Re: [LAU] Daemons, daemons...kill those daemons.

On Mon, Jun 15, 2009 at 05:14:46AM +0000, Fernando Lopez-Lezcano wrote: First thing I do when installig Fedora is remove sendmail and install postfix - for the sole reason that I'm capable of configuring postfix in less than 48 hours :-) Seriously now, talking about ALSA let's just consider resampling. Assume your hardware is running at 48 kHz and you want to use 44.1 kHz. Typical audio hardware will interrupt whenever it has N samples available for reading and writing, where in many cases N will be a power of 2. Now if you resample that you have basically two choices: 1. Keep the 'period time' generated by the HW, i.e. provide an interface that will trigger its client at the same rate. The consequence is that you don't have a fixed number of samples (in this example, if the HW period is 1024, you would have eiter 940 or 941 samples in a period). This would be unacceptable to some clients, e.g. Jack. 2. Try to create an interface that will offer a fixed 2^n period size. This requires extra buffering, in other words increased latency. And the timing of such periods will be irregular. In some cases a HW capture buffer will not be sufficent to generate a resampled buffer, you have to wait an extra period and you better have some samples buffered if the client wants them. In a similar way, a single playback buffer could complete two HW buffers, forcing you to store one of them. If you want the period times to be more or less regular, this again requires more latency. A client would quite legitimately expect the two resampled streams (capture and playback) to have the same period timing, as they would have for direct HW access. But in ALSA dmix and share are independent plugins, so anything goes. And even if they were coupled, synchronising the two streams could again imply more latency (it's possible but not simple to avoid this). The conclusion is that resampling at the sound card API, if this is in any way 'period based' breaks things, at least in the sense that it introduces more latency than the resampling operation (which implies filtering and hence delay) does on its own. And there is no good reason for ever doing that. Applications like soft synths should be capable of generating their output at the sample rate imposed by the hardware. The *only* cases were resampling is really needed is when dealing with *stored* data (either being read or written). But in that caes there will be buffering between the storage and the RT audio code, and *that* is the right place to apply resampling, it costs nothing (in terms of timing) if done there. Ciao, -- FA Io lo dico sempre: l'Italia è troppo stretta e lunga.