Re: [mpg123-devel] fixed point decoders

[Thomas O. <tho...@or...>]

Am Tue, 02 Jun 2009 15:20:46 +0900
schrieb Taihei Monma <tm...@ma...>: 

> I made a quick comparison between mpg123 and MAD. The RMS error of the  
> compliance bitstreams and the decoding speed of our benchmark set were  
> measured. Both use asm optimization for intel processors. Here is a  
> result:
> http://tmkk.hp.infoseek.co.jp/mpg123_mad.png

One has to look twice, at least, to get the hang of the graphs, but it really looks ... _nice_!

> Generally our decoder is performing very nice, on both accuracy and  
> speed. Note that the 16-bit outputs of MAD seem to be suffered from  
> the rounding issue we had before. With proper rounding (or 24-bit  
> output) they will be fully compliant, at least for "accuracy" and  
> "default" settings.

You hacked rounding into MAD? Or, did you not disable dithering?
For speed comparisons I used -i -d for madplay to disable CRC checking and dithering.
But I can confirm

> Anyway, our fixed point decoders are now accurate and fast (on x86)  
> enough.

You say it... you are human, you are sure? At least one very happily hacking the heck out of our old code.
I'm becoming accustomed to the pace of your revolving optimizations. That could be dangerous;-)
Is it rude to just become better as MAD at what they are so feverishly are proud of -- integer performance _and_ quality?
I don't count the 24bit integer output though, that can go and compare itself to our floating point decoder;-)

> Remaining task will be an optimization for other platforms  
> like ARM (I need iPod touch, hehe)...

Down-to-earth: That is what really matters for fixed point decoding... low-end on x86 should be i486 anyway (realtime on my i386DX40 ? a hard call...) and there we have specific code. Other i386 machines have more fun with MMX, 3Dnow, SSE ...
ARM is the platform that needs fixed point... for testing, I should be able to get mpg123 on my eBook (though that one actually has a decoder chip for MPEG audio...). Actually, one should be able to set up a VM for ARM with debian on it... 
... and get the folks on the little handhelds on track!

Hey, you Sony PSP and nintendo ds users -- I'm talking to you. Anyone want to toss in a comment about libmpg123 performance there and wishes to hack for some ARMs?


Alrighty then,

Thomas.