[GPU] Fallback to kernel caching in the case of dynamic models (#15842)

* use kernel caching for dynamic models

* replaced cl_cache with blob

* updated to serialize dims info of input and output

* updated to skip unicode tests in Windows

src/plugins/intel_gpu/include/intel_gpu/plugin/compiled_model.hpp

@@ -46,9 +46,6 @@ public:
     ExecutionConfig m_config;
     InferenceEngine::ITaskExecutor::Ptr m_taskExecutor;
     InferenceEngine::ITaskExecutor::Ptr m_waitExecutor;
-
-private:
-    bool is_serializable();
 };
 
 }  // namespace intel_gpu

src/plugins/intel_gpu/include/intel_gpu/plugin/plugin.hpp

@@ -18,7 +18,7 @@ namespace intel_gpu {
 class Plugin : public InferenceEngine::IInferencePlugin {
     struct impl;
     std::shared_ptr<impl> _impl;
-    bool isModelCachingEnabled = false;
+    bool isModelCachingEnabled = true;
 
     std::string default_device_id = "0";
     // key: device_id, value: cldnn device

src/plugins/intel_gpu/src/graph/impls/onednn/primitive_onednn_base.h

@@ -324,10 +324,8 @@ private:
     void build_primitive(const ExecutionConfig& config) {
         auto cache_outpath = get_cache_directory(config);
 
-        if (const char* env_p = std::getenv("OV_GPU_CACHE_MODEL")) {
-            if (env_p[0] == '1') {
-                cache_outpath = "";
-            }
+        if (!config.get_property(ov::intel_gpu::allow_new_shape_infer)) {
+            cache_outpath = "";
         }
 
         if (cache_outpath.empty()) {

src/plugins/intel_gpu/src/plugin/compiled_model.cpp

@@ -6,6 +6,7 @@
 #include "intel_gpu/graph/serialization/binary_buffer.hpp"
 #include "intel_gpu/graph/serialization/string_serializer.hpp"
 #include "intel_gpu/graph/serialization/utils.hpp"
+#include "intel_gpu/graph/serialization/vector_serializer.hpp"
 #include "intel_gpu/plugin/graph.hpp"
 #include "intel_gpu/runtime/itt.hpp"
 #include "intel_gpu/plugin/infer_request.hpp"
@@ -96,11 +97,14 @@ CompiledModel::CompiledModel(std::istream& networkModel, InferenceEngine::Remote
             std::string name;
             std::string precision;
             std::string layout;
+            InferenceEngine::SizeVector dims;
             ib >> name;
             ib >> precision;
             ib >> layout;
+            ib >> dims;
 
             DataPtr input = std::make_shared<Data>(name, Precision::FromStr(precision), cldnn::serial_util::layout_from_string(layout));
+            input->setDims(dims);
             InputInfo::Ptr infoNew = std::make_shared<InputInfo>();
             infoNew->setInputData(input);
             inputs.emplace(std::make_pair(name, infoNew));
@@ -115,11 +119,14 @@ CompiledModel::CompiledModel(std::istream& networkModel, InferenceEngine::Remote
             std::string name;
             std::string precision;
             std::string layout;
+            InferenceEngine::SizeVector dims;
             ib >> name;
             ib >> precision;
             ib >> layout;
+            ib >> dims;
 
             DataPtr output = std::make_shared<Data>(name, Precision::FromStr(precision), cldnn::serial_util::layout_from_string(layout));
+            output->setDims(dims);
             outputs.emplace(std::make_pair(name, output));
         }
 
@@ -317,14 +324,6 @@ IInferRequestInternal::Ptr CompiledModel::CreateInferRequest() {
                                                _callbackExecutor);
 }
 
-bool CompiledModel::is_serializable() {
-    // Dynamic model serialization is not yet supported.
-    if (m_graphs[0]->GetNetwork()->is_dynamic())
-        return false;
-
-    return true;
-}
-
 // Cache blob format:
 //     [ ConstInputsDataMap / ConstOutputsDataMap ]
 //     [ ov::Node::Input/ ov::Node::Output ]
@@ -334,9 +333,6 @@ void CompiledModel::Export(std::ostream& networkModel) {
     if (m_graphs.empty())
         IE_THROW(NetworkNotLoaded);
 
-    if (!is_serializable())
-        return;
-
     cldnn::BinaryOutputBuffer ob(networkModel);
 
     // InputsInfo and OutputsInfo for CNNNetwork
@@ -350,6 +346,7 @@ void CompiledModel::Export(std::ostream& networkModel) {
             std::stringstream ss;
             ss << in.second->getInputData()->getLayout();
             ob << ss.str();
+            ob << in.second->getTensorDesc().getDims();
         }
 
         ob << GetOutputsInfo().size();
@@ -361,6 +358,7 @@ void CompiledModel::Export(std::ostream& networkModel) {
             std::stringstream ss;
             ss << out.second->getLayout();
             ob << ss.str();
+            ob << out.second->getTensorDesc().getDims();
         }
     }
 

src/plugins/intel_gpu/src/plugin/plugin.cpp

@@ -142,12 +142,6 @@ Plugin::Plugin() : m_default_contexts({}) {
             m_default_contexts.insert({device.first, ctx});
         }
     }
-
-    if (const char* env_p = std::getenv("OV_GPU_CACHE_MODEL")) {
-        if (env_p[0] == '1') {
-            isModelCachingEnabled = true;
-        }
-    }
 }
 
 auto check_inputs = [](InferenceEngine::InputsDataMap _networkInputs) {
@@ -204,6 +198,9 @@ IExecutableNetworkInternal::Ptr Plugin::LoadExeNetworkImpl(const InferenceEngine
     {
         OV_ITT_SCOPED_TASK(itt::domains::intel_gpu_plugin, "Plugin::LoadExeNetworkImpl::CreateExeNetwork");
         CompiledModel::Ptr exeNetwork = std::make_shared<CompiledModel>(transformedNetwork, context, config);
+        if (exeNetwork->m_graphs[0]->GetNetwork()->is_dynamic()) {
+            isModelCachingEnabled = false;
+        }
         update_memory_statistics(context->get_impl());
         return exeNetwork;
     }

src/plugins/intel_gpu/src/runtime/kernels_cache.cpp

@@ -70,10 +70,8 @@ std::string kernels_cache::get_cache_path() const {
 }
 
 bool kernels_cache::is_cache_enabled() const {
-    if (const char* env_p = std::getenv("OV_GPU_CACHE_MODEL")) {
-        if (env_p[0] == '1') {
-            return false;
-        }
+    if (!_config.get_property(ov::intel_gpu::allow_new_shape_infer)) {
+        return false;
     }
 
     return !_config.get_property(ov::cache_dir).empty();