[OpenMP][Offload][AMDGPU] Add a new envar to enable copy regardless the value of HSA_XNACK

This patch is for SWDEV-494967. When this new envar OMPX_DGPU_MAPS=1, it will always perform copy no matter the value of HSA_XNACK when running on APUs. This envar will increase the portability of the code developed on MI300A for running on MI250X, MI300X, etc.

Change-Id: I70edb8682b96bb845a6c8f0269eec910e671ab94

Author: Kewen12

offload/plugins-nextgen/amdgpu/src/rtl.cpp

@@ -2685,6 +2685,7 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
         OMPX_APUPrefaultMemcopy("LIBOMPTARGET_APU_PREFAULT_MEMCOPY", "true"),
         OMPX_APUPrefaultMemcopySize("LIBOMPTARGET_APU_PREFAULT_MEMCOPY_SIZE",
                                     1 * 1024 * 1024), // 1MB
+        OMPX_DGPUMaps("OMPX_DGPU_MAPS", false),
         AMDGPUStreamManager(*this, Agent), AMDGPUEventManager(*this),
         AMDGPUSignalManager(*this), Agent(Agent), HostDevice(HostDevice) {}
 
@@ -3823,7 +3824,8 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
   /// XNACK can be enabled with a kernel boot parameter or with
   /// the HSA_XNACK environment variable.
   bool useAutoZeroCopyImpl() override {
-    return ((IsAPU || OMPX_ApuMaps) && IsXnackEnabled);
+    return !(OMPX_DGPUMaps && IsAPU) &&
+           ((IsAPU || OMPX_ApuMaps) && IsXnackEnabled);
   }
 
   /// Performs sanity checks on the selected zero-copy configuration and prints
@@ -4233,6 +4235,10 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
   /// is applied.
   UInt32Envar OMPX_APUPrefaultMemcopySize;
 
+  /// Value of OMPX_DGPU_MAPS. When enabled, it will always perform
+  /// copy on APUs regardless of the setting of HSA_XNACK.
+  BoolEnvar OMPX_DGPUMaps;
+
   /// Stream manager for AMDGPU streams.
   AMDGPUStreamManagerTy AMDGPUStreamManager;