SWDEV-232807

[ROCm][TCT][HIP] cooperative stream test case is failing.

Make sure lockXfer() in the blit manager returns a valid value.
Port the latest PAL backend logic into the ROCr backend.
This change doesn't fix the issue, reported in the ticket.

Change-Id: I54101a824f49a2dcfbbf5414cb5b3af41745306d

rocclr/device/rocm/rocblit.cpp

@@ -707,7 +707,7 @@ KernelBlitManager::KernelBlitManager(VirtualGPU& gpu, Setup setup)
       program_(nullptr),
       constantBuffer_(nullptr),
       xferBufferSize_(0),
-      lockXferOps_(nullptr) {
+      lockXferOps_("Transfer Ops Lock", true) {
   for (uint i = 0; i < BlitTotal; ++i) {
     kernels_[i] = nullptr;
   }
@@ -733,8 +733,6 @@ KernelBlitManager::~KernelBlitManager() {
   if (nullptr != constantBuffer_) {
     constantBuffer_->release();
   }
-
-  delete lockXferOps_;
 }
 
 bool KernelBlitManager::create(amd::Device& device) {
@@ -797,11 +795,6 @@ bool KernelBlitManager::createProgram(Device& device) {
     return false;
   }
 
-  lockXferOps_ = new amd::Monitor("Transfer Ops Lock", true);
-  if (nullptr == lockXferOps_) {
-    return false;
-  }
-
   return result;
 }
 

rocclr/device/rocm/rocblit.hpp

@@ -396,6 +396,8 @@ class KernelBlitManager : public DmaBlitManager {
   bool RunGwsInit(uint32_t value             //!< Initial value for GWS resource
                   ) const;
 
+  virtual amd::Monitor* lockXfer() const { return &lockXferOps_; }
+
  private:
   static const size_t MaxXferBuffers = 2;
   static const uint TransferSplitSize = 1;
@@ -448,7 +450,7 @@ class KernelBlitManager : public DmaBlitManager {
   amd::Kernel* kernels_[BlitTotal];           //!< GPU kernels for blit
   amd::Memory* constantBuffer_;               //!< An internal CB for blits
   size_t xferBufferSize_;                     //!< Transfer buffer size
-  amd::Monitor* lockXferOps_;                 //!< Lock transfer operation
+  mutable amd::Monitor  lockXferOps_;         //!< Lock transfer operation
 };
 
 static const char* BlitName[KernelBlitManager::BlitTotal] = {

rocclr/device/rocm/rocvirtual.cpp

@@ -2300,15 +2300,15 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
  */
 void VirtualGPU::submitKernel(amd::NDRangeKernelCommand& vcmd) {
   if (vcmd.cooperativeGroups() || vcmd.cooperativeMultiDeviceGroups()) {
+    // Wait for the execution on the current queue, since the coop groups will use the device queue
+    releaseGpuMemoryFence();
+
     // Get device queue for exclusive GPU access
     VirtualGPU* queue = dev().xferQueue();
 
     // Lock the queue, using the blit manager lock
     amd::ScopedLock lock(queue->blitMgr().lockXfer());
 
-    // Wait for the execution on the current queue, since the coop groups will use the device queue
-    releaseGpuMemoryFence();
-
     queue->profilingBegin(vcmd);
 
     if (vcmd.cooperativeGroups()) {