diff --git a/rocclr/runtime/device/comgrctx.cpp b/rocclr/runtime/device/comgrctx.cpp
index 5e71bf91a3..30b2acab1b 100644
--- a/rocclr/runtime/device/comgrctx.cpp
+++ b/rocclr/runtime/device/comgrctx.cpp
@@ -3,6 +3,7 @@
 //
 #if defined(USE_COMGR_LIBRARY)
 #include "os/os.hpp"
+#include "utils/flags.hpp"
 #include "comgrctx.hpp"
 
 namespace amd {
@@ -12,6 +13,7 @@ bool Comgr::is_ready_ = false;
 
 bool Comgr::LoadLib() {
 #if defined(COMGR_DYN_DLL)
+  LogInfo("Loading COMGR library.");
   static const char* ComgrLibName =
     LP64_SWITCH(WINDOWS_SWITCH("amdcomgr.dll", "libamdcomgr32.so"),
                 WINDOWS_SWITCH("amdcomgr64.dll", "libamdcomgr64.so"));
@@ -69,4 +71,4 @@ bool Comgr::LoadLib() {
 }
 
 }
-#endif
\ No newline at end of file
+#endif
diff --git a/rocclr/runtime/device/device.cpp b/rocclr/runtime/device/device.cpp
index be3a95727f..951d944605 100644
--- a/rocclr/runtime/device/device.cpp
+++ b/rocclr/runtime/device/device.cpp
@@ -6,6 +6,7 @@
 #include "thread/atomic.hpp"
 #include "thread/monitor.hpp"
 #include "utils/options.hpp"
+#include "comgrctx.hpp"
 
 #if defined(WITH_HSA_DEVICE)
 #include "device/rocm/rocdevice.hpp"
@@ -232,7 +233,6 @@ bool Device::ValidateComgr() {
 #if defined(USE_COMGR_LIBRARY)
   // Check if Lightning compiler was requested
   if (settings_->useLightning_) {
-    LogInfo("Loading COMGR library.");
     std::call_once(amd::Comgr::initialized, amd::Comgr::LoadLib);
     // Use Lightning only if it's available
     settings_->useLightning_ = amd::Comgr::IsReady();
diff --git a/rocclr/runtime/device/devkernel.cpp b/rocclr/runtime/device/devkernel.cpp
index e10bcf2301..b7f5d5a83b 100644
--- a/rocclr/runtime/device/devkernel.cpp
+++ b/rocclr/runtime/device/devkernel.cpp
@@ -9,6 +9,7 @@
 #include "utils/options.hpp"
 #include "utils/bif_section_labels.hpp"
 #include "utils/libUtils.h"
+#include "comgrctx.hpp"
 
 #include <map>
 #include <string>
@@ -30,6 +31,19 @@ using llvm::AMDGPU::HSAMD::ValueType;
 namespace device {
 
 #if defined(USE_COMGR_LIBRARY)
+amd_comgr_status_t getMetaBuf(const amd_comgr_metadata_node_t meta,
+                   std::string* str) {
+  size_t size = 0;
+  amd_comgr_status_t status = amd::Comgr::get_metadata_string(meta, &size, NULL);
+
+  if (status == AMD_COMGR_STATUS_SUCCESS) {
+    str->resize(size-1);    // minus one to discount the null character
+    status = amd::Comgr::get_metadata_string(meta, &size, &((*str)[0]));
+  }
+
+  return status;
+}
+
 static amd_comgr_status_t populateArgs(const amd_comgr_metadata_node_t key,
                                        const amd_comgr_metadata_node_t value,
                                        void *data) {
diff --git a/rocclr/runtime/device/devkernel.hpp b/rocclr/runtime/device/devkernel.hpp
index acff0adcec..561dfc087d 100644
--- a/rocclr/runtime/device/devkernel.hpp
+++ b/rocclr/runtime/device/devkernel.hpp
@@ -26,6 +26,7 @@ struct RuntimeHandle {
 };
 
 #if defined(USE_COMGR_LIBRARY)
+#include "amd_comgr.h"
 #include "llvm/Support/AMDGPUMetadata.h"
 typedef llvm::AMDGPU::HSAMD::Kernel::Arg::Metadata KernelArgMD;
 
@@ -383,6 +384,7 @@ class Kernel : public amd::HeapObject {
     std::string compileVecTypeHint_;  //!< kernel compiled vector type hint
     bool uniformWorkGroupSize_;       //!< uniform work group size option
     size_t wavesPerSimdHint_;         //!< waves per simd hit
+    int maxOccupancyPerCu_;           //!< Max occupancy per compute unit in threads
   };
 
   //! Default constructor
@@ -560,17 +562,6 @@ class Kernel : public amd::HeapObject {
 };
 
 #if defined(USE_COMGR_LIBRARY)
-static amd_comgr_status_t getMetaBuf(const amd_comgr_metadata_node_t meta,
-                                     std::string* str) {
-  size_t size = 0;
-  amd_comgr_status_t status = amd::Comgr::get_metadata_string(meta, &size, NULL);
-
-  if (status == AMD_COMGR_STATUS_SUCCESS) {
-    str->resize(size-1);    // minus one to discount the null character
-    status = amd::Comgr::get_metadata_string(meta, &size, &((*str)[0]));
-  }
-
-  return status;
-}
+amd_comgr_status_t getMetaBuf(const amd_comgr_metadata_node_t meta, std::string* str);
 #endif // defined(USE_COMGR_LIBRARY)
 } // namespace device
diff --git a/rocclr/runtime/device/devprogram.cpp b/rocclr/runtime/device/devprogram.cpp
index 499bee36b3..fb3d4b202a 100644
--- a/rocclr/runtime/device/devprogram.cpp
+++ b/rocclr/runtime/device/devprogram.cpp
@@ -10,6 +10,7 @@
 #include "utils/options.hpp"
 #include "utils/bif_section_labels.hpp"
 #include "utils/libUtils.h"
+#include "comgrctx.hpp"
 
 #if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
 #ifndef USE_COMGR_LIBRARY
diff --git a/rocclr/runtime/device/devprogram.hpp b/rocclr/runtime/device/devprogram.hpp
index b35cb0771e..67af239ced 100644
--- a/rocclr/runtime/device/devprogram.hpp
+++ b/rocclr/runtime/device/devprogram.hpp
@@ -8,11 +8,12 @@
 #include "platform/object.hpp"
 #include "platform/memory.hpp"
 #include "devwavelimiter.hpp"
-#include "comgrctx.hpp"
 
 #if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
 #ifndef USE_COMGR_LIBRARY
 #include "driver/AmdCompiler.h"
+#else
+#include "amd_comgr.h"
 #endif
 //#include "llvm/Support/AMDGPUMetadata.h"