SWDEV-436796 - Enable device memory for kernel arguments

Extra CPU read back will be performed before every submission to make sure
previous writes over PCIE reached GPU. HDP flush is done by CP.

Change-Id: I402d28ca26c8cee4a3920feb3599af8c285d0889

rocclr/device/pal/palsettings.cpp

@@ -143,7 +143,7 @@ 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;
+  useDeviceKernelArg_ = flagIsDefault(HIP_FORCE_DEV_KERNARG) ? true : HIP_FORCE_DEV_KERNARG;
 
   limit_blit_wg_ = 16;
 }