Merge pull request #1929 from oneapi-src/revert-1880-l0-native-enqueue

Revert "[L0] L0 impl for enqueue native command"

Author: omarahmed1111

source/adapters/level_zero/device.cpp

@@ -877,7 +877,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(
   }
   case UR_DEVICE_INFO_ENQUEUE_NATIVE_COMMAND_SUPPORT_EXP: {
     // L0 doesn't support enqueueing native work through the urNativeEnqueueExp
-    return ReturnValue(static_cast<ur_bool_t>(true));
+    return ReturnValue(static_cast<ur_bool_t>(false));
   }
 
   case UR_DEVICE_INFO_ESIMD_SUPPORT: {

source/adapters/level_zero/enqueue_native.cpp

@@ -10,72 +10,11 @@
 
 #include <ur_api.h>
 
-#include "logger/ur_logger.hpp"
 #include "queue.hpp"
-#include "ur_level_zero.hpp"
 
 ur_result_t ur_queue_handle_legacy_t_::enqueueNativeCommandExp(
-    ur_exp_enqueue_native_command_function_t pfnNativeEnqueue, void *data,
-    uint32_t, const ur_mem_handle_t *,
-    const ur_exp_enqueue_native_command_properties_t *,
-    uint32_t NumEventsInWaitList, const ur_event_handle_t *phEventList,
-    ur_event_handle_t *phEvent) {
-  auto Queue = this;
-  std::scoped_lock<ur_shared_mutex> lock(Queue->Mutex);
-
-  bool UseCopyEngine = false;
-
-  // Please note that the following code should be run before the
-  // subsequent getAvailableCommandList() call so that there is no
-  // dead-lock from waiting unsubmitted events in an open batch.
-  // The createAndRetainUrZeEventList() has the proper side-effect
-  // of submitting batches with dependent events.
-  //
-  _ur_ze_event_list_t TmpWaitList;
-  UR_CALL(TmpWaitList.createAndRetainUrZeEventList(
-      NumEventsInWaitList, phEventList, Queue, UseCopyEngine));
-
-  // Get a new command list to be used on this call
-  ur_command_list_ptr_t CommandList{};
-  // TODO: Change UseCopyEngine argument to 'true' once L0 backend
-  // support is added
-  UR_CALL(Queue->Context->getAvailableCommandList(
-      Queue, CommandList, UseCopyEngine, NumEventsInWaitList, phEventList));
-
-  // TODO: do we need to create a unique command type for this?
-  ze_event_handle_t ZeEvent = nullptr;
-  ur_event_handle_t InternalEvent;
-  bool IsInternal = phEvent == nullptr;
-  ur_event_handle_t *Event = phEvent ? phEvent : &InternalEvent;
-  UR_CALL(createEventAndAssociateQueue(Queue, Event,
-                                       UR_COMMAND_ENQUEUE_NATIVE_EXP,
-                                       CommandList, IsInternal, false));
-  ZeEvent = (*Event)->ZeEvent;
-  (*Event)->WaitList = TmpWaitList;
-
-  const auto &WaitList = (*Event)->WaitList;
-  if (WaitList.Length) {
-    ZE2UR_CALL(zeCommandListAppendWaitOnEvents,
-               (CommandList->first, WaitList.Length, WaitList.ZeEventList));
-  }
-
-  UR_CALL(Queue->executeCommandList(CommandList, false, false));
-  UR_CALL(Queue->Context->getAvailableCommandList(Queue, CommandList,
-                                                  UseCopyEngine, 0, nullptr));
-
-  {
-    ScopedCommandList Active{Queue, CommandList->first};
-
-    // Call interop func which enqueues native async work
-    pfnNativeEnqueue(Queue, data);
-  }
-
-  UR_CALL(Queue->executeCommandList(CommandList, false, false));
-  UR_CALL(Queue->Context->getAvailableCommandList(Queue, CommandList,
-                                                  UseCopyEngine, 0, nullptr));
-
-  ZE2UR_CALL(zeCommandListAppendSignalEvent, (CommandList->first, ZeEvent));
-
-  UR_CALL(Queue->executeCommandList(CommandList, false));
-  return UR_RESULT_SUCCESS;
+    ur_exp_enqueue_native_command_function_t, void *, uint32_t,
+    const ur_mem_handle_t *, const ur_exp_enqueue_native_command_properties_t *,
+    uint32_t, const ur_event_handle_t *, ur_event_handle_t *) {
+  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
 }

