diff --git a/rocclr/runtime/device/gpu/gpukernel.cpp b/rocclr/runtime/device/gpu/gpukernel.cpp
index 46a11f5a01..780d458ba9 100644
--- a/rocclr/runtime/device/gpu/gpukernel.cpp
+++ b/rocclr/runtime/device/gpu/gpukernel.cpp
@@ -3850,6 +3850,7 @@ HSAILKernel::loadArguments(
                     mem = amd::SvmManager::FindSvmBuffer(*reinterpret_cast<void* const*>(paramaddr));
                     if (mem != NULL) {
                         gpuMem = dev().getGpuMemory(mem);
+                        gpuMem->wait(gpu, WaitOnBusyEngine);
                         memList.push_back(gpuMem);
                     }
                     else {
@@ -3861,10 +3862,10 @@ HSAILKernel::loadArguments(
                     gpuMem = *reinterpret_cast<Memory* const*>(paramaddr);
                 }
                 else {
-                    mem = *reinterpret_cast<amd::Memory* const*>(paramaddr);
-                    if (mem != NULL) {
-                        gpuMem = dev().getGpuMemory(mem);
-                    }
+                        mem = *reinterpret_cast<amd::Memory* const*>(paramaddr);
+                        if (mem != NULL) {
+                             gpuMem = dev().getGpuMemory(mem);
+                        }
                 }
                 if (gpuMem == NULL) {
                     WriteAqlArg(&aqlArgBuf, &gpuMem, sizeof(void*));
diff --git a/rocclr/runtime/device/gpu/gpuvirtual.cpp b/rocclr/runtime/device/gpu/gpuvirtual.cpp
index 9a4497be2c..dcb73b18a8 100644
--- a/rocclr/runtime/device/gpu/gpuvirtual.cpp
+++ b/rocclr/runtime/device/gpu/gpuvirtual.cpp
@@ -3280,13 +3280,18 @@ VirtualGPU::processMemObjectsHSA(
         const HSAILKernel::Argument*  arg = hsaKernel.argument(i);
         Memory* memory = NULL;
         bool    readOnly = false;
+        amd::Memory* svmMem = NULL;
 
         // Find if current argument is a buffer
         if ((desc.type_ == T_POINTER) && (arg->addrQual_ != HSAIL_ADDRESS_LOCAL)) {
             if (kernelParams.boundToSvmPointer(dev(), params, i)) {
-                //!\todo Do we have to sync cache coherency or wait for SDMA?
-                flushL1Cache();
-                break;
+                svmMem = amd::SvmManager::FindSvmBuffer(
+                    *reinterpret_cast<void* const*>(params + desc.offset_));
+                if (!svmMem) {
+                    //!\todo Do we have to sync cache coherency or wait for SDMA?
+                    flushL1Cache();
+                    break;
+                }
             }
 
             if (nativeMem) {
@@ -3294,8 +3299,6 @@ VirtualGPU::processMemObjectsHSA(
             }
             else if (*reinterpret_cast<amd::Memory* const*>
                     (params + desc.offset_) != NULL) {
-                amd::Memory* svmMem = amd::SvmManager::FindSvmBuffer(
-                    *reinterpret_cast<void* const*>(params + desc.offset_));
                 if (NULL == svmMem) {
                     memory = dev().getGpuMemory(*reinterpret_cast<amd::Memory* const*>
                             (params + desc.offset_));