More flexible input q

source/Lib/CommonLib/NNFilterHOP.cpp

@@ -162,7 +162,9 @@ void NNFilterHOP::resizeInputs(int width, int height)
     cerr << "[ERROR] issue init model NNFilterHOP " << endl;
     exit(-1);
   }
-  m_input_quantizer = m_model->getInputsTemplate()[0].quantizer;   // assume all image inputs have same quantizer
+  assert(nb_inputs==m_inputs.size());
+  for(int i=0;i<nb_inputs;++i)
+    m_input_quantizer[i] = m_model->getInputsTemplate()[i].quantizer;
 
 }
 #if NN_HOP_UNIFIED_TEMPORAL_FILTERING
@@ -379,17 +381,17 @@ void NNFilterHOP::filterBlock(Picture &pic, UnitArea inferArea, int extLeft, int
   const double inputScaleIpb     = (1 << log2InputIbpScale);
 
   std::vector<InputData> listInputData;
-  listInputData.push_back({ NN_INPUT_REC, 0, inputScalePred, m_input_quantizer - log2InputBitdepth, true, true });
-  listInputData.push_back({ NN_INPUT_PRED, 1, inputScalePred, m_input_quantizer - log2InputBitdepth, true, true });
-  listInputData.push_back({ NN_INPUT_BS, 2, inputScalePred, m_input_quantizer - log2InputBitdepth, true, true });
-  listInputData.push_back({ NN_INPUT_GLOBAL_QP, 3, inputScaleQp, m_input_quantizer - log2InputQpScale, true, false });
-  listInputData.push_back({ NN_INPUT_LOCAL_QP, 4, inputScaleQp, m_input_quantizer - log2InputQpScale, true, false });
+  listInputData.push_back({ NN_INPUT_REC, 0, inputScalePred, m_input_quantizer[0] - log2InputBitdepth, true, true });
+  listInputData.push_back({ NN_INPUT_PRED, 1, inputScalePred, m_input_quantizer[1] - log2InputBitdepth, true, true });
+  listInputData.push_back({ NN_INPUT_BS, 2, inputScalePred, m_input_quantizer[2] - log2InputBitdepth, true, true });
+  listInputData.push_back({ NN_INPUT_GLOBAL_QP, 3, inputScaleQp, m_input_quantizer[3] - log2InputQpScale, true, false });
+  listInputData.push_back({ NN_INPUT_LOCAL_QP, 4, inputScaleQp, m_input_quantizer[4] - log2InputQpScale, true, false });
 #if NN_HOP_UNIFIED_FORCE_USE
   if (m_forceIntraType) {
-    listInputData.push_back({ NN_INPUT_ZERO, 5, inputScaleIpb, m_input_quantizer - log2InputIbpScale, true, false });
+    listInputData.push_back({ NN_INPUT_ZERO, 5, inputScaleIpb, m_input_quantizer[5] - log2InputIbpScale, true, false });
   } else
 #endif
-  listInputData.push_back({ NN_INPUT_IPB, 5, inputScaleIpb, m_input_quantizer - log2InputIbpScale, true, false });
+  listInputData.push_back({ NN_INPUT_IPB, 5, inputScaleIpb, m_input_quantizer[5] - log2InputIbpScale, true, false });
 
   NNInference::prepareInputs<TypeSadlHOP>(&pic, inferArea, m_inputs, seqQp, sliceQp, -1 /* sliceType */, listInputData);
 

source/Lib/CommonLib/NNFilterHOP.h

@@ -54,6 +54,7 @@ public:
   static constexpr float nnResidueScaleDerivationUpBound = 1.25f;
   static constexpr float nnResidueScaleDerivationLowBound = 0.0625f;
   static constexpr int max_scale=(1<<log2ResidueScale);
+  static constexpr int nb_inputs=6;
 #if NN_HOP_UNIFIED_TEMPORAL_FILTERING
   static constexpr int minimumTidUseTemporalFiltering = 3;
 #endif
@@ -130,7 +131,7 @@ public:
 #endif
 private:
   int  m_blocksize[2];   // current inputs size
-  int  m_input_quantizer = 0;
+  int  m_input_quantizer[nb_inputs] = {};
   void resizeInputs(int width, int height);
   std::unique_ptr<sadl::Model<TypeSadlHOP>> m_model;
   std::vector<sadl::Tensor<TypeSadlHOP>>    m_inputs;