[llvm-commits] [PATCH] Adding the Linpack test to the test suite

[Renato Golin]

On 26 November 2012 13:37, Duncan Sands <baldrick at free.fr> wrote:
> But perhaps you mean that initializing to zero is wrong and it should be set
> to something else?

No, that was it, I was wrong.

There isn't a right or wrong in those tests. The numbers are just
silly accumulations of the vectors, just to make sure our
optimizations are not destroying floating point accuracy. Each kernel
performs FP operations from arrays to arrays and, in the end, I get
one of the final arrays (the destination) and add up its elements and
print. Since each FP operation inside the loop is sensitive to
rounding, small differences will yield big ones in the accumulation
(if they don't cancel each other out, this is why I've setup FP
precision to 10^-4).

It's not the best strategy, but it's a simple one. It might be that
the heuristics is not a good one for most kernels, and we should
possibly turn them off and only rely on the timings, although, would
be good to have some kind of confirmation that we're not destroying FP
accuracy in the process.



--
cheers,
--renato

http://systemcall.org/