Waveform design enables cognitive radar to adapt to its working environment, which requires an effective approach to synthesize radar waveform with desired ambiguity function. The optimization problem of radar waveform design to shape its ambiguity function is considered in this letter. Precisely, a Riemannian manifold optimization is employed to transform the constant modulus constrained problem into an unconstrained one in Riemannian space. Moreover, a Riemannian conjugate gradient descent algorithm is developed to tackle the non-convex optimization problem. Simulation results demonstrated the effectiveness our proposed approach.