Just my option, but I'd recommend the CBLAS interfaces as the ones most likely to be portable.
BTW, ACML should work just fine on any other brand of X86 processor.
Hey. Check out Libra SDK from a company called GPU Systems. They support BLAS (dense, sparse),complex numbers, single&double precision etc... full portability across CPUs and GPUs and progrmming languages...I also heard they support AMD GPUs!!!! ...=) yihaa!