[SYCL][L0] Report more informative error at command execution failure

sycl/include/CL/sycl/detail/pi.h

@@ -115,6 +115,9 @@ typedef enum {
   PI_IMAGE_FORMAT_NOT_SUPPORTED = CL_IMAGE_FORMAT_NOT_SUPPORTED,
   PI_MEM_OBJECT_ALLOCATION_FAILURE = CL_MEM_OBJECT_ALLOCATION_FAILURE,
   PI_LINK_PROGRAM_FAILURE = CL_LINK_PROGRAM_FAILURE,
+  PI_COMMAND_EXECUTION_FAILURE =
+      -997, ///< PI_COMMAND_EXECUTION_FAILURE indicates an error occurred
+            ///< during command enqueue or execution.
   PI_FUNCTION_ADDRESS_IS_NOT_AVAILABLE =
       -998, ///< PI_FUNCTION_ADDRESS_IS_NOT_AVAILABLE indicates a fallback
             ///< method determines the function exists but its address cannot be

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -1376,8 +1376,17 @@ pi_result _pi_queue::executeCommandList(pi_command_list_ptr_t CommandList,
   // Offload command list to the GPU for asynchronous execution
   auto ZeCommandList = CommandList->first;
   zePrint("Calling zeCommandQueueExecuteCommandLists with Index = %d\n", Index);
-  ZE_CALL(zeCommandQueueExecuteCommandLists,
-          (ZeCommandQueue, 1, &ZeCommandList, CommandList->second.ZeFence));
+  auto ZeResult = ZE_CALL_NOCHECK(
+      zeCommandQueueExecuteCommandLists,
+      (ZeCommandQueue, 1, &ZeCommandList, CommandList->second.ZeFence));
+  if (ZeResult != ZE_RESULT_SUCCESS) {
+    this->Healthy = false;
+    if (ZeResult == ZE_RESULT_ERROR_UNKNOWN) {
+      // Turn into a more informative end-user error.
+      return PI_COMMAND_EXECUTION_FAILURE;
+    }
+    return mapError(ZeResult);
+  }
 
   // Check global control to make every command blocking for debugging.
   if (IsBlocking || (ZeSerialize & ZeSerializeBlock) != 0) {
@@ -3176,10 +3185,13 @@ pi_result piQueueRelease(pi_queue Queue) {
         return Res;
 
       // Make sure all commands get executed.
-      ZE_CALL(zeHostSynchronize, (Queue->ZeComputeCommandQueue));
-      for (uint32_t i = 0; i < Queue->ZeCopyCommandQueues.size(); ++i) {
-        if (Queue->ZeCopyCommandQueues[i])
-          ZE_CALL(zeHostSynchronize, (Queue->ZeCopyCommandQueues[i]));
+      // Only do so for a healthy queue as otherwise sync may not be valid.
+      if (Queue->Healthy) {
+        ZE_CALL(zeHostSynchronize, (Queue->ZeComputeCommandQueue));
+        for (uint32_t i = 0; i < Queue->ZeCopyCommandQueues.size(); ++i) {
+          if (Queue->ZeCopyCommandQueues[i])
+            ZE_CALL(zeHostSynchronize, (Queue->ZeCopyCommandQueues[i]));
+        }
       }
 
       // Destroy all the fences created associated with this queue.

sycl/plugins/level_zero/pi_level_zero.hpp

@@ -803,6 +803,9 @@ struct _pi_queue : _pi_object {
   // externally, and can wait for internal references to complete, and do proper
   // cleanup of the queue.
   std::atomic<pi_uint32> RefCountExternal{1};
+
+  // Indicates that the queue is healthy and all operations on it are OK.
+  bool Healthy{true};
 };
 
 struct _pi_mem : _pi_object {