!3815 compute threshold only once in post training quantization

Merge pull request !3815 from xutianchun/quant_0731

mindspore/lite/src/common/anf_exporter/anf_exporter.cc

@@ -98,7 +98,6 @@ schema::MetaGraphT *AnfExporter::Export(const FuncGraphPtr &funcGraph) {
     }
 
     node->primitive = std::unique_ptr<schema::PrimitiveT>(primitiveT_value->GetPrimitiveT());
-    primitiveT_value->SetPrimitiveT(nullptr);
     std::vector<schema::TensorT *> outputs;
     SetOpInputNode(cnode, metaGraphT.get(), node.get());
     SetOpOutputNode(outputs, metaGraphT.get(), node.get());
@@ -113,24 +112,22 @@ schema::MetaGraphT *AnfExporter::Export(const FuncGraphPtr &funcGraph) {
       auto input_quant_params = primitiveT_value->GetInputQuantParams();
       if (input_quant_params.empty()) {
         MS_LOG(WARNING) << "node: " << node->name << " input quant params is empty";
-        continue;
+      } else {
+        std::unique_ptr<schema::QuantParamT> input_quant_param =
+          std::make_unique<schema::QuantParamT>(input_quant_params[0]);
+        tensor_input->quantParams.emplace_back(std::move(input_quant_param));
       }
-
-      std::unique_ptr<schema::QuantParamT> input_quant_param =
-        std::make_unique<schema::QuantParamT>(input_quant_params[0]);
-      tensor_input->quantParams.emplace_back(std::move(input_quant_param));
       // output
       auto output_index = node->outputIndex[0];
       auto tensor_output = metaGraphT->allTensors[output_index].get();
       auto output_quant_params = primitiveT_value->GetOutputQuantParams();
       if (output_quant_params.empty()) {
         MS_LOG(WARNING) << "node: " << node->name << " output quant params is empty";
-        continue;
+      } else {
+        std::unique_ptr<schema::QuantParamT> output_quant_param =
+          std::make_unique<schema::QuantParamT>(output_quant_params[0]);
+        tensor_output->quantParams.emplace_back(std::move(output_quant_param));
       }
-
-      std::unique_ptr<schema::QuantParamT> output_quant_param =
-        std::make_unique<schema::QuantParamT>(output_quant_params[0]);
-      tensor_output->quantParams.emplace_back(std::move(output_quant_param));
       //      // TensorType
       //      valuePtr = primitive->GetAttr(kInputTensorDataType);
       //      if (valuePtr != nullptr) {

mindspore/lite/src/ir/primitive_t_value.h

@@ -26,8 +26,8 @@ namespace mindspore::lite {
 class PrimitiveTValue : public Value {
  public:
   explicit PrimitiveTValue(schema::PrimitiveT *primt) : primitive(primt) {}
-
-  ~PrimitiveTValue() override { delete this->primitive; }
+  // not responsible to free primitive, the one created the dynamic memory is responsible to free it.
+  ~PrimitiveTValue() override = default;
 
   MS_DECLARE_PARENT(PrimitiveTValue, Value)
 

mindspore/lite/tools/converter/optimizer/node/weight_format_pass.cc

@@ -27,7 +27,7 @@ int WeightFormatPass::Run(GraphNode *graphNode) {
     MS_LOG(ERROR) << "ShapeFormatTrans failed: " << status;
     return status;
   }
-  if (this->quantType == QuantType_AwareTrainning) {
+  if (this->quantType == QuantType_AwareTrainning || this->quantType == QuantType_PostTraining) {
     status = QuantDataFormatTrans(graphNode);
     if (status != 0) {
       MS_LOG(ERROR) << "QuantDataFormatTrans failed: " << status;
@@ -147,7 +147,8 @@ int WeightFormatPass::ShapeFormatTrans(GraphNode *graphNode) {
   } else if (fmkType == converter::FmkType_TFLITE) {
     switch (node->quantType) {
       case QuantType_QUANT_NONE:
-      case QuantType_AwareTrainning: {
+      case QuantType_AwareTrainning:
+      case QuantType_PostTraining: {
         if (opType == schema::PrimitiveType_Conv2D) {
           weightTensor->format = schema::Format_KHWC;
         } else if (opType == schema::PrimitiveType_DepthwiseConv2D) {

mindspore/lite/tools/converter/quantizer/post_training.cc

@@ -292,13 +292,32 @@ STATUS Calibrator::RecordMaxValue(std::string opName, vector<float> data,
 }
 
 STATUS Calibrator::ComputeThreshold() {
-  for (auto iter = this->input_diverg_info_.begin(); iter != this->input_diverg_info_.end(); iter++) {
+  for (auto iter = this->output_diverg_info_.begin(); iter != this->output_diverg_info_.end(); iter++) {
     DivergInfo *info = iter->second.get();
     info->ComputeThreshold();
   }
-  for (auto iter = this->output_diverg_info_.begin(); iter != this->output_diverg_info_.end(); iter++) {
+  // node A's input may be node B's output, no need to re-compute the node A's input quant param which is the same as
+  for (auto iter = this->input_diverg_info_.begin(); iter != this->input_diverg_info_.end(); iter++) {
     DivergInfo *info = iter->second.get();
-    info->ComputeThreshold();
+    auto cnode = info->cnode;
+
+    bool already_computed = false;
+    auto input = cnode->input(1);
+    if (input->isa<mindspore::CNode>()) {
+      auto input_cnode = std::dynamic_pointer_cast<mindspore::CNode>(input);
+      for (const auto &output_diverg_info : output_diverg_info_) {
+        auto output_diverg_cnode = output_diverg_info.second->cnode;
+        if (output_diverg_cnode == input_cnode) {
+          *info = *(output_diverg_info.second);
+          info->cnode = cnode;
+          already_computed = true;
+          break;
+        }
+      }
+    }
+    if (!already_computed) {
+      info->ComputeThreshold();
+    }
   }
   return RET_OK;
 }