SWDEV-376453 - Changes to fix crash seen with HIP and recent compiler change with -O0

Change-Id: I37de83b7449461ce2378741e97ff8fb8da736523


[ROCm/clr commit: c585aefdb7]

projects/clr/rocclr/device/devkernel.cpp

@@ -1343,7 +1343,8 @@ void Kernel::InitParameters(const amd_comgr_metadata_node_t kernelMD) {
 
     // Allocate the hidden arguments, but abstraction layer will skip them
     if (desc.info_.hidden_) {
-      if (desc.info_.oclObject_ == amd::KernelParameterDescriptor::HiddenCompletionAction) {
+      if (desc.info_.oclObject_ == amd::KernelParameterDescriptor::HiddenCompletionAction &&
+          !amd::IS_HIP) {
         setDynamicParallelFlag(true);
       }
       if (codeObjectVer() == 2) {

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

@@ -348,12 +348,12 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(VirtualGPU& gpu, const
         }
         break;
       case amd::KernelParameterDescriptor::HiddenDefaultQueue:
-        if (vmDefQueue != 0) {
+        if (vmDefQueue != 0 && dynamicParallelism()) {
           WriteAqlArgAt(hidden_arguments, vmDefQueue, it.size_, it.offset_);
         }
         break;
       case amd::KernelParameterDescriptor::HiddenCompletionAction:
-        if (*vmParentWrap != 0) {
+        if (*vmParentWrap != 0 && dynamicParallelism()) {
           WriteAqlArgAt(hidden_arguments, *vmParentWrap, it.size_, it.offset_);
         }
         break;

projects/clr/rocclr/device/rocm/rocvirtual.cpp

@@ -2908,7 +2908,7 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
         case amd::KernelParameterDescriptor::HiddenDefaultQueue: {
           uint64_t vqVA = 0;
           amd::DeviceQueue* defQueue = kernel.program().context().defDeviceQueue(dev());
-          if (nullptr != defQueue) {
+          if (nullptr != defQueue && devKernel->dynamicParallelism()) {
             if (!createVirtualQueue(defQueue->size()) || !createSchedulerParam()) {
               return false;
             }
@@ -2919,7 +2919,7 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
         }
         case amd::KernelParameterDescriptor::HiddenCompletionAction: {
           uint64_t spVA = 0;
-          if (nullptr != schedulerParam_) {
+          if (nullptr != schedulerParam_ && devKernel->dynamicParallelism()) {
             Memory* schedulerMem = dev().getRocMemory(schedulerParam_);
             AmdAqlWrap* wrap = reinterpret_cast<AmdAqlWrap*>(
                                reinterpret_cast<uint64_t>(schedulerParam_->getHostMem()) + sizeof(SchedulerParam));