WIP

Author: mateuszpn

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

@@ -8,11 +8,11 @@
 //
 //===----------------------------------------------------------------------===//
 
-#include "command_list_manager.hpp"
 #include "../helpers/kernel_helpers.hpp"
 #include "../helpers/memory_helpers.hpp"
 #include "../ur_interface_loader.hpp"
 #include "command_buffer.hpp"
+#include "command_list_manager.hpp"
 #include "context.hpp"
 #include "kernel.hpp"
 #include "memory.hpp"
@@ -37,7 +37,6 @@ ur_result_t ur_command_list_manager::appendGenericFillUnlocked(
       size, zeCommandList.get(), waitListView));
 
   // Store pattern in event if this is an async operation
-  // This prevents use-after-return bug when pPattern points to stack memory
   const void *patternPtr = pPattern;
   if (phEvent && zeSignalEvent) {
     phEvent->retainFillPattern(pPattern, patternSize);
@@ -46,8 +45,6 @@ ur_result_t ur_command_list_manager::appendGenericFillUnlocked(
 
   // PatternSize must be a power of two for zeCommandListAppendMemoryFill.
   // When it's not, the fill is emulated with zeCommandListAppendMemoryCopy.
-  // WORKAROUND: Level Zero driver rejects zeCommandListAppendMemoryFill when
-  // patternSize == size, returning ZE_RESULT_ERROR_INVALID_SIZE (0x78000008).
   if (isPowerOf2(patternSize) && patternSize != size) {
     ZE2UR_CALL(zeCommandListAppendMemoryFill,
                (zeCommandList.get(), pDst, patternPtr, patternSize, size,

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

@@ -141,14 +141,15 @@ uint64_t ur_event_handle_t_::getEventEndTimestamp() {
   return profilingData.getEventEndTimestamp();
 }
 
-void ur_event_handle_t_::retainFillPattern(const void *pPattern, size_t patternSize) {
+void ur_event_handle_t_::retainFillPattern(const void *pPattern,
+                                           size_t patternSize) {
   if (!fillPattern) {
     fillPattern.emplace();
   }
-  
+
   auto &storage = fillPattern.value();
   storage.size = patternSize;
-  
+
   // Small buffer optimization: use inline buffer for patterns <= 16 bytes
   if (patternSize <= FillPatternStorage::INLINE_SIZE) {
     std::memcpy(storage.inlineBuffer.data(), pPattern, patternSize);

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

@@ -154,8 +154,8 @@ struct ur_event_handle_t_ : ur_object {
     std::vector<uint8_t> heapBuffer;
     size_t size = 0;
     bool useHeap = false;
-    
-    const void* data() const {
+
+    const void *data() const {
       return useHeap ? heapBuffer.data() : inlineBuffer.data();
     }
   };

unified-runtime/source/loader/layers/sanitizer/sanitizer_common/sanitizer_utils.hpp

@@ -75,28 +75,8 @@ ur_result_t EnqueueUSMSet(ur_queue_handle_t Queue, void *Ptr, T Value,
                           const ur_event_handle_t *EventWaitList = nullptr,
                           ur_event_handle_t *OutEvent = nullptr) {
   assert(Size % sizeof(T) == 0);
-  ur_event_handle_t LocalEvent = nullptr;
-  ur_event_handle_t *EventStorage = OutEvent ? OutEvent : &LocalEvent;
-
-  auto Result = getContext()->urDdiTable.Enqueue.pfnUSMFill(
-      Queue, Ptr, sizeof(T), &Value, Size, NumEvents, EventWaitList,
-      EventStorage);
-  if (Result != UR_RESULT_SUCCESS) {
-    return Result;
-  }
-
-  if (!OutEvent && LocalEvent) {
-    auto WaitResult = getContext()->urDdiTable.Event.pfnWait(1, &LocalEvent);
-    auto ReleaseResult = getContext()->urDdiTable.Event.pfnRelease(LocalEvent);
-    if (WaitResult != UR_RESULT_SUCCESS) {
-      return WaitResult;
-    }
-    if (ReleaseResult != UR_RESULT_SUCCESS) {
-      return ReleaseResult;
-    }
-  }
-
-  return Result;
+  return getContext()->urDdiTable.Enqueue.pfnUSMFill(
+      Queue, Ptr, sizeof(T), &Value, Size, NumEvents, EventWaitList, OutEvent);
 }
 
 void PrintUrBuildLogIfError(ur_result_t Result, ur_program_handle_t Program,