Discussion:
Nexus S and feature android.hardware.audio.low_latency
(too old to reply)
Glenn Kasten
2010-12-16 20:51:45 UTC
Permalink
The Nexus S phone does not claim the
android.hardware.audio.low_latency feature, as seen by adb shell pm
list features. The upcoming Android Compatibility Definition Document
(CDD) will describe the requirements for this feature more precisely.

http://developer.android.com/reference/android/content/pm/PackageManager.html#FEATURE_AUDIO_LOW_LATENCY

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Glenn Kasten
2010-12-18 00:50:43 UTC
Permalink
The Android Compatibility Definition for 2.3 (gingerbread) is now
published at
http://source.android.com/compatibility/index.html
(click on Getting Started / Current CDD). Though the CDD is directed
primarily at device manufacturers, section 5.3 Audio Latency should
also be of interest to app developers using the Android native audio
APIs based on Khronos Group OpenSL ES.

On Dec 16, 12:51 pm, Glenn Kasten <***@android.com> wrote:
> The Nexus S phone does not claim the
> android.hardware.audio.low_latency feature, as seen by adb shell pm
> list features. The upcoming Android Compatibility Definition Document
> (CDD) will describe the requirements for this feature more precisely.
>
> http://developer.android.com/reference/android/content/pm/PackageMana...

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
dario
2010-12-18 08:56:23 UTC
Permalink
Thanks Glenn for the info.

Any chance that there would be a list of devices that would be likely
candidates for having the low_latency feature upon upgrading to 2.3?

-Dario

On Dec 17, 7:50 pm, Glenn Kasten <***@android.com> wrote:
> The Android Compatibility Definition for 2.3 (gingerbread) is now
> published athttp://source.android.com/compatibility/index.html
> (click on Getting Started / Current CDD). Though the CDD is directed
> primarily at device manufacturers, section 5.3 Audio Latency should
> also be of interest to app developers using the Android native audio
> APIs based on Khronos Group OpenSL ES.
>
> On Dec 16, 12:51 pm, Glenn Kasten <***@android.com> wrote:
>
>
>
>
>
>
>
> > The Nexus S phone does not claim the
> > android.hardware.audio.low_latency feature, as seen by adb shell pm
> > list features. The upcoming Android Compatibility Definition Document
> > (CDD) will describe the requirements for this feature more precisely.
>
> >http://developer.android.com/reference/android/content/pm/PackageMana...

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
mic _
2010-12-18 15:09:31 UTC
Permalink
The first batch of devices running Gingerbread are likely not to support
this feature since the hardware for those devices was finalized a long time
ago, and meeting these requirements on a platform that wasn't developed with
the requirements in mind can be quite hard if not impossible.

Personally I think these requirements need to be more specific. What's the
intended sound source / destination? Do you only need to meet these
requirements for the internal mic(s) and the internal loudspeaker? What
about wired accessories? Or A2DP? (that would probably be a hopeless task).

/Michael

On Sat, Dec 18, 2010 at 9:56 AM, dario <***@gmail.com> wrote:

> Thanks Glenn for the info.
>
> Any chance that there would be a list of devices that would be likely
> candidates for having the low_latency feature upon upgrading to 2.3?
>
> -Dario
>
> On Dec 17, 7:50 pm, Glenn Kasten <***@android.com> wrote:
> > The Android Compatibility Definition for 2.3 (gingerbread) is now
> > published athttp://source.android.com/compatibility/index.html
> > (click on Getting Started / Current CDD). Though the CDD is directed
> > primarily at device manufacturers, section 5.3 Audio Latency should
> > also be of interest to app developers using the Android native audio
> > APIs based on Khronos Group OpenSL ES.
> >
> > On Dec 16, 12:51 pm, Glenn Kasten <***@android.com> wrote:
> >
> >
> >
> >
> >
> >
> >
> > > The Nexus S phone does not claim the
> > > android.hardware.audio.low_latency feature, as seen by adb shell pm
> > > list features. The upcoming Android Compatibility Definition Document
> > > (CDD) will describe the requirements for this feature more precisely.
> >
> > >http://developer.android.com/reference/android/content/pm/PackageMana.
> ..
>
> --
> You received this message because you are subscribed to the Google Groups
> "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com.
> To unsubscribe from this group, send email to
> android-ndk+***@googlegroups.com<android-ndk%***@googlegroups.com>
> .
> For more options, visit this group at
> http://groups.google.com/group/android-ndk?hl=en.
>
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-19 09:49:33 UTC
Permalink
Hi,

I looked at the CDD section 5.3, and it says:

"continuous output latency of 45 milliseconds or less"

45ms is not low latency.

Implementing a drum pad, or some synth which react to touch events, is not
realistic over 10ms latency, unless you want users to laugh at your product.

Olivier

On 12/18/2010 01:50 AM, Glenn Kasten wrote:
> The Android Compatibility Definition for 2.3 (gingerbread) is now
> published at
> http://source.android.com/compatibility/index.html
> (click on Getting Started / Current CDD). Though the CDD is directed
> primarily at device manufacturers, section 5.3 Audio Latency should
> also be of interest to app developers using the Android native audio
> APIs based on Khronos Group OpenSL ES.
>
> On Dec 16, 12:51 pm, Glenn Kasten <***@android.com> wrote:
>> The Nexus S phone does not claim the
>> android.hardware.audio.low_latency feature, as seen by adb shell pm
>> list features. The upcoming Android Compatibility Definition Document
>> (CDD) will describe the requirements for this feature more precisely.
>>
>> http://developer.android.com/reference/android/content/pm/PackageMana...
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
niko20
2010-12-19 23:23:55 UTC
Permalink
Ageed. Low latency should be defined as 15ms or less. At most 20ms but
45 is definitely not low latency.

On Dec 19, 3:49 am, Olivier Guilyardi <***@samalyse.com> wrote:
> Hi,
>
> I looked at the CDD section 5.3, and it says:
>
> "continuous output latency of 45 milliseconds or less"
>
> 45ms is not low latency.
>
> Implementing a drum pad, or some synth which react to touch events, is not
> realistic over 10ms latency, unless you want users to laugh at your product.
>
> Olivier
>
> On 12/18/2010 01:50 AM, Glenn Kasten wrote:
>
>
>
> > The Android Compatibility Definition for 2.3 (gingerbread) is now
> > published at
> >http://source.android.com/compatibility/index.html
> > (click on Getting Started / Current CDD). Though the CDD is directed
> > primarily at device manufacturers, section 5.3 Audio Latency should
> > also be of interest to app developers using the Android native audio
> > APIs based on Khronos Group OpenSL ES.
>
> > On Dec 16, 12:51 pm, Glenn Kasten <***@android.com> wrote:
> >> The Nexus S phone does not claim the
> >> android.hardware.audio.low_latency feature, as seen by adb shell pm
> >> list features. The upcoming Android Compatibility Definition Document
> >> (CDD) will describe the requirements for this feature more precisely.
>
> >>http://developer.android.com/reference/android/content/pm/PackageMana...

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
niko20
2010-12-19 23:27:39 UTC
Permalink
Actually the warm output latency in the ccd is 10ms so i think it does
look reasonable already.

