Yeah, fair enough. True blurs can be done faster by downsampling, blurring the smaller image, then linear filtering the small image up to the original size again, but the rendering engine in the runtime can't support this right now - it'll be included in 2.0 though.
I just read this can be done sampling with tex2Dbias(s, t), where s is a sampler. t is a 4D vector. The mip level is biased by t.w before the lookup takes place.
I'll give it a whirl later, I'm just putting it up here so I don't forget