diff --git a/python/akg/composite/topi.py b/python/akg/composite/topi.py
index 3bcff8760888f54831a5b785e2ec3ff43c42acc9..4ab8b251fde3a30458ea92b759c497813b62f592 100644
--- a/python/akg/composite/topi.py
+++ b/python/akg/composite/topi.py
@@ -17,9 +17,15 @@
 """composite topi"""
 import akg.topi as topi
 from akg import tvm
+from akg.tvm.hybrid import script
 from akg.utils.format_transform import get_const
 from akg.utils import validation_check as vc_util
 
+import logging
+import os
+from pathlib import Path
+import importlib.util
+
 
 @tvm.register_func("ElemAny")
 def elem_any(inputs, attrs):
@@ -405,4 +411,77 @@ def cumprod(inputs, attrs):
                     ib.store(dst, idx_cur, ib.load(dst, idx_pre) * ib.load(data, idx_pre if exclusive else idx_cur))
         return ib.get()
     return tvm.extern(shape, [in_tensor], lambda ins, outs : kernel_ir(ins[0], outs[0]), name=output_name,
-                          dtype=in_tensor.dtype)
\ No newline at end of file
+                          dtype=in_tensor.dtype)
+
+@tvm.register_func("UserDefined")
+def user_defined(inputs, attrs):
+
+    op_desc_attr = []
+    source_str = ""
+    op_imply_path = ""
+    func_name = ""
+    func_type = ""
+
+    for ext_arg in attrs.items():
+        attr_name = ext_arg[0]
+        if attr_name == "func_source_str":
+            source_str = ext_arg[1].value
+        elif attr_name == "op_imply_path":
+            op_imply_path = ext_arg[1].value
+        elif attr_name == "func_name":
+            func_name = ext_arg[1].value
+        elif attr_name == "func_type":
+            func_type = ext_arg[1].value
+        elif not (attr_name == "akg" or "_format" in attr_name):
+            # store the rest of op attr for op build
+            op_desc_attr.append(ext_arg)
+
+    op_attrs = []
+    ir_builder_attrs = {}
+    for ext_arg in op_desc_attr:
+        if func_type == "ir_builder":
+            # ir_builder functions take attrs as a dict/tvm.Map
+            ir_builder_attrs[ext_arg[0]] = ext_arg[1]
+        else:
+            op_attrs.append(ext_arg[1])
+
+    func_kernel = None
+    if len(source_str) > 0:
+        capture = locals()
+        capture["source_str"] = source_str
+
+        func_mod = compile(source_str, "", "exec")
+        exec(func_mod)
+        func_kernel = locals()[func_name]
+
+    elif len(op_imply_path) > 0:
+        if os.path.isfile(op_imply_path):
+            custom_mod_name = Path(op_imply_path).resolve().stem
+            mod_spec = importlib.util.spec_from_file_location(
+                custom_mod_name, op_imply_path)
+            custom_mod = importlib.util.module_from_spec(mod_spec)
+            mod_spec.loader.exec_module(custom_mod)
+            func_kernel = getattr(custom_mod, func_name, None)
+        else:
+            logging.error("Can't find file under path: %s",
+                          str(op_imply_path))
+    else:
+        logging.error(
+            "Neither source_str nor op_imply_path is provided in the json file")
+
+    if func_kernel is None:
+        logging.error(
+            "Failed in compiling op function from userdefine op")
+
+    output = None
+    if func_type == "hybrid":
+        hybrid_func = script(func_kernel, capture=capture)
+        inputs = list(inputs) + op_attrs
+        output = hybrid_func(*inputs)
+    elif func_type == "ir_builder":
+        output = func_kernel(inputs, ir_builder_attrs)
+    else:
+        inputs = list(inputs) + op_attrs
+        output = func_kernel(*inputs)
+
+    return output
\ No newline at end of file
diff --git a/src/composite/util.cc b/src/composite/util.cc
index 0022986640398d6859f6e2e5bd94241b2da8d856..7ff50d90f13a73e27677f1be01d2d131f5f02840 100644
--- a/src/composite/util.cc
+++ b/src/composite/util.cc
@@ -77,7 +77,7 @@ bool IsOtherOp(const std::string &op_name) {
   std::unordered_set<std::string> elems = {"MatMul",      "BatchMatMul",   "Conv",        "Transpose", "Tile",
                                            "Assign",      "InplaceAssign", "EquivFormat", "TransData", "AddMinValue",
                                            "BroadcastTo", "PadAkg",        "UnPadAkg",    "Conv2D", "CumSum",
-                                           "CumProd",     "StridedSlice"};
+                                           "CumProd",     "StridedSlice",  "UserDefined"};
   return elems.find(op_name) != elems.end();
 }
 bool IsElemwise(const std::string &op_name) {
diff --git a/third_party/incubator-tvm/python/tvm/hybrid/__init__.py b/third_party/incubator-tvm/python/tvm/hybrid/__init__.py
index abe59a062f54a36e1c14d027b82d1450bb5b1326..594457d9d4706c80a34d988b3cddd0a85f97ecca 100644
--- a/third_party/incubator-tvm/python/tvm/hybrid/__init__.py
+++ b/third_party/incubator-tvm/python/tvm/hybrid/__init__.py
@@ -26,6 +26,7 @@ HalideIR.
 """
 
 # 2019.12.30 - Enhance function script.
+# 2021.08.17 - Update getting source logic to support UserDefined ops
 
 # TODO(@were): Make this module more complete.
 # 1. Support HalideIR dumping to Hybrid Script
@@ -95,7 +96,10 @@ def script(pyfunc=None, intrinsics=None, capture=None):
             if _is_tvm_arg_types(args):
                 _patch_intrins_to_calls(intrinsics=intrinsics)
                 _patch_intrins_to_runtime(intrinsics=intrinsics)
-                src = _pruned_source(func)
+                if capture.get("source_str"):
+                    src = capture["source_str"]
+                else:
+                    src = _pruned_source(func)
                 op = source_to_op(src, args, func.__globals__, closure_vars)
                 _unpatch_intrins_from_runtime(intrinsics=intrinsics)
                 _unpatch_intrins_from_calls(intrinsics=intrinsics)