On Dec 19, 5:23 pm, niko20 <***@yahoo.com> wrote:
> Ageed. Low latency should be defined as 15ms or less. At most 20ms but
> 45 is definitely not low latency.
>
> On Dec 19, 3:49 am, Olivier Guilyardi <***@samalyse.com> wrote:
>
>
>
> > Hi,
>
> > I looked at the CDD section 5.3, and it says:
>
> > "continuous output latency of 45 milliseconds or less"
>
> > 45ms is not low latency.
>
> > Implementing a drum pad, or some synth which react to touch events, is not
> > realistic over 10ms latency, unless you want users to laugh at your product.
>
> > Olivier
>
> > On 12/18/2010 01:50 AM, Glenn Kasten wrote:
>
> > > The Android Compatibility Definition for 2.3 (gingerbread) is now
> > > published at
> > >http://source.android.com/compatibility/index.html
> > > (click on Getting Started / Current CDD). Though the CDD is directed
> > > primarily at device manufacturers, section 5.3 Audio Latency should
> > > also be of interest to app developers using the Android native audio
> > > APIs based on Khronos Group OpenSL ES.
>
> > > On Dec 16, 12:51 pm, Glenn Kasten <***@android.com> wrote:
> > >> The Nexus S phone does not claim the
> > >> android.hardware.audio.low_latency feature, as seen by adb shell pm
> > >> list features. The upcoming Android Compatibility Definition Document
> > >> (CDD) will describe the requirements for this feature more precisely.
>
> > >>http://developer.android.com/reference/android/content/pm/PackageMana...

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-20 08:37:24 UTC
Permalink
I don't think so:

------------
"warm output latency" is defined to be the interval between when an application
requests audio playback and when sound begins playing, when the
audio system has been recently used but is currently idle (that is, silent)

"continuous output latency" is defined to be the interval between when an
application issues a sample to be played and when the speaker physically
plays the corresponding sound, while the device is currently playing back audio
------------

IIUC, "continuous" latency is the delay between the moment audio data is passed
to the OpenSL API, and the moment it gets out of the speaker.

"warm" latency just seems to be some startup time. It seems to be about resuming
playback, in which case an internal buffer may already carry some data.
Therefore these 10ms could be the real hardware latency, while the extra
continuous latency may come from intermediary software layers, I suppose.

Unless I'm wrong, continuous latency is the main one for us app developers.

Olivier

On 12/20/2010 12:27 AM, niko20 wrote:
> Actually the warm output latency in the ccd is 10ms so i think it does
> look reasonable already.
>
> On Dec 19, 5:23 pm, niko20 <***@yahoo.com> wrote:
>> Ageed. Low latency should be defined as 15ms or less. At most 20ms but
>> 45 is definitely not low latency.
>>
>> On Dec 19, 3:49 am, Olivier Guilyardi <***@samalyse.com> wrote:
>>
>>
>>
>>> Hi,
>>> I looked at the CDD section 5.3, and it says:
>>> "continuous output latency of 45 milliseconds or less"
>>> 45ms is not low latency.
>>> Implementing a drum pad, or some synth which react to touch events, is not
>>> realistic over 10ms latency, unless you want users to laugh at your product.
>>> Olivier
>>> On 12/18/2010 01:50 AM, Glenn Kasten wrote:
>>>> The Android Compatibility Definition for 2.3 (gingerbread) is now
>>>> published at
>>>> http://source.android.com/compatibility/index.html
>>>> (click on Getting Started / Current CDD). Though the CDD is directed
>>>> primarily at device manufacturers, section 5.3 Audio Latency should
>>>> also be of interest to app developers using the Android native audio
>>>> APIs based on Khronos Group OpenSL ES.
>>>> On Dec 16, 12:51 pm, Glenn Kasten <***@android.com> wrote:
>>>>> The Nexus S phone does not claim the
>>>>> android.hardware.audio.low_latency feature, as seen by adb shell pm
>>>>> list features. The upcoming Android Compatibility Definition Document
>>>>> (CDD) will describe the requirements for this feature more precisely.
>>>>> http://developer.android.com/reference/android/content/pm/PackageMana...
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Robert Green
2010-12-20 09:30:25 UTC
Permalink
Low latency should be defined as an amount which allows for accurate
and believable synchronization with video. Most 2D apps can, in
theory, run at 60FPS, assuming some degree of efficiency of execution
on all parts. 60FPS = 0.016666ms = 16ms per frame. I believe 16ms
would be a more practical definition of "low latency" as it would
allow for accurate sync with video. 45ms seems like a number made up,
hoping that it's low enough to entice OEMs to target.

Comparisons:

Windows Audio (NOT DirectSound) has a latency lower than 45ms on my PC
DirectSound often is around 20ms for most AC97 hardware (so I've
read), but is still too high for real time audio manipulation.
Core Audio (apple) is advertised at being able to be .2 to 1.5ms!
ASIO drivers for audio production usually are below 11ms, typically
hitting 5ms.

These new devices have _plenty_ of horsepower to copy a small array of
bytes (2,822 for stereo 44.1k) over to a hardware buffer every 16ms
without interruption. Apple has been doing it since the first iPhone
without problem, and that is now only a fraction of the speed of the
current batch of hardware.

Also I must note that the current generation of phones is very, very
sad with how big the smallest continuous buffer is.

My Samsung Galaxy S (vibrant) is one of my highest latency devices,
despite being one of the fastest I test on. (sounds like 100ms)
My Nexus One is also very high. (sounds like 75ms)
My G1 actually has _less_ latency then both of those!
My Droid is decent. (sounds like 40ms)

When I run my own mixer, I've found that running 44.1khz stereo gives
the lowest latency via AudioTrack for most devices. It's because the
smallest allowable buffer size is the same for stereo and mono, so
obviously stereo will be half the latency. Still, the results I get
are unsuitable for most real-time games and ensuring a 45ms buffer
would _barely_ get in the ballpark of what would look like a realistic
sound response time for what you're seeing on-screen. I can't imagine
it would work well enough for a good rhythm game or a drum synth.

Just my 2 cents.

On Dec 20, 12:37 am, Olivier Guilyardi <***@samalyse.com> wrote:
> I don't think so:
>
> ------------
> "warm output latency" is defined to be the interval between when an application
> requests audio playback and when sound begins playing, when the
> audio system has been recently used but is currently idle (that is, silent)
>
> "continuous output latency" is defined to be the interval between when an
> application issues a sample to be played and when the speaker physically
> plays the corresponding sound, while the device is currently playing back audio
> ------------
>
> IIUC, "continuous" latency is the delay between the moment audio data is passed
> to the OpenSL API, and the moment it gets out of the speaker.
>
> "warm" latency just seems to be some startup time. It seems to be about resuming
> playback, in which case an internal buffer may already carry some data.
> Therefore these 10ms could be the real hardware latency, while the extra
> continuous latency may come from intermediary software layers, I suppose.
>
> Unless I'm wrong, continuous latency is the main one for us app developers.
>
> Olivier
>
> On 12/20/2010 12:27 AM, niko20 wrote:
>
>
>
>
>
>
>
> > Actually the warm output latency in the ccd is 10ms so i think it does
> > look reasonable already.
>
> > On Dec 19, 5:23 pm, niko20 <***@yahoo.com> wrote:
> >> Ageed. Low latency should be defined as 15ms or less. At most 20ms but
> >> 45 is definitely not low latency.
>
> >> On Dec 19, 3:49 am, Olivier Guilyardi <***@samalyse.com> wrote:
>
> >>> Hi,
> >>> I looked at the CDD section 5.3, and it says:
> >>> "continuous output latency of 45 milliseconds or less"
> >>> 45ms is not low latency.
> >>> Implementing a drum pad, or some synth which react to touch events, is not
> >>> realistic over 10ms latency, unless you want users to laugh at your product.
> >>> Olivier
> >>> On 12/18/2010 01:50 AM, Glenn Kasten wrote:
> >>>> The Android Compatibility Definition for 2.3 (gingerbread) is now
> >>>> published at
> >>>>http://source.android.com/compatibility/index.html
> >>>> (click on Getting Started / Current CDD). Though the CDD is directed
> >>>> primarily at device manufacturers, section 5.3 Audio Latency should
> >>>> also be of interest to app developers using the Android native audio
> >>>> APIs based on Khronos Group OpenSL ES.
> >>>> On Dec 16, 12:51 pm, Glenn Kasten <***@android.com> wrote:
> >>>>> The Nexus S phone does not claim the
> >>>>> android.hardware.audio.low_latency feature, as seen by adb shell pm
> >>>>> list features. The upcoming Android Compatibility Definition Document
> >>>>> (CDD) will describe the requirements for this feature more precisely.
> >>>>>http://developer.android.com/reference/android/content/pm/PackageMana...

