initial

mryzhov · Feb 21, 2023 · d7bff1f · d7bff1f
1 parent 86b5004
commit d7bff1f
Show file tree

Hide file tree

Showing 3 changed files with 47 additions and 1 deletion.
diff --git a/src/plugins/intel_gna/src/gna_data_types.hpp b/src/plugins/intel_gna/src/gna_data_types.hpp
@@ -45,6 +45,7 @@ using ConcatConnection = std::unordered_map<std::string, GNAConcatLayer>;
 using SplitConnection = std::unordered_map<std::string, GNASplitLayer>;
 using CropConnection = std::unordered_map<std::string, GNACropLayer>;
 using ConstConnections = std::unordered_map<std::string, void*>;
+using SubgraphCPUMap = std::unordered_map<std::string, std::shared_ptr<ov::Model>>;
 
 }  // namespace intel_gna
 }  // namespace ov
diff --git a/src/plugins/intel_gna/src/gna_plugin.cpp b/src/plugins/intel_gna/src/gna_plugin.cpp
@@ -1143,6 +1143,29 @@ uint32_t GNAPlugin::QueueInference(const InferenceEngine::BlobMap& inputs, Infer
                      importedElements,
                      importedElements);
 
+        {
+            auto subgraph_it = subgraph_cpu_map.find(input.first);
+            if (subgraph_it != subgraph_cpu_map.end()) {
+                std::shared_ptr<ov::Model> model = subgraph_it->second;
+                ov::TensorVector model_result(1);
+
+                const ngraph::Shape & model_input_shape = model->get_parameters()[0]->get_output_shape(0);
+                const ngraph::element::Type & model_input_type = model->get_parameters()[0]->get_element_type();
+
+                void * input_ptr = inputs_ptr_->at(input.first).ptrs[index];
+                ov::Tensor input_tensor(model_input_type, model_input_shape, input_ptr);
+
+                if (model->evaluate(model_result, ov::TensorVector{input_tensor})) {
+                    const size_t model_result_size = model_result[0].get_byte_size();
+                    ie_memcpy(input_ptr, model_result_size, model_result[0].data(), model_result_size);
+                } else {
+                    THROW_GNA_EXCEPTION << "Error evalutate ngraph::Function for output " << input.first;
+                }
+            } else {
+                std::cout << "input " << input.first << " model NOT found " << std::endl;
+            }
+        }
+
         auto transpose_info = transpose_inputs_info.find(input.first);
         if (transpose_info != std::end(transpose_inputs_info)) {
             size_t batchSize = (dims.size() > 1) ? dims[0] : 1;
@@ -1252,7 +1275,27 @@ RequestStatus GNAPlugin::WaitFor(uint32_t request_idx, int64_t millisTimeout) {
         auto elementsPerBatch =
             isScalar ? 1
                      : (is1D ? dims.front() : InferenceEngine::details::product(++std::begin(dims), std::end(dims)));
-
+        {
+            auto subgraph_it = subgraph_cpu_map.find(outputBlobIt.first);
+            if (subgraph_it != subgraph_cpu_map.end()) {
+                std::shared_ptr<ov::Model> model = subgraph_it->second;
+                ov::TensorVector model_result(1);
+
+                void * output_ptr = outputDesc.ptrs[request_idx];
+                const ngraph::Shape & model_input_shape = model->get_parameters()[0]->get_output_shape(0);
+                const ngraph::element::Type & model_input_type = model->get_parameters()[0]->get_element_type();
+                ov::Tensor input_tensor(model_input_type, model_input_shape, output_ptr);
+
+                if (model->evaluate(model_result, ov::TensorVector{input_tensor})) {
+                    const size_t model_result_size = model_result[0].get_byte_size();
+                    ie_memcpy(output_ptr, model_result_size, model_result[0].data(), model_result_size);
+                } else {
+                    THROW_GNA_EXCEPTION << "Error evalutate ngraph::Function for output " << outputBlobIt.first;
+                }
+            } else {
+                std::cout << "output " << outputBlobIt.first << " model NOT found " << std::endl;
+            }
+        }
         auto transpose_output_info = transpose_outputs_info.find(outputBlobIt.first);
         if (transpose_output_info != std::end(transpose_outputs_info) &&
             FoundPartToTranspose(transpose_output_info->second)) {

diff --git a/src/plugins/intel_gna/src/gna_plugin.hpp b/src/plugins/intel_gna/src/gna_plugin.hpp
@@ -54,6 +54,8 @@ class GNAPlugin : public InferenceEngine::IInferencePlugin {
     TranspositionInfoMap transpose_inputs_info;
     TranspositionInfoMap transpose_outputs_info;
 
+    SubgraphCPUMap subgraph_cpu_map;
+
     uint32_t dnn_dump_write_index = 0;
     intel_dnn_number_type_t output_type = kDnnInt;