Re: [hatari-devel] FPU update

[ Thread Index | Date Index | More Archives ]

Am 21.01.2017 um 23:04 schrieb Andreas Grabher <andreas.grabher@xxxxxxxxxxxx>:

> Am 21.01.2017 um 22:08 schrieb Toni Wilen <twilen@xxxxxxxxxx>:
>>>>> Attached is screenshot that shows errors, what do they mean? (Is
>>>>> it correct or incorrect value or something else?)
>>>> FSGLMUL and FSGLDIV error cause most likely is this (which is not 
>>>> handled yet):
>>>> "The input operand mantissas truncate to single precision before
>>>> the multiply operation" (from M68000PRM.pdf)
>>>> (How to emulate this correctly?)
>>> These SGL functions are quite weird. I wrote some special functions
>>> assuming this: - The input values are truncated (not rounded) to
>>> single precision mantissa (after normalization, if necessary) - The
>>> output value is rounded to single precision mantissa with these
>>> exceptions: - Output overflows: Infinity or biggest normal number
>>> with extended mantissa - Output severely underflows: Zero or smallest
>>> denormalized number with extended mantissa
>>> Please try the appended patch, but be aware this is highly
>>> experimental. I'll have a look at FNEG and FABS tomorrow. But on
>>> first view this can only be caused by missing NaN handling in those
>>> functions. Other than this these functions are quite simple.
>> I merged it quickly but now all FSGL tests return error, including
>> FSGLDIV tests that did pass previously.
>> As usual, official documentation is mostly useless.. Perhaps real
>> hardware tests are again needed?
> Yes, maybe. Or maybe it would help if we knew the input values and what is expected as an output. That might point us to the issue.
> I noticed there is a little mistake from merging my previous code. In the function normalizeFloatx80Subnormal there is an extra line at the end of the function, that should have been removed.
I may have missed something, but can anyone point me to the source code of the test utility that Toni used for his test?

Mail converted by MHonArc 2.6.19+