soc-dsp questions

soc-dsp questions

[pl bossart]

To prepare next week's ALSA-Asoc meeting, I reviewed Liam's
dsp-upstream code, and I am a bit unclear on the 'no_host_mode'
supported by some DAIs. Apparently these are regular ALSA PCM
substreams, except that there are no data exchanges to/from the host.
- If these substreams are known, an application can in theory
open/close them. How would it know that no data is to be written/read?
I would think additional changes are required in alsa-lib?
- is this really important that such PCM devices be known in
userspace? Or are they declared only so that ALSA controls are enabled
for these streams?
- along the same lines, how useful is the .pointer routine for such
devices? Since no data will be provided by the host, is this really
needed? I see it's only meaningful for the low-power playback mode,
but I am unsure if the information provided would actually be used by
anyone?
- what is the meaning of the SND_SOC_DAI_LINK_OPT_HOST definition? It
doesn't seem to be updated depending on which back-end is used, so it
this really needed?
Thanks
-Pierre

Re: soc-dsp questions

[Liam Girdwood]

On Mon, 2011-04-25 at 17:01 -0500, pl bossart wrote:
> To prepare next week's ALSA-Asoc meeting, I reviewed Liam's
> dsp-upstream code, and I am a bit unclear on the 'no_host_mode'
> supported by some DAIs. Apparently these are regular ALSA PCM
> substreams, except that there are no data exchanges to/from the host.

Ah, this part is WIP and will not be part of the initial submission. We
are currently using this for audio between the MODEM and ABE that does
not pass through the CPU. 

> - If these substreams are known, an application can in theory
> open/close them. How would it know that no data is to be written/read?
> I would think additional changes are required in alsa-lib?
> - is this really important that such PCM devices be known in
> userspace? Or are they declared only so that ALSA controls are enabled
> for these streams?

The intention is that userspace apps would know there is no data, but
this is probably better discussed at conference. It may also require an
alsa-lib update too.

> - along the same lines, how useful is the .pointer routine for such
> devices? Since no data will be provided by the host, is this really
> needed? I see it's only meaningful for the low-power playback mode,
> but I am unsure if the information provided would actually be used by
> anyone?

I think in most cases pointer() would not be required here.

> - what is the meaning of the SND_SOC_DAI_LINK_OPT_HOST definition? It
> doesn't seem to be updated depending on which back-end is used, so it
> this really needed?

Again, this part is WIP and was used to signal that the DAI could be
optionally be used in no host mode.

Liam 

> Thanks
> -Pierre
> _______________________________________________
> Alsa-devel mailing list
> Alsa-devel@alsa-project.org
> http://mailman.alsa-project.org/mailman/listinfo/alsa-devel

Re: soc-dsp questions

[Mark Brown]

On Tue, Apr 26, 2011 at 10:41:05AM +0100, Liam Girdwood wrote:
> On Mon, 2011-04-25 at 17:01 -0500, pl bossart wrote:
> > To prepare next week's ALSA-Asoc meeting, I reviewed Liam's
> > dsp-upstream code, and I am a bit unclear on the 'no_host_mode'
> > supported by some DAIs. Apparently these are regular ALSA PCM
> > substreams, except that there are no data exchanges to/from the host.

> Ah, this part is WIP and will not be part of the initial submission. We
> are currently using this for audio between the MODEM and ABE that does
> not pass through the CPU. 

This is roughly the same thing I've been talking about for digital DAPM
links.  I've got code which runs at the minute but the implementation
sucks too much, should be able to pull out some of the preparation work
in the next day or so.

> > - is this really important that such PCM devices be known in
> > userspace? Or are they declared only so that ALSA controls are enabled
> > for these streams?

> The intention is that userspace apps would know there is no data, but
> this is probably better discussed at conference. It may also require an
> alsa-lib update too.

Personally I think we should just hide them, but it's fairly painful to
do that immediately due to the ALSA core infrastructure we're using.  If
we refactor to reduce this or to support masking PCMs in the core this
should be less of an issue.

