[GPU] Add fusing false condition for fs_b_yx_fsv32_network at eltwise and quantize pattern

Author: kelvinchoi-intel

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

@@ -12,17 +12,18 @@
 
 using namespace cldnn;
 
-static bool eltwise_supports_fusings(eltwise_node& node) {
-    auto out_layout = node.get_output_layout();
-    // This condition refers to optimizied kernel EltwiseKernel_fs_b_yx_fsv32
-    if (out_layout.data_type == data_types::f16 && out_layout.batch() > 1 && out_layout.format == format::fs_b_yx_fsv32) {
-        return false;
-    }
+void fuse_primitives_with_layout::run(program& p) {
+    auto eltwise_supports_fusings = [&](eltwise_node& node) -> bool {
+        auto out_layout = node.get_output_layout();
+        // This condition refers to optimizied kernel EltwiseKernel_fs_b_yx_fsv32
+        if (out_layout.data_type == data_types::f16 && out_layout.batch() > 1 && 
+            (_lo.get_optimization_attributes().fs_b_yx_fsv32_network || out_layout.format == format::fs_b_yx_fsv32)) {
+            return false;
+        }
 
-    return true;
-}
+        return true;
+    };
 
-void fuse_primitives_with_layout::run(program& p) {
     bool need_recalc_processing_order = false;
     std::map<primitive_id, std::vector<std::pair<primitive_id, size_t>>> fusing_history;
 
@@ -35,7 +36,7 @@ void fuse_primitives_with_layout::run(program& p) {
             continue;
 
         // No optimized Eltwise kernel supports fused-operation for fs_b_yx_fsv32
-        // Check fusing quantize to eltwsise for this case
+        // Check fusing quantize to eltwise for this case
         auto func_fuse_quantize = [&](quantize_node& node) {
             bool should_fuse = false;
             auto out_layout = node.get_output_layout();
@@ -49,7 +50,6 @@ void fuse_primitives_with_layout::run(program& p) {
                 return;
 
             should_fuse |= input_node.is_type<eltwise>() && eltwise_supports_fusings(input_node.as<eltwise>());
-
             if (!should_fuse)
                 return;
 

src/plugins/intel_gpu/src/graph/include/pass_manager.h

@@ -397,10 +397,12 @@ class mark_runtime_skippable_nodes : public base_pass {
 
 class fuse_primitives_with_layout : public base_pass {
 public:
-    fuse_primitives_with_layout() : base_pass("fuse_primitives_with_layout") {}
+    explicit fuse_primitives_with_layout(layout_optimizer& lo_ref) :
+        base_pass("fuse_primitives_with_layout"), _lo(lo_ref) {}
 
 private:
     void run(program& p) override;
+    layout_optimizer& _lo;
 };
 
 }  // namespace cldnn

src/plugins/intel_gpu/src/graph/program.cpp

@@ -594,7 +594,7 @@ void program::pre_optimize_graph(bool is_internal) {
 
     // Check fusing primitives based on preferred format or layout optimization
     if (optimize_data) {
-        apply_opt_pass<fuse_primitives_with_layout>();
+        apply_opt_pass<fuse_primitives_with_layout>(lo);
     }
 
     // add optimization attributes for onednn primitives