Merge branch 'main' into memleak-fix

Author: lslusarczyk

cmake/helpers.cmake

@@ -70,6 +70,7 @@ function(add_ur_target_compile_options name)
         )
         if (CMAKE_BUILD_TYPE STREQUAL "Release")
             target_compile_definitions(${name} PRIVATE -D_FORTIFY_SOURCE=2)
+            target_compile_options(${name} PRIVATE -fvisibility=hidden)
         endif()
         if(UR_DEVELOPER_MODE)
             target_compile_options(${name} PRIVATE

examples/collector/CMakeLists.txt

@@ -17,6 +17,6 @@ target_link_libraries(${TARGET_NAME} PRIVATE ${TARGET_XPTI})
 target_include_directories(${TARGET_NAME} PRIVATE ${xpti_SOURCE_DIR}/include)
 
 if(MSVC)
-    target_compile_definitions(${TARGET_NAME} PRIVATE
-        XPTI_STATIC_LIBRARY XPTI_CALLBACK_API_EXPORTS)
+    target_compile_definitions(${TARGET_NAME} PRIVATE XPTI_STATIC_LIBRARY)
 endif()
+target_compile_definitions(${TARGET_NAME} PRIVATE XPTI_CALLBACK_API_EXPORTS)

include/ur_api.h

@@ -332,9 +332,17 @@ typedef enum ur_structure_type_t {
 #if defined(_WIN32)
 /// @brief Microsoft-specific dllexport storage-class attribute
 #define UR_APIEXPORT __declspec(dllexport)
+#endif // defined(_WIN32)
+#endif // UR_APIEXPORT
+
+///////////////////////////////////////////////////////////////////////////////
+#ifndef UR_APIEXPORT
+#if __GNUC__ >= 4
+/// @brief GCC-specific dllexport storage-class attribute
+#define UR_APIEXPORT __attribute__((visibility("default")))
 #else
 #define UR_APIEXPORT
-#endif // defined(_WIN32)
+#endif // __GNUC__ >= 4
 #endif // UR_APIEXPORT
 
 ///////////////////////////////////////////////////////////////////////////////

scripts/core/common.yml

@@ -39,6 +39,12 @@ desc: "Microsoft-specific dllexport storage-class attribute"
 condition: "defined(_WIN32)"
 name: $X_APIEXPORT
 value: __declspec(dllexport)
+--- #--------------------------------------------------------------------------
+type: macro
+desc: "GCC-specific dllexport storage-class attribute"
+condition: "__GNUC__ >= 4"
+name: $X_APIEXPORT
+value: __attribute__ ((visibility ("default")))
 altvalue: ""
 --- #--------------------------------------------------------------------------
 type: macro

source/adapters/level_zero/context.cpp

@@ -530,6 +530,8 @@ ur_result_t ur_context_handle_t_::getFreeSlotInExistingOrNewPool(
         counterBasedExt.flags =
             ZE_EVENT_POOL_COUNTER_BASED_EXP_FLAG_NON_IMMEDIATE;
       }
+      logger::debug("ze_event_pool_desc_t counter based flags set to: {}",
+                    counterBasedExt.flags);
       ZeEventPoolDesc.pNext = &counterBasedExt;
     }
 

source/adapters/level_zero/event.cpp

@@ -196,7 +196,9 @@ ur_result_t urEnqueueEventsWaitWithBarrier(
         //
         if (Queue->isInOrderQueue() && InOrderBarrierBySignal &&
             !Queue->isProfilingEnabled()) {
-          if (EventWaitList.Length) {
+          // If we are using driver in order lists, then append wait on events
+          // is unnecessary and we can signal the event created.
+          if (EventWaitList.Length && !CmdList->second.IsInOrderList) {
             ZE2UR_CALL(zeCommandListAppendWaitOnEvents,
                        (CmdList->first, EventWaitList.Length,
                         EventWaitList.ZeEventList));
@@ -1546,8 +1548,13 @@ ur_result_t _ur_ze_event_list_t::createAndRetainUrZeEventList(
 
           ZE2UR_CALL(zeCommandListAppendWaitOnEvents,
                      (ZeCommandList, 1u, &EventList[I]->ZeEvent));
-          if (!MultiDeviceEvent->CounterBasedEventsEnabled)
+          if (!MultiDeviceEvent->CounterBasedEventsEnabled) {
             ZE2UR_CALL(zeEventHostSignal, (MultiDeviceZeEvent));
+          } else {
+            ZE2UR_CALL(zeCommandListAppendSignalEvent,
+                       (ZeCommandList, MultiDeviceZeEvent));
+          }
+          MultiDeviceEvent->Completed = true;
 
