diff --git a/projects/clr/rocclr/runtime/device/pal/palconstbuf.cpp b/projects/clr/rocclr/runtime/device/pal/palconstbuf.cpp
index fe4b265fe7..b5a266fe43 100644
--- a/projects/clr/rocclr/runtime/device/pal/palconstbuf.cpp
+++ b/projects/clr/rocclr/runtime/device/pal/palconstbuf.cpp
@@ -53,7 +53,8 @@ bool ManagedBuffer::create(Resource::MemoryType type) {
 
 // ================================================================================================
 address ManagedBuffer::reserve(uint32_t size, uint64_t* gpu_address) {
-  static constexpr uint32_t MemAlignment = 256;
+  // Align to the maximum data size available in OpenCL
+  static constexpr uint32_t MemAlignment = sizeof(cl_double16);
 
   // Align reserve size on the vector's boundary
   uint32_t count = amd::alignUp(size, MemAlignment);
@@ -103,8 +104,7 @@ bool ConstantBuffer::Create() {
 }
 
 // ================================================================================================
-uint64_t ConstantBuffer::UploadDataToHw(uint32_t size) const
-{
+uint64_t ConstantBuffer::UploadDataToHw(uint32_t size) const {
   uint64_t  vm_address;
   address   cpu_address = mbuf_.reserve(size, &vm_address);
   // Update memory with new CB data
@@ -112,4 +112,13 @@ uint64_t ConstantBuffer::UploadDataToHw(uint32_t size) const
   return vm_address;
 }
 
+// ================================================================================================
+uint64_t ConstantBuffer::UploadDataToHw(const void* sysmem, uint32_t size) const {
+  uint64_t  vm_address;
+  address   cpu_address = mbuf_.reserve(size, &vm_address);
+  // Update memory with new CB data
+  memcpy(cpu_address, sysmem, size);
+  return vm_address;
+}
+
 }  // namespace pal
diff --git a/projects/clr/rocclr/runtime/device/pal/palconstbuf.hpp b/projects/clr/rocclr/runtime/device/pal/palconstbuf.hpp
index d2595036a1..5b6cb2af1b 100644
--- a/projects/clr/rocclr/runtime/device/pal/palconstbuf.hpp
+++ b/projects/clr/rocclr/runtime/device/pal/palconstbuf.hpp
@@ -79,6 +79,14 @@ public:
   uint64_t UploadDataToHw(uint32_t size  //!< real data size for upload
                           ) const;
 
+  /*! \brief Uploads current constant buffer data from sysMemCopy_ to HW
+  *
+  *  \return GPU address for the uploaded data
+  */
+  uint64_t UploadDataToHw(const void* sysmem,  //!< Pointer to the data for upload
+                          uint32_t    size     //!< Real data size for upload
+                          ) const;
+
   //! Returns a pointer to the system memory copy for CB
   address SysMemCopy(uint32_t size = 0) const { return sys_mem_copy_; }
 
diff --git a/projects/clr/rocclr/runtime/device/pal/palkernel.cpp b/projects/clr/rocclr/runtime/device/pal/palkernel.cpp
index d8a290ffba..18d03debf5 100644
--- a/projects/clr/rocclr/runtime/device/pal/palkernel.cpp
+++ b/projects/clr/rocclr/runtime/device/pal/palkernel.cpp
@@ -921,16 +921,14 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
   static const bool WaitOnBusyEngine = true;
   uint64_t ldsAddress = ldsSize();
   address aqlArgBuf = gpu.cb(0)->SysMemCopy();
-  address aqlStruct = gpu.cb(1)->SysMemCopy();
   bool srdResource = false;
 
   if (dynamicParallelism()) {
     // Provide the host parent AQL wrap object to the kernel
-    AmdAqlWrap* wrap = reinterpret_cast<AmdAqlWrap*>(aqlStruct);
-    memset(wrap, 0, sizeof(AmdAqlWrap));
-    wrap->state = AQL_WRAP_BUSY;
+    AmdAqlWrap wrap = {};
+    wrap.state = AQL_WRAP_BUSY;
     const ConstantBuffer* cb = gpu.cb(1);
-    *vmParentWrap = cb->UploadDataToHw(sizeof(AmdAqlWrap));
+    *vmParentWrap = cb->UploadDataToHw(&wrap, sizeof(AmdAqlWrap));
     gpu.addVmMemory(cb->ActiveMemory());
   }
 
