Add option to compact event L3 flush packet

Related-To: NEO-7434

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

Author: zzdanowicz

level_zero/core/source/cmdlist/cmdlist_hw.h

@@ -287,10 +287,18 @@ struct CommandListCoreFamily : CommandListImp {
                                   size_t dstSize,
                                   CmdListFillKernelArguments &outArguments,
                                   Kernel *kernel);
+    bool compactL3FlushEvent(bool dcFlush) const {
+        return this->compactL3FlushEventPacket && dcFlush;
+    }
+    bool eventSignalPipeControl(bool splitKernel, bool dcFlush) const {
+        return (this->pipeControlMultiKernelEventSync && splitKernel) ||
+               compactL3FlushEvent(dcFlush);
+    }
 
     size_t cmdListCurrentStartOffset = 0;
     bool containsAnyKernel = false;
     bool pipeControlMultiKernelEventSync = false;
+    bool compactL3FlushEventPacket = false;
 };
 
 template <PRODUCT_FAMILY gfxProductFamily>

level_zero/core/source/cmdlist/cmdlist_hw.inl

@@ -138,6 +138,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::initialize(Device *device, NEO
     this->frontEndStateTracking = L0HwHelper::enableFrontEndStateTracking(hwInfo);
     this->pipelineSelectStateTracking = L0HwHelper::enablePipelineSelectStateTracking(hwInfo);
     this->pipeControlMultiKernelEventSync = L0HwHelper::usePipeControlMultiKernelEventSync(hwInfo);
+    this->compactL3FlushEventPacket = L0HwHelper::useCompactL3FlushEventPacket(hwInfo);
 
     if (device->isImplicitScalingCapable() && !this->internalUsage && !isCopyOnly()) {
         this->partitionCount = static_cast<uint32_t>(this->device->getNEODevice()->getDeviceBitfield().count());
@@ -1187,19 +1188,20 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryCopy(void *dstptr,
     }
 
     CmdListKernelLaunchParams launchParams = {};
-
+    bool dcFlush = false;
     Event *signalEvent = nullptr;
     if (hSignalEvent) {
         signalEvent = Event::fromHandle(hSignalEvent);
         launchParams.isHostSignalScopeEvent = !!(signalEvent->signalScope & ZE_EVENT_SCOPE_FLAG_HOST);
+        dcFlush = getDcFlushRequired(!!signalEvent->signalScope);
     }
 
     uint32_t kernelCounter = leftSize > 0 ? 1 : 0;
     kernelCounter += middleSizeBytes > 0 ? 1 : 0;
     kernelCounter += rightSize > 0 ? 1 : 0;
 
     launchParams.isKernelSplitOperation = kernelCounter > 1;
-    bool singlePipeControlPacket = this->pipeControlMultiKernelEventSync && launchParams.isKernelSplitOperation;
+    bool singlePipeControlPacket = eventSignalPipeControl(launchParams.isKernelSplitOperation, dcFlush);
 
     appendEventForProfilingAllWalkers(signalEvent, true, singlePipeControlPacket);
 
@@ -1551,9 +1553,11 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryFill(void *ptr,
     CmdListKernelLaunchParams launchParams = {};
 
     Event *signalEvent = nullptr;
+    bool dcFlush = false;
     if (hSignalEvent) {
         signalEvent = Event::fromHandle(hSignalEvent);
         launchParams.isHostSignalScopeEvent = !!(signalEvent->signalScope & ZE_EVENT_SCOPE_FLAG_HOST);
+        dcFlush = getDcFlushRequired(!!signalEvent->signalScope);
     }
 
     if (isCopyOnly()) {
@@ -1610,7 +1614,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendMemoryFill(void *ptr,
     setupFillKernelArguments(dstAllocation.offset, patternSize, size, fillArguments, builtinKernel);
 
     launchParams.isKernelSplitOperation = (fillArguments.leftRemainingBytes > 0 || fillArguments.rightRemainingBytes > 0);
-    bool singlePipeControlPacket = this->pipeControlMultiKernelEventSync && launchParams.isKernelSplitOperation;
+    bool singlePipeControlPacket = eventSignalPipeControl(launchParams.isKernelSplitOperation, dcFlush);
 
     appendEventForProfilingAllWalkers(signalEvent, true, singlePipeControlPacket);
 

level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl

@@ -163,20 +163,26 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
                               threadGroupDimensions->groupCountY,
                               threadGroupDimensions->groupCountZ);
     }
-    NEO::GraphicsAllocation *eventAlloc = nullptr;
+
     uint64_t eventAddress = 0;
     bool isTimestampEvent = false;
     bool l3FlushEnable = false;
     bool isHostSignalScopeEvent = launchParams.isHostSignalScopeEvent;
+    Event *compactEvent = nullptr;
     if (event) {
-        eventAlloc = &event->getAllocation(this->device);
-        commandContainer.addToResidencyContainer(eventAlloc);
-        bool flushRequired = !!event->signalScope &&
-                             !launchParams.isKernelSplitOperation;
-        l3FlushEnable = getDcFlushRequired(flushRequired);
-        isTimestampEvent = event->isUsingContextEndOffset();
-        eventAddress = event->getPacketAddress(this->device);
         isHostSignalScopeEvent = !!(event->signalScope & ZE_EVENT_SCOPE_FLAG_HOST);
+        if (compactL3FlushEvent(getDcFlushRequired(!!event->signalScope))) {
+            compactEvent = event;
+            event = nullptr;
+        } else {
+            NEO::GraphicsAllocation *eventAlloc = &event->getAllocation(this->device);
+            commandContainer.addToResidencyContainer(eventAlloc);
+            bool flushRequired = !!event->signalScope &&
+                                 !launchParams.isKernelSplitOperation;
+            l3FlushEnable = getDcFlushRequired(flushRequired);
+            isTimestampEvent = event->isUsingContextEndOffset();
+            eventAddress = event->getPacketAddress(this->device);
+        }
     }
 
     bool isKernelUsingSystemAllocation = false;
@@ -249,6 +255,10 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
 
     std::list<void *> additionalCommands;
 
