Fix benchmark

[zasdfgbnm]

Fixes #96

[naoyam]

Why did it cause the error message? Can we detect the error and give a better error message?

[zasdfgbnm]

Why did it cause the error message? Can we detect the error and give a better error message?

When the inputs is given to compileFusion, it will try to infer launch parameters and will try to compute shared memory usage. However, the fusion is scheduled like split(i, blockDim.x), which can not be infered without the launch constraints.

[naoyam]

LGTM, though still unclear why this was working.

[naoyam]

Looks like those benchmarks are not the only ones that do compileFusion a scheduled fusion without passing launch constraints.

[liqiangxl]

here is another one, #102
after this, all the benchmarks work fine.

[naoyam]

Thanks @liqiangxl. Could you please do a quick check to find where this started to happen? I wonder if my PR of cleaning up index types might be the reason, but I'm pretty sure I ran all the benchmarks. Looking at the recent merged PRs, I don't see anything particularly suspicious.

[liqiangxl]

It failed at least on the initial commit of this new Fuser repo. Let me check the old one.