diff --git a/src/core/CL/CLCompileContext.cpp b/src/core/CL/CLCompileContext.cpp
index 3db0fe5..bf3a866 100644
--- a/src/core/CL/CLCompileContext.cpp
+++ b/src/core/CL/CLCompileContext.cpp
@@ -29,6 +29,8 @@
 #include "arm_compute/core/Utils.h"
 #include "support/StringSupport.h"
 
+#include <regex>
+
 namespace arm_compute
 {
 CLBuildOptions::CLBuildOptions()
@@ -263,6 +265,19 @@
         ARM_COMPUTE_ERROR("Non uniform workgroup size is not supported!!");
     }
 
+    const GPUTarget arch = get_arch_from_target(_device.target());
+    if(arch != GPUTarget::UNKNOWN && arch != GPUTarget::MIDGARD)
+    {
+        const std::string device_vers = _device.device_version();
+        const std::regex  ddk_regex("r([0-9]*)p[0-9]");
+        std::smatch       ddk_match;
+
+        if(std::regex_search(device_vers, ddk_match, ddk_regex) && std::stoi(ddk_match[1]) >= 9)
+        {
+            concat_str += " -DUNROLL_WITH_PRAGMA ";
+        }
+    }
+
     std::string build_options = stringify_set(build_options_set, kernel_path) + concat_str;
 
     return build_options;
diff --git a/src/core/CL/cl_kernels/tile_helpers.h b/src/core/CL/cl_kernels/tile_helpers.h
index 4959c04..f2d2f26 100644
--- a/src/core/CL/cl_kernels/tile_helpers.h
+++ b/src/core/CL/cl_kernels/tile_helpers.h
@@ -70,6 +70,7 @@
 #define TENSOR4D_STR(name, type) TENSOR4D_##type(name)
 #define TENSOR4D(name, type) TENSOR4D_STR(name, type)
 
+#if !defined(UNROLL_WITH_PRAGMA)
 #define UNROLL_INCR(idx, step, macro) idx += (step); (macro)
 
 #define LOOP_UNROLLING_1(idx, step, macro) (macro)
@@ -201,12 +202,22 @@
 #define LOOP_UNROLLING_127(idx, step, macro) LOOP_UNROLLING_126(idx, step, macro); UNROLL_INCR(idx, step, macro)
 #define LOOP_UNROLLING_128(idx, step, macro) LOOP_UNROLLING_127(idx, step, macro); UNROLL_INCR(idx, step, macro)
 
-#define LOOP_UNROLLING(type, idx, start, step, num, macro) LOOP_UNROLLING_STR(type, idx, start, step, num, macro)
 #define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \
     {                                                          \
         type idx = start;                                      \
         LOOP_UNROLLING_##num(idx, step, macro);                \
     }
+#else // !defined(UNROLL_WITH_PRAGMA)
+#define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \
+    {                                                          \
+        _Pragma("unroll")                                      \
+        for(type idx = start; idx < (num * step); idx += step) \
+        {                                                      \
+            (macro);                                           \
+        }                                                      \
+    }
+#endif // !defined(UNROLL_WITH_PRAGMA)
+#define LOOP_UNROLLING(type, idx, start, step, num, macro) LOOP_UNROLLING_STR(type, idx, start, step, num, macro)
 
 /** Get the get_global_id with partial N0. This function is useful when the dimension is not multiple of N0 and we need to use a partial N0
  *  to avoid out-of-bound read/write