Re: soc-dsp questions

[Patrick Lai]

On 4/26/2011 3:18 AM, Mark Brown wrote:
> On Tue, Apr 26, 2011 at 10:41:05AM +0100, Liam Girdwood wrote:
>> On Mon, 2011-04-25 at 17:01 -0500, pl bossart wrote:
>>> To prepare next week's ALSA-Asoc meeting, I reviewed Liam's
>>> dsp-upstream code, and I am a bit unclear on the 'no_host_mode'
>>> supported by some DAIs. Apparently these are regular ALSA PCM
>>> substreams, except that there are no data exchanges to/from the host.
>
>> Ah, this part is WIP and will not be part of the initial submission. We
>> are currently using this for audio between the MODEM and ABE that does
>> not pass through the CPU.

This mode is also deployed on QC MSM as well.
>
> This is roughly the same thing I've been talking about for digital DAPM
> links.  I've got code which runs at the minute but the implementation
> sucks too much, should be able to pull out some of the preparation work
> in the next day or so.
>
Is it possible I can take an early glimpse of implementation? I do have
use cases that PCM is exchanged between two back-ends. Right now, I
need to define DUMMY hostless front-end DAI links to bring up the
back-ends. Another query is how hardware parameters is passed to
back-end with your design? Not able to choose back-end channel mode
independent of front-end channel mode is a big problem especially if
channel mode is more than stereo. DSP is handles upmixing/downmixing
happens in our design. Right now, we force channel mode to stereo. So,
for the scenario which we just want mono input, we have codec
configured to pick up single mic input to both left and right channel.
DSP takes average of two channels into mono stream. Once we need to
support > 2 channel recording, it's wasteful to go with the same
approach if all we want is mono input. Did we talk about this topic
during workshop? Why is my problem also a concern on OMAP4/ABE?
Suggestion?


-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

Re: soc-dsp questions

[Mark Brown]

On Thu, Jun 09, 2011 at 11:58:23PM -0700, Patrick Lai wrote:
> On 4/26/2011 3:18 AM, Mark Brown wrote:

> >This is roughly the same thing I've been talking about for digital DAPM
> >links.  I've got code which runs at the minute but the implementation
> >sucks too much, should be able to pull out some of the preparation work
> >in the next day or so.

> Is it possible I can take an early glimpse of implementation? I do have

No, the code ended up colliding with something someone else had done and
wasn't worth saving so I just threw it away.  However I did get the
external interface upstream before I did that - that was the code to
support not providing a platform driver for a DAI.  What the code did
was to look to see if there was a platform driver and if there wasn't
it'd add some DAPM nodes and links which would bring up the DAI with no
userspace involvement.

> use cases that PCM is exchanged between two back-ends. Right now, I
> need to define DUMMY hostless front-end DAI links to bring up the
> back-ends. Another query is how hardware parameters is passed to
> back-end with your design? Not able to choose back-end channel mode

For the initial code they were just inferred from the capabilities of
the DAI - all the cases I'm interested in are for interoperation with
something that's fixed format at one end of the link so I could punt on
that issue.  I'd thought about allowing the dai_link to have a set of
hw_params settings stored in it which the user would be given an
enumeration to select from but hadn't actually done anything concrete
with it.

> independent of front-end channel mode is a big problem especially if
> channel mode is more than stereo. DSP is handles upmixing/downmixing
> happens in our design. Right now, we force channel mode to stereo. So,
> for the scenario which we just want mono input, we have codec
> configured to pick up single mic input to both left and right channel.
> DSP takes average of two channels into mono stream. Once we need to
> support > 2 channel recording, it's wasteful to go with the same
> approach if all we want is mono input. Did we talk about this topic

That might be handlable by either of the methods I was suggesting above.
Of course depending on the algorithms you're running the DSP may want
more mics than it's producing output channels - beam forming or noise
cancellation are the obvious examples there.

> during workshop? Why is my problem also a concern on OMAP4/ABE?

Do you mean not also a concern?  I *believe* OMAP is passing the
configuration through to the external DAI using the front end/back end
connection so the format gets selected by the app when it does a record,
possibly with rewriting through the hook functions in the machine driver.

Re: soc-dsp questions

[Patrick Lai]

On 6/10/2011 2:42 AM, Mark Brown wrote:

>
> That might be handlable by either of the methods I was suggesting above.
> Of course depending on the algorithms you're running the DSP may want
> more mics than it's producing output channels - beam forming or noise
> cancellation are the obvious examples there.