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-20 11:00:48 UTC
Permalink
On 12/20/2010 10:30 AM, Robert Green wrote:
> Low latency should be defined as an amount which allows for accurate
> and believable synchronization with video. Most 2D apps can, in
> theory, run at 60FPS, assuming some degree of efficiency of execution
> on all parts. 60FPS = 0.016666ms = 16ms per frame. I believe 16ms
> would be a more practical definition of "low latency" as it would
> allow for accurate sync with video. 45ms seems like a number made up,
> hoping that it's low enough to entice OEMs to target.

Video (visual) sync is not that crucial to me. For example, with a drum pad,
what's important is that the user doesn't notice any delay between the moment he
or she tap the screen and the moment the sounds gets out of the speaker. 10ms is
a *maximum* for this.

I'm not a game developer, but I suppose that similar constraints apply.

16ms, that's about 700 frames at 44.1Khz. Again, that really can't be called low
latency to me.

> Comparisons:

[...]

> Apple has been doing it since the first iPhone
> without problem, and that is now only a fraction of the speed of the
> current batch of hardware.

>From what I read, you can get 5ms on the iPhone. That's certainly why serious
audio software companies such as Izotope are targeting this platform. Check
their iDrum app to understand what can't be done on Android currently.

[...]
> When I run my own mixer, I've found that running 44.1khz stereo gives
> the lowest latency via AudioTrack for most devices. It's because the
> smallest allowable buffer size is the same for stereo and mono, so
> obviously stereo will be half the latency. Still, the results I get
> are unsuitable for most real-time games and ensuring a 45ms buffer
> would _barely_ get in the ballpark of what would look like a realistic
> sound response time for what you're seeing on-screen. I can't imagine
> it would work well enough for a good rhythm game or a drum synth.

I agree.

Is there an OS design problem?

--
Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
mic _
2010-12-20 14:53:25 UTC
Permalink
I haven't looked at the OpenSL implementation, but one thing I'll point out
is this: avoid starting and stopping audio outputs frequently. The
AudioPolicyManager will assume that it needs to reroute the audio when
output starts, and each reroute can take a relatively long time (as for why
it assumes this; that's partly because applications implicitly can make
changes in the AudioHardware state without the policy manager's knowledge).

/Michael

On Mon, Dec 20, 2010 at 12:00 PM, Olivier Guilyardi <***@samalyse.com>wrote:

> On 12/20/2010 10:30 AM, Robert Green wrote:
> > Low latency should be defined as an amount which allows for accurate
> > and believable synchronization with video. Most 2D apps can, in
> > theory, run at 60FPS, assuming some degree of efficiency of execution
> > on all parts. 60FPS = 0.016666ms = 16ms per frame. I believe 16ms
> > would be a more practical definition of "low latency" as it would
> > allow for accurate sync with video. 45ms seems like a number made up,
> > hoping that it's low enough to entice OEMs to target.
>
> Video (visual) sync is not that crucial to me. For example, with a drum
> pad,
> what's important is that the user doesn't notice any delay between the
> moment he
> or she tap the screen and the moment the sounds gets out of the speaker.
> 10ms is
> a *maximum* for this.
>
> I'm not a game developer, but I suppose that similar constraints apply.
>
> 16ms, that's about 700 frames at 44.1Khz. Again, that really can't be
> called low
> latency to me.
>
> > Comparisons:
>
> [...]
>
> > Apple has been doing it since the first iPhone
> > without problem, and that is now only a fraction of the speed of the
> > current batch of hardware.
>
> From what I read, you can get 5ms on the iPhone. That's certainly why
> serious
> audio software companies such as Izotope are targeting this platform. Check
> their iDrum app to understand what can't be done on Android currently.
>
> [...]
> > When I run my own mixer, I've found that running 44.1khz stereo gives
> > the lowest latency via AudioTrack for most devices. It's because the
> > smallest allowable buffer size is the same for stereo and mono, so
> > obviously stereo will be half the latency. Still, the results I get
> > are unsuitable for most real-time games and ensuring a 45ms buffer
> > would _barely_ get in the ballpark of what would look like a realistic
> > sound response time for what you're seeing on-screen. I can't imagine
> > it would work well enough for a good rhythm game or a drum synth.
>
> I agree.
>
> Is there an OS design problem?
>
> --
> Olivier
>
> --
> You received this message because you are subscribed to the Google Groups
> "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com.
> To unsubscribe from this group, send email to
> android-ndk+***@googlegroups.com<android-ndk%***@googlegroups.com>
> .
> For more options, visit this group at
> http://groups.google.com/group/android-ndk?hl=en.
>
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
niko20
2010-12-20 15:37:26 UTC
Permalink
Unfortunately this is a problem I think we'll have to live with for a
while. There is no drum pad app where users don't say "too much
latency", and it's not in our control. Even if we find a device which
has low latency the next one may not meet the same standard. It's just
the nature of the beast (Open SL , open standards). To target the
largest audience for audio apps that require low latency we'd probably
have to go with iPhone for quite a while yet.

-niko

On Dec 20, 8:53 am, mic _ <***@gmail.com> wrote:
> I haven't looked at the OpenSL implementation, but one thing I'll point out
> is this: avoid starting and stopping audio outputs frequently. The
> AudioPolicyManager will assume that it needs to reroute the audio when
> output starts, and each reroute can take a relatively long time (as for why
> it assumes this; that's partly because applications implicitly can make
> changes in the AudioHardware state without the policy manager's knowledge).
>
> /Michael
>
> On Mon, Dec 20, 2010 at 12:00 PM, Olivier Guilyardi <***@samalyse.com>wrote:
>
>
>
>
>
>
>
> > On 12/20/2010 10:30 AM, Robert Green wrote:
> > > Low latency should be defined as an amount which allows for accurate
> > > and believable synchronization with video.  Most 2D apps can, in
> > > theory, run at 60FPS, assuming some degree of efficiency of execution
> > > on all parts.  60FPS = 0.016666ms = 16ms per frame.  I believe 16ms
> > > would be a more practical definition of "low latency" as it would
> > > allow for accurate sync with video.  45ms seems like a number made up,
> > > hoping that it's low enough to entice OEMs to target.
>
> > Video (visual) sync is not that crucial to me. For example, with a drum
> > pad,
> > what's important is that the user doesn't notice any delay between the
> > moment he
> > or she tap the screen and the moment the sounds gets out of the speaker.
> > 10ms is
> > a *maximum* for this.
>
> > I'm not a game developer, but I suppose that similar constraints apply.
>
> > 16ms, that's about 700 frames at 44.1Khz. Again, that really can't be
> > called low
> > latency to me.
>
> > > Comparisons:
>
> > [...]
>
> > > Apple has been doing it since the first iPhone
> > > without problem, and that is now only a fraction of the speed of the
> > > current batch of hardware.
>
> > From what I read, you can get 5ms on the iPhone. That's certainly why
> > serious
> > audio software companies such as Izotope are targeting this platform. Check
> > their iDrum app to understand what can't be done on Android currently.
>
> > [...]
> > > When I run my own mixer, I've found that running 44.1khz stereo gives
> > > the lowest latency via AudioTrack for most devices.  It's because the
> > > smallest allowable buffer size is the same for stereo and mono, so
> > > obviously stereo will be half the latency.  Still, the results I get
> > > are unsuitable for most real-time games and ensuring a 45ms buffer
> > > would _barely_ get in the ballpark of what would look like a realistic
> > > sound response time for what you're seeing on-screen.  I can't imagine
> > > it would work well enough for a good rhythm game or a drum synth.
>
> > I agree.
>
> > Is there an OS design problem?
>
> > --
> >   Olivier
>
> > --
> > You received this message because you are subscribed to the Google Groups
> > "android-ndk" group.
> > To post to this group, send email to android-***@googlegroups.com.
> > To unsubscribe from this group, send email to
> > android-ndk+***@googlegroups.com<android-ndk%***@googlegr oups.com>
> > .
> > For more options, visit this group at
> >http://groups.google.com/group/android-ndk?hl=en.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-20 16:24:15 UTC
Permalink
I quite agree. Considering both this CDD and device fragmentation, one must be
careful about which app and features to develop when working on Android.

