feature: Experimental support of immediate cmd list in-order execution [4/n]

Related-To: LOCI-4332

- Simplify CmdList-Event dependency
- Add waiting on in-order dependency
- Prepare Event for in-order synchronization
- Adjust downloading sync allocation in TBX mode

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>

level_zero/core/source/event/event.cpp

@@ -335,10 +335,6 @@ ze_result_t EventPool::openEventPoolIpcHandle(const ze_ipc_event_pool_handle_t &
 }
 
 ze_result_t Event::destroy() {
-    if (latestUsedInOrderCmdList) {
-        latestUsedInOrderCmdList->unsetLastInOrderOutEvent(this->toHandle());
-    }
-
     delete this;
     return ZE_RESULT_SUCCESS;
 }
@@ -391,12 +387,13 @@ void Event::resetPackets(bool resetAllPackets) {
 void Event::setIsCompleted() {
     if (this->isCompleted.load() == STATE_CLEARED) {
         this->isCompleted = STATE_SIGNALED;
-
-        if (latestUsedInOrderCmdList) {
-            latestUsedInOrderCmdList->unsetLastInOrderOutEvent(this->toHandle());
-            latestUsedInOrderCmdList = nullptr;
-        }
     }
 }
 
+void Event::enableInOrderExecMode(NEO::GraphicsAllocation &inOrderDependenciesAllocation, uint32_t signalValue) {
+    inOrderExecEvent = true;
+    inOrderExecSignalValue = signalValue;
+    inOrderExecDataAllocation = &inOrderDependenciesAllocation;
+}
+
 } // namespace L0

level_zero/core/source/event/event.h

@@ -39,7 +39,6 @@ struct DriverHandle;
 struct DriverHandleImp;
 struct Device;
 struct Kernel;
-struct CommandListImp;
 
 #pragma pack(1)
 struct IpcEventPoolData {
@@ -204,10 +203,9 @@ struct Event : _ze_event_handle_t {
     void setMetricStreamer(MetricStreamer *metricStreamer) {
         this->metricStreamer = metricStreamer;
     }
-
-    void setLatestUsedInOrderCmdList(CommandListImp *newCmdList) {
-        latestUsedInOrderCmdList = newCmdList;
-    }
+    void enableInOrderExecMode(NEO::GraphicsAllocation &inOrderDependenciesAllocation, uint32_t signalValue);
+    bool isInOrderExecEvent() const { return inOrderExecEvent; }
+    uint32_t getInOrderExecSignalValue() const { return inOrderExecSignalValue; }
 
   protected:
     Event(EventPool *eventPool, int index, Device *device) : device(device), eventPool(eventPool), index(index) {}
@@ -239,12 +237,13 @@ struct Event : _ze_event_handle_t {
     Device *device = nullptr;
     EventPool *eventPool = nullptr;
     Kernel *kernelWithPrintf = nullptr;
-    CommandListImp *latestUsedInOrderCmdList = nullptr;
+    NEO::GraphicsAllocation *inOrderExecDataAllocation = nullptr;
 
     uint32_t maxKernelCount = 0;
     uint32_t kernelCount = 1u;
     uint32_t maxPacketCount = 0;
     uint32_t totalEventSize = 0;
+    uint32_t inOrderExecSignalValue = 0;
 
     ze_event_scope_flags_t signalScope = 0u;
     ze_event_scope_flags_t waitScope = 0u;
@@ -257,6 +256,7 @@ struct Event : _ze_event_handle_t {
     bool usingContextEndOffset = false;
     bool signalAllEventPackets = false;
     bool isFromIpcPool = false;
+    bool inOrderExecEvent = false;
 };
 
 struct EventPool : _ze_event_pool_handle_t {

level_zero/core/source/event/event_impl.inl

@@ -186,6 +186,9 @@ ze_result_t EventImp<TagSizeT>::queryStatus() {
     if (this->downloadAllocationRequired) {
         for (auto &csr : csrs) {
             csr->downloadAllocation(this->getAllocation(this->device));
+            if (inOrderExecEvent) {
+                csr->downloadAllocation(*this->inOrderExecDataAllocation);
+            }
         }
     }
 
@@ -352,9 +355,10 @@ ze_result_t EventImp<TagSizeT>::hostSynchronize(uint64_t timeout) {
 
 template <typename TagSizeT>
 ze_result_t EventImp<TagSizeT>::reset() {
-    if (latestUsedInOrderCmdList) {
-        latestUsedInOrderCmdList->unsetLastInOrderOutEvent(this->toHandle());
-        latestUsedInOrderCmdList = nullptr;
+    if (inOrderExecEvent) {
+        inOrderExecDataAllocation = nullptr;
+        inOrderExecSignalValue = 0;
+        inOrderExecEvent = false;
     }
     this->resetCompletionStatus();
     this->resetDeviceCompletionData(false);