Add debug flag to select event field for event completion

Related-To: NEO-6871

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>

Author: zzdanowicz

level_zero/core/source/event/event_impl.inl

@@ -5,6 +5,7 @@
  *
  */
 
+#include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/memory_manager/internal_allocation_storage.h"
 
 #include "level_zero/core/source/event/event.h"
@@ -30,7 +31,12 @@ Event *Event::create(EventPool *eventPool, const ze_event_desc_t *desc, Device *
     event->signalScope = desc->signal;
     event->waitScope = desc->wait;
     event->csr = neoDevice->getDefaultEngine().commandStreamReceiver;
-    event->setUsingContextEndOffset(L0HwHelper::get(neoDevice->getHardwareInfo().platform.eRenderCoreFamily).multiTileCapablePlatform());
+    bool useContextEndOffset = L0HwHelper::get(neoDevice->getHardwareInfo().platform.eRenderCoreFamily).multiTileCapablePlatform();
+    int32_t overrideUseContextEndOffset = NEO::DebugManager.flags.UseContextEndOffsetForEventCompletion.get();
+    if (overrideUseContextEndOffset != -1) {
+        useContextEndOffset = !!overrideUseContextEndOffset;
+    }
+    event->setUsingContextEndOffset(useContextEndOffset);
 
     EventPoolImp *EventPoolImp = static_cast<struct EventPoolImp *>(eventPool);
     // do not reset even if it has been imported, since event pool

level_zero/core/test/unit_tests/sources/event/test_event.cpp

@@ -603,6 +603,80 @@ TEST_F(EventCreate, givenAnEventCreateWithInvalidIndexUsingThisEventPoolThenErro
     ASSERT_EQ(ZE_RESULT_ERROR_INVALID_ARGUMENT, value);
 }
 
+HWTEST2_F(EventCreate, givenPlatformSupportMultTileWhenDebugKeyIsSetToNotUseContextEndThenDoNotUseContextEndOffset, isXeHpOrXeHpcCore) {
+    DebugManagerStateRestore restorer;
+    NEO::DebugManager.flags.UseContextEndOffsetForEventCompletion.set(0);
+
+    bool useContextEndOffset = L0HwHelper::get(neoDevice->getHardwareInfo().platform.eRenderCoreFamily).multiTileCapablePlatform();
+    EXPECT_TRUE(useContextEndOffset);
+
+    ze_event_pool_desc_t eventPoolDesc = {
+        ZE_STRUCTURE_TYPE_EVENT_POOL_DESC,
+        nullptr,
+        0,
+        1};
+    const ze_event_desc_t eventDesc = {
+        ZE_STRUCTURE_TYPE_EVENT_DESC,
+        nullptr,
+        0,
+        0,
+        0};
+
+    ze_event_handle_t eventHandle = nullptr;
+
+    ze_result_t result = ZE_RESULT_SUCCESS;
+    std::unique_ptr<L0::EventPool> eventPool(EventPool::create(driverHandle.get(), context, 0, nullptr, &eventPoolDesc, result));
+    EXPECT_EQ(ZE_RESULT_SUCCESS, result);
+    ASSERT_NE(nullptr, eventPool);
+
+    ze_result_t value = eventPool->createEvent(&eventDesc, &eventHandle);
+    ASSERT_NE(nullptr, eventHandle);
+    EXPECT_EQ(ZE_RESULT_SUCCESS, value);
+
+    auto event = Event::fromHandle(eventHandle);
+    EXPECT_FALSE(event->isEventTimestampFlagSet());
+    EXPECT_FALSE(event->isUsingContextEndOffset());
+
+    event->destroy();
+}
+
+HWTEST2_F(EventCreate, givenPlatformNotSupportsMultTileWhenDebugKeyIsSetToUseContextEndThenUseContextEndOffset, isNotXeHpOrXeHpcCore) {
+    DebugManagerStateRestore restorer;
+    NEO::DebugManager.flags.UseContextEndOffsetForEventCompletion.set(1);
+
+    bool useContextEndOffset = L0HwHelper::get(neoDevice->getHardwareInfo().platform.eRenderCoreFamily).multiTileCapablePlatform();
+    EXPECT_FALSE(useContextEndOffset);
+
+    ze_event_pool_desc_t eventPoolDesc = {
+        ZE_STRUCTURE_TYPE_EVENT_POOL_DESC,
+        nullptr,
+        0,
+        1};
+    const ze_event_desc_t eventDesc = {
+        ZE_STRUCTURE_TYPE_EVENT_DESC,
+        nullptr,
+        0,
+        0,
+        0};
+
+    ze_event_handle_t eventHandle = nullptr;
+
+    ze_result_t result = ZE_RESULT_SUCCESS;
+    std::unique_ptr<L0::EventPool> eventPool(EventPool::create(driverHandle.get(), context, 0, nullptr, &eventPoolDesc, result));
+    EXPECT_EQ(ZE_RESULT_SUCCESS, result);
+    ASSERT_NE(nullptr, eventPool);
+
+    ze_result_t value = eventPool->createEvent(&eventDesc, &eventHandle);
+    ASSERT_NE(nullptr, eventHandle);
+    EXPECT_EQ(ZE_RESULT_SUCCESS, value);
+
+    auto event = Event::fromHandle(eventHandle);
+    EXPECT_FALSE(event->isEventTimestampFlagSet());
+    EXPECT_TRUE(event->isUsingContextEndOffset());
+
+    event->destroy();
+}
+
 class EventSynchronizeTest : public Test<DeviceFixture> {
   public:
     void SetUp() override {

