CUTLASS 2.1 (#83)

CUTLASS 2.1 contributes:
- BLAS-style host-side API added to CUTLASS Library
- Planar Complex GEMM kernels targeting Volta and Turing Tensor Cores
- Minor enhancements and bug fixes

include/cutlass/gemm/threadblock/mma_pipelined.h

@@ -118,6 +118,15 @@ public:
   /// Warp-level Mma
   using Operator = typename Policy::Operator;
 
+  /// Obtain the arch tag from the warp-level operator
+  using ArchTag = typename Policy::Operator::ArchTag;
+
+  /// Complex transform on A operand
+  static ComplexTransform const kTransformA = Operator::kTransformA;
+
+  /// Complex transform on B operand
+  static ComplexTransform const kTransformB = Operator::kTransformB;
+
   // staticaly assert kStages for MmaPipelined is two (Double-buffered pipeline)
   static_assert((Base::kStages==2), "MmaPipelined requires kStages set to value 2");
 
@@ -256,8 +265,8 @@ public:
 
           __syncthreads();
           
-          ++this->smem_iterator_B_;
           ++this->smem_iterator_A_;
+          ++this->smem_iterator_B_;
 
           // Add negative offsets to return iterators to the 'start' of the circular buffer in shared memory
           if (smem_write_stage_idx == 1) {
@@ -299,7 +308,8 @@ public:
           }
         }
 
-        warp_mma(accum, warp_frag_A[warp_mma_k % 2], warp_frag_B[warp_mma_k % 2], accum);
+        warp_mma(accum, warp_frag_A[warp_mma_k % 2],
+                 warp_frag_B[warp_mma_k % 2], accum);
       }
     }
 
@@ -311,3 +321,5 @@ public:
 } // namespace threadblock
 } // namespace gemm
 } // namespace cutlass
+
+/////////////////////////////////////////////////////////////////////////////////////////////////