Accelerating GPU kernels for dense linear algebra