None of these libraries unfortunately allows making good use of CPU vectorized units. Xla might produce some SIMD code but it pales in (performance) comparison to routines written explicitly for SIMD on GPU. ISPC is a good example of this.