SWDEV-300515 - Apply CU granularity on the CU count

Report granularity for possible app query

Change-Id: I98857c6f4cc7ae590927ea35ce57d181abe7860b


[ROCm/clr commit: f613831471]

projects/clr/rocclr/device/device.hpp

@@ -521,6 +521,8 @@ struct Info : public amd::EmbeddedObject {
   uint32_t numRTQueues_;
   //! Number of available real time compute units
   uint32_t numRTCUs_;
+  //! The granularity at which compute units can be dedicated to a queue
+  uint32_t granularityRTCUs_;
   //! Thread trace enable
   uint32_t threadTraceEnable_;
 

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

@@ -596,7 +596,10 @@ void NullDevice::fillDeviceInfo(const Pal::DeviceProperties& palProp,
     info_.numAsyncQueues_ = numComputeRings;
 
     info_.numRTQueues_ = numExclusiveComputeRings;
-    info_.numRTCUs_ = palProp.engineProperties[Pal::EngineTypeCompute].maxNumDedicatedCu;
+
+    const auto& engineProp = palProp.engineProperties[Pal::EngineTypeCompute];
+    info_.numRTCUs_ = engineProp.maxNumDedicatedCu;
+    info_.granularityRTCUs_ = engineProp.dedicatedCuGranularity;
 
     info_.threadTraceEnable_ = settings().threadTraceEnable_;
 

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

@@ -87,7 +87,8 @@ VirtualGPU::Queue* VirtualGPU::Queue::Create(const VirtualGPU& gpu, Pal::QueueTy
     cmdCreateInfo.engineType = qCreateInfo.engineType = Pal::EngineTypeCompute;
     qCreateInfo.priority = Pal::QueuePriority::Medium;
   } else if (amd::CommandQueue::RealTimeDisabled != rtCU) {
-    qCreateInfo.numReservedCu = rtCU;
+    qCreateInfo.numReservedCu = amd::alignDown(rtCU,
+      gpu.dev().properties().engineProperties[Pal::EngineTypeCompute].dedicatedCuGranularity);
     if ((priority == amd::CommandQueue::Priority::Medium)  &&
          // If Windows HWS is enabled, then the both real time queues are allocated
          // on the same engine