[GPU] Add intermediate weight reorder including siblings without impl

yet

Author: kelvinchoi-intel

src/plugins/intel_gpu/src/graph/graph_optimizer/post_optimize_weights.cpp

@@ -129,16 +129,31 @@ void post_optimize_weights::optimize_weights(T& node, program& p) {
                     auto& weights_reorder_node = node.get_dependency(i);
                     weights_reorder_node.get_output_layout(false);
                 } else {
+                    auto siblings = prev_node.get_users();
                     auto weights_reorder = _rf.get_weights_reorder(prev_node.id(), weights_reorder_params);
                     // insert new weights reorder node to topology
                     p.add_intermediate(weights_reorder.first, node, i, !weights_reorder.second);
                     // set weights reorder's node output layout and implementation
                     auto& weights_reorder_node = node.get_dependency(i);
                     weights_reorder_node.get_output_layout(false);
 
+                    // apply to other siblings
+                    for (auto sib : siblings) {
+                        auto sib_impl = sib->get_selected_impl();
+                        if (sib->id().compare(node.id()) != 0 && sib->id().compare(weights_reorder.first->id) != 0 && !sib_impl) {
+                            auto weights_reorder2 = _rf.get_weights_reorder(prev_node.id(), weights_reorder_params);
+                            p.add_intermediate(weights_reorder2.first, *sib, i, !weights_reorder2.second);
+
+                            // auto& weights_reorder_node = node.get_dependency(i);
+                            // weights_reorder_node.get_output_layout(false);
+                        }
+                    }
+
                     if (!weights_reorder.second) {
                         set_implementation(weights_reorder_node);
                     }
+
+
                 }
             }
         }

src/plugins/intel_gpu/tests/unit/passes/post_optimize_weights.cpp

@@ -2,7 +2,9 @@
 // SPDX-License-Identifier: Apache-2.0
 //
 
+#include "intel_gpu/primitives/permute.hpp"
 #include "test_utils.h"
+#include "random_generator.hpp"
 #include "program_wrapper.h"
 #include "fully_connected_inst.h"
 #include "convolution_inst.h"
@@ -242,3 +244,49 @@ TEST(post_optimize_weights, onednn_group_conv_weights_reorder_test) {
     ASSERT_TRUE(onednn_weights_params->_out_desc.get_size() == prog->get_node("weights_weights_reorder_0").get_output_layout().bytes_count());
 #endif
 }
+
+TEST(post_optimize_weights, fuse_constant_transposes_removal_and_add_intermediate_including_siblings) {
+    auto& engine = get_test_engine();
+
+    auto input2_layout_dyn = layout{ ov::PartialShape{ -1, 32 }, data_types::f16, format::bfyx };
+
+    auto input = engine.allocate_memory({ { 2, 32 }, data_types::f16, format::bfyx });
+    auto input2 = engine.allocate_memory({ { 2, 32 }, data_types::f16, format::bfyx });
+    auto weights = engine.allocate_memory({{ 32, 2 }, data_types::f32, format::bfyx });
+
+    tests::random_generator rg(GET_SUITE_NAME);
+    auto input_data = rg.generate_random_2d<ov::float16>(2, 32, -1, 1);
+    auto input2_data = rg.generate_random_2d<ov::float16>(2, 32, -1, -1);
+    auto weights_data = rg.generate_random_2d<float>(32, 2, -1, 1);
+
+    set_values(input, flatten_2d(format::bfyx, input_data));
+    set_values(input2, input2_data);
+    set_values(weights, flatten_2d(format::bfyx, weights_data));
+
+    topology topology(
+        input_layout("input", input->get_layout()),
+        input_layout("input2", input2_layout_dyn),
+        data("weights", weights),
+        permute("permute_test", input_info("weights"), {1, 0}),
+        reorder("reorder_dt", input_info("permute_test"), format::any, data_types::f16, std::vector<float>()),
+        fully_connected("fc1", input_info("input"), { "reorder_dt" }, "", data_types::f16),
+        fully_connected("fc2", input_info("input2"), { "reorder_dt" }, "", data_types::f16)
+    );
+
+    ExecutionConfig config = get_test_default_config(engine);
+    config.set_property(ov::intel_gpu::optimize_data(true));
+    config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
+
+    if (engine.get_device_info().supports_immad) {
+        ov::intel_gpu::ImplementationDesc fc_impl = { format::bfyx, "", impl_types::onednn };
+        config.set_property(ov::intel_gpu::force_implementations(ov::intel_gpu::ImplForcingMap{ {"fc1", fc_impl} }));
+    }
+
+    cldnn::network network(engine, topology, config);
+    network.set_input_data("input", input);
+    network.set_input_data("input2", input2);
+
+    auto outputs = network.execute();
+    auto output = outputs.at("fc1").get_memory();
+    cldnn::mem_lock<ov::float16> output_ptr(output, get_test_stream());
+}