[FFmpeg-devel] suggestions for debugging ff_fft_calc_altivec?

[Pavel Koshevoy]

On 8/13/13 2:58 AM, Michael Niedermayer wrote:
> On Sat, Aug 10, 2013 at 10:27:27PM -0600, Pavel Koshevoy wrote:
>> Hi,
>>
>> I've decided to take a closer look at OSX 10.5 PPC build with
>> gas-preprocessor.pl.  The last time I tried it more than half a year
>> ago the result was that some audio decoders didn't sound right (AAC,
>> AC3 -- both sound like loud white noise).
>>
>> I tried again today and compared decoder output from builds with and
>> without gas-preprocessor.  I've narrowed it down to output of
>> fft_calc_c (called in ff_imdct_half_c) being different from output of
>> ff_fft_calc_altivec (called in imdct_half_altivec).
>>
>> I am not a PPC assembly expert, nor am I familiar with FFT
>> implementation details to be able to easily spot what the problem is
>> with ff_fft_calc_altivec.  Can anyone offer some suggestions to
>> proceed further?
> If you havent yet found the issue
> you could test the working and non working with various test patterns
> like all 0 input with one coefficient non zero
> If this still doesnt hint to where the issue is you can comment parts
> of the asm out to see when the difference disappears
> the idea would be sort of like bisecting the lines of code to narrow
> down where a difference is
>
> also make sure you use a gas preprocessor with martins fixes:
> https://github.com/michaelni/gas-preprocessor
> and when it helps, pester yuvi to merge them
>

It appears the problem was with gas-preprocessor -- the one recommended 
at http://ffmpeg.org/platform.html doesn't work, the version you've 
mentioned works.  I am not sure how I would go about pestering yuvi to 
merge the fixes...  I am just glad to finally have assembly optimization 
on ppc.

Thanks very much,
     Pavel.