@@ -1063,11 +1061,10 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
         break;
       }
       case HSAIL_ARGTYPE_REFERENCE: {
-        // Copy the current structure into CB1
-        memcpy(aqlStruct, paramaddr, arg->size_);
         const ConstantBuffer* cb = gpu.cb(1);
+        // Copy the current structure into CB1
+        size_t gpuPtr = static_cast<size_t>(cb->UploadDataToHw(paramaddr, arg->size_));
         // Then use a pointer in aqlArgBuffer to CB1
-        size_t gpuPtr = static_cast<size_t>(cb->UploadDataToHw(arg->size_));
         WriteAqlArg(&aqlArgBuf, &gpuPtr, sizeof(size_t), sizeof(size_t));
         gpu.addVmMemory(cb->ActiveMemory());
         break;
@@ -1105,11 +1102,10 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
         //! Copy SRD to CB1, so blit manager will be able to release
         //! this view without a wait for SRD resource.
         if (image->memoryType() == Resource::ImageView) {
-          // Copy the current structre into CB1
-          memcpy(aqlStruct, image->hwState(), HsaImageObjectSize);
+          // Copy the current image SRD into CB1
           const ConstantBuffer* cb = gpu.cb(1);
+          uint64_t srd = cb->UploadDataToHw(image->hwState(), HsaImageObjectSize);
           // Then use a pointer in aqlArgBuffer to CB1
-          uint64_t srd = cb->UploadDataToHw(HsaImageObjectSize);
           WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd), sizeof(srd));
           gpu.addVmMemory(cb->ActiveMemory());
         } else {
diff --git a/projects/clr/rocclr/runtime/device/pal/palvirtual.cpp b/projects/clr/rocclr/runtime/device/pal/palvirtual.cpp
index 4d07346f49..e24ddc5373 100644
--- a/projects/clr/rocclr/runtime/device/pal/palvirtual.cpp
+++ b/projects/clr/rocclr/runtime/device/pal/palvirtual.cpp
@@ -360,42 +360,45 @@ void VirtualGPU::Queue::DumpMemoryReferences() const {
   if (dump.is_open()) {
     dump << start << " Queue: ";
     switch (iQueue_->Type()) {
-      case Pal::QueueTypeCompute:
-        dump << "Compute";
-        break;
-      case Pal::QueueTypeDma:
-        dump << "SDMA";
-        break;
-      default:
-        dump << "unknown";
-        break;
+    case Pal::QueueTypeCompute:
+      dump << "Compute";
+      break;
+    case Pal::QueueTypeDma:
+      dump << "SDMA";
+      break;
+    default:
+      dump << "unknown";
+      break;
     }
     dump << "\n"
-         << "Resident memory resources:\n";
+        << "Resident memory resources:\n";
     uint idx = 0;
     for (auto it : memReferences_) {
       dump << " " << idx << "\t[";
       dump.setf(std::ios::hex, std::ios::basefield);
       dump.setf(std::ios::showbase);
       dump << (it.first)->iMem()->Desc().gpuVirtAddr << ", "
-           << (it.first)->iMem()->Desc().gpuVirtAddr + (it.first)->iMem()->Desc().size;
+          << (it.first)->iMem()->Desc().gpuVirtAddr + (it.first)->iMem()->Desc().size;
       dump.setf(std::ios::dec);
-      dump << "] CbId:" << it.second << "\n";
+      dump << "] CbId:" << it.second <<
+          ", Heap: " << (it.first)->iMem()->Desc().preferredHeap << "\n";
       idx++;
     }
-  }
-  if (last_kernel_ != nullptr) {
-    const amd::KernelSignature& signature = last_kernel_->signature();
-    dump << last_kernel_->name() << std::endl;
-    for (size_t i = 0; i < signature.numParameters(); ++i) {
-      const amd::KernelParameterDescriptor& desc = signature.at(i);
-      // Find if the current argument is a memory object
-      if ((desc.type_ == T_POINTER) && (desc.addressQualifier_ != CL_KERNEL_ARG_ADDRESS_LOCAL)) {
-        dump << " " << desc.name_ << ": " << std::endl;
+
+    if (last_kernel_ != nullptr) {
+      const amd::KernelSignature& signature = last_kernel_->signature();
+      dump << last_kernel_->name() << std::endl;
+      for (size_t i = 0; i < signature.numParameters(); ++i) {
+        const amd::KernelParameterDescriptor& desc = signature.at(i);
+        // Find if the current argument is a memory object
+        if ((desc.type_ == T_POINTER) &&
+            (desc.addressQualifier_ != CL_KERNEL_ARG_ADDRESS_LOCAL)) {
+          dump << " " << desc.name_ << ": " << std::endl;
+        }
       }
     }
+    dump.close();
   }
-  dump.close();
 }
 
 bool VirtualGPU::MemoryDependency::create(size_t numMemObj) {