If you take the iPhone iDrum app which I mentioned, most of it can actually be
implemented on Android. The only feature which is problematic is the pad which
allows to play over, or record, a drum sequence in real time.

And that's quite a central feature, but the app would be usable without this
pad. Now, whether it would be popular is another question... I suppose it would.

Olivier

On 12/20/2010 04:37 PM, niko20 wrote:
> Unfortunately this is a problem I think we'll have to live with for a
> while. There is no drum pad app where users don't say "too much
> latency", and it's not in our control. Even if we find a device which
> has low latency the next one may not meet the same standard. It's just
> the nature of the beast (Open SL , open standards). To target the
> largest audience for audio apps that require low latency we'd probably
> have to go with iPhone for quite a while yet.
>
> -niko
>
> On Dec 20, 8:53 am, mic _ <***@gmail.com> wrote:
>> I haven't looked at the OpenSL implementation, but one thing I'll point out
>> is this: avoid starting and stopping audio outputs frequently. The
>> AudioPolicyManager will assume that it needs to reroute the audio when
>> output starts, and each reroute can take a relatively long time (as for why
>> it assumes this; that's partly because applications implicitly can make
>> changes in the AudioHardware state without the policy manager's knowledge).
>>
>> /Michael
>>
>> On Mon, Dec 20, 2010 at 12:00 PM, Olivier Guilyardi <***@samalyse.com>wrote:
>>
>>
>>
>>
>>
>>
>>
>>> On 12/20/2010 10:30 AM, Robert Green wrote:
>>>> Low latency should be defined as an amount which allows for accurate
>>>> and believable synchronization with video. Most 2D apps can, in
>>>> theory, run at 60FPS, assuming some degree of efficiency of execution
>>>> on all parts. 60FPS = 0.016666ms = 16ms per frame. I believe 16ms
>>>> would be a more practical definition of "low latency" as it would
>>>> allow for accurate sync with video. 45ms seems like a number made up,
>>>> hoping that it's low enough to entice OEMs to target.
>>> Video (visual) sync is not that crucial to me. For example, with a drum
>>> pad,
>>> what's important is that the user doesn't notice any delay between the
>>> moment he
>>> or she tap the screen and the moment the sounds gets out of the speaker.
>>> 10ms is
>>> a *maximum* for this.
>>> I'm not a game developer, but I suppose that similar constraints apply.
>>> 16ms, that's about 700 frames at 44.1Khz. Again, that really can't be
>>> called low
>>> latency to me.
>>>> Comparisons:
>>> [...]
>>>> Apple has been doing it since the first iPhone
>>>> without problem, and that is now only a fraction of the speed of the
>>>> current batch of hardware.
>>> From what I read, you can get 5ms on the iPhone. That's certainly why
>>> serious
>>> audio software companies such as Izotope are targeting this platform. Check
>>> their iDrum app to understand what can't be done on Android currently.
>>> [...]
>>>> When I run my own mixer, I've found that running 44.1khz stereo gives
>>>> the lowest latency via AudioTrack for most devices. It's because the
>>>> smallest allowable buffer size is the same for stereo and mono, so
>>>> obviously stereo will be half the latency. Still, the results I get
>>>> are unsuitable for most real-time games and ensuring a 45ms buffer
>>>> would _barely_ get in the ballpark of what would look like a realistic
>>>> sound response time for what you're seeing on-screen. I can't imagine
>>>> it would work well enough for a good rhythm game or a drum synth.
>>> I agree.
>>> Is there an OS design problem?
>>> --
>>> Olivier
>>> --
>>> You received this message because you are subscribed to the Google Groups
>>> "android-ndk" group.
>>> To post to this group, send email to android-***@googlegroups.com.
>>> To unsubscribe from this group, send email to
>>> android-ndk+***@googlegroups.com<android-ndk%***@googlegr oups.com>
>>> .
>>> For more options, visit this group at
>>> http://groups.google.com/group/android-ndk?hl=en.
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Ross Bencina
2010-12-21 12:01:52 UTC
Permalink
On Dec 21, 3:24 am, Olivier Guilyardi <***@samalyse.com> wrote:
> I quite agree. Considering both this CDD and device fragmentation, one must be
> careful about which app and features to develop when working on Android.

Well the main problem is the CDD no? Google should fix it to something
realistic like 2ms or 5ms for each direction.

The latency is most likely not in the hardware anyway. It may be in
the drivers, and it could easily be in the intervening software stack
(remember 10ms latency was possible on Windows 95, but then MS
inserted "kmixer" and added 40ms of latency until Windows Vista)..
this is something Google could show leadership on and make sure the
Android audio path was low-latency (like Apple does).

I'd be interested if anyone has tried to bypass AudioFlinger and deal
direct with AudioHardwareInterface as a test to see where the latency
is.



--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-21 21:31:19 UTC
Permalink
On 12/21/2010 01:01 PM, Ross Bencina wrote:
> On Dec 21, 3:24 am, Olivier Guilyardi <***@samalyse.com> wrote:
>> I quite agree. Considering both this CDD and device fragmentation, one must be
>> careful about which app and features to develop when working on Android.
>
> Well the main problem is the CDD no? Google should fix it to something
> realistic like 2ms or 5ms for each direction.

This CDD makes it hard to deal with device fragmentation.

It says that with 45ms or less, the device can be considered to feature
android.hardware.audio.low_latency.

Therefore, you may have some devices which support much lower latency, suitable
for advanced games and audio apps. But some device may be around the 45ms limit
and still report themselves as featuring low latency.

In this situation, the corresponding <uses-feature> tag is unreliable.

I believe that things should be called by their name, and 45ms is not low
latency. Maybe it was in the 50's, but I can't tell, I wasn't born.

10ms sounds ok to me. You can't expect a professional zero-latency DAW anyway.

--
Olivier
.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Glenn Kasten
2010-12-22 00:44:53 UTC
Permalink
I can't comment on unreleased devices.

The CDD is an evolving document with each Android release. The
requirements are adjusted as the platform and devices advance.

