SWDEV-505833 - Provide functionality to avoid L2 flush for CPX mode for dispatch packets

- Added DEBUG_CLR_SKIP_RELEASE_SCOPE flag to force release scope to
   SCOPE_NONE in AQL packet header

Change-Id: Ife02cddb9d5cd4749103ce585d3d5fe9024c6868

rocclr/device/rocm/rocvirtual.cpp

@@ -906,6 +906,12 @@ bool VirtualGPU::dispatchGenericAqlPacket(
     addSystemScope_ = false;
   }
 
+  if (DEBUG_CLR_SKIP_RELEASE_SCOPE) {
+    header &= ~(HSA_FENCE_SCOPE_AGENT << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE |
+                HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
+    header |= (HSA_FENCE_SCOPE_NONE << HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE);
+  }
+
   auto expected_fence_state = extractAqlBits(header, HSA_PACKET_HEADER_SCRELEASE_FENCE_SCOPE,
                          HSA_PACKET_HEADER_WIDTH_SCRELEASE_FENCE_SCOPE);
 

rocclr/utils/flags.hpp

@@ -269,6 +269,8 @@ release(bool, DEBUG_CLR_USE_STDMUTEX_IN_AMD_MONITOR, false,                   \
         "Use std::mutex in amd::monitor")                                     \
 release(bool, DEBUG_CLR_KERNARG_HDP_FLUSH_WA, false,                          \
         "Toggle kernel arg copy workaround")                                  \
+release(bool, DEBUG_CLR_SKIP_RELEASE_SCOPE, false,                            \
+        "Forces release scope to SCOPE_NONE for aql packets")                 \
 release(uint, DEBUG_HIP_7_PREVIEW, 0,                                         \
         "Enables specific backward incompatible changes support before 7.0,"  \
         "using the mask. By default the changes are disabled and is set to 0")\