diff --git a/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp b/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp index 2ce08e7044..0125aa086e 100644 --- a/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp +++ b/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp @@ -1937,7 +1937,7 @@ struct XdlopsGemm template __device__ static auto GetGfx11InputBlkIdx() { - const auto laneId = GetLaneId() % mfma_instr.num_threads_per_blk; + auto laneId = GetLaneId() % mfma_instr.num_threads_per_blk; if constexpr(SwizzleA) { laneId = ((laneId & 1) << 3) | (laneId >> 1);