P4 to Git Change 1521567 by gandryey@gera-w8 on 2018/03/01 18:29:47

SWDEV-145716 - dGMA via OpenCL not working properly on AMD EPYC Ethanol platform with two WX9100/W9100s
	- Add a stall and cache flushes before a barrier write in GSL and PAL
	- Place temporary workarounds for CP prefetch of WRITE_DATA and the lack of SDMA marker in PAL

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#414 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#76 edit


[ROCm/clr commit: 30eacbe243]

projects/clr/rocclr/runtime/device/gpu/gpuvirtual.cpp

@@ -2380,6 +2380,8 @@ void VirtualGPU::submitSignal(amd::SignalCommand& vcmd) {
     uint64_t markerOffset = markerAddr - surfAddr;
     cs()->p2pMarkerOp(gpuMemory->gslResource(), vcmd.markerValue(), markerOffset, false);
   } else if (vcmd.type() == CL_COMMAND_WRITE_SIGNAL_AMD) {
+    static constexpr bool FlushL2 = true;
+    flushCUCaches(FlushL2);
     cs()->p2pMarkerOp(gpuMemory->gslResource(), vcmd.markerValue(), vcmd.markerOffset(), true);
   }
   eventEnd(MainEngine, gpuEvent);

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

@@ -2537,6 +2537,17 @@ void VirtualGPU::submitSignal(amd::SignalCommand& vcmd) {
     iCmd()->CmdWaitBusAddressableMemoryMarker(*(pGpuMemory->iMem()), value, 0xFFFFFFFF,
                                               Pal::CompareFunc::GreaterEqual);
   } else if (vcmd.type() == CL_COMMAND_WRITE_SIGNAL_AMD) {
+    // Make sure GPU finished operation and data reached memory before the marker write
+    static constexpr bool FlushL2 = true;
+    addBarrier(FlushL2);
+    // \todo: Implement the right changes in PAL
+    // Workarounds: for CP overfetch issues and the lack of SDMA sync
+    {
+      // Flush CB associated with the DGMA buffer
+      isDone(pGpuMemory->getGpuEvent(*this));
+      // Make sure SDMA is done on the DGMA buffer
+      pGpuMemory->wait(*this, true);
+    }
     iCmd()->CmdUpdateBusAddressableMemoryMarker(*(pGpuMemory->iMem()), value);
   }
   eventEnd(MainEngine, gpuEvent);