IntelGPU backend: Function no cache usage control (#1959)

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

@@ -349,6 +349,15 @@ runtime::intelgpu::IntelGPUBackend::IntelGPUBackend()
         m_cldnn_dump_enable = true;
     }
 
+    // Delete compiled Function from the cache after execution.
+    // It helps in cases where a lot of small functions used
+    // in case of memory consumption. It slow overall execution
+    // because Function compilation required every time
+    if (getenv("NGRAPH_INTELGPU_FUNCTION_CACHE_DISABLE") != nullptr)
+    {
+        m_function_cache_disabled = true;
+    }
+
     cldnn::engine_configuration cldnn_configuration(profiling);
     ocl_engine = make_shared<cldnn::engine>(cldnn_configuration);
 }
@@ -1614,6 +1623,11 @@ bool runtime::intelgpu::IntelGPUBackend::call(shared_ptr<Function> func,
                                mem_after_call);
     }
 
+    if (m_function_cache_disabled)
+    {
+        remove_compiled_function(func);
+    }
+
     return true;
 }
 

src/ngraph/runtime/intelgpu/intelgpu_backend.hpp

@@ -85,6 +85,7 @@ private:
     bool m_dump_graph_enable = false;
     bool m_cldnn_graph_optimize = true;
     bool m_cldnn_dump_enable = false;
+    bool m_function_cache_disabled = false;
     std::string m_cldnn_dump_dir = std::string("intelgpu_codegen");
     std::string delim = std::string(":");
 };