Use MiStoreDataImm for event signals

Signed-off-by: Aravind Gopalakrishnan <[email protected]>

Author: aravindksg

level_zero/core/source/cmdlist/cmdlist_hw.inl

@@ -1450,7 +1450,26 @@ void CommandListCoreFamily<gfxCoreFamily>::appendSignalEventPostWalker(ze_event_
     if (event->isEventTimestampFlagSet()) {
         appendEventForProfiling(hEvent, false);
     } else {
-        CommandListCoreFamily<gfxCoreFamily>::appendSignalEvent(hEvent);
+        using POST_SYNC_OPERATION = typename GfxFamily::PIPE_CONTROL::POST_SYNC_OPERATION;
+        using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
+        auto event = Event::fromHandle(hEvent);
+
+        commandContainer.addToResidencyContainer(&event->getAllocation(this->device));
+        uint64_t baseAddr = event->getGpuAddress(this->device);
+
+        if (isCopyOnly()) {
+            NEO::MiFlushArgs args;
+            args.commandWithPostSync = true;
+            NEO::EncodeMiFlushDW<GfxFamily>::programMiFlushDw(*commandContainer.getCommandStream(), baseAddr, Event::STATE_SIGNALED, args);
+        } else {
+            NEO::PipeControlArgs args;
+            args.dcFlushEnable = (!event->signalScope) ? false : true;
+            NEO::MemorySynchronizationCommands<GfxFamily>::addPipeControlAndProgramPostSyncOperation(
+                *commandContainer.getCommandStream(), POST_SYNC_OPERATION::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA,
+                baseAddr, Event::STATE_SIGNALED,
+                commandContainer.getDevice()->getHardwareInfo(),
+                args);
+        }
     }
 }
 
@@ -1560,6 +1579,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendSignalEvent(ze_event_han
     using POST_SYNC_OPERATION = typename GfxFamily::PIPE_CONTROL::POST_SYNC_OPERATION;
     using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
     auto event = Event::fromHandle(hEvent);
+    bool applyScope = false;
 
     commandContainer.addToResidencyContainer(&event->getAllocation(this->device));
     uint64_t baseAddr = event->getGpuAddress(this->device);
@@ -1574,12 +1594,23 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendSignalEvent(ze_event_han
         NEO::EncodeMiFlushDW<GfxFamily>::programMiFlushDw(*commandContainer.getCommandStream(), ptrOffset(baseAddr, eventSignalOffset), Event::STATE_SIGNALED, args);
     } else {
         NEO::PipeControlArgs args;
-        args.dcFlushEnable = (!event->signalScope) ? false : true;
-        NEO::MemorySynchronizationCommands<GfxFamily>::addPipeControlAndProgramPostSyncOperation(
-            *commandContainer.getCommandStream(), POST_SYNC_OPERATION::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA,
-            ptrOffset(baseAddr, eventSignalOffset), Event::STATE_SIGNALED,
-            commandContainer.getDevice()->getHardwareInfo(),
-            args);
+        applyScope = (!event->signalScope) ? false : true;
+        args.dcFlushEnable = applyScope;
+        if (applyScope || event->isEventTimestampFlagSet()) {
+            NEO::MemorySynchronizationCommands<GfxFamily>::addPipeControlAndProgramPostSyncOperation(
+                *commandContainer.getCommandStream(), POST_SYNC_OPERATION::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA,
+                ptrOffset(baseAddr, eventSignalOffset), Event::STATE_SIGNALED,
+                commandContainer.getDevice()->getHardwareInfo(),
+                args);
+        } else {
+            using MI_STORE_DATA_IMM = typename GfxFamily::MI_STORE_DATA_IMM;
+            MI_STORE_DATA_IMM storeDataImmediate = GfxFamily::cmdInitStoreDataImm;
+            storeDataImmediate.setStoreQword(false);
+            storeDataImmediate.setDwordLength(MI_STORE_DATA_IMM::DWORD_LENGTH::DWORD_LENGTH_STORE_QWORD);
+            storeDataImmediate.setDataDword0(Event::STATE_SIGNALED);
+            auto buffer = commandContainer.getCommandStream()->template getSpaceForCmd<MI_STORE_DATA_IMM>();
+            *buffer = storeDataImmediate;
+        }
     }
     return ZE_RESULT_SUCCESS;
 }
@@ -1752,7 +1783,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendWriteGlobalTimestamp(
     }
 
     if (hSignalEvent) {
-        CommandListCoreFamily<gfxCoreFamily>::appendSignalEvent(hSignalEvent);
+        CommandListCoreFamily<gfxCoreFamily>::appendSignalEventPostWalker(hSignalEvent);
     }
 
     auto allocationStruct = getAlignedAllocation(this->device, dstptr, sizeof(uint64_t));

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

@@ -1515,7 +1515,7 @@ HWTEST_F(CommandListCreate, givenCommandListWithCopyOnlyWhenAppendSignalEventThe
     EXPECT_NE(cmdList.end(), itor);
 }
 
