Removed trivial copy constructors on parameter classes to enable devi… (#366)

* Removed trivial copy constructors on parameter classes to enable device-side launch of CUTLASS kernels

* Added SFINAE to the `TensorRef(NonConstTensorRef const&)` constructor to avoid making it a copy-constructor for device code

* std => platform

* fix affine2

* really fix affine2

Co-authored-by: Haicheng Wu <haichengw@nvidia.com>

include/cutlass/coord.h

@@ -94,14 +94,6 @@ public:
     }
   }
 
-  /// Copy constructor
-  CUTLASS_HOST_DEVICE
-  Coord(Coord<kRank, Index, LongIndex> const &coord) {
-    for (int i = 0; i < kRank; ++i) {
-      idx[i] = coord[i];
-    }
-  }
-
   /// Returns a slice of the Coord which may be larger or smaller in rank
   /// than this.
   template <int Slice>

include/cutlass/gemm/kernel/gemm_universal.h

@@ -162,6 +162,10 @@ public:
       ptr_A(ptr_A), ptr_B(ptr_B), ptr_C(ptr_C), ptr_D(ptr_D), 
       batch_stride_A(batch_stride_A), batch_stride_B(batch_stride_B), batch_stride_C(batch_stride_C), batch_stride_D(batch_stride_D), 
       stride_a(stride_a), stride_b(stride_b), stride_c(stride_c), stride_d(stride_d) {
+      lda = 0;
+      ldb = 0;
+      ldc = 0;
+      ldd = 0;
 
       CUTLASS_TRACE_HOST("GemmUniversal::Arguments::Arguments() - problem_size: " << problem_size);
       }

include/cutlass/tensor_ref.h

@@ -219,9 +219,12 @@ class TensorRef {
   }
 
   /// Converting constructor from TensorRef to non-constant data.
+  template<typename _Magic = int>
   CUTLASS_HOST_DEVICE
   TensorRef(
-    NonConstTensorRef const &ref              ///< TensorRef to non-const data
+    NonConstTensorRef const &ref,             ///< TensorRef to non-const data
+    ///SFINAE trick to avoid creating a copy-constructor when Element_ is already non-const
+    _Magic magic = (typename platform::enable_if< ! platform::is_same<NonConstTensorRef, TensorRef<Element_, Layout_> >::value, _Magic>::type)0
   ):
     ptr_(ref.data()), layout_(ref.layout()) { }