}
+#endif
+
 #ifdef PADDLE_WITH_CUDA
 template class PaddleInferenceAnakinPredictor<
     anakin::NV, anakin::Precision::FP32, ::anakin::OpRunType::ASYNC>;
@@ -338,6 +360,10 @@ template class PaddleInferenceAnakinPredictor<
 template class PaddleInferenceAnakinMLUPredictor<anakin::Precision::FP32,
                                                  ::anakin::OpRunType::SYNC>;
 #endif
+#ifdef ANAKIN_BM_PLACE
+template class PaddleInferenceAnakinBMPredictor<anakin::Precision::FP32,
+                                                ::anakin::OpRunType::ASYNC>;
+#endif
 
 // A factory to help create difference predictor.
 template <>
@@ -365,6 +391,14 @@ CreatePaddlePredictor<contrib::AnakinConfig, PaddleEngineKind::kAnakin>(
                                               ::anakin::OpRunType::SYNC>(
             config));
   }
+#endif
+#ifdef ANAKIN_BM_PLACE
+  if (config.target_type == contrib::AnakinConfig::BM) {
+    return std::unique_ptr<PaddlePredictor>(
+        new PaddleInferenceAnakinBMPredictor<anakin::Precision::FP32,
+                                             ::anakin::OpRunType::ASYNC>(
+            config));
+  }
 #endif
   LOG(FATAL) << "Anakin Predictor create on unknown platform.";
   return nullptr;

paddle/fluid/inference/api/api_anakin_engine.h

@@ -92,4 +92,19 @@ class PaddleInferenceAnakinMLUPredictor final
   void Predict() override;
 };
 #endif
+
+#ifdef ANAKIN_BM_PLACE
+template <Precision P, OpRunType R>
+class PaddleInferenceAnakinBMPredictor final
+    : public PaddleInferenceAnakinPredictor<anakin::BM, P, R> {
+ public:
+  explicit PaddleInferenceAnakinBMPredictor(const AnakinConfig& config) {
+    this->ResetConfig(config);
+    this->InitPredictor();
+  }
+  void OptimizeGraph() override;
+  void InitNet() override;
+  void Predict() override;
+};
+#endif
 }  // namespace paddle

paddle/fluid/inference/api/paddle_anakin_config.h

@@ -25,7 +25,7 @@ namespace paddle {
 namespace contrib {
 // Configurations for Anakin engine.
 struct AnakinConfig : public PaddlePredictor::Config {
-  enum TargetType { NVGPU = 0, X86, MLU };
+  enum TargetType { NVGPU = 0, X86, MLU, BM };
   int device_id{0};
   std::string model_file;
   std::map<std::string, std::vector<int>> init_inputs_shape;