comment

NVIDIA · Sep 23, 2024 · 3bb1b95 · 3bb1b95
1 parent c319e5c
commit 3bb1b95
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/csrc/device_lower/pass/circular_buffer.cpp b/csrc/device_lower/pass/circular_buffer.cpp
@@ -582,8 +582,8 @@ class TmaCircularBufferLoopCloner : public CircularBufferLoopCloner {
 
     LoadStoreOp* ldst = expr->as<LoadStoreOp>();
 
-    // There should be a single mbarrier_arrive_tx_ for all ldst in current
-    // stage.
+    // There should be a single mbarrier_arrive_tx_ for each cpAsynBulk load
+    // expression.
     NVF_ERROR(mbarrier_arrive_tx_ == nullptr);
     mbarrier_arrive_tx_ =
         createMbarrierArriveExpectTx(ldst, current_load_stage_);