Have you also looked at its memory consumption? Note that quicksort itself is worse than linear, with a quite bad worse case runtime. This multiplies with the worse than linear runtime behaviour of certain steps of template compilation and instantiation (sometimes those are exponentional). You should maybe graph your compiletime for various datasets to observe the real complexity class for your code. Usually template metaprogramming with such large datasets is not feasible.
Edit:
Out of curiosity I tried the code out and found that up until ~500 it follows roughly the formula pow(N*log(N),1.47)*0.0004+0.6 but then starts to get incredibly much slower, with 155 seconds for 700 items. Also at around that it starts eating very much ram (3GiB for 600) which leads me to the conclusion that for 1000 elements it will need more ram than most people have and will take hours to compile.
Further note that the code does not work when not every element is unique.