move utilty functions to mma_utils

NVIDIA · Dec 10, 2024 · 20a4b5e · 20a4b5e
1 parent 4b1c5bb
commit 20a4b5e
Show file tree

Hide file tree

Showing 7 changed files with 357 additions and 639 deletions.
diff --git a/csrc/scheduler/hopper_multi_matmul.cpp b/csrc/scheduler/hopper_multi_matmul.cpp
diff --git a/csrc/scheduler/hopper_multi_matmul.h b/csrc/scheduler/hopper_multi_matmul.h
@@ -178,20 +178,6 @@ class HopperMultipleMatmulScheduler : public MultipleMatmulScheduler {
 
   void setUpCircularBuffering();
 
-  //! Schedules the copy operation of output of a Mma op which resided in the
-  //! registers to shared memory.
-  void scheduleStMatrixForMmaOutput(
-      TensorView* tv,
-      MmaInputSmemSwizzle swizzle,
-      int64_t tile_m,
-      int64_t tile_n);
-
-  //! Schedules the copy operation of output of a Mma op which resided in the
-  //! shared memory to global memory.
-  void scheduleTMAStoreForMmaOutput(
-      TensorView* tv,
-      MmaInputSmemSwizzle swizzle);
-
   // Map TensorView's iterDomain to its ValGroup.
   // Then, find the MatmulDimRole for the ValGroup.
   // Return MatmulDimRole for IterDomain