[Shootout-list] What people say about shootout

[Sebastien Loisel]

> (...) why are all of
> the floating point benchmarks I see out there (not just the Shootout) dou=
ble
> precision when in reality single precision (...)

The problem I faced in implicitode is that a large fraction of
languages don't even have a single precision data type (e.g., python.)

A possible test would be one that does a matrix-vector multiplication,
where the sizes are tuned so that the matrix will fit in cache if
floats are used, but not if doubles are used. This way, languages that
fake floats using doubles will actually get penalized. However, the
interest is to write tests that work in all language, so we can't
actually fail those languages so the output would have to be the same
in float or double calculations.

The biggest problem is that it's very hard to guess what size matrix
will fit in cache, because cache size varies widely across cpu's.

Sebastien Loisel