P4 to Git Change 1363243 by jatang@jatang-opencl-hsa-stg1 on 2017/01/18 09:49:05

SWDEV-95919 - Fix num-of-waves PerfCounter is 0 on Baffin.

	Counters are clockgated on Fiji and above. Need to program the RLC_PERFMON_CLK_CNTL register thru IDeviceSetClockMode(), so that meaningful data is recorded. We leave kmd/pplib to keep track of the OCL process to restore the RLC_PERFMON_CLK_CNTL register when the OCL process is killed.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#41 edit


[ROCm/clr commit: e826e296d4]

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

@@ -2507,7 +2507,15 @@ VirtualGPU::submitPerfCounter(amd::PerfCounterCommand& vcmd)
             palPerf = counter->iPerf();
             // Find the state and sends the command to PAL
             if (vcmd.getState() == amd::PerfCounterCommand::Begin) {
+                Pal::SetClockModeInput  input;
+                Pal::SetClockModeOutput output = {};
+                input.clockMode = Pal::DeviceClockMode::Profiling;
+                dev().iDev()->SetClockMode(input, &output);
+                GpuEvent event;
+                eventBegin(MainEngine);
                 iCmd()->CmdBeginPerfExperiment(palPerf);
+                eventEnd(MainEngine, event);
+                setGpuEvent(event);
             }
             else if (vcmd.getState() == amd::PerfCounterCommand::End) {
                 GpuEvent event;