           UR_CALL(Queue->executeCommandList(CommandList, /* IsBlocking */ false,
                                             /* OkToBatchCommand */ true));

source/adapters/level_zero/kernel.cpp

@@ -494,11 +494,19 @@ ur_result_t urEnqueueDeviceGlobalVariableWrite(
 ) {
   std::scoped_lock<ur_shared_mutex> lock(Queue->Mutex);
 
+  ze_module_handle_t ZeModule{};
+  auto It = Program->ZeModuleMap.find(Queue->Device->ZeDevice);
+  if (It != Program->ZeModuleMap.end()) {
+    ZeModule = It->second;
+  } else {
+    ZeModule = Program->ZeModule;
+  }
+
   // Find global variable pointer
   size_t GlobalVarSize = 0;
   void *GlobalVarPtr = nullptr;
   ZE2UR_CALL(zeModuleGetGlobalPointer,
-             (Program->ZeModule, Name, &GlobalVarSize, &GlobalVarPtr));
+             (ZeModule, Name, &GlobalVarSize, &GlobalVarPtr));
   if (GlobalVarSize < Offset + Count) {
     setErrorMessage("Write device global variable is out of range.",
                     UR_RESULT_ERROR_INVALID_VALUE,

source/adapters/level_zero/queue.cpp

@@ -2386,6 +2386,7 @@ ur_command_list_ptr_t &ur_queue_handle_t_::ur_queue_group_t::getImmCmdList() {
   ZeCommandQueueDesc.ordinal = QueueOrdinal;
   ZeCommandQueueDesc.index = QueueIndex;
   ZeCommandQueueDesc.mode = ZE_COMMAND_QUEUE_MODE_ASYNCHRONOUS;
+  bool isInOrderList = false;
   const char *Priority = "Normal";
   if (Queue->isPriorityLow()) {
     ZeCommandQueueDesc.priority = ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_LOW;
@@ -2401,6 +2402,7 @@ ur_command_list_ptr_t &ur_queue_handle_t_::ur_queue_group_t::getImmCmdList() {
   }
 
   if (Queue->Device->useDriverInOrderLists() && Queue->isInOrderQueue()) {
+    isInOrderList = true;
     ZeCommandQueueDesc.flags |= ZE_COMMAND_QUEUE_FLAG_IN_ORDER;
   }
 
@@ -2449,7 +2451,7 @@ ur_command_list_ptr_t &ur_queue_handle_t_::ur_queue_group_t::getImmCmdList() {
               ZeCommandList,
               ur_command_list_info_t(
                   nullptr, true, false, nullptr, ZeCommandQueueDesc,
-                  Queue->useCompletionBatching(), true, false, true)})
+                  Queue->useCompletionBatching(), true, isInOrderList, true)})
           .first;
 
   return ImmCmdLists[Index];

source/loader/ur_loader.cpp

@@ -17,6 +17,18 @@ namespace ur_loader {
 context_t *getContext() { return context_t::get_direct(); }
 
 ur_result_t context_t::init() {
+#ifdef _WIN32
+    // Suppress system errors.
+    // Tells the system to not display the critical-error-handler message box.
+    // Instead, the system sends the error to the calling process.
+    // This is crucial for graceful handling of adapters that couldn't be
+    // loaded, e.g. due to missing native run-times.
+    // TODO: add reporting in case of an error.
+    // NOTE: we restore the old mode to not affect user app behavior.
+    // See https://github.com/intel/llvm/blob/sycl/sycl/ur_win_proxy_loader/ur_win_proxy_loader.cpp (preloadLibraries())
+    UINT SavedMode = SetErrorMode(SEM_FAILCRITICALERRORS);
+#endif
+
 #ifdef UR_STATIC_ADAPTER_LEVEL_ZERO
     // If the adapters were force loaded, it means the user wants to use
     // a specific adapter library. Don't load any static adapters.
@@ -35,6 +47,10 @@ ur_result_t context_t::init() {
             }
         }
     }
+#ifdef _WIN32
+    // Restore system error handling.
+    (void)SetErrorMode(SavedMode);
+#endif
 
     forceIntercept = getenv_tobool("UR_ENABLE_LOADER_INTERCEPT");
 

test/layers/tracing/CMakeLists.txt

@@ -15,9 +15,9 @@ target_link_libraries(test_collector PRIVATE ${TARGET_XPTI})
 target_include_directories(test_collector PRIVATE ${xpti_SOURCE_DIR}/include)
 
 if(MSVC)
-    target_compile_definitions(test_collector PRIVATE
-        XPTI_STATIC_LIBRARY XPTI_CALLBACK_API_EXPORTS)
+    target_compile_definitions(test_collector PRIVATE XPTI_STATIC_LIBRARY)
 endif()
+target_compile_definitions(test_collector PRIVATE XPTI_CALLBACK_API_EXPORTS)
 
 function(set_tracing_test_props target_name collector_name)
     set_tests_properties(${target_name} PROPERTIES