Merge pull request oneapi-src#1802 from nrspruit/fix_immediate_cmdlis…

…t_reuse

[L0] Fix immediate command list use in Command Queues

scripts/templates/ldrddi.cpp.mako

@@ -56,7 +56,7 @@ namespace ur_loader
         size_t adapterIndex = 0;
         if( nullptr != ${obj['params'][1]['name']} && ${obj['params'][0]['name']} !=0)
         {
-            for( auto& platform : context->platforms )
+            for( auto& platform : getContext()->platforms )
             {
                 if(platform.initStatus != ${X}_RESULT_SUCCESS)
                     continue;
@@ -81,7 +81,7 @@ namespace ur_loader
 
         if( ${obj['params'][2]['name']} != nullptr )
         {
-            *${obj['params'][2]['name']} = static_cast<uint32_t>(context->platforms.size());
+            *${obj['params'][2]['name']} = static_cast<uint32_t>(getContext()->platforms.size());
         }
 
         %elif re.match(r"\w+PlatformGet$", th.make_func_name(n, tags, obj)):
@@ -360,13 +360,13 @@ ${tbl['export']['name']}(
     if( nullptr == pDdiTable )
         return ${X}_RESULT_ERROR_INVALID_NULL_POINTER;
 
-    if( ur_loader::context->version < version )
+    if( ur_loader::getContext()->version < version )
         return ${X}_RESULT_ERROR_UNSUPPORTED_VERSION;
 
     ${x}_result_t result = ${X}_RESULT_SUCCESS;
 
     // Load the device-platform DDI tables
-    for( auto& platform : ur_loader::context->platforms )
+    for( auto& platform : ur_loader::getContext()->platforms )
     {
         if(platform.initStatus != ${X}_RESULT_SUCCESS)
             continue;
@@ -379,7 +379,7 @@ ${tbl['export']['name']}(
 
     if( ${X}_RESULT_SUCCESS == result )
     {
-        if( ur_loader::context->platforms.size() != 1 || ur_loader::context->forceIntercept )
+        if( ur_loader::getContext()->platforms.size() != 1 || ur_loader::getContext()->forceIntercept )
         {
             // return pointers to loader's DDIs
             %for obj in tbl['functions']:
@@ -397,7 +397,7 @@ ${tbl['export']['name']}(
         else
         {
             // return pointers directly to platform's DDIs
-            *pDdiTable = ur_loader::context->platforms.front().dditable.${n}.${tbl['name']};
+            *pDdiTable = ur_loader::getContext()->platforms.front().dditable.${n}.${tbl['name']};
         }
     }
 

scripts/templates/libapi.cpp.mako

@@ -69,15 +69,15 @@ try {
     %endfor
 
     static ${x}_result_t result = ${X}_RESULT_SUCCESS;
-    std::call_once(${x}_lib::context->initOnce, [device_flags, hLoaderConfig]() {
-        result = ${x}_lib::context->Init(device_flags, hLoaderConfig);
+    std::call_once(${x}_lib::getContext()->initOnce, [device_flags, hLoaderConfig]() {
+        result = ${x}_lib::getContext()->Init(device_flags, hLoaderConfig);
     });
 
     return result;
 %elif th.obj_traits.is_loader_only(obj):
     return ur_lib::${th.make_func_name(n, tags, obj)}(${", ".join(th.make_param_lines(n, tags, obj, format=["name"]))} );
 %else:
-    auto ${th.make_pfn_name(n, tags, obj)} = ${x}_lib::context->${n}DdiTable.${th.get_table_name(n, tags, obj)}.${th.make_pfn_name(n, tags, obj)};
+    auto ${th.make_pfn_name(n, tags, obj)} = ${x}_lib::getContext()->${n}DdiTable.${th.get_table_name(n, tags, obj)}.${th.make_pfn_name(n, tags, obj)};
     if( nullptr == ${th.make_pfn_name(n, tags, obj)} )
         return ${X}_RESULT_ERROR_UNINITIALIZED;
 

source/adapters/level_zero/context.cpp

@@ -733,6 +733,10 @@ ur_result_t ur_context_handle_t_::getAvailableCommandList(
           !(ZeCommandListIt->second.InOrderList)) {
         continue;
       }
+      // Only allow to reuse Regular Command Lists
+      if (ZeCommandListIt->second.IsImmediate) {
+        continue;
+      }
       auto &ZeCommandList = ZeCommandListIt->first;
       auto it = Queue->CommandListMap.find(ZeCommandList);
       if (it != Queue->CommandListMap.end()) {