-HWTEST_F(CommandListCreate, givenCommandListWhenAppendSignalEventThenPipeControlIsProgrammed) {
+HWTEST_F(CommandListCreate, givenCommandListWhenAppendSignalEventWithScopeThenPipeControlIsProgrammed) {
     using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
     ze_result_t returnValue;
     std::unique_ptr<L0::CommandList> commandList(CommandList::create(productFamily, device, NEO::EngineGroupType::RenderCompute, returnValue));

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

@@ -20,9 +20,10 @@ namespace L0 {
 namespace ult {
 
 using CommandListAppendSignalEvent = Test<CommandListFixture>;
-HWTEST_F(CommandListAppendSignalEvent, WhenAppendingSignalEventThenPipeControlIsGenerated) {
+HWTEST_F(CommandListAppendSignalEvent, WhenAppendingSignalEventWithoutScopeThenMiStoreImmIsGenerated) {
     using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
     using POST_SYNC_OPERATION = typename PIPE_CONTROL::POST_SYNC_OPERATION;
+    using MI_STORE_DATA_IMM = typename FamilyType::MI_STORE_DATA_IMM;
 
     auto usedSpaceBefore = commandList->commandContainer.getCommandStream()->getUsed();
     auto result = commandList->appendSignalEvent(event->toHandle());
@@ -35,22 +36,8 @@ HWTEST_F(CommandListAppendSignalEvent, WhenAppendingSignalEventThenPipeControlIs
     ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(
         cmdList, ptrOffset(commandList->commandContainer.getCommandStream()->getCpuBase(), 0), usedSpaceAfter));
 
-    // Find a PC w/ a WriteImmediateData and CS stall
-    auto itorPC = findAll<PIPE_CONTROL *>(cmdList.begin(), cmdList.end());
-    ASSERT_NE(0u, itorPC.size());
-    bool postSyncFound = false;
-    for (auto it : itorPC) {
-        auto cmd = genCmdCast<PIPE_CONTROL *>(*it);
-        if (cmd->getPostSyncOperation() == POST_SYNC_OPERATION::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA) {
-            EXPECT_TRUE(cmd->getCommandStreamerStallEnable());
-            EXPECT_EQ(cmd->getImmediateData(), Event::STATE_SIGNALED);
-            auto gpuAddress = event->getGpuAddress(device);
-            EXPECT_EQ(cmd->getAddressHigh(), gpuAddress >> 32u);
-            EXPECT_EQ(cmd->getAddress(), uint32_t(gpuAddress));
-            postSyncFound = true;
-        }
-    }
-    ASSERT_TRUE(postSyncFound);
+    auto itor = findAll<MI_STORE_DATA_IMM *>(cmdList.begin(), cmdList.end());
+    ASSERT_NE(0u, itor.size());
 }
 
 HWTEST_F(CommandListAppendSignalEvent, givenCmdlistWhenAppendingSignalEventThenEventPoolGraphicsAllocationIsAddedToResidencyContainer) {
@@ -66,37 +53,6 @@ HWTEST_F(CommandListAppendSignalEvent, givenCmdlistWhenAppendingSignalEventThenE
     }
 }
 
-HWTEST_F(CommandListAppendSignalEvent, givenEventWithScopeFlagNoneWhenAppendingSignalEventThenPipeControlHasNoDcFlush) {
-    using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
-    using POST_SYNC_OPERATION = typename PIPE_CONTROL::POST_SYNC_OPERATION;
-
-    auto usedSpaceBefore = commandList->commandContainer.getCommandStream()->getUsed();
-    auto result = commandList->appendSignalEvent(event->toHandle());
-    ASSERT_EQ(ZE_RESULT_SUCCESS, result);
-
-    auto usedSpaceAfter = commandList->commandContainer.getCommandStream()->getUsed();
-    ASSERT_GT(usedSpaceAfter, usedSpaceBefore);
-
-    GenCmdList cmdList;
-    ASSERT_TRUE(FamilyType::PARSE::parseCommandBuffer(cmdList,
-                                                      ptrOffset(commandList->commandContainer.getCommandStream()->getCpuBase(), 0),
-                                                      usedSpaceAfter));
-
-    auto itorPC = findAll<PIPE_CONTROL *>(cmdList.begin(), cmdList.end());
-    ASSERT_NE(0u, itorPC.size());
-    bool postSyncFound = false;
-    for (auto it : itorPC) {
-        auto cmd = genCmdCast<PIPE_CONTROL *>(*it);
-        if (cmd->getPostSyncOperation() == POST_SYNC_OPERATION::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA) {
-            EXPECT_EQ(cmd->getImmediateData(), Event::STATE_SIGNALED);
-            EXPECT_TRUE(cmd->getCommandStreamerStallEnable());
-            EXPECT_FALSE(cmd->getDcFlushEnable());
-            postSyncFound = true;
-        }
-    }
-    ASSERT_TRUE(postSyncFound);
-}
-
 HWTEST_F(CommandListAppendSignalEvent, givenEventWithScopeFlagDeviceWhenAppendingSignalEventThenPipeControlHasNoDcFlush) {
     using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
     using POST_SYNC_OPERATION = typename PIPE_CONTROL::POST_SYNC_OPERATION;