SWDEV-79445 - HIP improvement

- Make sure only one GPU barrier is issued per dispatch
when memory tracking is disabled

Change-Id: I974569ab42a8835304a2930eef87b561a3750327


[ROCm/clr commit: 481cecec78]

projects/clr/rocclr/device/pal/palvirtual.cpp

@@ -563,8 +563,7 @@ void VirtualGPU::MemoryDependency::validate(VirtualGPU& gpu, const Memory* memor
   bool flushL1Cache = false;
 
   if (maxMemObjectsInQueue_ == 0) {
-    // Flush cache
-    gpu.addBarrier(RgpSqqtBarrierReason::MemDependency);
+    // Return earlier if tracking is disabled
     return;
   }
 
@@ -3549,6 +3548,9 @@ bool VirtualGPU::processMemObjectsHSA(const amd::Kernel& kernel, const_address p
     addVmMemory(scratch->memObj_);
   }
 
+  // Synchronize dispatches unconditionally in case memory tracking is disabled
+  memoryDependency().sync(*this);
+
   return true;
 }
 

projects/clr/rocclr/device/pal/palvirtual.hpp

@@ -251,6 +251,13 @@ class VirtualGPU : public device::VirtualDevice {
     //! Validates memory object on dependency
     void validate(VirtualGPU& gpu, const Memory* memory, bool readOnly);
 
+    //! Invalidates GPU caches if memory dependency tracking is disabled
+    void sync(VirtualGPU& gpu) const {
+      if (maxMemObjectsInQueue_ == 0) {
+        gpu.addBarrier(RgpSqqtBarrierReason::MemDependency);
+      }
+    }
+
     //! Clear memory dependency
     void clear(bool all = true);