Use latest sent task count for Linux flush stamp

Related-To: NEO-5845

Signed-off-by: Zbigniew Zdanowicz <[email protected]>

Author: zzdanowicz

opencl/source/os_interface/linux/drm_command_stream.h

@@ -32,6 +32,7 @@ class DrmCommandStreamReceiver : public DeviceCommandStreamReceiver<GfxFamily> {
     using BaseClass::requiredScratchSize;
     using CommandStreamReceiverHw<GfxFamily>::CommandStreamReceiver::getTagAddress;
     using CommandStreamReceiverHw<GfxFamily>::CommandStreamReceiver::getTagAllocation;
+    using CommandStreamReceiverHw<GfxFamily>::CommandStreamReceiver::latestSentTaskCount;
     using CommandStreamReceiverHw<GfxFamily>::CommandStreamReceiver::taskCount;
     using CommandStreamReceiverHw<GfxFamily>::CommandStreamReceiver::useNotifyEnableForPostSync;
 

opencl/source/os_interface/linux/drm_command_stream.inl

@@ -116,7 +116,7 @@ bool DrmCommandStreamReceiver<GfxFamily>::flush(BatchBuffer &batchBuffer, Reside
     }
 
     if (isUserFenceWaitActive()) {
-        this->flushStamp->setStamp(taskCount);
+        this->flushStamp->setStamp(latestSentTaskCount);
     } else {
         this->flushStamp->setStamp(bb->peekHandle());
     }

opencl/test/unit_test/command_stream/command_stream_receiver_flush_task_2_tests.cpp

@@ -32,7 +32,7 @@
 
 using namespace NEO;
 
-typedef UltCommandStreamReceiverTest CommandStreamReceiverFlushTaskTests;
+using CommandStreamReceiverFlushTaskTests = UltCommandStreamReceiverTest;
 
 HWTEST_F(CommandStreamReceiverFlushTaskTests, GivenBlockedKernelNotRequiringDCFlushWhenUnblockedThenDCFlushIsNotAdded) {
     typedef typename FamilyType::PIPE_CONTROL PIPE_CONTROL;
@@ -491,8 +491,12 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, givenCommandStreamReceiverWhenFenc
     EXPECT_FALSE(csr.isMadeResident(csr.globalFenceAllocation));
     EXPECT_FALSE(csr.isMadeNonResident(csr.globalFenceAllocation));
 
+    csr.taskCount = 2u;
     flushSmallTask(csr);
 
+    EXPECT_EQ(3u, csr.latestSentTaskCount);
+    EXPECT_EQ(3u, csr.latestFlushedTaskCount);
+
     EXPECT_TRUE(csr.isMadeResident(csr.globalFenceAllocation));
     EXPECT_TRUE(csr.isMadeNonResident(csr.globalFenceAllocation));
 

opencl/test/unit_test/mocks/linux/mock_drm_command_stream_receiver.h

@@ -21,6 +21,7 @@ class TestedDrmCommandStreamReceiver : public DrmCommandStreamReceiver<GfxFamily
     using CommandStreamReceiver::flushStamp;
     using CommandStreamReceiver::getTagAddress;
     using CommandStreamReceiver::globalFenceAllocation;
+    using CommandStreamReceiver::latestSentTaskCount;
     using CommandStreamReceiver::makeResident;
     using CommandStreamReceiver::taskCount;
     using CommandStreamReceiver::useGpuIdleImplicitFlush;

opencl/test/unit_test/mocks/mock_aub_csr.h

@@ -78,10 +78,6 @@ struct MockAubCsr : public AUBCommandStreamReceiverHw<GfxFamily> {
         return this->tagAllocation;
     }
 
-    void setLatestSentTaskCount(uint32_t latestSentTaskCount) {
-        this->latestSentTaskCount = latestSentTaskCount;
-    }
-
     bool flushBatchedSubmissions() override {
         flushBatchedSubmissionsCalled = true;
         return true;

opencl/test/unit_test/os_interface/linux/drm_command_stream_tests_2.cpp

@@ -789,7 +789,7 @@ HWTEST_TEMPLATED_F(DrmCommandStreamEnhancedTest, givenAllocationWithSingleBuffer
 }
 
 HWTEST_TEMPLATED_F(DrmCommandStreamEnhancedTest,
-                   givenWaitUserFenceFlagAndVmBindAvailableSetWhenDrmCsrFlushedThenExpectTaskCountStoredAsFlushStamp) {
+                   givenWaitUserFenceFlagAndVmBindAvailableSetWhenDrmCsrFlushedThenExpectLatestSentTaskCountStoredAsFlushStamp) {
     DebugManagerStateRestore restorer;
     DebugManager.flags.EnableUserFenceForCompletionWait.set(1);
 
@@ -811,7 +811,7 @@ HWTEST_TEMPLATED_F(DrmCommandStreamEnhancedTest,
     CommandStreamReceiverHw<FamilyType>::alignToCacheLine(cs);
     BatchBuffer batchBuffer{cs.getGraphicsAllocation(), 0, 0, nullptr, false, false, QueueThrottle::MEDIUM, QueueSliceCount::defaultSliceCount, cs.getUsed(), &cs, nullptr, false};
 
-    testedCsr->taskCount = 160u;
+    testedCsr->latestSentTaskCount = 160u;
     testedCsr->flush(batchBuffer, testedCsr->getResidencyAllocations());
 
     EXPECT_EQ(160u, testedCsr->flushStamp->peekStamp());
@@ -841,7 +841,7 @@ HWTEST_TEMPLATED_F(DrmCommandStreamEnhancedTest,
 
     DrmAllocation *alloc = static_cast<DrmAllocation *>(cs.getGraphicsAllocation());
     auto boHandle = static_cast<FlushStamp>(alloc->getBO()->peekHandle());
-    testedCsr->taskCount = 160u;
+    testedCsr->latestSentTaskCount = 160u;
 
     testedCsr->flush(batchBuffer, testedCsr->getResidencyAllocations());
 
@@ -875,7 +875,7 @@ HWTEST_TEMPLATED_F(DrmCommandStreamEnhancedTest,
 
     DrmAllocation *alloc = static_cast<DrmAllocation *>(cs.getGraphicsAllocation());
     auto boHandle = static_cast<FlushStamp>(alloc->getBO()->peekHandle());
-    testedCsr->taskCount = 160u;
+    testedCsr->latestSentTaskCount = 160u;
 
     testedCsr->flush(batchBuffer, testedCsr->getResidencyAllocations());
 

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -1199,8 +1199,6 @@ void CommandStreamReceiverHw<GfxFamily>::flushPipeControl() {
 
     this->flushSmallTask(commandStream, commandStreamStart);
 
-    this->latestFlushedTaskCount = taskCount + 1;
-    this->latestSentTaskCount = taskCount + 1;
     taskCount++;
 }
 
@@ -1291,6 +1289,9 @@ void CommandStreamReceiverHw<GfxFamily>::flushSmallTask(LinearStream &commandStr
         makeResident(*globalFenceAllocation);
     }
 
+    this->latestFlushedTaskCount = taskCount + 1;
+    this->latestSentTaskCount = taskCount + 1;
+
     BatchBuffer batchBuffer{commandStreamTask.getGraphicsAllocation(), commandStreamStartTask, 0, nullptr, false, false, QueueThrottle::MEDIUM, QueueSliceCount::defaultSliceCount,
                             commandStreamTask.getUsed(), &commandStreamTask, endingCmdPtr, false};