Yes, exactly.

> Do you mean not also a concern?  I *believe* OMAP is passing the
> configuration through to the external DAI using the front end/back end
> connection so the format gets selected by the app when it does a record,

Can you elaborate more on "configuration through external DAI" ?
Is there an example?

> possibly with rewriting through the hook functions in the machine driver.

Are you referring to fixup function in the machine driver? It works for
hardware parameter that is fixed per machine. For example, regardless
sample rate of front-ends that are routed to same back-end,
back-end sample rate is fixed to 48KHz. I am already taking advantage
of the hook.

Another query I have is how to handle back-end error. The audio bus
which is running on my machine requires close coordination between CPU
and CODEC. Essentially, if one side is unable to respond to incoming
data in time, data exchange halts. I am looking for way to reset both
CPU and CODEC back to fresh state. One approach I am thinking is to
generate XRUN error(snd_pcm_stop(SNDRV_PCM_STATE_XRUN) and have
application call prepare() to reset CPU and CODEC back to good state.
I see each back-end is registered as PCM device so it's possible that
application can read /dev/snd/timer to get notified. However, do I call
prepare() on one of FE PCM devices that are routed to the back end in
question? Would this approach work? Any suggestion?

Thanks
Patrick

-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

Re: soc-dsp questions

[Mark Brown]

On Fri, Jun 10, 2011 at 06:19:57PM -0700, Patrick Lai wrote:
> On 6/10/2011 2:42 AM, Mark Brown wrote:

> >Do you mean not also a concern?  I *believe* OMAP is passing the
> >configuration through to the external DAI using the front end/back end
> >connection so the format gets selected by the app when it does a record,

> Can you elaborate more on "configuration through external DAI" ?
> Is there an example?

That's the whole soc-dsp front end/back end connection thing so the OMAP
drivers should provide an example.

> >possibly with rewriting through the hook functions in the machine driver.

> Are you referring to fixup function in the machine driver? It works for
> hardware parameter that is fixed per machine. For example, regardless
> sample rate of front-ends that are routed to same back-end,
> back-end sample rate is fixed to 48KHz. I am already taking advantage
> of the hook.

Yes.  Since it's code it *could* do conditional things based on some
setting if it needs to.

> Another query I have is how to handle back-end error. The audio bus
> which is running on my machine requires close coordination between CPU
> and CODEC. Essentially, if one side is unable to respond to incoming
> data in time, data exchange halts. I am looking for way to reset both
> CPU and CODEC back to fresh state. One approach I am thinking is to
> generate XRUN error(snd_pcm_stop(SNDRV_PCM_STATE_XRUN) and have
> application call prepare() to reset CPU and CODEC back to good state.
> I see each back-end is registered as PCM device so it's possible that
> application can read /dev/snd/timer to get notified. However, do I call
> prepare() on one of FE PCM devices that are routed to the back end in
> question? Would this approach work? Any suggestion?

I'd expect that from an application point of view this will just work
already?  The application will just operate on the PCM it's operating on
and will notice a stall in the same way it does for any other device
then the front/back end machinery will connect everything up in the same
way it does for every operation when (if!) the application tries to
recover.

Re: soc-dsp questions

[Patrick Lai]

>> Are you referring to fixup function in the machine driver? It works for
>> hardware parameter that is fixed per machine. For example, regardless
>> sample rate of front-ends that are routed to same back-end,
>> back-end sample rate is fixed to 48KHz. I am already taking advantage
>> of the hook.
>
> Yes.  Since it's code it *could* do conditional things based on some
> setting if it needs to.
>
Yes, I don't see any better way unless propagating knowledge of 
front-end/back-end all the way to user-space. I plan on adding 
enumeration of
channel mode in machine driver which sets variable read by the fix up
function. So, CPU and CODEC drivers do not need to provide enumeration
of channel modes. However, same control would have to be added to every
machine driver.

