let's pretend it all has been done in one commit

Author: EuphoricThinking

unified-runtime/source/adapters/level_zero/CMakeLists.txt

@@ -165,6 +165,7 @@ if(UR_BUILD_ADAPTER_L0_V2)
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/memory.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/lockable.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_api.hpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_batched.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_immediate_in_order.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_immediate_out_of_order.hpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/usm.hpp
@@ -181,6 +182,7 @@ if(UR_BUILD_ADAPTER_L0_V2)
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/kernel.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/memory.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_api.cpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_batched.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_create.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_immediate_in_order.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/v2/queue_immediate_out_of_order.cpp

unified-runtime/source/adapters/level_zero/v2/command_list_manager.cpp

@@ -61,6 +61,16 @@ ur_result_t ur_command_list_manager::appendGenericFillUnlocked(
   return UR_RESULT_SUCCESS;
 }
 
+v2::raii::command_list_unique_handle &&
+ur_command_list_manager::releaseCommandList() {
+  return std::move(zeCommandList);
+}
+
+void ur_command_list_manager::replaceCommandList(
+    v2::raii::command_list_unique_handle &&cmdlist) {
+  zeCommandList = std::move(cmdlist);
+}
+
 ur_result_t ur_command_list_manager::appendGenericCopyUnlocked(
     ur_mem_buffer_t *src, ur_mem_buffer_t *dst, bool blocking, size_t srcOffset,
     size_t dstOffset, size_t size, uint32_t numEventsInWaitList,
@@ -129,6 +139,7 @@ wait_list_view ur_command_list_manager::getWaitListView(
       numWaitEvents + (additionalWaitEvent != nullptr ? 1 : 0);
   waitList.resize(totalNumWaitEvents);
   for (uint32_t i = 0; i < numWaitEvents; i++) {
+    phWaitEvents[i]->runBatch();
     waitList[i] = phWaitEvents[i]->getZeEvent();
   }
   if (additionalWaitEvent != nullptr) {

unified-runtime/source/adapters/level_zero/v2/command_list_manager.hpp

@@ -231,6 +231,10 @@ struct ur_command_list_manager {
                                const ur_event_handle_t *phEventWaitList,
                                ur_event_handle_t phEvent);
 
+  v2::raii::command_list_unique_handle &&releaseCommandList();
+
+  void replaceCommandList(v2::raii::command_list_unique_handle &&cmdlist);
+
 private:
   ur_result_t appendGenericCommandListsExp(
       uint32_t numCommandLists, ze_command_list_handle_t *phCommandLists,

unified-runtime/source/adapters/level_zero/v2/event.cpp

@@ -19,6 +19,8 @@
 
 #include "../ur_interface_loader.hpp"
 
+ur_event_generation_t unbatchedQueue = 1;
+
 static uint64_t adjustEndEventTimestamp(uint64_t adjustedStartTimestamp,
                                         uint64_t endTimestamp,
                                         uint64_t timestampMaxValue,
@@ -123,10 +125,20 @@ void ur_event_handle_t_::setQueue(ur_queue_t_ *hQueue) {
   profilingData.reset();
 }
 
+void ur_event_handle_t_::setBatch(ur_event_generation_t batch_generation) {
+  this->batchGeneration = batch_generation;
+}
+
 void ur_event_handle_t_::setCommandType(ur_command_t commandType) {
   this->commandType = commandType;
 }
 
+void ur_event_handle_t_::runBatch() {
+  if (batchGeneration != unbatchedQueue) {
+    hQueue->runBatchIfActive(batchGeneration);
+  }
+}
+
 void ur_event_handle_t_::recordStartTimestamp() {
   // queue and device must be set before calling this
   assert(hQueue);
@@ -192,6 +204,10 @@ ur_event_handle_t_::getEventEndTimestampAndHandle() {
 
 ur_queue_t_ *ur_event_handle_t_::getQueue() const { return hQueue; }
 
+ur_event_generation_t ur_event_handle_t_::getBatch() const {
+  return batchGeneration;
+}
+
 ur_context_handle_t ur_event_handle_t_::getContext() const { return hContext; }
 
 ur_command_t ur_event_handle_t_::getCommandType() const { return commandType; }

unified-runtime/source/adapters/level_zero/v2/event.hpp

@@ -20,6 +20,9 @@
 #include "common/ur_ref_count.hpp"
 #include "event_provider.hpp"
 
+using ur_event_generation_t = int64_t;
+extern ur_event_generation_t unbatchedQueue;
+
 namespace v2 {
 class event_pool;
 }
@@ -68,7 +71,9 @@ struct ur_event_handle_t_ : ur_object {
 
   // Set the queue and command that this event is associated with
   void setQueue(ur_queue_t_ *hQueue);
+  void setBatch(ur_event_generation_t batch_generation);
   void setCommandType(ur_command_t commandType);
+  void runBatch();
 
   void reset();
   ze_event_handle_t getZeEvent() const;
@@ -98,6 +103,8 @@ struct ur_event_handle_t_ : ur_object {
   // Get the type of the command that this event is associated with
   ur_command_t getCommandType() const;
 
+  ur_event_generation_t getBatch() const;
+
   // Get the device associated with this event
   ur_device_handle_t getDevice() const;
 
@@ -129,6 +136,8 @@ struct ur_event_handle_t_ : ur_object {
   // queue and commandType that this event is associated with, set by enqueue
   // commands
   ur_queue_t_ *hQueue = nullptr;
+  // default for non-batched queues
+  ur_event_generation_t batchGeneration = unbatchedQueue; //-1;
   ur_command_t commandType = UR_COMMAND_FORCE_UINT32;
   ur_device_handle_t hDevice = nullptr;
 

unified-runtime/source/adapters/level_zero/v2/event_pool.cpp

@@ -50,6 +50,7 @@ void event_pool::free(ur_event_handle_t event) {
 
   event->reset();
   freelist.push_back(event);
+  event->setBatch(unbatchedQueue); //-1);
 
   // The event is still in the pool, so we need to increment the refcount
   assert(event->RefCount.getCount() == 0);