diff --git a/rocclr/compiler/lib/utils/amdilUtils.cpp b/rocclr/compiler/lib/utils/amdilUtils.cpp
new file mode 100644
index 0000000000..babdd4a5ec
--- /dev/null
+++ b/rocclr/compiler/lib/utils/amdilUtils.cpp
@@ -0,0 +1,14 @@
+#include "amdilUtils.hpp"
+#include <regex>
+#include <string>
+#include <sstream>
+
+// Change all private uav length in a kernel
+void amdilUtils::changePrivateUAVLength(std::string& kernel, unsigned length) {
+  std::regex pattern("dcl_typeless_uav_id\\(([[:digit:]]+)\\)_stride"
+      "\\(([[:digit:]]+)\\)_length\\([[:digit:]]+\\)_access\\(private\\)");
+  std::stringstream ss;
+  ss << "dcl_typeless_uav_id($1)_stride($2)_length(" << length <<
+      ")_access(private)";
+  kernel = std::regex_replace(kernel, pattern, ss.str());
+}
diff --git a/rocclr/compiler/lib/utils/amdilUtils.hpp b/rocclr/compiler/lib/utils/amdilUtils.hpp
new file mode 100644
index 0000000000..f28bcc3ee7
--- /dev/null
+++ b/rocclr/compiler/lib/utils/amdilUtils.hpp
@@ -0,0 +1,11 @@
+#ifndef AMDILUTILS_H_
+#define AMDILUTILS_H_
+
+#include <string>
+
+namespace amdilUtils {
+// Change all private uav length in a kernel
+void changePrivateUAVLength(std::string& kernel, unsigned length);
+
+}
+#endif /* AMDILUTILS_H_ */
diff --git a/rocclr/runtime/device/gpu/gpuprogram.cpp b/rocclr/runtime/device/gpu/gpuprogram.cpp
index 896e307d8c..09b3afb24c 100644
--- a/rocclr/runtime/device/gpu/gpuprogram.cpp
+++ b/rocclr/runtime/device/gpu/gpuprogram.cpp
@@ -5,6 +5,7 @@
 #include "os/os.hpp"
 #include "utils/flags.hpp"
 #include "include/aclTypes.h"
+#include "utils/amdilUtils.hpp"
 #include "utils/bif_section_labels.hpp"
 #include "device/gpu/gpuprogram.hpp"
 #include "device/gpu/gpublit.hpp"
@@ -232,6 +233,21 @@ NullProgram::isCalled(const ILFunc* base, const ILFunc* func)
     return false;
 }
 
+uint
+ILFunc::totalHwPrivateUsage() {
+  if (totalHwPrivateSize_ >= 0)
+    return totalHwPrivateSize_;
+
+  uint maxChildUsage = 0;
+  for (size_t i = 0; i < calls_.size(); ++i) {
+    uint childUsage = calls_[i]->totalHwPrivateUsage();
+    if (childUsage > maxChildUsage)
+      maxChildUsage = childUsage;
+  }
+  totalHwPrivateSize_ = hwPrivateSize_ + maxChildUsage;
+  return totalHwPrivateSize_;
+}
+
 void
 NullProgram::patchMain(std::string& kernel, uint index)
 {
@@ -540,7 +556,6 @@ NullProgram::linkImpl(amd::option::Options* options)
                     // Accumulate all emulated local and private sizes,
                     // necessary for the kernel execution
                     initData.localSize_   += func->localSize_;
-                    initData.privateSize_ += func->privateSize_;
 
                     // Accumulate all HW local and private sizes,
                     // necessary for the kernel execution
@@ -548,6 +563,9 @@ NullProgram::linkImpl(amd::option::Options* options)
                     initData.hwPrivateSize_ += func->hwPrivateSize_;
                     initData.flags_         |= func->flags_;
                 }
+                initData.privateSize_ = baseFunc->totalHwPrivateUsage();
+                amdilUtils::changePrivateUAVLength(kernel,
+                    initData.privateSize_);
 
                 // Create a GPU kernel
                 bool    created;
@@ -913,7 +931,6 @@ NullProgram::linkImpl(const std::vector<device::Program*>& inputPrograms,
                     // Accumulate all emulated local and private sizes,
                     // necessary for the kernel execution
                     initData.localSize_   += func->localSize_;
-                    initData.privateSize_ += func->privateSize_;
 
                     // Accumulate all HW local and private sizes,
                     // necessary for the kernel execution
@@ -921,6 +938,9 @@ NullProgram::linkImpl(const std::vector<device::Program*>& inputPrograms,
                     initData.hwPrivateSize_ += func->hwPrivateSize_;
                     initData.flags_         |= func->flags_;
                 }
+                initData.privateSize_ = baseFunc->totalHwPrivateUsage();
+                amdilUtils::changePrivateUAVLength(kernel,
+                    initData.privateSize_);
 
                 // Create a GPU kernel
                 bool    created;
diff --git a/rocclr/runtime/device/gpu/gpuprogram.hpp b/rocclr/runtime/device/gpu/gpuprogram.hpp
index 6cc151b353..17b9850546 100644
--- a/rocclr/runtime/device/gpu/gpuprogram.hpp
+++ b/rocclr/runtime/device/gpu/gpuprogram.hpp
@@ -50,6 +50,7 @@ public:
         , hwPrivateSize_(0)
         , hwLocalSize_(0)
         , flags_(0)
+        , totalHwPrivateSize_(-1)
     {
         code_.begin_ = code_.end_ = 0;
         metadata_.begin_ = metadata_.end_ = 0;
@@ -74,6 +75,7 @@ public:
         hwPrivateSize_  = func.hwPrivateSize_;
         hwLocalSize_    = func.hwLocalSize_;
         flags_          = func.flags_;
+        totalHwPrivateSize_ = func.totalHwPrivateSize_;
 
         // Note: we don't copy calls_ and macros_
         return *this;
@@ -89,9 +91,11 @@ public:
     uint        hwPrivateSize_; //!< HW private ring allocation by the function
     uint        hwLocalSize_;   //!< HW local ring allocation by the function
     uint        flags_;         //!< The IL func flags/properties
-
-    std::vector<const ILFunc*>  calls_; //! Functions called from the current
+    long long   totalHwPrivateSize_; //!< total HW private usage including called functions
+    std::vector<ILFunc*>  calls_; //! Functions called from the current
     std::vector<uint>  macros_;         //! Macros, used in the IL function
+
+    uint totalHwPrivateUsage();   //!< total HW private usage including called functions
 };
 
 //! \class empty program