Merge remote-tracking branch 'origin/develop' into aosewski/ggemm_multi_d2

include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp

@@ -795,11 +795,6 @@ struct BlockwiseGemmXdlops_v2
                       "wrong!");
     }
 
-    __host__ __device__ BlockwiseGemmXdlops_v2(const BlockwiseGemmXdlops_v2& other)
-        : a_thread_copy_(other.a_origin), b_thread_copy_(other.b_origin)
-    {
-    }
-
     // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
     __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
     {

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp

@@ -587,7 +587,7 @@ struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
                                          BatchStrideD1s,
                                          BatchStrideE1}
         {
-            if(ck::EnvIsEnabled(ENV(CK_LOGGING)))
+            if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
             {
                 std::cout << "a0_grid_desc_m_k_{" << a0_grid_desc_m_k_.GetLength(I0) << ", "
                           << a0_grid_desc_m_k_.GetLength(I1) << "}" << std::endl;