feature: add waitWithKmdNotifyFallback to immediate cmdlist

- hostSynchronize() with infinite timeout may fallback to KmdNotify when kmdNotify is enabled Related-To: NEO-2024 Signed-off-by: Mateusz Hoppe <mateusz.hoppe@intel.com>

Author: HoppeMateusz

level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl

@@ -1242,9 +1242,20 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::hostSynchronize(uint6
  if (inOrderWaitAllowed) {
  status = synchronizeInOrderExecution(timeout, (waitQueue == this->cmdQImmediateCopyOffload));
  } else {
- const int64_t timeoutInMicroSeconds = timeout / 1000;
+
  const auto indefinitelyPoll = timeout == std::numeric_limits<uint64_t>::max();
- const auto waitStatus = waitCsr->waitForCompletionWithTimeout(NEO::WaitParams{indefinitelyPoll, !indefinitelyPoll, false, timeoutInMicroSeconds}, waitTaskCount);
+ auto waitStatus = NEO::WaitStatus::notReady;
+
+ if (indefinitelyPoll) {
+ waitStatus = waitCsr->waitForTaskCountWithKmdNotifyFallback(waitTaskCount,
+ waitCsr->obtainCurrentFlushStamp(),
+ true,
+ NEO::QueueThrottle::MEDIUM);
+ } else {
+ const int64_t timeoutInMicroSeconds = timeout / 1000;
+ waitStatus = waitCsr->waitForCompletionWithTimeout(NEO::WaitParams{indefinitelyPoll, !indefinitelyPoll, false, timeoutInMicroSeconds}, waitTaskCount);
+ }
+
  if (waitStatus == NEO::WaitStatus::gpuHang) {
  status = ZE_RESULT_ERROR_DEVICE_LOST;
  } else if (waitStatus == NEO::WaitStatus::notReady) {

level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_1.cpp

@@ -2414,6 +2414,7 @@ TEST_F(CommandListCreateTests, GivenGpuHangWhenCreatingImmCmdListWithSyncModeAnd
 
  MockCommandStreamReceiver mockCommandStreamReceiver(*neoDevice->executionEnvironment, neoDevice->getRootDeviceIndex(), neoDevice->getDeviceBitfield());
  mockCommandStreamReceiver.waitForCompletionWithTimeoutReturnValue = WaitStatus::gpuHang;
+ mockCommandStreamReceiver.waitForCompletionWithKmdNotifyFallbackReturnValue = WaitStatus::gpuHang;
 
  auto queue = static_cast<WhiteBox<::L0::CommandQueue> *>(whiteBoxCmdList->cmdQImmediate);
 
@@ -2486,6 +2487,7 @@ HWTEST_F(CommandListCreateTests, GivenGpuHangWhenCreatingImmediateCommandListAnd
 
  MockCommandStreamReceiver mockCommandStreamReceiver(*neoDevice->executionEnvironment, neoDevice->getRootDeviceIndex(), neoDevice->getDeviceBitfield());
  mockCommandStreamReceiver.waitForCompletionWithTimeoutReturnValue = WaitStatus::gpuHang;
+ mockCommandStreamReceiver.waitForCompletionWithKmdNotifyFallbackReturnValue = WaitStatus::gpuHang;
 
  auto queue = static_cast<WhiteBox<::L0::CommandQueue> *>(whiteBoxCmdList->cmdQImmediate);
 
@@ -2605,6 +2607,7 @@ HWTEST_F(CommandListCreateTests, GivenGpuHangWhenCreatingImmediateCommandListAnd
 
  MockCommandStreamReceiver mockCommandStreamReceiver(*neoDevice->executionEnvironment, neoDevice->getRootDeviceIndex(), neoDevice->getDeviceBitfield());
  mockCommandStreamReceiver.waitForCompletionWithTimeoutReturnValue = WaitStatus::gpuHang;
+ mockCommandStreamReceiver.waitForCompletionWithKmdNotifyFallbackReturnValue = WaitStatus::gpuHang;
 
  auto queue = static_cast<WhiteBox<::L0::CommandQueue> *>(whiteBoxCmdList->cmdQImmediate);
 
