Re: [hatari-devel] FPU update

[Andreas Grabher <andreas.grabher@xxxxxxxxxxxx>]

Great! Looking forward to the new version! Is it possible to add some special values to the test cases? I'm thinking of stuff like smallest normal for each precision, inputs that will overflow, underflow, overflow or underflow through rounding, maybe some zeros, infinities and (signaling) NaNs and different bit patterns, that do not always cause inexact exceptions (e.g. powers of 2).

We recently found some bugs, that did not show with the test utility, because the affected values were not used. e.g. we discovered, that overflow in SoftFloat always sets the inexact bit (fixed), while this should only happen, if rounded off bits are non-zero.

The problems you are describing should be gone with SoftFloat.

Am 06.02.2017 um 17:52 schrieb Douglas Little <doug694@xxxxxxxxxxxxxx>:

Hi,

I have recently been updating the FPU diagnostic tool to support extended precision format - while this is not quite finished yet, I have already noticed a high failure rate for Hatari vs HW and it *seems* to be related to unused bits set to 1 in source data using the 'memory exploded' extended format (resulting in INF or NaN output instead of real values). This probably isn't serious for compatibility purposes but its an example of something that only shows up when testing this format...

Anyway I'll post the new tool soon...

D

On 1 February 2017 at 16:41, Toni Wilen <twilen@xxxxxxxxxx> wrote:

> Yes it can be added - but the expectation data is going to get bigger
> again. It will also need some changes to accommodate the larger types.
>
> It would benefit from loading separate expectation data for each test,
> instead of one giant blob at the start. At some point it won't be be
> possible to confirm it on real HW due to the size. This would also allow
> the individual tests to be more exhaustive.

At least trigonometric and logarithmic functions can be dropped
temporarily/moved to separate test program. They are too different and
most of them don't have softfloat algorithm anyway.