I appreciate everyone's feedback about the CDD audio latency section
in particular -- the numbers, definitions, test conditions,
performance of various devices, application use cases, etc.

To clarify a point in the CDD ... the difference between warm output
latency and continuous output latency represents the maximum time that
can ascribed to user-mode buffers. The warm output latency represents
the maximum time that can be ascribed to device driver and hardware
buffers alone, that is when the user-mode buffers are empty.

Audio latency is clearly a topic of great interest, and your inputs
are valuable.

On Dec 21, 1:31 pm, Olivier Guilyardi <***@samalyse.com> wrote:
> On 12/21/2010 01:01 PM, Ross Bencina wrote:
>
> > On Dec 21, 3:24 am, Olivier Guilyardi <***@samalyse.com> wrote:
> >> I quite agree. Considering both this CDD and device fragmentation, one must be
> >> careful about which app and features to develop when working on Android.
>
> > Well the main problem is the CDD no? Google should fix it to something
> > realistic like 2ms or 5ms for each direction.
>
> This CDD makes it hard to deal with device fragmentation.
>
> It says that with 45ms or less, the device can be considered to feature
> android.hardware.audio.low_latency.
>
> Therefore, you may have some devices which support much lower latency, suitable
> for advanced games and audio apps. But some device may be around the 45ms limit
> and still report themselves as featuring low latency.
>
> In this situation, the corresponding <uses-feature> tag is unreliable.
>
> I believe that things should be called by their name, and 45ms is not low
> latency. Maybe it was in the 50's, but I can't tell, I wasn't born.
>
> 10ms sounds ok to me. You can't expect a professional zero-latency DAW anyway.
>
> --
>   Olivier
> .

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Ross Bencina
2010-12-22 10:17:22 UTC
Permalink
On Dec 22, 11:44 am, Glenn Kasten <***@android.com> wrote:
> To clarify a point in the CDD ... the difference between warm output
> latency and continuous output latency represents the maximum time that
> can ascribed to user-mode buffers. The warm output latency represents
> the maximum time that can be ascribed to device driver and hardware
> buffers alone, that is when the user-mode buffers are empty.

Hi Glenn

Nice to hear from you here.

I'm not really sure I understand the buffering model that underlies
your statements above (or the CDD) is this documented somewhere? I'm
used to thinking of audio buffering as a ring buffer or double buffer
(some kind of shared buffer anyway) where the user is writing new data
into the buffer and the hardware is playing it back out of the other
end of the buffer (at least conceptually, perhaps there is a driver
thread copying it into a hardware buffer). Latency is usually
determined by how big the buffer needs to be to mask client and driver
scheduling jitter. I'm not sure where your concept of user-mode
buffers fits into this picture -- are they some immutable concept
within the AudioFlinger architecture?

Thanks :-)

Ross.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
niko20
2010-12-22 20:49:19 UTC
Permalink
The hardware should just allow reading straight from a user buffer.
Ill take care of keeping it full. The would be the lowest latency.

On Dec 22, 4:17 am, Ross Bencina <***@audiomulch.com> wrote:
> On Dec 22, 11:44 am, Glenn Kasten <***@android.com> wrote:
>
> > To clarify a point in the CDD ... the difference between warm output
> > latency and continuous output latency represents the maximum time that
> > can ascribed to user-mode buffers. The warm output latency represents
> > the maximum time that can be ascribed to device driver and hardware
> > buffers alone, that is when the user-mode buffers are empty.
>
> Hi Glenn
>
> Nice to hear from you here.
>
> I'm not really sure I understand the buffering model that underlies
> your statements above (or the CDD) is this documented somewhere? I'm
> used to thinking of audio buffering as a ring buffer or double buffer
> (some kind of shared buffer anyway) where the user is writing new data
> into the buffer and the hardware is playing it back out of the other
> end of the buffer (at least conceptually, perhaps there is a driver
> thread copying it into a hardware buffer). Latency is usually
> determined by how big the buffer needs to be to mask client and driver
> scheduling jitter. I'm not sure where your concept of user-mode
> buffers fits into this picture -- are they some immutable concept
> within the AudioFlinger architecture?
>
> Thanks :-)
>
> Ross.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
niko20
2010-12-22 20:54:18 UTC
Permalink
I mean I don't think its rocketscience . Just turn on the audio
hardware when a play command comes in. It then just starts reading
from a user mode buffer and outputting straight to the d/a converter
to the speaker. If we have access directly to the last buffer in the
layers we should be able to get low latency

On Dec 22, 2:49 pm, niko20 <***@yahoo.com> wrote:
> The hardware should just allow reading straight from a user buffer.
> Ill take care of keeping it full. The would be the lowest latency.
>
> On Dec 22, 4:17 am, Ross Bencina <***@audiomulch.com> wrote:
>
>
>
> > On Dec 22, 11:44 am, Glenn Kasten <***@android.com> wrote:
>
> > > To clarify a point in the CDD ... the difference between warm output
> > > latency and continuous output latency represents the maximum time that
> > > can ascribed to user-mode buffers. The warm output latency represents
> > > the maximum time that can be ascribed to device driver and hardware
> > > buffers alone, that is when the user-mode buffers are empty.
>
> > Hi Glenn
>
> > Nice to hear from you here.
>
> > I'm not really sure I understand the buffering model that underlies
> > your statements above (or the CDD) is this documented somewhere? I'm
> > used to thinking of audio buffering as a ring buffer or double buffer
> > (some kind of shared buffer anyway) where the user is writing new data
> > into the buffer and the hardware is playing it back out of the other
> > end of the buffer (at least conceptually, perhaps there is a driver
> > thread copying it into a hardware buffer). Latency is usually
> > determined by how big the buffer needs to be to mask client and driver
> > scheduling jitter. I'm not sure where your concept of user-mode
> > buffers fits into this picture -- are they some immutable concept
> > within the AudioFlinger architecture?
>
> > Thanks :-)
>
> > Ross.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
niko20
2010-12-22 20:57:19 UTC
Permalink
Don't mean to spam butas an example at 44khz if I ask for audiobuffer
min size I get 16384 samples. That's 326ms. That value to me seems
like an intermediate buffer somewhere in audioflinger or something.

On Dec 22, 2:54 pm, niko20 <***@yahoo.com> wrote:
> I mean I don't think its rocketscience . Just turn on the audio
> hardware when a play command comes in. It then just starts reading
> from a user mode buffer and outputting straight to the d/a converter
> to the speaker. If we have access directly to the last buffer in the
> layers we should be able to get low latency
>
> On Dec 22, 2:49 pm, niko20 <***@yahoo.com> wrote:
>
>
>
> > The hardware should just allow reading straight from a user buffer.
> > Ill take care of keeping it full. The would be the lowest latency.
>
> > On Dec 22, 4:17 am, Ross Bencina <***@audiomulch.com> wrote:
>
> > > On Dec 22, 11:44 am, Glenn Kasten <***@android.com> wrote:
>
> > > > To clarify a point in the CDD ... the difference between warm output
> > > > latency and continuous output latency represents the maximum time that
> > > > can ascribed to user-mode buffers. The warm output latency represents
> > > > the maximum time that can be ascribed to device driver and hardware
> > > > buffers alone, that is when the user-mode buffers are empty.
>
> > > Hi Glenn
>
> > > Nice to hear from you here.
>
> > > I'm not really sure I understand the buffering model that underlies
> > > your statements above (or the CDD) is this documented somewhere? I'm
> > > used to thinking of audio buffering as a ring buffer or double buffer
> > > (some kind of shared buffer anyway) where the user is writing new data
> > > into the buffer and the hardware is playing it back out of the other
> > > end of the buffer (at least conceptually, perhaps there is a driver
> > > thread copying it into a hardware buffer). Latency is usually
> > > determined by how big the buffer needs to be to mask client and driver
> > > scheduling jitter. I'm not sure where your concept of user-mode
> > > buffers fits into this picture -- are they some immutable concept
> > > within the AudioFlinger architecture?
>
> > > Thanks :-)
>
> > > Ross.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Ross Bencina
2010-12-23 05:22:42 UTC
Permalink
On Dec 23, 7:54 am, niko20 <***@yahoo.com> wrote:
> I mean I don't think its rocketscience . Just turn on the audio
> hardware when a play command comes in. It then just starts reading
> from a user mode buffer and outputting straight to the d/a converter
> to the speaker. If we have access directly to the last buffer in the
> layers we should be able to get low latency

