Workaround for nested fused op: run FusedOpDecomposition twice.

src/ngraph/runtime/gpu/gpu_compiled_function.cpp

@@ -172,6 +172,7 @@ void runtime::gpu::GPUCompiledFunction::compile()
     pass_manager.register_pass<runtime::gpu::pass::BatchNormCache>();
     pass_manager.register_pass<ngraph::pass::LikeReplacement>();
     pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>();
+    pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>();
     pass_manager.register_pass<runtime::gpu::pass::GPULayout>(this);
     pass_manager.register_pass<ngraph::pass::AssignLayout<descriptor::layout::DenseTensorLayout>>();
     pass_manager.register_pass<ngraph::pass::GetOutputElementElimination>();

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

@@ -425,6 +425,8 @@ shared_ptr<runtime::Executable>
     {
         pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>(
             IntelGPUBackend::is_supported_impl);
+        pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>(
+            IntelGPUBackend::is_supported_impl);
     }
 
     if (m_disable_backend_optimizations < 1)