source/adapters/level_zero/queue.cpp

@@ -705,15 +705,6 @@ ur_result_t ur_queue_handle_legacy_t_::queueGetNativeHandle(
 ) {
   auto Queue = this;
 
-  // Needed for EnqueueNativeCommandExp, so that the native queue 'got' in the
-  // interop func is the as the native queue used to manage dependencies
-  // before the interop func invocation
-  if (Queue->getThreadLocalCommandList() != ze_command_list_handle_t{0}) {
-    auto ZeCmdList = ur_cast<ze_command_list_handle_t *>(NativeQueue);
-    *ZeCmdList = Queue->getThreadLocalCommandList();
-    return UR_RESULT_SUCCESS;
-  }
-
   // Lock automatically releases when this goes out of scope.
   std::shared_lock<ur_shared_mutex> lock(Queue->Mutex);
 

source/adapters/level_zero/queue.hpp

@@ -423,12 +423,6 @@ struct ur_queue_handle_legacy_t_ : _ur_object, public ur_queue_handle_t_ {
                           uint32_t, const ur_event_handle_t *,
                           ur_event_handle_t *) override;
 
-  // Thread local stream will be used if ScopedStream is active
-  static ze_command_list_handle_t &getThreadLocalCommandList() {
-    static thread_local ze_command_list_handle_t CommandList{0};
-    return CommandList;
-  }
-
   using queue_type = ur_device_handle_t_::queue_group_info_t::type;
   // PI queue is in general a one to many mapping to L0 native queues.
   struct ur_queue_group_t {
@@ -947,23 +941,3 @@ ur_result_t setSignalEvent(ur_queue_handle_legacy_t Queue, bool UseCopyEngine,
 ur_result_t CleanupEventListFromResetCmdList(
     std::vector<ur_event_handle_t> &EventListToCleanup,
     bool QueueLocked = false);
-
-// RAII object to make hQueue command list getter methods all return the same
-// command list within the lifetime of this object.
-//
-// This is useful for urEnqueueNativeCommandExp where we want guarantees that
-// the user submitted native calls will be dispatched to a known command list,
-// which must be "got" within the user submitted fuction.
-class ScopedCommandList {
-  ur_queue_handle_legacy_t hQueue;
-
-public:
-  ScopedCommandList(ur_queue_handle_legacy_t hQueue,
-                    ze_command_list_handle_t CommandList)
-      : hQueue{hQueue} {
-    hQueue->getThreadLocalCommandList() = CommandList;
-  }
-  ~ScopedCommandList() {
-    hQueue->getThreadLocalCommandList() = ze_command_list_handle_t{0};
-  }
-};

test/conformance/exp_enqueue_native/CMakeLists.txt

@@ -15,21 +15,4 @@ if (UR_BUILD_ADAPTER_CUDA)
   target_link_libraries(test-exp_enqueue_native PRIVATE cudadrv)
 endif()
 
-if (UR_BUILD_ADAPTER_L0)
-  add_conformance_test_with_kernels_environment(
-    exp_enqueue_native
-    enqueue_native_level_zero.cpp
-  )
-  target_link_libraries(test-exp_enqueue_native PRIVATE
-      LevelZeroLoader
-      LevelZeroLoader-Headers
-  )
-
-  target_include_directories(test-exp_enqueue_native PRIVATE
-      ${PROJECT_SOURCE_DIR}/source
-      ${PROJECT_SOURCE_DIR}/source/adapters/level_zero
-      LevelZeroLoader-Headers
-  )
-endif()
-
 # TODO: Add more tests for different triples