I think we need to be careful about what we're talking about here.
You're talking about exclusive direct access to the audio hardware.
That's a great way to get lowest latency, but you loose services that
a phone needs to provide (like ringing when a call comes in).
Generally OSes these days have a mixer layer that sits between apps
and the hardware. This can still be done with low latency.

I'm not asking for direct hardware access. I'm just asking for a low
latency raw audio buffer api.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Ross Bencina
2010-12-23 05:24:23 UTC
Permalink
On Dec 23, 7:49 am, niko20 <***@yahoo.com> wrote:
> The hardware should just allow reading straight from a user buffer.
> Ill take care of keeping it full. The would be the lowest latency.

You probably want to OS to take care of scheduling your feeder thread
too (ie an asynchronous callback interface for buffer fill/process)
otherwise you're going to have portability issues trying to manage the
scheduling yourself.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Robert Green
2010-12-23 09:29:52 UTC
Permalink
So far my favorite style of API is one in which you implement a
callback and tell the OS to start pulling audio and how much to pull
each time (so maybe 2048 bytes).. Then the OS runs its own thread
calling your "fillBuffer" method for chunks of 2048 bytes.

It's a super easy way to make for a reliable system in which you don't
depend on the 3rd party app to push any data, you just pull from it.

After that it's got the bytes from your app and any other system
services so after a quick additive mix/resample/interleave, it's
straight to /dev/pcm and we're done.

On Dec 22, 9:24 pm, Ross Bencina <***@audiomulch.com> wrote:
> On Dec 23, 7:49 am, niko20 <***@yahoo.com> wrote:
>
> > The hardware should just allow reading straight from a user buffer.
> > Ill take care of keeping it full. The would be the lowest latency.
>
> You probably want to OS to take care of scheduling your feeder thread
> too (ie an asynchronous callback interface for buffer fill/process)
> otherwise you're going to have portability issues trying to manage the
> scheduling yourself.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Glenn Kasten
2010-12-24 03:45:04 UTC
Permalink
The lowest level APIs in Android native audio are based on the "buffer
queue" concept from OpenSL ES, which itself was influenced by other
audio APIs. The documentation for Android native audio, in android-ndk-
r5/docs/opensles/index.html, discusses audio output and input via PCM
buffer queues. For an example of low-level PCM audio output, please
see the buffer queue parts of the sample code at android-ndk-r5/
samples/native-audio. In addition to the Android-specific
documentation and sample code, you'll also want to read about buffer
queues in the Khronos Group OpenSL ES 1.0.1 specification.

As someone else noted, audio output via PCM buffer queues is routed
through the system mixer aka "AudioFlinger" prior to output to the
device driver layers, but it does not pass through a decode layer.

This group android-ndk is primarily for application developers using
the Android NDK. However, since there is a lot of interest here in
improving the platform, I'll make a few comments here on that
topic ...

Regarding "patches not accepted", I'm not aware of any previous
policy, so I can't comment on that.

However, to quote a well-worn phrase, the current state is "quality
patches welcomed :-)". Generally "quality" means good code that is
easy to review, well tested, won't break the platform portability or
compatibility, etc. I suggest that contributors read
http://source.android.com/source/index.html to become familiar with
building the platform and the patch process. Also see
http://source.android.com/community/ especially "Open Source Project
discussions". Then find a group (e.g. android-contrib), post your
proposal there and review it with other contributors before investing
a large amount of time. When it makes sense, submit a smallish patch
to start. There's no guarantee that a given patch will be accepted,
but if you have reviewed your proposal early, gotten positive
feedback, and then submit a series of small high quality patches, it
is more likely that they can be accepted.

I can't comment on the audio stack roadmap, but if you post a proposal
early on android-contrib, someone should alert you if the proposal is
in an area that is likely to conflict with other work.

I encourage contributors who want to make Android low-level audio
better to make improvements that use or extend the OpenSL ES based
APIs, and the parts of the stack below those APIs (i.e. AudioFlinger).
We're unlikely to accept a contribution that introduces yet another
native audio API.

On Dec 23, 1:29 am, Robert Green <***@gmail.com> wrote:
> So far my favorite style of API is one in which you implement a
> callback and tell the OS to start pulling audio and how much to pull
> each time (so maybe 2048 bytes).. Then the OS runs its own thread
> calling your "fillBuffer" method for chunks of 2048 bytes.
>
> It's a super easy way to make for a reliable system in which you don't
> depend on the 3rd party app to push any data, you just pull from it.
>
> After that it's got the bytes from your app and any other system
> services so after a quick additive mix/resample/interleave, it's
> straight to /dev/pcm and we're done.
>
> On Dec 22, 9:24 pm, Ross Bencina <***@audiomulch.com> wrote:
>
> > On Dec 23, 7:49 am, niko20 <***@yahoo.com> wrote:
>
> > > The hardware should just allow reading straight from a user buffer.
> > > Ill take care of keeping it full. The would be the lowest latency.
>
> > You probably want to OS to take care of scheduling your feeder thread
> > too (ie an asynchronous callback interface for buffer fill/process)
> > otherwise you're going to have portability issues trying to manage the
> > scheduling yourself.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Angus Lees
2010-12-24 04:14:32 UTC
Permalink
Glenn, any comment on reducing the CDD requirement for declaring a
device as "low latency" from the current 45ms to (strawman) 8ms?
Since we don't have any such devices on the market yet, now seems the
right time to address this particular aspect of the problem.

Otherwise we are designing in a need for a rather silly
"android.hardware.audio.actually_low_latency" so app authors can
differentiate genuinely low latency audio devices from these 45ms
ones.

How was the current 45ms threshold chosen?
Choosing the limit around some human perception threshold rather than
comparison to competing platforms seems the right thing to do - since
then the classes of features it enables will be constant over time,
regardless of hardware improvements.

