Generating High-Performance Tensor Operators with Hardware Primitives