diff --git a/rocclr/device/pal/palresource.cpp b/rocclr/device/pal/palresource.cpp
index 370c585ede..45b2df8bb4 100644
--- a/rocclr/device/pal/palresource.cpp
+++ b/rocclr/device/pal/palresource.cpp
@@ -1828,9 +1828,7 @@ void* Resource::gpuMemoryMap(size_t* pitch, uint flags, Pal::IGpuMemory* resourc
 // ================================================================================================
 void Resource::gpuMemoryUnmap(Pal::IGpuMemory* resource) const {
   if (desc_.cardMemory_ && !isPersistentDirectMap()) {
-    // @todo remove const cast
     Unimplemented();
-    //        const_cast<Device&>(dev()).resUnmapLocal(resource);
   } else {
     Pal::Result result = resource->Unmap();
     if (Pal::Result::Success != result) {
@@ -1928,6 +1926,10 @@ bool Resource::isPersistentDirectMap(bool writeMap) const {
     directMap = IS_WINDOWS && !dev().settings().linearPersistentImage_;
   }
 
+  if (memoryType() == View) {
+    directMap = viewOwner_->isPersistentDirectMap();
+  }
+
   return directMap;
 }
 
diff --git a/rocclr/device/pal/palsettings.cpp b/rocclr/device/pal/palsettings.cpp
index 877498059a..aecb4fde84 100644
--- a/rocclr/device/pal/palsettings.cpp
+++ b/rocclr/device/pal/palsettings.cpp
@@ -147,6 +147,7 @@ Settings::Settings() {
   alwaysResident_ = amd::IS_HIP ? true : false;
   prepinnedMinSize_ = 0;
   cpDmaCopySizeMax_ = GPU_CP_DMA_COPY_SIZE * Ki;
+  useDeviceKernelArg_ = flagIsDefault(HIP_FORCE_DEV_KERNARG) ? false : HIP_FORCE_DEV_KERNARG;
 }
 
 bool Settings::create(const Pal::DeviceProperties& palProp,
diff --git a/rocclr/device/pal/palsettings.hpp b/rocclr/device/pal/palsettings.hpp
index 32d3ad2a14..cb8e3ce890 100644
--- a/rocclr/device/pal/palsettings.hpp
+++ b/rocclr/device/pal/palsettings.hpp
@@ -80,7 +80,8 @@ class Settings : public device::Settings {
       uint imageBufferWar_ : 1;         //!< Image buffer workaround for Gfx10
       uint disableSdma_ : 1;            //!< Disable SDMA support
       uint alwaysResident_ : 1;         //!< Make resources resident at allocation time
-      uint reserved_ : 9;
+      uint useDeviceKernelArg_ : 1;     //!< Use persistent memory for kernel arguments
+      uint reserved_ : 8;
     };
     uint value_;
   };
diff --git a/rocclr/device/pal/palvirtual.cpp b/rocclr/device/pal/palvirtual.cpp
index 96a199d9eb..2e4d23c72d 100644
--- a/rocclr/device/pal/palvirtual.cpp
+++ b/rocclr/device/pal/palvirtual.cpp
@@ -949,8 +949,17 @@ bool VirtualGPU::create(bool profiling, uint deviceQueueSize, uint rtCUs,
     return false;
   }
 
-  if (!managedBuffer_.create(Resource::RemoteUSWC)) {
-    return false;
+  // Create buffers for kernel arg management
+  if (!managedBuffer_.create(
+      dev().settings().useDeviceKernelArg_ ? Resource::Persistent : Resource::RemoteUSWC)) {
+    // Try just USWC if persistent memory failed
+    if (dev().settings().useDeviceKernelArg_) {
+      if (!managedBuffer_.create(Resource::RemoteUSWC)) {
+        return false;
+      }
+    } else {
+      return false;
+    }
   }
 
   // Diable double copy optimization,