source/adapters/level_zero/context.hpp

@@ -30,6 +30,7 @@
 struct l0_command_list_cache_info {
   ZeStruct<ze_command_queue_desc_t> ZeQueueDesc;
   bool InOrderList = false;
+  bool IsImmediate = false;
 };
 
 struct ur_context_handle_t_ : _ur_object {

source/adapters/level_zero/queue.cpp

@@ -665,6 +665,7 @@ ur_result_t ur_queue_handle_legacy_t_::queueRelease() {
           struct l0_command_list_cache_info ListInfo;
           ListInfo.ZeQueueDesc = it->second.ZeQueueDesc;
           ListInfo.InOrderList = it->second.IsInOrderList;
+          ListInfo.IsImmediate = it->second.IsImmediate;
           ZeCommandListCache.push_back({it->first, ListInfo});
         } else {
           // A non-reusable comamnd list that came from a make_queue call is
@@ -745,7 +746,8 @@ void ur_queue_handle_legacy_t_::ur_queue_group_t::setImmCmdList(
           .insert(std::pair<ze_command_list_handle_t, ur_command_list_info_t>{
               ZeCommandList,
               ur_command_list_info_t(nullptr, true, false, nullptr, ZeQueueDesc,
-                                     queue->useCompletionBatching(), false)})
+                                     queue->useCompletionBatching(), false,
+                                     false, true)})
           .first);
 }
 
@@ -2080,6 +2082,7 @@ ur_result_t ur_queue_handle_legacy_t_::resetCommandList(
     struct l0_command_list_cache_info ListInfo;
     ListInfo.ZeQueueDesc = CommandList->second.ZeQueueDesc;
     ListInfo.InOrderList = CommandList->second.IsInOrderList;
+    ListInfo.IsImmediate = CommandList->second.IsImmediate;
     ZeCommandListCache.push_back({CommandList->first, ListInfo});
   }
 
@@ -2430,9 +2433,9 @@ ur_queue_handle_legacy_t_::ur_queue_group_t::getImmCmdList() {
       Queue->CommandListMap
           .insert(std::pair<ze_command_list_handle_t, ur_command_list_info_t>{
               ZeCommandList,
-              ur_command_list_info_t(nullptr, true, false, nullptr,
-                                     ZeCommandQueueDesc,
-                                     Queue->useCompletionBatching())})
+              ur_command_list_info_t(
+                  nullptr, true, false, nullptr, ZeCommandQueueDesc,
+                  Queue->useCompletionBatching(), true, false, true)})
           .first;
 
   return ImmCmdLists[Index];

source/adapters/level_zero/queue.hpp

@@ -168,10 +168,11 @@ struct ur_command_list_info_t {
                          bool IsClosed, ze_command_queue_handle_t ZeQueue,
                          ZeStruct<ze_command_queue_desc_t> ZeQueueDesc,
                          bool UseCompletionBatching, bool CanReuse = true,
-                         bool IsInOrderList = false)
+                         bool IsInOrderList = false, bool IsImmediate = false)
       : ZeFence(ZeFence), ZeFenceInUse(ZeFenceInUse), IsClosed(IsClosed),
         ZeQueue(ZeQueue), ZeQueueDesc(ZeQueueDesc),
-        IsInOrderList(IsInOrderList), CanReuse(CanReuse) {
+        IsInOrderList(IsInOrderList), CanReuse(CanReuse),
+        IsImmediate(IsImmediate) {
     if (UseCompletionBatching) {
       completions = ur_completion_batches();
     }
@@ -204,6 +205,7 @@ struct ur_command_list_info_t {
   // Indicates if this is an inorder list
   bool IsInOrderList;
   bool CanReuse;
+  bool IsImmediate;
 
   // Helper functions to tell if this is a copy command-list.
   bool isCopy(ur_queue_handle_legacy_t Queue) const;

source/loader/CMakeLists.txt

@@ -187,14 +187,10 @@ if(WIN32)
     target_sources(ur_loader
         PRIVATE
             ${CMAKE_CURRENT_SOURCE_DIR}/windows/adapter_search.cpp
-            ${CMAKE_CURRENT_SOURCE_DIR}/windows/lib_init.cpp
-            ${CMAKE_CURRENT_SOURCE_DIR}/windows/loader_init.cpp
     )
 else()
     target_sources(ur_loader
         PRIVATE
             ${CMAKE_CURRENT_SOURCE_DIR}/linux/adapter_search.cpp
-            ${CMAKE_CURRENT_SOURCE_DIR}/linux/lib_init.cpp
-            ${CMAKE_CURRENT_SOURCE_DIR}/linux/loader_init.cpp
     )
 endif()