opencl/test/unit_test/test_files/igdrcl.config

@@ -411,5 +411,6 @@ DirectSubmissionReadBackCommandBuffer = -1
 DirectSubmissionReadBackRingBuffer = -1
 ReadBackCommandBufferAllocation = -1
 PrintImageBlitBlockCopyCmdDetails = 0
+UseContextEndOffsetForEventCompletion = -1
 DirectSubmissionInsertExtraMiMemFenceCommands = -1
 DirectSubmissionInsertSfenceInstructionPriorToSubmission = -1

shared/source/debug_settings/debug_variables_base.inl

@@ -200,6 +200,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, ForcePreParserEnabledForMiArbCheck, -1, "-1: def
 DECLARE_DEBUG_VARIABLE(int32_t, BatchBufferStartPrepatchingWaEnabled, -1, "-1: default , 0: disabled, 1: enabled. WA applies valid VA pointing to 'self' instead of 0x0. This mitigates incorrect VA preparsing.")
 DECLARE_DEBUG_VARIABLE(int32_t, SetVmAdviseAtomicAttribute, -1, "-1: default - atomic system, 0: atomic none, 1: atomic device, 2: atomic system)")
 DECLARE_DEBUG_VARIABLE(int32_t, ReadBackCommandBufferAllocation, -1, "Read command buffer allocation back on the host side. -1: default, 0 - disabled, 1 - local memory only, 2 - local and system memory")
+DECLARE_DEBUG_VARIABLE(int32_t, UseContextEndOffsetForEventCompletion, -1, "Use Context End or Context Start for event completion signalling. -1: default: platform dependent, 0 - Use Context Start, 1 - Use Context End")
 DECLARE_DEBUG_VARIABLE(bool, DisableScratchPages, false, "Disable scratch pages during VM creations")
 /*LOGGING FLAGS*/
 DECLARE_DEBUG_VARIABLE(int32_t, PrintDriverDiagnostics, -1, "prints driver diagnostics messages to standard output, value corresponds to hint level")

shared/test/common/test_macros/header/common_matchers.h

@@ -37,6 +37,8 @@ using IsAtMostXeHpcCore = IsAtMostGfxCore<IGFX_XE_HPC_CORE>;
 using isXeHpOrXeHpcCore = IsAnyGfxCores<IGFX_XE_HP_CORE, IGFX_XE_HPC_CORE>;
 using isXeHpcOrXeHpgCore = IsAnyGfxCores<IGFX_XE_HPC_CORE, IGFX_XE_HPG_CORE>;
 
+using isNotXeHpOrXeHpcCore = IsNotAnyGfxCores<IGFX_XE_HP_CORE, IGFX_XE_HPC_CORE>;
+
 using IsSKL = IsProduct<IGFX_SKYLAKE>;
 using IsKBL = IsProduct<IGFX_KABYLAKE>;
 using IsCFL = IsProduct<IGFX_COFFEELAKE>;

shared/test/common/test_macros/test.h

@@ -830,6 +830,14 @@ struct IsAnyGfxCores {
     }
 };
 
+template <GFXCORE_FAMILY... args>
+struct IsNotAnyGfxCores {
+    template <PRODUCT_FAMILY productFamily>
+    static constexpr bool isMatched() {
+        return (... && IsNotGfxCore<args>::template isMatched<productFamily>());
+    }
+};
+
 template <PRODUCT_FAMILY product>
 struct IsProduct {
     template <PRODUCT_FAMILY productFamily>