Add PowerSaving debug variable.

- It allows to easily turn ON the power saving mode for the driver.
- In that mode, whenever GPU is not ready it will put current thread to sleep,
thread is resumed when GPU completes.
- PowerSaving mode reduces CPU utilization on blocking calls, while increasing
completion latencies.

Change-Id: I3de83713687952ac31a1ec8c397f48aa4212781d

Author: MichalMrozek

runtime/helpers/kmd_notify_properties.cpp

@@ -7,6 +7,7 @@
 
 #include <cstdint>
 #include "runtime/helpers/kmd_notify_properties.h"
+#include "runtime/os_interface/debug_settings_manager.h"
 
 using namespace OCLRT;
 
@@ -15,6 +16,15 @@ bool KmdNotifyHelper::obtainTimeoutParams(int64_t &timeoutValueOutput,
                                           uint32_t currentHwTag,
                                           uint32_t taskCountToWait,
                                           FlushStamp flushStampToWait) {
+    if (flushStampToWait == 0) {
+        return false;
+    }
+
+    if (DebugManager.flags.PowerSavingMode.get()) {
+        timeoutValueOutput = 1;
+        return true;
+    }
+
     int64_t multiplier = (currentHwTag < taskCountToWait) ? static_cast<int64_t>(taskCountToWait - currentHwTag) : 1;
     if (!properties->enableKmdNotify && multiplier > KmdNotifyConstants::minimumTaskCountDiffToCheckAcLine) {
         updateAcLineStatus();
@@ -32,7 +42,7 @@ bool KmdNotifyHelper::obtainTimeoutParams(int64_t &timeoutValueOutput,
         timeoutValueOutput = getBaseTimeout(multiplier);
     }
 
-    return flushStampToWait != 0 && (properties->enableKmdNotify || !acLineConnected);
+    return (properties->enableKmdNotify || !acLineConnected);
 }
 
 bool KmdNotifyHelper::applyQuickKmdSleepForSporadicWait() const {

runtime/os_interface/debug_variables_base.inl

@@ -100,6 +100,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, OverrideEnableQuickKmdSleep, -1, "-1: dont overr
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideQuickKmdSleepDelayMicroseconds, -1, "-1: dont override, 0: infinite timeout, >0: timeout in microseconds")
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideEnableQuickKmdSleepForSporadicWaits, -1, "-1: dont override, 0: disable, 1: enable. It works only when QuickKmdSleep is enabled.")
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideDelayQuickKmdSleepForSporadicWaitsMicroseconds, -1, "-1: dont override, >0: timeout in microseconds")
+DECLARE_DEBUG_VARIABLE(int32_t, PowerSavingMode, 0, "0: default 1: enable. Whenever driver waits on GPU and its not ready, put waiting thread to sleep and wait for notification.")
 DECLARE_DEBUG_VARIABLE(int32_t, CsrDispatchMode, 0, "Chooses DispatchMode for Csr")
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideDefaultFP64Settings, -1, "-1: dont override, 0: disable, 1: enable.")
 

unit_tests/event/event_tests.cpp

@@ -1389,6 +1389,7 @@ HWTEST_F(EventTest, givenQuickKmdSleepRequestWhenWaitIsCalledThenPassRequestToWa
     pDevice->resetCommandStreamReceiver(csr);
 
     Event event(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, 0, 0);
+    event.updateCompletionStamp(1u, 1u, 1u);
 
     EXPECT_CALL(*csr, waitForCompletionWithTimeout(::testing::_,
                                                    localHwInfo.capabilityTable.kmdNotifyProperties.delayQuickKmdSleepMicroseconds, ::testing::_))
@@ -1414,6 +1415,7 @@ HWTEST_F(EventTest, givenNonQuickKmdSleepRequestWhenWaitIsCalledThenPassRequestT
     pDevice->resetCommandStreamReceiver(csr);
 
     Event event(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, 0, 0);
+    event.updateCompletionStamp(1u, 1u, 1u);
 
     EXPECT_CALL(*csr, waitForCompletionWithTimeout(::testing::_,
                                                    localHwInfo.capabilityTable.kmdNotifyProperties.delayKmdNotifyMicroseconds, ::testing::_))

unit_tests/helpers/kmd_notify_tests.cpp

@@ -8,6 +8,7 @@
 #include "runtime/command_queue/command_queue.h"
 #include "runtime/os_interface/os_context.h"
 
+#include "unit_tests/helpers/debug_manager_state_restore.h"
 #include "unit_tests/mocks/mock_device.h"
 #include "unit_tests/mocks/mock_context.h"
 #include "test.h"
@@ -366,6 +367,34 @@ TEST_F(KmdNotifyTests, givenDisabledKmdNotifyMechanismAndFlushStampIsZeroWhenAcL
     EXPECT_FALSE(timeoutEnabled);
 }
 
+TEST_F(KmdNotifyTests, givenDisabledKmdNotifyMechanismWhenPowerSavingModeIsSetThenKmdNotifyMechanismIsUsedAndReturnsShortestWaitingTimePossible) {
+    DebugManagerStateRestore stateRestore;
+    DebugManager.flags.PowerSavingMode.set(1u);
+    localHwInfo.capabilityTable.kmdNotifyProperties.enableKmdNotify = false;
+    MockKmdNotifyHelper helper(&(localHwInfo.capabilityTable.kmdNotifyProperties));
+    helper.acLineConnected = false;
+
+    int64_t timeout = 0;
+    FlushStamp flushStampToWait = 1;
+    bool timeoutEnabled = helper.obtainTimeoutParams(timeout, false, 1, 2, flushStampToWait);
+    EXPECT_TRUE(timeoutEnabled);
+    EXPECT_EQ(1, timeout);
+}
+
+TEST_F(KmdNotifyTests, givenEnabledKmdNotifyMechanismWhenPowerSavingModeIsSetAndNoFlushStampProvidedWhenParametersAreObtainedThenFalseIsReturned) {
+    DebugManagerStateRestore stateRestore;
+    DebugManager.flags.PowerSavingMode.set(1u);
+    localHwInfo.capabilityTable.kmdNotifyProperties.enableKmdNotify = true;
+    MockKmdNotifyHelper helper(&(localHwInfo.capabilityTable.kmdNotifyProperties));
+    helper.acLineConnected = false;
+
+    int64_t timeout = 0;
+    FlushStamp flushStampToWait = 0;
+    bool timeoutEnabled = helper.obtainTimeoutParams(timeout, false, 1, 2, flushStampToWait);
+    EXPECT_FALSE(timeoutEnabled);
+    EXPECT_EQ(0, timeout);
+}
+
 #if defined(__clang__)
 #pragma clang diagnostic pop
 #endif

unit_tests/test_files/igdrcl.config

@@ -95,4 +95,5 @@ UseAubStream = false
 AubDumpOverrideMmioRegister = 0
 AubDumpOverrideMmioRegisterValue = 0
 AubDumpAddMmioRegister = 0
-AubDumpAddMmioRegisterValue = 0 
+AubDumpAddMmioRegisterValue = 0 
+PowerSavingMode = 0