diff --git a/projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp b/projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp
index edab454abd..b332749399 100644
--- a/projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp
+++ b/projects/clr/rocclr/runtime/device/gpu/gpukernel.cpp
@@ -3902,6 +3902,9 @@ HSAILKernel::loadArguments(
                 }
                 if (nativeMem) {
                     gpuMem = *reinterpret_cast<Memory* const*>(paramaddr);
+                    if (NULL != gpuMem) {
+                        mem = gpuMem->owner();
+                    }
                 }
                 else {
                         mem = *reinterpret_cast<amd::Memory* const*>(paramaddr);
diff --git a/projects/clr/rocclr/runtime/device/gpu/gpumemory.cpp b/projects/clr/rocclr/runtime/device/gpu/gpumemory.cpp
index e9a4590329..7a659c15d2 100644
--- a/projects/clr/rocclr/runtime/device/gpu/gpumemory.cpp
+++ b/projects/clr/rocclr/runtime/device/gpu/gpumemory.cpp
@@ -922,8 +922,22 @@ Memory::allocMapTarget(
     address mapAddress = NULL;
     size_t  offset = origin[0];
 
-    incIndMapCount();
+    //For SVM implementation, we cannot use cached map. if svm space, use the svm host pointer
+    void *initHostPtr = owner()->getSvmPtr();
+    if (NULL != initHostPtr) {
+        owner()->commitSvmMemory();
+    }
 
+    if (owner()->numDevices() > 1) {
+        if ((NULL == initHostPtr) && (owner()->getHostMem() == NULL)) {
+            static const bool forceAllocHostMem = true;
+            if (!owner()->allocHostMemory(NULL, forceAllocHostMem)) {
+                return NULL;
+            }
+        }
+    }
+
+    incIndMapCount();
     // If host memory exists, use it
     if ((owner()->getHostMem() != NULL) && isDirectMap()) {
         mapAddress = reinterpret_cast<address>(owner()->getHostMem());
@@ -953,8 +967,6 @@ Memory::allocMapTarget(
             bool    failed = false;
             amd::Memory*   memory = NULL;
             // Search for a possible indirect resource
-            //For SVM implementation, we cannot use cached map. if svm space, use the svm host pointer
-            void *initHostPtr = owner()->getSvmPtr();
             cl_mem_flags flag = 0;
             bool canBeCached = true;
             if (NULL != initHostPtr) {
diff --git a/projects/clr/rocclr/runtime/platform/memory.cpp b/projects/clr/rocclr/runtime/platform/memory.cpp
index 2b4088997a..8e35fe9764 100644
--- a/projects/clr/rocclr/runtime/platform/memory.cpp
+++ b/projects/clr/rocclr/runtime/platform/memory.cpp
@@ -251,7 +251,7 @@ Memory::allocHostMemory(void* initFrom, bool allocHostMem, bool forceCopy)
         pipe->end_idx = asPipe()->getMaxNumPackets();
     }
 
-    if (flags_ & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) {
+    if ((flags_ & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) && (NULL == lastWriter_)) {
         // Signal write, so coherency mechanism will initialize
         // memory on all devices
         signalWrite(NULL);