@@ -2674,6 +2677,7 @@ HWTEST_F(CommandListCreateTests, GivenGpuHangAndEnabledFlushTaskSubmissionFlagWh
 
  MockCommandStreamReceiver mockCommandStreamReceiver(*neoDevice->executionEnvironment, neoDevice->getRootDeviceIndex(), neoDevice->getDeviceBitfield());
  mockCommandStreamReceiver.waitForCompletionWithTimeoutReturnValue = WaitStatus::gpuHang;
+ mockCommandStreamReceiver.waitForCompletionWithKmdNotifyFallbackReturnValue = WaitStatus::gpuHang;
 
  auto queue = static_cast<WhiteBox<::L0::CommandQueue> *>(whiteBoxCmdList->cmdQImmediate);
 

level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_8.cpp

@@ -1616,14 +1616,18 @@ HWTEST_F(ImmediateCommandListHostSynchronize, givenMaxTimeoutIsProvidedWaitParam
 
  auto cmdList = createCmdList<FamilyType::gfxCoreFamily>(csr);
 
+ csr->captureWaitForTaskCountWithKmdNotifyInputParams = true;
+
  csr->callBaseWaitForCompletionWithTimeout = false;
  csr->returnWaitForCompletionWithTimeout = WaitStatus::ready;
 
  EXPECT_EQ(cmdList->hostSynchronize(std::numeric_limits<uint64_t>::max()), ZE_RESULT_SUCCESS);
 
  auto waitParams = csr->latestWaitForCompletionWithTimeoutWaitParams;
  EXPECT_FALSE(waitParams.enableTimeout);
- EXPECT_TRUE(waitParams.indefinitelyPoll);
+ EXPECT_FALSE(waitParams.indefinitelyPoll);
+
+ EXPECT_NE(0u, csr->waitForTaskCountWithKmdNotifyInputParams.size());
 }
 
 using CommandListHostSynchronize = Test<DeviceFixture>;

shared/test/common/mocks/mock_command_stream_receiver.h

@@ -185,11 +185,11 @@ class MockCommandStreamReceiver : public CommandStreamReceiver {
  }
 
  WaitStatus waitForTaskCountWithKmdNotifyFallback(TaskCountType taskCountToWait, FlushStamp flushStampToWait, bool quickKmdSleep, QueueThrottle throttle) override {
- return WaitStatus::ready;
+ return waitForCompletionWithKmdNotifyFallbackReturnValue;
  }
 
  WaitStatus waitForTaskCountWithKmdNotifyFallback(TaskCountType taskCountToWait, FlushStamp flushStampToWait, bool quickKmdSleep, bool forcePowerSavingMode) {
- return WaitStatus::ready;
+ return waitForCompletionWithKmdNotifyFallbackReturnValue;
  }
 
  TaskCountType flushBcsTask(const BlitPropertiesContainer &blitPropertiesContainer, bool blocking, Device &device) override { return taskCount; };
@@ -320,6 +320,7 @@ class MockCommandStreamReceiver : public CommandStreamReceiver {
  std::optional<bool> isGpuHangDetectedReturnValue{};
  std::optional<bool> testTaskCountReadyReturnValue{};
  WaitStatus waitForCompletionWithTimeoutReturnValue{WaitStatus::ready};
+ WaitStatus waitForCompletionWithKmdNotifyFallbackReturnValue{WaitStatus::ready};
  CommandStreamReceiverType commandStreamReceiverType = CommandStreamReceiverType::hardware;
  BatchBuffer latestFlushedBatchBuffer = {};
  QueueThrottle getLastDirectSubmissionThrottleReturnValue = QueueThrottle::MEDIUM;