- Gus

On Fri, Dec 24, 2010 at 14:45, Glenn Kasten <***@android.com> wrote:
> The lowest level APIs in Android native audio are based on the "buffer
> queue" concept from OpenSL ES, which itself was influenced by other
> audio APIs. The documentation for Android native audio, in android-ndk-
> r5/docs/opensles/index.html, discusses audio output and input via PCM
> buffer queues. For an example of low-level PCM audio output, please
> see the buffer queue parts of the sample code at android-ndk-r5/
> samples/native-audio. In addition to the Android-specific
> documentation and sample code, you'll also want to read about buffer
> queues in the Khronos Group OpenSL ES 1.0.1 specification.
>
> As someone else noted, audio output via PCM buffer queues is routed
> through the system mixer aka "AudioFlinger" prior to output to the
> device driver layers, but it does not pass through a decode layer.
>
> This group android-ndk is primarily for application developers using
> the Android NDK. However, since there is a lot of interest here in
> improving the platform, I'll make a few comments here on that
> topic ...
>
> Regarding "patches not accepted", I'm not aware of any previous
> policy, so I can't comment on that.
>
> However, to quote a well-worn phrase, the current state is "quality
> patches welcomed :-)". Generally "quality" means good code that is
> easy to review, well tested, won't break the platform portability or
> compatibility, etc. I suggest that contributors read
> http://source.android.com/source/index.html to become familiar with
> building the platform and the patch process. Also see
> http://source.android.com/community/ especially "Open Source Project
> discussions". Then find a group (e.g. android-contrib), post your
> proposal there and review it with other contributors before investing
> a large amount of time. When it makes sense, submit a smallish patch
> to start. There's no guarantee that a given patch will be accepted,
> but if you have reviewed your proposal early, gotten positive
> feedback, and then submit a series of small high quality patches, it
> is more likely that they can be accepted.
>
> I can't comment on the audio stack roadmap, but if you post a proposal
> early on android-contrib, someone should alert you if the proposal is
> in an area that is likely to conflict with other work.
>
> I encourage contributors who want to make Android low-level audio
> better to make improvements that use or extend the OpenSL ES based
> APIs, and the parts of the stack below those APIs (i.e. AudioFlinger).
> We're unlikely to accept a contribution that introduces yet another
> native audio API.
>
> On Dec 23, 1:29 am, Robert Green <***@gmail.com> wrote:
>> So far my favorite style of API is one in which you implement a
>> callback and tell the OS to start pulling audio and how much to pull
>> each time (so maybe 2048 bytes).. Then the OS runs its own thread
>> calling your "fillBuffer" method for chunks of 2048 bytes.
>>
>> It's a super easy way to make for a reliable system in which you don't
>> depend on the 3rd party app to push any data, you just pull from it.
>>
>> After that it's got the bytes from your app and any other system
>> services so after a quick additive mix/resample/interleave, it's
>> straight to /dev/pcm and we're done.
>>
>> On Dec 22, 9:24 pm, Ross Bencina <***@audiomulch.com> wrote:
>>
>> > On Dec 23, 7:49 am, niko20 <***@yahoo.com> wrote:
>>
>> > > The hardware should just allow reading straight from a user buffer.
>> > > Ill take care of keeping it full. The would be the lowest latency.
>>
>> > You probably want to OS to take care of scheduling your feeder thread
>> > too (ie an asynchronous callback interface for buffer fill/process)
>> > otherwise you're going to have portability issues trying to manage the
>> > scheduling yourself.
>
> --
> You received this message because you are subscribed to the Google Groups "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com.
> To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
>
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-24 13:22:10 UTC
Permalink
On 12/24/2010 04:45 AM, Glenn Kasten wrote:

> Regarding "patches not accepted", I'm not aware of any previous
> policy, so I can't comment on that.

Here we go:
http://groups.google.com/group/android-ndk/msg/9f6e46b39f4f1fae

> However, to quote a well-worn phrase, the current state is "quality
> patches welcomed :-)". Generally "quality" means good code that is
> easy to review, well tested, won't break the platform portability or
> compatibility, etc.

Thanks for the update. That's interesting. I know we're not on android-contrib,
but is the audio stack git head up-to-date?

--
Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Robert Green
2010-12-25 20:25:58 UTC
Permalink
Just to throw my 2 cents in here again.. I agree that bells and
whistles mean nothing without low latency audio. If there were only 1
priority of what to do with Android audio, it would be to guarantee
developers a low latency spot to dump PCM data. End of story!
Everything else can be developed on top of that, but without it, all
the effects in the world are mute because users will be experiencing
unsynchronized AV.

Also whoever posted that his audio was high latency because of Java
was incorrect IMO. It's not java's fault that the OEM put a 100ms min
buffer on the audio track for their device. My sound mixer's latency
is acceptable on a Droid and horribly bad on a Galaxy S and somewhere
in between on a Nexus One, showing that it's mostly the OEM's
implementations and decisions. In the current Audiotrack
implementation, all Java has to do is copy a byte array back to native
and that certainly doesn't take over 1ms, so I don't accept that as
the defining problem, though of course it would be better to keep it
native in the first place.

Anyway, point is. Latency matters more than anything IMO. With low
latency PCM input, I don't need an effect library, a mixer, a
soundfont player, or anything. I can use open source stuff, write my
own, or someone can add it later to the project. Without it, nothing
will sound right, no matter how hard we try! I'm not trying to do
anything particularly special with audio, either. I just want to
produce high quality games!

On Dec 24, 5:22 am, Olivier Guilyardi <***@samalyse.com> wrote:
> On 12/24/2010 04:45 AM, Glenn Kasten wrote:
>
> > Regarding "patches not accepted", I'm not aware of any previous
> > policy, so I can't comment on that.
>
> Here we go:http://groups.google.com/group/android-ndk/msg/9f6e46b39f4f1fae
>
> > However, to quote a well-worn phrase, the current state is "quality
> > patches welcomed :-)". Generally "quality" means good code that is
> > easy to review, well tested, won't break the platform portability or
> > compatibility, etc.
>
> Thanks for the update. That's interesting. I know we're not on android-contrib,
> but is the audio stack git head up-to-date?
>
> --
>   Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Robert Green
2010-12-26 11:11:13 UTC
Permalink
And when I said "mute," I meant moot. All the audio-talk was messing
with my vocab. :)

On Dec 25, 12:25 pm, Robert Green <***@gmail.com> wrote:
> Just to throw my 2 cents in here again.. I agree that bells and
> whistles mean nothing without low latency audio.  If there were only 1
> priority of what to do with Android audio, it would be to guarantee
> developers a low latency spot to dump PCM data.  End of story!
> Everything else can be developed on top of that, but without it, all
> the effects in the world are mute because users will be experiencing
> unsynchronized AV.
>
> Also whoever posted that his audio was high latency because of Java
> was incorrect IMO.  It's not java's fault that the OEM put a 100ms min
> buffer on the audio track for their device.  My sound mixer's latency
> is acceptable on a Droid and horribly bad on a Galaxy S and somewhere
> in between on a Nexus One, showing that it's mostly the OEM's
> implementations and decisions.  In the current Audiotrack
> implementation, all Java has to do is copy a byte array back to native
> and that certainly doesn't take over 1ms, so I don't accept that as
> the defining problem, though of course it would be better to keep it
> native in the first place.
>
> Anyway, point is. Latency matters more than anything IMO.  With low
> latency PCM input, I don't need an effect library, a mixer, a
> soundfont player, or anything.  I can use open source stuff, write my
> own, or someone can add it later to the project.  Without it, nothing
> will sound right, no matter how hard we try!  I'm not trying to do
> anything particularly special with audio, either.  I just want to
> produce high quality games!
>
> On Dec 24, 5:22 am, Olivier Guilyardi <***@samalyse.com> wrote:
>
>
>
>
>
>
>
> > On 12/24/2010 04:45 AM, Glenn Kasten wrote:
>
> > > Regarding "patches not accepted", I'm not aware of any previous
> > > policy, so I can't comment on that.
>
> > Here we go:http://groups.google.com/group/android-ndk/msg/9f6e46b39f4f1fae
>
> > > However, to quote a well-worn phrase, the current state is "quality
> > > patches welcomed :-)". Generally "quality" means good code that is
> > > easy to review, well tested, won't break the platform portability or
> > > compatibility, etc.
>
> > Thanks for the update. That's interesting. I know we're not on android-contrib,
> > but is the audio stack git head up-to-date?
>
> > --
> >   Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-26 11:55:59 UTC
Permalink
Robert, I posted this link so that Glenn /learns/ about the "previous policy"
applied by his own team, regarding patches to the audio stack...

