P4 to Git Change 1396011 by todli@todli-win-opencl-kv1 on 2017/04/09 02:28:44

SWDEV-96241 - Unlike in ORCA, in PAL the actual memory allocated for BusAddressable memory is sometimes larger than "Pow2Align(size, pageSize) + pageSize".
	So virtual address of marker should be calculated in another way for WaitMarker command.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#47 edit

rocclr/runtime/device/pal/palvirtual.cpp

@@ -2794,9 +2794,11 @@ VirtualGPU::submitSignal(amd::SignalCommand & vcmd)
     uint32_t size = vcmd.memory().getSize();
 
     addVmMemory(pGpuMemory);
+    
+    uint32_t offset = pGpuMemory->iMem()->Desc().markerBusAddr - pGpuMemory->iMem()->Desc().surfaceBusAddr;
 
     if (vcmd.type() == CL_COMMAND_WAIT_SIGNAL_AMD) {
-        iCmd()->CmdWaitMemoryValue(*(pGpuMemory->iMem()), size, value, 0xFFFFFFFF, Pal::CompareFunc::GreaterEqual);
+        iCmd()->CmdWaitMemoryValue(*(pGpuMemory->iMem()), offset, value, 0xFFFFFFFF, Pal::CompareFunc::GreaterEqual);
     }
     else if (vcmd.type() == CL_COMMAND_WRITE_SIGNAL_AMD) {
         iCmd()->CmdUpdateMemory(*(pGpuMemory->iMem()), size, 4, &value);