More refactoring

src/plugins/intel_gpu/src/plugin/transformations/fc_horizontal_fusion.cpp

@@ -13,6 +13,7 @@
 #include "openvino/pass/pattern/op/wrap_type.hpp"
 #include "transformations/utils/utils.hpp"
 #include "intel_gpu/op/placeholder.hpp"
+#include "intel_gpu/runtime/debug_configuration.hpp"
 
 namespace ov {
 namespace intel_gpu {
@@ -140,38 +141,63 @@ FullyConnectedHorizontalFusion::FullyConnectedHorizontalFusion() {
             }
         }
 
+        size_t bias_concat_axis = 0;
         if (bias_nodes.empty() && n_bias_users == fc_nodes.size()) {
-            // set user as bias
+            // Set Add user as bias input to FC
             for (size_t i = 0; i < fc_nodes.size(); ++i) {
                 auto orig_fc = fc_nodes[i];
                 auto bias_node = orig_fc->get_users()[0];
-                auto bias_const = orig_fc->get_users()[0]->input_value(1);
-                auto orig_users_of_bias_user = bias_node->get_users();
-                ov::OutputVector fc_inputs = orig_fc->input_values();
-                fc_inputs[2] = bias_const;
-                auto new_fc = orig_fc->clone_with_new_inputs(fc_inputs);
-                new_fc->set_friendly_name(orig_fc->get_friendly_name() + "_with_bias");
-                ov::copy_runtime_info(orig_fc, new_fc);
-                for (auto u : orig_users_of_bias_user) {
-                    for (size_t idx = 0; idx < u->inputs().size(); ++idx) {
-                        if (u->get_input_node_shared_ptr(idx) == bias_node) {
-                            u->input(idx).replace_source_output(new_fc->output(0));
+                auto bias_const_ptr = orig_fc->get_users()[0]->get_input_node_shared_ptr(1);
+                bias_nodes.push_back(bias_const_ptr);
+            }
+            // Check shape and find axis
+            const auto bias_rank = bias_nodes[0]->get_output_partial_shape(0).size();
+            std::vector<int32_t> bias_add_shape_diffs(bias_rank, 0);
+            for (size_t i = 1; i < bias_nodes.size(); ++i) {
+                for (size_t j = 0; j < bias_rank; ++j) {
+                    bias_add_shape_diffs[j] += (bias_nodes[i]->get_output_shape(0)[j] - bias_nodes[i - 1]->get_output_shape(0)[j]);
+                }
+            }
+            auto non_zero_diffs = std::count_if(bias_add_shape_diffs.begin(), bias_add_shape_diffs.end(), [](int32_t diff) { return diff != 0; });
+            if (non_zero_diffs <= 1) {
+                for (size_t i = 0; i < bias_rank; ++i) {
+                    if (bias_add_shape_diffs[i] != 0)
+                        bias_concat_axis = i;
+                }
+                for (size_t i = 0; i < fc_nodes.size(); ++i) {
+                    auto orig_fc = fc_nodes[i];
+                    auto bias_node = orig_fc->get_users()[0];
+                    GPU_DEBUG_TRACE_DETAIL << "Set Add op user " << bias_node->get_friendly_name() << " as the FC "
+                                           << orig_fc->get_friendly_name() << "'s bias input" << std::endl;
+                    auto bias_const = orig_fc->get_users()[0]->input_value(1);
+                    auto orig_users_of_bias_user = bias_node->get_users();
+                    ov::OutputVector fc_inputs = orig_fc->input_values();
+                    fc_inputs[2] = bias_const;
+                    auto new_fc = orig_fc->clone_with_new_inputs(fc_inputs);
+                    new_fc->set_friendly_name(orig_fc->get_friendly_name() + "_with_bias");
+                    ov::copy_runtime_info(orig_fc, new_fc);
+                    for (auto u : orig_users_of_bias_user) {
+                        for (size_t idx = 0; idx < u->inputs().size(); ++idx) {
+                            if (u->get_input_node_shared_ptr(idx) == bias_node) {
+                                u->input(idx).replace_source_output(new_fc->output(0));
+                            }
                         }
                     }
+                    fc_nodes[i] = std::dynamic_pointer_cast<op::FullyConnectedCompressed>(new_fc);
+                    bias_node->clear_control_dependencies();
+                    orig_fc->clear_control_dependencies();
                 }
-                fc_nodes[i] = std::dynamic_pointer_cast<op::FullyConnectedCompressed>(new_fc);
-                bias_nodes.push_back(fc_nodes[i]->get_input_node_shared_ptr(2));
-                bias_node->clear_control_dependencies();
-                orig_fc->clear_control_dependencies();
+            } else {
+                // biases cannot be fusable. Not to set users as bias input
+                bias_nodes.clear();
             }
         }
         std::shared_ptr<ov::Node> fused_bias;
-        if (bias_nodes.size() > 1) {
+        if (bias_nodes.size() == fc_nodes.size()) {
             ov::OutputVector bias_nodes_as_output_vector;
             for (size_t i = 0; i < bias_nodes.size(); ++i) {
                 bias_nodes_as_output_vector.push_back(bias_nodes[i]);
             }
-            const auto bias_concat_axis = 2;
             fused_bias = std::make_shared<ov::op::v0::Concat>(bias_nodes_as_output_vector, bias_concat_axis);
             fused_bias->set_friendly_name(bias_nodes[0]->get_friendly_name() + "_fused_bias");
             ov::copy_runtime_info(bias_nodes, fused_bias);
@@ -261,7 +287,7 @@ FullyConnectedHorizontalFusion::FullyConnectedHorizontalFusion() {
             }
             org_fc->clear_control_dependencies();
         }
-        std::cout << "new fc: " << new_fc_name << std::endl;
+        GPU_DEBUG_TRACE_DETAIL << "Created a new fused FC " << new_fc_name << std::endl;
         return true;
     };