On Tue, Sep 30, 2008 at 09:10:57AM -0700, Troy Kisky wrote:

Currently, we have the line

runtime->hw.formats = codec_dai->playback.formats & cpu_dai->playback.formats;

This seems to force the two to support a common subset. For example some codec support only big-endian or little-endian

You'll find that for historical reasons all current codec drivers claim to support little endian only though actually all they're really declaring is the word length - they don't care about the memory layout since by the time it hits the DAI and they can see it the endianness is gone.

My codec expects the most significant bit to be shifted 1st. I would call that big-endian. But the codec marks itself as only doing little endian, because the cpu_dai converts from little-endian to big-endian. Do all codecs expect the msb 1st? If so, then endianness would seem irrelevant to codecs and both should be set.

Yes, pretty much all codecs expect that format. Either the codecs should be setting both endianesses or this should be fudged for them in the core before doing the DAI capability matching. I'm tending towards the latter since it makes the codec drivers less verbose, it's on my list of patches to cook up since it's an existing problem for AVR32 systems.

The same goes with number of channels. The davinci dma engine can convert a playback stream from mono to stereo, or a capture stream from stereo

At the minute ASoC has a restricted view of the digital domain - up until fairly recently there hasn't been much need to do a lot there since the complexity was all been in the analogue side. The digital side usually just provided fairly simple data transfer. That is changing with modern audio hub codecs working much more in the digital domain but dealing well with them will require some core work - at present I can't immediately think of a clean way of handling things like that.

I think off the top of my head what'd be best for handling the channel conversion is support for representing digital mixing and paths in DAPM. The audio hub codecs certainly need that. You'd probably then end up with something like a memory interface widget that knows about the memory format, a DAI widget and a digital mixer widget between the two. Like I say, that's off the top of my head rather than a proper design.

to mono. The davinci McBsp can also be more efficient if allowed to swap left and right channels. Is there a method to let the codec know to swap them back?

No, that's not possible at present - this is the first time it's been suggested that I can remember.

It's easy enough to do, though: add a lr_swap operation to the DAIs similar to the existing digital mute operation which codecs can then implement (the codecs that have this at the minute do it by exposing a control to user space). We can then either have machine drivers do the configuration or add something to the core to figure out when to set it up automatically. If the core does do it automatically it ought to be overridable by the machine driver - safer to just put it in the machine driver only, at least initially.