Fix AUTO plugin mismatch issue for get_tensor function

src/plugins/intel_cpu/src/graph.cpp

@@ -996,7 +996,9 @@ void Graph::PushInputData(const std::string& name, const ov::Tensor &in) {
     }
 }
 
-void Graph::PullOutputData(std::unordered_map<std::string, ov::Tensor> &out) {
+void Graph::PullOutputData(std::unordered_map<std::string, ov::Tensor>& out,
+                           std::unordered_map<std::string, bool>& out_precision_changed,
+                           std::unordered_map<std::string, ov::Tensor>& aux_tensors) {
     if (!IsReady())
         IE_THROW() << "Wrong state. Topology not ready.";
 
@@ -1060,6 +1062,25 @@ void Graph::PullOutputData(std::unordered_map<std::string, ov::Tensor> &out) {
         void *ext_blob_ptr = ext_blob.data();
         void *intr_blob_ptr = intr_blob.GetData();
 
+        // If output precision has been changed comparing to original model's output, it must be copied to aux tensor
+        if (out_precision_changed[name]) {
+            auto it = aux_tensors.find(name);
+            if (it == aux_tensors.end()) {
+                OPENVINO_THROW("Output precision has been changed, but cannot find its aux tensor.");
+            }
+            void* ext_blob_ptr = it->second.data();
+            if ((intr_blob_ptr == nullptr) || (ext_blob_ptr == nullptr)) {
+                OPENVINO_THROW("Get tensor has no allocated memory");
+            }
+
+            cpu_convert(intr_blob_ptr,
+                        ext_blob_ptr,
+                        srcPrec,
+                        InferenceEngine::details::convertPrecision(it->second.get_element_type()),
+                        it->second.get_size());
+            continue;
+        }
+
         // That is the same memory. No need to copy
         if (ext_blob_ptr == intr_blob_ptr) continue;
 

src/plugins/intel_cpu/src/graph.h

@@ -59,7 +59,9 @@ public:
     }
 
     void PushInputData(const std::string& name, const ov::Tensor& in);
-    void PullOutputData(std::unordered_map<std::string, ov::Tensor>& out);
+    void PullOutputData(std::unordered_map<std::string, ov::Tensor>& out,
+                        std::unordered_map<std::string, bool>& out_precision_changed,
+                        std::unordered_map<std::string, ov::Tensor>& aux_tensors);
 
     void Infer(SyncInferRequest* request = nullptr);
 

src/plugins/intel_cpu/src/infer_request.cpp

@@ -80,6 +80,8 @@ void SyncInferRequest::create_infer_request() {
     }
     for (const auto& it : _output_ports_map) {
         init_tensor(it.first);
+        // allocate aux tensor for output if output precision has been changed
+        get_tensor(it.second);
     }
 
     // Save all MemoryLayer data tensors. Will use insight about mechanics
@@ -233,7 +235,7 @@ void SyncInferRequest::infer() {
 
     throw_if_canceled();
 
-    graph->PullOutputData(_outputs);
+    graph->PullOutputData(_outputs, _port_precision_changed, _aux_tensors);
 }
 
 std::vector<ov::ProfilingInfo> SyncInferRequest::get_profiling_info() const {
@@ -486,25 +488,7 @@ ov::Tensor SyncInferRequest::get_tensor(const ov::Output<const ov::Node>& _port)
         _aux_tensors[port_name] = ov::Tensor(_orig_ports_map[port_name].get_element_type(), external_shape);
     }
 
-    // input tensor is in aux tensors, don't need copy anything
-    auto& aux_tensor = _aux_tensors[port_name];
-    if (is_input) {
-        return aux_tensor;
-    }
-
-    // output tensor need copy data from compiled tensor
-    const void* srcData = compiled_tensor.data();
-    void* dstData = aux_tensor.data();
-    if ((dstData == nullptr) || (srcData == nullptr)) {
-        OPENVINO_THROW("Get tensor has no allocated memory");
-    }
-
-    cpu_convert(srcData,
-                dstData,
-                InferenceEngine::details::convertPrecision(compiled_tensor.get_element_type()),
-                InferenceEngine::details::convertPrecision(aux_tensor.get_element_type()),
-                compiled_tensor.get_size());
-    return aux_tensor;
+    return _aux_tensors[port_name];
 }
 
 std::vector<ov::Tensor> SyncInferRequest::get_tensors(const ov::Output<const ov::Node>& _port) const {

src/plugins/intel_cpu/src/plugin.cpp

@@ -32,6 +32,7 @@
 
 #include <cpu/x64/cpu_isa_traits.hpp>
 #include <itt.h>
+#include <unordered_set>
 
 using namespace InferenceEngine;
 

src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp

@@ -209,9 +209,6 @@ std::vector<std::string> disabledTestPatterns() {
         R"(.*smoke_Transpose(2|4|5|6)D/TransposeLayerTest.CompareWithRefs.*)",
         // Issue: 113704 - Layout information maybe incorrect when covert tensor to blob
         R"(.*smoke_.*BehaviorTests/InferRequestPreprocessConversionTest.*)",
-        // Issue: There are some problem when work with AUTO plugin API 2.0
-        R"(.*smoke_Auto_CachingSupportCase_CPU/CompileModelCacheTestBase.CompareWithRefImpl.*)",
-        R"(.*smoke_Auto_CachingSupportCase_CPU_Float/CompileModelCacheTestBase.CompareWithRefImpl.*)",
         // Issue: JIT choose error issue, don't know why not choose jit_avx512_BF16
         R"(.*smoke_PSROIPoolingAverageLayoutTest/PSROIPoolingLayerCPUTest.*)",
         R"(.*smoke_PSROIPoolingBilinearLayoutTest/PSROIPoolingLayerCPUTest.*)",