+    if (compactEvent) {
+        appendEventForProfilingAllWalkers(compactEvent, true, true);
+    }
+
     NEO::EncodeDispatchKernelArgs dispatchKernelArgs{
         eventAddress,                                             // eventAddress
         neoDevice,                                                // device
@@ -273,7 +283,9 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
     NEO::EncodeDispatchKernel<GfxFamily>::encode(commandContainer, dispatchKernelArgs, getLogicalStateHelper());
     this->containsStatelessUncachedResource = dispatchKernelArgs.requiresUncachedMocs;
 
-    if (event) {
+    if (compactEvent) {
+        appendEventForProfilingAllWalkers(compactEvent, false, true);
+    } else if (event) {
         if (partitionCount > 1) {
             event->setPacketsInUse(partitionCount);
         }
@@ -404,7 +416,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelSplit(Kernel
                                                                           Event *event,
                                                                           const CmdListKernelLaunchParams &launchParams) {
     if (event) {
-        if (this->pipeControlMultiKernelEventSync && launchParams.isKernelSplitOperation) {
+        if (eventSignalPipeControl(launchParams.isKernelSplitOperation, getDcFlushRequired(!!event->signalScope))) {
             event = nullptr;
         } else {
             event->increaseKernelCount();

level_zero/core/source/hw_helpers/l0_hw_helper.cpp

@@ -53,4 +53,11 @@ bool L0HwHelper::usePipeControlMultiKernelEventSync(const NEO::HardwareInfo &hwI
     return false;
 }
 
+bool L0HwHelper::useCompactL3FlushEventPacket(const NEO::HardwareInfo &hwInfo) {
+    if (NEO::DebugManager.flags.CompactL3FlushEventPacket.get() != -1) {
+        return !!NEO::DebugManager.flags.CompactL3FlushEventPacket.get();
+    }
+    return false;
+}
+
 } // namespace L0

level_zero/core/source/hw_helpers/l0_hw_helper.h

@@ -35,6 +35,7 @@ class L0HwHelper {
     static bool enableStateComputeModeTracking(const NEO::HardwareInfo &hwInfo);
     static bool enableImmediateCmdListHeapSharing(const NEO::HardwareInfo &hwInfo, bool cmdlistSupport);
     static bool usePipeControlMultiKernelEventSync(const NEO::HardwareInfo &hwInfo);
+    static bool useCompactL3FlushEventPacket(const NEO::HardwareInfo &hwInfo);
     virtual void setAdditionalGroupProperty(ze_command_queue_group_properties_t &groupProperty, NEO::EngineGroupT &group) const = 0;
     virtual L0::Event *createEvent(L0::EventPool *eventPool, const ze_event_desc_t *desc, L0::Device *device) const = 0;
 

level_zero/core/test/unit_tests/fixtures/cmdlist_fixture.h

@@ -188,7 +188,9 @@ struct TestExpectedValues {
     uint32_t expectedKernelCount = 0;
     uint32_t expectedWalkerPostSyncOp = 0;
     uint32_t expectedPostSyncPipeControls = 0;
+    uint32_t expectDcFlush = 0;
     bool postSyncAddressZero = false;
+    bool workloadPartition = false;
 };
 
 } // namespace ult

level_zero/core/test/unit_tests/mocks/mock_cmdlist.h

@@ -47,6 +47,7 @@ struct WhiteBox<::L0::CommandListCoreFamily<gfxCoreFamily>>
     using BaseClass::commandListPerThreadScratchSize;
     using BaseClass::commandListPreemptionMode;
     using BaseClass::commandsToPatch;
+    using BaseClass::compactL3FlushEventPacket;
     using BaseClass::containsAnyKernel;
     using BaseClass::containsCooperativeKernelsFlag;
     using BaseClass::csr;
@@ -123,6 +124,7 @@ struct WhiteBox<L0::CommandListCoreFamilyImmediate<gfxCoreFamily>>
     using BaseClass::clearCommandsToPatch;
     using BaseClass::cmdQImmediate;
     using BaseClass::commandsToPatch;
+    using BaseClass::compactL3FlushEventPacket;
     using BaseClass::csr;
     using BaseClass::finalStreamState;
     using BaseClass::frontEndStateTracking;
@@ -142,6 +144,7 @@ struct WhiteBox<L0::CommandListCoreFamilyImmediate<gfxCoreFamily>>
 template <GFXCORE_FAMILY gfxCoreFamily>
 struct MockCommandListImmediate : public CommandListCoreFamilyImmediate<gfxCoreFamily> {
     using BaseClass = CommandListCoreFamilyImmediate<gfxCoreFamily>;
+    using BaseClass::compactL3FlushEventPacket;
     using BaseClass::containsAnyKernel;
     using BaseClass::immediateCmdListHeapSharing;
     using BaseClass::indirectAllocationsAllowed;