Slightly problematic float->int conversions

[m8pple]

The docs explain that only round-nearest-even is supported, which
is fine for all the normal operators, but a bit more problematic for
float -> int conversion where it is likely we need more control.

Given the lack of math.h, people will reach for int x=(int)f;, but the implementation
applies RNE, so we lose compatibility with C which expects RTZ, and code will
silently break. This can be problematic for things like special functions, where
there is often table-lookup for coefficients based on an integer/fractional split.

If the altera float->fixed IP supports a dynamic rounding config, could that
be used to support both RTZ and RNE (and RDN & RDU)?

If not, it would be helpful to add modes to the existing elf checker somehow,
so that people know if an unsupported rounding mode had crept in.

[mn416]

Hi David,

Unfortunately, RNE is all the float -> int core gives us.

We could try the float -> 32.32 fixed-point converter and then implement the rounding manually.

Am open to a pull request for the ELF checker :)

In the meantime, I'll reference this issue in the limitations section.

Thanks,

Matt