CUTLASS 2.6.1 - functional and performance enhancements to strided DGRAD, fixes, and tuning

* cutlass 2.6 update

* remove debug prints

* cutlass 2.6.1 (minor update)

* Updated CHANGELOG.

* Minor edit to readme to indicate patch version.

* Minor edit to readme.

Co-authored-by:  Haicheng Wu <haichengw@nvidia.com>, Andrew Kerr <akerr@nvidia.com>

include/cutlass/gemm/kernel/default_gemm_with_k_reduction.h

@@ -103,8 +103,8 @@ template <
     int Stages,
     /// Operation performed by GEMM
     typename Operator,
-    /// Use zfill or predicate for SM80 out-of-bound cp.async
-    bool UseZfill = false,
+    /// Use zfill or predicate for out-of-bound cp.async
+    SharedMemoryClearOption SharedMemoryClear = SharedMemoryClearOption::kNone,
     ///
     typename Enable = void>
 struct DefaultGemmWithKReduction {
@@ -116,7 +116,7 @@ struct DefaultGemmWithKReduction {
       ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB,
       ElementAccumulator, layout::RowMajor, arch::OpClassTensorOp, kReduceKForA, arch::Sm80,
       ThreadblockShape, WarpShape, InstructionShape, Stages,
-      Operator, false, UseZfill>::ThreadblockMma;
+      Operator, false, SharedMemoryClear>::ThreadblockMma;
 
   static const int kPartitionsK = ThreadblockShape::kK / WarpShape::kK;
 

include/cutlass/gemm/kernel/gemm.h

@@ -34,6 +34,7 @@
 #include "cutlass/gemm/gemm.h"
 #include "cutlass/matrix_coord.h"
 #include "cutlass/semaphore.h"
+#include "cutlass/arch/arch.h"
 
 /////////////////////////////////////////////////////////////////////////////////////////////////