Re: [2/3] vcomp: Implement _vcomp_reduction_r4 and add tests.

[Henri Verbeet]

On 8 September 2016 at 08:21, Sebastian Lackner <sebastian(a)fds-team.de> wrote:
> Yes, there are really bool_and and bool_or reduction operations for float/double.
> I decided to explicitly write down the comparison with 0.0 in this case because
> it otherwise looks like a mistake. According to my tests all values != 0.0 are
> seen as true (without any small epsilon to avoid rounding errors... well).

I'd expect it to just look at the bit pattern, but just for laughs,
what about -0.0f?

[2/3] vcomp: Implement _vcomp_reduction_r4 and add tests.

[Sebastian Lackner]

On 08.09.2016 11:18, Henri Verbeet wrote:
> On 8 September 2016 at 08:21, Sebastian Lackner <sebastian(a)fds-team.de> wrote:
> > Yes, there are really bool_and and bool_or reduction operations for float/double.
> > I decided to explicitly write down the comparison with 0.0 in this case because
> > it otherwise looks like a mistake. According to my tests all values != 0.0 are
> > seen as true (without any small epsilon to avoid rounding errors... well).
>
> I'd expect it to just look at the bit pattern, but just for laughs,
> what about -0.0f?

Thanks for the feedback. I've sent an updated version including tests for -0.0,
which shows that the current implementation matches Windows behavior. Of course
that does not necessarily mean that they implement it the same way, but its at
least not just matching the bit pattern of +0.0.

Regards,
Sebastian