[Matroska-devel] Re: Timestamp precision in matroska files

[Steve Lhomme]

Ronald Bultje wrote:

> On Fri, 2004-01-09 at 16:05, Christian HJ Wiesner wrote:
> 
>>Moritz Bunkus wrote:
>>
>>>Heya,
>>>FLOAT WARNING! If we ever want Matroska to become a reality on hardware
>>>devices then using floats everywhere is a big no!
>>
>>ACK .......
> 
> 
> Eh, samplerate already is a float, so you might be a little late...
> 
> I do agree, though.

The sampling rate is not used anywhere else in the format. So it can be 
computed once and you're done. With a floating point library it's no big 
deal.

The problem is the precision of TrackTimecodeScale. I made the same 
(successful) test but with float instead of double. And the 
approximation fails (to estimate the actual timecode of a sample in ns) 
at the 372nd sample. So the single precision floating point is not 
enough for sample accurate seeking. But as this timecode complete 
accuracy will be needed only for seeking (during playback a delay of a 
few ns is not a big deal). It is only computed in rare cases. So using a 
third party library for that won't have much performance impact anyway.