>> Another query I have is how to handle back-end error. The audio bus
>> which is running on my machine requires close coordination between CPU
>> and CODEC. Essentially, if one side is unable to respond to incoming
>> data in time, data exchange halts. I am looking for way to reset both
>> CPU and CODEC back to fresh state. One approach I am thinking is to
>> generate XRUN error(snd_pcm_stop(SNDRV_PCM_STATE_XRUN) and have
>> application call prepare() to reset CPU and CODEC back to good state.
>> I see each back-end is registered as PCM device so it's possible that
>> application can read /dev/snd/timer to get notified. However, do I call
>> prepare() on one of FE PCM devices that are routed to the back end in
>> question? Would this approach work? Any suggestion?
>
> I'd expect that from an application point of view this will just work
> already?  The application will just operate on the PCM it's operating on
> and will notice a stall in the same way it does for any other device
> then the front/back end machinery will connect everything up in the same
> way it does for every operation when (if!) the application tries to
> recover.

Yes, I suppose existing mechanism would work if the front-end PCM is not 
hostless. Eventually, underrun/overrun would be detected. If front-end 
PCM is hostless, host processor for my case would not even know that 
Front-end is stalled even though back-end knows. I suppose
we could enhance framework to propagate XRUN error from back-end to 
front-end and have front end generate XRUN signal back to user-space. 
However, for multiple Front-end to same back end case, all clients of 
front-end PCM get notified of XRUN error and all of them call prepare(). 
I can see there can be adverse effect with such behavior.
Maybe it's easier to recover under the hood without user-space knowing it.

Thanks
Patrick

-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

Re: soc-dsp questions

[Liam Girdwood]

On 11/06/11 02:19, Patrick Lai wrote:
> On 6/10/2011 2:42 AM, Mark Brown wrote:
> 
>>
>> That might be handlable by either of the methods I was suggesting above.
>> Of course depending on the algorithms you're running the DSP may want
>> more mics than it's producing output channels - beam forming or noise
>> cancellation are the obvious examples there.
> 
> Yes, exactly.
> 
>> Do you mean not also a concern?  I *believe* OMAP is passing the
>> configuration through to the external DAI using the front end/back end
>> connection so the format gets selected by the app when it does a record,
> 

This is correct. We pass configuration from the FE PCM to all the connected BE DAIs.

Liam

Re: soc-dsp questions

[Liam Girdwood]

On 11/06/11 12:48, Mark Brown wrote:
> On Fri, Jun 10, 2011 at 06:19:57PM -0700, Patrick Lai wrote:

> 
>> Another query I have is how to handle back-end error. The audio bus
>> which is running on my machine requires close coordination between CPU
>> and CODEC. Essentially, if one side is unable to respond to incoming
>> data in time, data exchange halts. I am looking for way to reset both
>> CPU and CODEC back to fresh state. One approach I am thinking is to
>> generate XRUN error(snd_pcm_stop(SNDRV_PCM_STATE_XRUN) and have
>> application call prepare() to reset CPU and CODEC back to good state.
>> I see each back-end is registered as PCM device so it's possible that
>> application can read /dev/snd/timer to get notified. However, do I call
>> prepare() on one of FE PCM devices that are routed to the back end in
>> question? Would this approach work? Any suggestion?
> 
> I'd expect that from an application point of view this will just work
> already?  The application will just operate on the PCM it's operating on
> and will notice a stall in the same way it does for any other device
> then the front/back end machinery will connect everything up in the same
> way it does for every operation when (if!) the application tries to
> recover.

This is how it works on OMAP4. 

The ABE handles and recovers most BE errors internally and can also interrupt the CPU for serious errors (to signal XRUNs etc).

Liam

Re: soc-dsp questions

[Liam Girdwood]

On 13/06/11 05:55, Patrick Lai wrote:
>>> Are you referring to fixup function in the machine driver? It works for
>>> hardware parameter that is fixed per machine. For example, regardless
>>> sample rate of front-ends that are routed to same back-end,
>>> back-end sample rate is fixed to 48KHz. I am already taking advantage
>>> of the hook.
>>
>> Yes.  Since it's code it *could* do conditional things based on some
>> setting if it needs to.
>>
> Yes, I don't see any better way unless propagating knowledge of 
> front-end/back-end all the way to user-space. I plan on adding 
> enumeration of
> channel mode in machine driver which sets variable read by the fix up
> function. So, CPU and CODEC drivers do not need to provide enumeration
> of channel modes. However, same control would have to be added to every
> machine driver.

Ok, sounds like this may be useful code for other DSPs too.

