[PATCH] D10991: [LNT] Reduce I/O execution time for Polybench

[Renato Golin]

In http://reviews.llvm.org/D10991#200265, @grosser wrote:

> thank you for working on this. This is goes definitely in the right direction. We could possibly even compute a single output hash in Polybench and just print this hash. In the end, we only want to know if the output changed.


That was the original plan, but I'd need to add a hash function to Polybench.

So I did the intermediate solution: a simple dump function that would be quick enough to avoid being the hottest function in the benchmark, but concise enough to not rely on I/O speed or to invoke printf-like functions for floating point data.

In the end, we have one malloc+free, ~4000 inlined calls to a bunch of op+write and one memcpy. All of them easy to optimise on almost every target.

The speed up is already impressive, I think over-engineering this would be counter-productive. :)

cheers,
--renato


Repository:
  rL LLVM

http://reviews.llvm.org/D10991