[FFmpeg-devel] [PATCH] ARM: NEON optimised simple_idct

[Alexander Strange]

On Aug 25, 2008, at 4:04 PM, M?ns Rullg?rd wrote:

> Michael Niedermayer <michaelni at gmx.at> writes:
>
>> On Mon, Aug 25, 2008 at 07:47:16PM +0100, M?ns Rullg?rd wrote:
>>> Michael Niedermayer <michaelni at gmx.at> writes:
>> [...]
>>>> 2. depending on the pattern of non zero / all zero rows one of 8
>>>> optimized column transforms is used.  This may be a bad idea though
>>>> for a CPU with a small code cache ...
>>>>
>>>> also maybe it would make sense to look at i386/idct_sse2_xvid.c
>>>> which uses SSE2 (128bit registers), this one uses only 16bit  
>>>> operations
>>>> for the column transform so it may be faster when the tricks of  
>>>> the simple
>>>> idct arent applicable
>>>
>>> Do you expect any sane person to be able to read that?
>>
>> well, a little insanity may be needed
>>
>>> That's also
>>> not bitexact, right?
>>
>> it is supposed to be bitexact, and i cannot remember a case where any
>> input lead to different output. Also the MMX one is used in the
>> regression tests and they match between MMX and non x86 cpus ...
>
> All the different IDCT variants (int, simple, simplemmx, libmpeg2mmx,
> xvidmmx, faani) give different output on my machine with current
> FFmpeg.  Which one is correct?

All of them are correct; none of the IDCT-using codecs specify exact  
rounding.
simple* and xvid* should be the same as their C versions, though.
It's best to stick with simpleidct so we can at least have bit-exact  
compatibility with ffmpeg-encoded files.

>>>> also
>>>>
>>>>    Intel 64 and IA-32 Architectures
>>>>    Software Developers Manual
>>>>                              Volume 2A (and B)
>>>>           Instruction Set Reference
>>>>
>>>> contains very readable and unambigious explanations of what all the
>>>> MMX, SSE* instruction do, if you ever want to decypher mmx or sse  
>>>> code
>>>
>>> I have those documents, and reading Chinese is easier.
>>
>> This is great, so you can help me communicate with zhentan who is a  
>> SOC
>> student and IIRC chinese.
>
> No, but maybe he can explain mmx to me.

Googling the intrinsic names sometimes turns up better or at least  
different documentation, but MS's website is really not accessible...