Fwiw, OMAP4 fixups are based on physical BE DAI so we always use the same config for each BE DAI atm, but flexibility would be good.

Liam 

Re: soc-dsp questions

[Vinod Koul]

On Tue, 2011-11-08 at 20:20 +0000, Girdwood, Liam wrote:
> Hi Vinod,
> 
> Sorry, just about to travel again so my replies might be delayed.
> 
> On 8 November 2011 08:22, Vinod Koul <vinod.koul@linux.intel.com> wrote:
> > Hey Liam,
> >
> > I had few questions on your dsp work, would be great if you could answer
> > them and help me in understanding the approach.
> >
> > - Is it possible to have one FE connected to multiple BEs (voice FE to
> > modem and headset bBE), if so how?
> 
> yes, the DAPM graph is extended to link the AIF of each FE PCM to each
> supported BE DAI (please look at omap-abe-dsp.c).
> 
> This allows (via kcontrol settings) the routing between FE PCMs and BE
> DAIs to be controlled at stream runtime. The Dynamic PCM core takes
> care of all the PCM ops and hw_params for each DAI (and re-writes any
> params that require changes for a particular DAI).
Good, that is what i was hoping for :)
> 
> > - SND_SOC_DAI_LINK_OPT_HOST, what does this mean, does not seem to be
> > used anywhere in core?
> 
> It's historical and can be removed.
> 
> > - Why comment need dummy codec for some BEs like BT etc, when soc-core
> > has already one?
> 
> ditto.
> 
> > - any reason why be haves a number associated with a BE, than usual
> > convention of referring things by string name in asoc?
> >
> 
> Sorry, not sure what part do you mean here ?
you are referring to BEs using a number of be_id in the dai_link, like
as OMAP_ABE_DAI_XXX. Not usually by the name of the BE DAI. Any reason
for this departure from usual asoc convention of using string names?

> > and last one, when would this be merged to mainline, 3.3?
> >
> 
> Ah yes, the important question ;) I've been traveling the last 3
> months or so and have not had a chance to finish the upstreaming. I
> have one more trip planned this year and my intention is 3.3 when I
> return.
Okay, thanks for the info


-- 
~Vinod

Re: soc-dsp questions

[Girdwood, Liam]

Hi Vinod,

Sorry, just about to travel again so my replies might be delayed.

On 8 November 2011 08:22, Vinod Koul <vinod.koul@linux.intel.com> wrote:
> Hey Liam,
>
> I had few questions on your dsp work, would be great if you could answer
> them and help me in understanding the approach.
>
> - Is it possible to have one FE connected to multiple BEs (voice FE to
> modem and headset bBE), if so how?

yes, the DAPM graph is extended to link the AIF of each FE PCM to each
supported BE DAI (please look at omap-abe-dsp.c).

This allows (via kcontrol settings) the routing between FE PCMs and BE
DAIs to be controlled at stream runtime. The Dynamic PCM core takes
care of all the PCM ops and hw_params for each DAI (and re-writes any
params that require changes for a particular DAI).

> - SND_SOC_DAI_LINK_OPT_HOST, what does this mean, does not seem to be
> used anywhere in core?

It's historical and can be removed.

> - Why comment need dummy codec for some BEs like BT etc, when soc-core
> has already one?

ditto.

> - any reason why be haves a number associated with a BE, than usual
> convention of referring things by string name in asoc?
>

Sorry, not sure what part do you mean here ?

> and last one, when would this be merged to mainline, 3.3?
>

Ah yes, the important question ;) I've been traveling the last 3
months or so and have not had a chance to finish the upstreaming. I
have one more trip planned this year and my intention is 3.3 when I
return.

Regards

Liam

soc-dsp questions

[Vinod Koul]

Hey Liam,

I had few questions on your dsp work, would be great if you could answer
them and help me in understanding the approach.

- Is it possible to have one FE connected to multiple BEs (voice FE to
modem and headset bBE), if so how?
- SND_SOC_DAI_LINK_OPT_HOST, what does this mean, does not seem to be
used anywhere in core?
- Why comment need dummy codec for some BEs like BT etc, when soc-core
has already one?
- any reason why be haves a number associated with a BE, than usual
convention of referring things by string name in asoc?

and last one, when would this be merged to mainline, 3.3?

-- 
~Vinod