Reduce runtime check

Author: Antonyvance

include/cutlass/epilogue/collective/xe_array_epilogue.hpp

@@ -430,8 +430,10 @@ class CollectiveEpilogue<
       CUTLASS_PRAGMA_UNROLL
       for (int epi_m = 0; epi_m < FragsM; epi_m++) {
 
-        if (is_C_load_needed) {
-          if constexpr (is_source_supported) {
+        //Instead of calling is_C_load_needed. We do heirachical check 
+        //so that runtime check not there when ElementC is void
+        if constexpr (is_source_supported) {
+          if (fusion_callbacks.is_C_load_needed()) {
             //cordinates for C and D are the same
             copy(params.xe_load_c.with(get<0>(load_store_tensors)), tCgD(_, epi_m, epi_n), trC);
           }

include/cutlass/epilogue/collective/xe_epilogue.hpp

@@ -410,10 +410,11 @@ class CollectiveEpilogue<
       CUTLASS_PRAGMA_UNROLL
       for (int epi_m = 0; epi_m < FragsM; epi_m++) {
         cst_callbacks.begin_loop(epi_m, epi_n);
-        
-        //avoid evaluating xe_load_c when ElementC is void during compilation
-        if (is_C_load_needed) {
-          if constexpr (is_source_supported) {
+
+        //Instead of calling is_C_load_needed. We do heirachical check 
+        //so that runtime check not there when ElementC is void
+        if constexpr (is_source_supported) {
+          if (fusion_callbacks.is_C_load_needed()) {
             copy(params.xe_load_c, tCgC(_, epi_m, epi_n), trC);
           }
         }