diff --git a/projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp b/projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp
index 414046eb5b..363a96c6e6 100644
--- a/projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp
+++ b/projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp
@@ -87,7 +87,7 @@ VirtualGPU::MemoryDependency::validate(
 
     if (flushL1Cache) {
         // Flush cache
-        gpu.flushL1Cache();
+        gpu.flushCUCaches();
 
         // Clear memory dependency state
         const static bool All = true;
@@ -1724,8 +1724,6 @@ VirtualGPU::submitKernelInternalHSA(
     if (hsaKernel.dynamicParallelism()) {
         // Make sure exculsive access to the device queue
         amd::ScopedLock(defQueue->lock());
-        //! \todo Remove flush. We start parent earlier.
-        flushDMA(MainEngine);
 
         if (GPU_PRINT_CHILD_KERNEL != 0) {
             waitForEvent(&gpuEvent);
@@ -1819,6 +1817,8 @@ VirtualGPU::submitKernelInternalHSA(
             *gpuDefQueue->virtualQueue_,
             *gpuDefQueue->schedParams_, gpuDefQueue->schedParamIdx_,
             gpuDefQueue->vqHeader_->aql_slot_num);
+        const static bool FlushL2 = true;
+        gpuDefQueue->flushCUCaches(FlushL2);
 
         // Get the address of PM4 template and add write it to params
         //! @note DMA flush must not occur between patch and the scheduler
@@ -3290,7 +3290,7 @@ VirtualGPU::processMemObjectsHSA(
                     *reinterpret_cast<void* const*>(params + desc.offset_));
                 if (!svmMem) {
                     //!\todo Do we have to sync cache coherency or wait for SDMA?
-                    flushL1Cache();
+                    flushCUCaches();
                     break;
                 }
             }
diff --git a/projects/clr/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.cpp b/projects/clr/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.cpp
index 540aee8329..46c078df59 100644
--- a/projects/clr/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.cpp
+++ b/projects/clr/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.cpp
@@ -464,9 +464,9 @@ CALGSLContext::flushIOCaches() const
 }
 
 void
-CALGSLContext::flushL1Cache() const
+CALGSLContext::flushCUCaches(bool flushL2) const
 {
-    m_cs->FlushL1Cache();
+    m_cs->FlushCUCaches(flushL2);
 }
 
 gslProgramObject
diff --git a/projects/clr/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.h b/projects/clr/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.h
index d7fd3c565b..849101423e 100644
--- a/projects/clr/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.h
+++ b/projects/clr/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.h
@@ -51,7 +51,7 @@ public:
     bool             isDone(GpuEvent* event);
     void             waitForEvent(GpuEvent* event);
     void             flushIOCaches() const;
-    void             flushL1Cache() const;
+    void             flushCUCaches(bool flushL2 = false) const;
     void             eventBegin(EngineType engId)
     {
         m_eventQueue[engId].begin();