SWDEV-430911 - Force SDMA only if explicitly specified & pick appropriate engine.

Change-Id: Ib34fa6b2782f74b753899fa8fddff646dc60c8ce


[ROCm/clr commit: 3daf8b334a]

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

@@ -691,7 +691,7 @@ bool DmaBlitManager::hsaCopy(const Memory& srcMemory, const Memory& dstMemory,
   // Determine engine and assign a copy mask for the new versatile ROCr API
   // If engine preferred is SDMA, assign the SdmaWrite path
   if ((srcAgent.handle == dev().getCpuAgent().handle) &&
-      (dstAgent.handle != dev().getCpuAgent().handle) || forceSDMA) {
+      (dstAgent.handle != dev().getCpuAgent().handle)) {
     engine = HwQueueEngine::SdmaWrite;
     copyMask = kUseRegularCopyApi ? 0 : dev().fetchSDMAMask(this, false);
   } else if ((srcAgent.handle != dev().getCpuAgent().handle) &&
@@ -700,6 +700,11 @@ bool DmaBlitManager::hsaCopy(const Memory& srcMemory, const Memory& dstMemory,
     copyMask = kUseRegularCopyApi ? 0 : dev().fetchSDMAMask(this, true);
   }
 
+  if (engine == HwQueueEngine::Unknown && forceSDMA) {
+    engine = HwQueueEngine::SdmaRead;
+    copyMask = kUseRegularCopyApi ? 0 : dev().fetchSDMAMask(this, true);
+  }
+
   // Check if host wait has to be forced
   bool forceHostWait = forceHostWaitFunc(size[0]);