Optimize profiling calls.

- do not call KMD query for submitted timestamp, we only need CPU
timestamp here.

Change-Id: Id60c1e367d9430d893fb3a253ffc058f45fc9609
Signed-off-by: Michal Mrozek <[email protected]>

Author: MichalMrozek

opencl/source/command_queue/enqueue_common.h

@@ -698,9 +698,9 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
         DEBUG_BREAK_IF(device->getDeviceInfo().preemptionSupported != false);
     }
 
-    TimeStampData submitTimeStamp;
+    TimeStampData submitTimeStamp = {};
     if (isProfilingEnabled() && eventBuilder.getEvent()) {
-        this->getDevice().getOSTime()->getCpuGpuTime(&submitTimeStamp);
+        this->getDevice().getOSTime()->getCpuTime(&submitTimeStamp.CPUTimeinNS);
         eventBuilder.getEvent()->setSubmitTimeStamp(&submitTimeStamp);
         getGpgpuCommandStreamReceiver().makeResident(*eventBuilder.getEvent()->getHwTimeStampNode()->getBaseGraphicsAllocation());
         if (isPerfCountersEnabled()) {

opencl/test/unit_test/mocks/mock_ostime.h

@@ -9,16 +9,15 @@
 #include "shared/source/os_interface/os_time.h"
 
 namespace NEO {
+static int PerfTicks = 0;
 class MockOSTime : public OSTime {
   public:
     bool getCpuGpuTime(TimeStampData *pGpuCpuTime) override {
-        static int PerfTicks = 0;
         pGpuCpuTime->GPUTimeStamp = ++PerfTicks;
         pGpuCpuTime->CPUTimeinNS = PerfTicks;
         return true;
     }
     bool getCpuTime(uint64_t *timeStamp) override {
-        static int PerfTicks = 0;
         *timeStamp = ++PerfTicks;
         return true;
     };

opencl/test/unit_test/profiling/profiling_tests.cpp

@@ -195,6 +195,35 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ProfilingTests, GivenCommandQueueWithProfolingWhenWa
     clReleaseEvent(event);
 }
 
+HWCMDTEST_F(IGFX_GEN8_CORE, ProfilingTests, GivenCommandQueueWithProfilingWhenNonBlockedEnqueueIsExecutedThenSubmittedTimestampDoesntHaveGPUTime) {
+    MockKernel kernel(program.get(), kernelInfo, *pClDevice);
+    ASSERT_EQ(CL_SUCCESS, kernel.initialize());
+
+    size_t globalOffsets[3] = {0, 0, 0};
+    size_t workItems[3] = {1, 1, 1};
+    uint32_t dimensions = 1;
+    cl_event event;
+    cl_kernel clKernel = &kernel;
+
+    static_cast<CommandQueueHw<FamilyType> *>(pCmdQ)->enqueueKernel(
+        clKernel,
+        dimensions,
+        globalOffsets,
+        workItems,
+        nullptr,
+        0,
+        nullptr,
+        &event);
+
+    auto mockEvent = static_cast<MockEvent<Event> *>(event);
+    EXPECT_NE(0u, mockEvent->queueTimeStamp.GPUTimeStamp);
+    EXPECT_NE(0u, mockEvent->queueTimeStamp.CPUTimeinNS);
+    EXPECT_LT(mockEvent->queueTimeStamp.CPUTimeinNS, mockEvent->submitTimeStamp.CPUTimeinNS);
+    EXPECT_EQ(0u, mockEvent->submitTimeStamp.GPUTimeStamp);
+
+    clReleaseEvent(event);
+}
+
 /*
 #   One additional MI_STORE_REGISTER_MEM is expected before and after GPGPU_WALKER.
 */