[EXP][Command-Buffer] Optimize L0 command buffer submission

- Adds command buffer property to explicitly enable profiling
- Add ability to enforce use of in-order command lists

include/ur_api.h

@@ -7892,6 +7892,9 @@ typedef struct ur_exp_command_buffer_desc_t {
                                ///< ::UR_STRUCTURE_TYPE_EXP_COMMAND_BUFFER_DESC
     const void *pNext;         ///< [in][optional] pointer to extension-specific structure
     ur_bool_t isUpdatable;     ///< [in] Commands in a finalized command-buffer can be updated.
+    ur_bool_t isInOrder;       ///< [in] Commands in a command-buffer may be executed in-order without
+                               ///< explicit dependencies.
+    ur_bool_t enableProfiling; ///< [in] Command-buffer profiling is enabled.
 
 } ur_exp_command_buffer_desc_t;
 

include/ur_print.hpp

@@ -9401,6 +9401,16 @@ inline std::ostream &operator<<(std::ostream &os, const struct ur_exp_command_bu
 
     os << (params.isUpdatable);
 
+    os << ", ";
+    os << ".isInOrder = ";
+
+    os << (params.isInOrder);
+
+    os << ", ";
+    os << ".enableProfiling = ";
+
+    os << (params.enableProfiling);
+
     os << "}";
     return os;
 }

scripts/core/EXP-COMMAND-BUFFER.rst

@@ -77,9 +77,12 @@ Command-Buffer Creation
 Command-Buffers are tied to a specific ${x}_context_handle_t and
 ${x}_device_handle_t. ${x}CommandBufferCreateExp optionally takes a descriptor
 to provide additional properties for how the command-buffer should be
-constructed. The only unique member defined in ${x}_exp_command_buffer_desc_t
-is ``isUpdatable``, which should be set to ``true`` to support :ref:`updating
+constructed. The members defined in ${x}_exp_command_buffer_desc_t are:
+* ``isUpdatable``, which should be set to ``true`` to support :ref:`updating
 command-buffer commands`.
+* ``isInOrder``, which should be set to ``true`` to enable the graph workload
+  to be submitted to an in-order command-list.
+* ``enableProfiling``, which should be set to ``true`` to enable graph profling.
 
 Command-buffers are reference counted and can be retained and released by
 calling ${x}CommandBufferRetainExp and ${x}CommandBufferReleaseExp respectively.

scripts/core/exp-command-buffer.yml

@@ -113,6 +113,12 @@ members:
     - type: $x_bool_t
       name: isUpdatable
       desc: "[in] Commands in a finalized command-buffer can be updated."
+    - type: $x_bool_t
+      name: isInOrder
+      desc: "[in] Commands in a command-buffer may be executed in-order without explicit dependencies."
+    - type: $x_bool_t
+      name: enableProfiling
+      desc: "[in] Command-buffer profiling is enabled."
 --- #--------------------------------------------------------------------------
 type: struct
 desc: "Descriptor type for updating a kernel command memobj argument."

source/adapters/level_zero/command_buffer.hpp

@@ -30,7 +30,7 @@ struct ur_exp_command_buffer_handle_t_ : public _ur_object {
       ze_command_list_handle_t CommandList,
       ze_command_list_handle_t CommandListResetEvents,
       ZeStruct<ze_command_list_desc_t> ZeDesc,
-      const ur_exp_command_buffer_desc_t *Desc);
+      const ur_exp_command_buffer_desc_t *Desc, const bool IsInOrderCmdList);
 
   ~ur_exp_command_buffer_handle_t_();
 
@@ -82,6 +82,14 @@ struct ur_exp_command_buffer_handle_t_ : public _ur_object {
   bool IsUpdatable = false;
   // Indicates if command buffer was finalized.
   bool IsFinalized = false;
+  // Command-buffer profiling is enabled.
+  bool IsProfilingEnabled = false;
+  // Command-buffer can be submitted to an in-order command-list.
+  bool IsInOrderCmdList;
+  // List of kernels.
+  // This list is needed to release all kernels retained by the
+  // command_buffer.
+  std::vector<ur_kernel_handle_t> KernelsList;
 };
 
 struct ur_exp_command_buffer_command_handle_t_ : public _ur_object {

source/adapters/level_zero/event.cpp

@@ -760,7 +760,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urExtEventCreate(
     ur_event_handle_t
         *Event ///< [out] pointer to the handle of the event object created.
 ) {
-  UR_CALL(EventCreate(Context, nullptr, false, true, Event));
+  UR_CALL(EventCreate(Context, nullptr, false, true, false, Event));
 
   (*Event)->RefCountExternal++;
   ZE2UR_CALL(zeEventHostSignal, ((*Event)->ZeEvent));
@@ -778,7 +778,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventCreateWithNativeHandle(
   // we dont have urEventCreate, so use this check for now to know that
   // the call comes from urEventCreate()
   if (NativeEvent == nullptr) {
-    UR_CALL(EventCreate(Context, nullptr, false, true, Event));
+    UR_CALL(EventCreate(Context, nullptr, false, true, false, Event));
 
     (*Event)->RefCountExternal++;
     ZE2UR_CALL(zeEventHostSignal, ((*Event)->ZeEvent));
@@ -1057,9 +1057,11 @@ ur_result_t CleanupCompletedEvent(ur_event_handle_t Event, bool QueueLocked,
 //
 ur_result_t EventCreate(ur_context_handle_t Context, ur_queue_handle_t Queue,
                         bool IsMultiDevice, bool HostVisible,
+                        bool ForceDisableProfiling,
                         ur_event_handle_t *RetEvent) {
 
-  bool ProfilingEnabled = !Queue || Queue->isProfilingEnabled();
+  bool ProfilingEnabled =
+      ForceDisableProfiling ? false : (!Queue || Queue->isProfilingEnabled());
 
   ur_device_handle_t Device = nullptr;
 

source/adapters/level_zero/event.hpp

@@ -31,6 +31,7 @@ extern "C" {
 ur_result_t urEventReleaseInternal(ur_event_handle_t Event);
 ur_result_t EventCreate(ur_context_handle_t Context, ur_queue_handle_t Queue,
                         bool IsMultiDevice, bool HostVisible,
+                        bool ForceDisableProfiling,
                         ur_event_handle_t *RetEvent);
 } // extern "C"
 

source/adapters/level_zero/queue.cpp

@@ -1519,7 +1519,7 @@ ur_result_t createEventAndAssociateQueue(ur_queue_handle_t Queue,
 
   if (*Event == nullptr)
     UR_CALL(EventCreate(Queue->Context, Queue, IsMultiDevice,
-                        HostVisible.value(), Event));
+                        HostVisible.value(), false, Event));
 
   (*Event)->UrQueue = Queue;
   (*Event)->CommandType = CommandType;