[GPU] Optimize Condition operation by better integrating its subnetwork primitives with the main network (#21062)

* [GPU] Optimize Condition operation by better integrating its subnetwork primitives with the main network

* Reuse main network's stream in sub-networks

src/plugins/intel_gpu/src/graph/condition.cpp

@@ -219,8 +219,8 @@ Condition primitive is reusing memory with the input.
 */
 condition_inst::typed_primitive_inst(network& network, condition_node const& node)
     : parent(network, node),
-      _net_true(network::allocate_network(node.get_program().get_engine(), node.get_branch_true().inner_program)),
-      _net_false(network::allocate_network(node.get_program().get_engine(), node.get_branch_false().inner_program)) {
+      _net_true(network::allocate_network(network.get_stream_ptr(), node.get_branch_true().inner_program)),
+      _net_false(network::allocate_network(network.get_stream_ptr(), node.get_branch_false().inner_program)) {
     this->set_inner_networks({_net_true, _net_false});
 }
 
@@ -257,7 +257,7 @@ void condition_inst::postprocess_output_memory(network::ptr executed_net, cldnn:
     for (auto out_mem_map : branch.output_map) {
         auto out_mem_idx = out_mem_map.first;
         auto inner_out_id = out_mem_map.second;
-        auto mem_ptr = executed_net->get_output(inner_out_id).get_memory();
+        auto mem_ptr = executed_net->get_output_memory(inner_out_id);
         if (mem_ptr) {
             auto layout = _impl_params->get_output_layout(out_mem_idx);
             GPU_DEBUG_LOG << "Reshape output from " << mem_ptr->get_layout().to_short_string()

src/plugins/intel_gpu/src/graph/impls/common/condition.cpp

@@ -30,13 +30,15 @@ struct condition_impl : typed_primitive_impl<condition> {
     }
 
     event::ptr execute_impl(const std::vector<event::ptr>& events, condition_inst& instance) override {
-        for (auto& a : events) {
-            a->wait();
-        }
-        auto ev = instance.get_network().get_stream().create_user_event(false);
+        // Wait for condition statement event only, and pass all other events to sub-network directly
+        if (!events.empty())
+            events[0]->wait();
+
+        auto& stream = instance.get_network().get_stream();
+        auto ev = stream.create_user_event(false);
         set_node_params(instance.get_node());
 
-        auto pred = condition_inst::get_pred_from_memory(instance.pred_memory_ptr(), instance.get_network().get_stream());
+        auto pred = condition_inst::get_pred_from_memory(instance.pred_memory_ptr(), stream);
         network::ptr executed_net = pred ? instance.get_net_true() : instance.get_net_false();
         auto branch = pred ? instance.get_branch_true() : instance.get_branch_false();
         executed_net->set_shape_predictor(instance.get_network().get_shape_predictor());
@@ -62,7 +64,7 @@ struct condition_impl : typed_primitive_impl<condition> {
             }
         }
 
-        executed_net->execute({});
+        auto sub_net_results = executed_net->execute(events);
 
         // Update output layout of impl_param in condition_inst
         instance.update_output_layout();
@@ -70,8 +72,12 @@ struct condition_impl : typed_primitive_impl<condition> {
         // Set output memory of condition_inst to inner network output memory after inner network execution
         instance.postprocess_output_memory(executed_net, branch);
 
-        ev->set();
-        return ev;
+        std::vector<event::ptr> output_events;
+        for (auto& output : sub_net_results)
+            if (output.second.get_event() != nullptr)
+                output_events.push_back(output.second.get_event());
+
+        return stream.group_events(output_events);
     }
 
     static std::unique_ptr<primitive_impl> create(const condition_node& arg, const kernel_impl_params&) {

src/plugins/intel_gpu/src/runtime/ocl/ocl_stream.cpp

@@ -61,7 +61,9 @@ cl_int set_kernel_arg(ocl_kernel_type& kernel, uint32_t idx, cldnn::memory::cptr
         return kernel.setArg(idx, buf);
     } else if (memory_capabilities::is_usm_type(mem->get_allocation_type())) {
         auto buf = std::dynamic_pointer_cast<const ocl::gpu_usm>(mem)->get_buffer();
-        GPU_DEBUG_TRACE_DETAIL << "kernel: " << kernel.get() << " set arg (usm) " << idx << " mem: " << buf.get() << " size: " << mem->size() << std::endl;
+        auto mem_type = std::dynamic_pointer_cast<const ocl::gpu_usm>(mem)->get_allocation_type();
+        GPU_DEBUG_TRACE_DETAIL << "kernel: " << kernel.get() << " set arg (" << mem_type << ") " << idx
+                               << " mem: " << buf.get() << " size: " << mem->size() << std::endl;
         return kernel.setArgUsm(idx, buf);
     } else {
         auto buf = std::dynamic_pointer_cast<const ocl::gpu_buffer>(mem)->get_buffer();