diff --git a/projects/clr/rocclr/runtime/device/appprofile.hpp b/projects/clr/rocclr/runtime/device/appprofile.hpp
index b81f1fe4cb..c328419447 100644
--- a/projects/clr/rocclr/runtime/device/appprofile.hpp
+++ b/projects/clr/rocclr/runtime/device/appprofile.hpp
@@ -17,6 +17,9 @@ public:
     bool init();
 
     const std::string& GetBuildOptsAppend() const { return buildOptsAppend_; }
+
+    const std::string& appFileName() const { return appFileName_; }
+
 protected:
     enum DataTypes
     {
diff --git a/projects/clr/rocclr/runtime/platform/program.cpp b/projects/clr/rocclr/runtime/platform/program.cpp
index f883deca97..a4c7b3238a 100644
--- a/projects/clr/rocclr/runtime/platform/program.cpp
+++ b/projects/clr/rocclr/runtime/platform/program.cpp
@@ -12,6 +12,9 @@
 
 #include <cstdlib> // for malloc
 #include <cstring> // for strcmp
+#include <sstream>
+#include <fstream>
+#include <iostream>
 #include <utility>
 
 namespace amd {
@@ -413,6 +416,46 @@ Program::link(
     return retval;
 }
 
+void Program::StubProgramSource(const std::string& app_name)
+{
+    static uint program_counter = 0;
+    std::fstream        stub_read;
+    std::stringstream   file_name;
+    std::string         app_name_no_ext;
+
+    std::size_t length = app_name.rfind(".exe");
+    if (length == std::string::npos) {
+        length = app_name.size();
+    }
+    app_name_no_ext.assign(app_name.c_str(), length);
+
+    // Construct a unique file name for the CL program
+    file_name << app_name_no_ext << "_program_" << program_counter << ".cl";
+
+    stub_read.open(file_name.str().c_str(), (std::fstream::in | std::fstream::binary));
+    // Check if we have OpenCL program
+    if (stub_read.is_open()) {
+        // Find the stream size
+        stub_read.seekg(0, std::fstream::end);
+        size_t size = stub_read.tellg();
+        stub_read.seekg(0, std::ios::beg);
+
+        char* data = new char[size];
+        stub_read.read(data, size);
+        stub_read.close();
+
+        sourceCode_.assign(data, size);
+        delete[] data;
+    }
+    else {
+        std::fstream    stub_write;
+        stub_write.open(file_name.str().c_str(), (std::fstream::out | std::fstream::binary));
+        stub_write << sourceCode_;
+        stub_write.close();
+    }
+    program_counter++;
+}
+
 cl_int
 Program::build(
     const std::vector<Device*>& devices,
@@ -431,6 +474,11 @@ Program::build(
         }
     }
 
+    if (OCL_STUB_PROGRAMS && !sourceCode_.empty()) {
+        // The app name should be the samme for all device
+        StubProgramSource(devices[0]->appProfile()->appFileName());
+    }
+
     // Clear the program object
     clear();
 
diff --git a/projects/clr/rocclr/runtime/platform/program.hpp b/projects/clr/rocclr/runtime/platform/program.hpp
index aa7638c59b..bcc89e0733 100644
--- a/projects/clr/rocclr/runtime/platform/program.hpp
+++ b/projects/clr/rocclr/runtime/platform/program.hpp
@@ -77,6 +77,9 @@ public:
     typedef std::map<std::string, Symbol> symbols_t;
 
 private:
+    //! Replaces the compiled program with the new version from HD
+    void StubProgramSource(const std::string& app_name);
+
     //! The context this program is part of.
     SharedReference<Context> context_;
 
diff --git a/projects/clr/rocclr/runtime/utils/flags.hpp b/projects/clr/rocclr/runtime/utils/flags.hpp
index 0dcbfdd03a..cbed0e4567 100644
--- a/projects/clr/rocclr/runtime/utils/flags.hpp
+++ b/projects/clr/rocclr/runtime/utils/flags.hpp
@@ -181,6 +181,8 @@ release(uint, GPU_WAVES_PER_SIMD, 0,                                          \
         "Force the number of waves per SIMD (1-10)")                          \
 release(bool, GPU_WAVE_LIMIT_ENABLE, false,                                   \
         "1 = Enable adaptive wave limiter")                                   \
+release(bool, OCL_STUB_PROGRAMS, false,                                       \
+        "1 = Enables OCL programs stubing")                                   \
 release_on_stg(uint, GPU_WAVE_LIMIT_CU_PER_SH, 0,                             \
         "Assume the number of CU per SH for wave limiter")                    \
 release_on_stg(uint, GPU_WAVE_LIMIT_MAX_WAVE, 10,                             \