I didn't mean to focus on the Java vs native discussion. It's OT here IMO.

About bells and whistles in OpenSL, effect libraries, etc.. I do agree with you.
All of these application-level features are out of scope, when the current audio
API fails to provide reliable core I/O functionality.

I am sure that working with so many manufacturers is a big challenge, and in
this context all efforts should IMO focus on providing reliable audio
input/output, instead of bringing extra complexity and high-level features to
the audio stack.

Olivier

Le 25/12/10 21:25, Robert Green a écrit :
> Just to throw my 2 cents in here again.. I agree that bells and
> whistles mean nothing without low latency audio. If there were only 1
> priority of what to do with Android audio, it would be to guarantee
> developers a low latency spot to dump PCM data. End of story!
> Everything else can be developed on top of that, but without it, all
> the effects in the world are mute because users will be experiencing
> unsynchronized AV.
>
> Also whoever posted that his audio was high latency because of Java
> was incorrect IMO. It's not java's fault that the OEM put a 100ms min
> buffer on the audio track for their device. My sound mixer's latency
> is acceptable on a Droid and horribly bad on a Galaxy S and somewhere
> in between on a Nexus One, showing that it's mostly the OEM's
> implementations and decisions. In the current Audiotrack
> implementation, all Java has to do is copy a byte array back to native
> and that certainly doesn't take over 1ms, so I don't accept that as
> the defining problem, though of course it would be better to keep it
> native in the first place.
>
> Anyway, point is. Latency matters more than anything IMO. With low
> latency PCM input, I don't need an effect library, a mixer, a
> soundfont player, or anything. I can use open source stuff, write my
> own, or someone can add it later to the project. Without it, nothing
> will sound right, no matter how hard we try! I'm not trying to do
> anything particularly special with audio, either. I just want to
> produce high quality games!
>
> On Dec 24, 5:22 am, Olivier Guilyardi<***@samalyse.com> wrote:
>> On 12/24/2010 04:45 AM, Glenn Kasten wrote:
>>
>>> Regarding "patches not accepted", I'm not aware of any previous
>>> policy, so I can't comment on that.
>>
>> Here we go:http://groups.google.com/group/android-ndk/msg/9f6e46b39f4f1fae
>>
>>> However, to quote a well-worn phrase, the current state is "quality
>>> patches welcomed :-)". Generally "quality" means good code that is
>>> easy to review, well tested, won't break the platform portability or
>>> compatibility, etc.
>>
>> Thanks for the update. That's interesting. I know we're not on android-contrib,
>> but is the audio stack git head up-to-date?
>>
>> --
>> Olivier
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Patrick Flanagan
2010-12-26 18:01:56 UTC
Permalink
Robert is right that the new effects are effectively "mute" and will
go unused without low latency, which ought to be top priority. I have
a vocal effects app 90% ready for release, but it's not going to see
the market until the latency problem is fixed.

Regarding Java vs. Native, a few researchers at ETH Zurich were able
to push latency in Java SE down to 0.5 ms. Paper here:
http://diuf.unifr.ch/pai/wiki/doku.php?id=Publications&page=publication&kind=single&ID=16.
They used RtAudio as an intermediary to talk to ALSA but only as a
means to establish portability with Windows and MacOS. In a purely
Linux environment, RtAudio wouldn't be necessary, according to the
authors. This suggests to me a native solution isn't strictly
necessary.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
mic _
2010-12-28 10:12:53 UTC
Permalink
I'll add some of my opinions, which are somewhat related to latency..

When starting a new output stream, the AudioPolicyManager will force a
re-route of the audio. Of course that's up to each vendor to change, but if
you don't force a re-route you risk having some weird issues since there are
a *lot* of possible use cases you need to handle - not to mention that the
AudioFlinger can go behind the policy manager's back and trigger some
routing in AudioHardware on its own.

To actually do the routing the AudioPolicyManager first picks a suitable
device to route the audio to based on the stream type, which accessories you
have attached etc. It then constructs a message (key/value pair) that it
sends to AudioHardware which is the class responsible for actually
enabling/disabling audio devices (among other things).

Now the problem is that the AudioPolicyManager doesn't have a direct line to
AudioHardware. Instead it sends the message to the AudioPolicyService, which
adds it to a vector that is processed by the AudioCommandThread's
threadLoop. The AudioCommandThread sends the message to the AudioSystem,
which in turn sends it to the AudioFlinger. The AudioFlinger then adds it to
another list/queue which is processed by the relevant mixer thread, which
finally sends it to the AudioHardware.

These are a lot of intermediaries for the message to pass through. And when
it finally gets to the AudioHardware you're still not done, because you
might have to enable/disable the relevant devices, set the proper
filters/volumes/etc, and in some cases turn on some amplifier (which incurs
additional delays because you want to avoid click & pops).

In my view there are too many layers in the audio chain. My main concern
isn't latency though; it's the lack of centralized resource management for
audio. *All* routing decisions, and *all* decisions about whether or not an
application should be allowed to start/continue audio playback should IMO be
centralized in some place at the lower levels, where all information about
ongoing playback, available tracks, phone state, attached accessories,
ringer mode etc is known. IIRC there used to be problems (and there may
still be) where applications could hog all the playback resources so that
when you got an incoming call there would be no ringtone. For an OS running
primarily on mobile phones that's not exactly the optimal behavior.


/Michael


On Sun, Dec 26, 2010 at 7:01 PM, Patrick Flanagan <***@gmail.com>wrote:

> Robert is right that the new effects are effectively "mute" and will
> go unused without low latency, which ought to be top priority. I have
> a vocal effects app 90% ready for release, but it's not going to see
> the market until the latency problem is fixed.
>
> Regarding Java vs. Native, a few researchers at ETH Zurich were able
> to push latency in Java SE down to 0.5 ms. Paper here:
>
> http://diuf.unifr.ch/pai/wiki/doku.php?id=Publications&page=publication&kind=single&ID=16
> .
> They used RtAudio as an intermediary to talk to ALSA but only as a
> means to establish portability with Windows and MacOS. In a purely
> Linux environment, RtAudio wouldn't be necessary, according to the
> authors. This suggests to me a native solution isn't strictly
> necessary.
>
> --
> You received this message because you are subscribed to the Google Groups
> "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com.
> To unsubscribe from this group, send email to
> android-ndk+***@googlegroups.com<android-ndk%***@googlegroups.com>
> .
> For more options, visit this group at
> http://groups.google.com/group/android-ndk?hl=en.
>
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-28 14:22:14 UTC
Permalink
On 12/28/2010 11:12 AM, mic _ wrote:
> I'll add some of my opinions, which are somewhat related to latency..
>
> When starting a new output stream, the AudioPolicyManager will force a
> re-route of the audio. Of course that's up to each vendor to change, but
> if you don't force a re-route you risk having some weird issues since
> there are a *lot* of possible use cases you need to handle - not to
> mention that the AudioFlinger can go behind the policy manager's back
> and trigger some routing in AudioHardware on its own.
>
> To actually do the routing the AudioPolicyManager first picks a suitable
> device to route the audio to based on the stream type, which accessories
> you have attached etc. It then constructs a message (key/value pair)
> that it sends to AudioHardware which is the class responsible for
> actually enabling/disabling audio devices (among other things).
>
> Now the problem is that the AudioPolicyManager doesn't have a direct
> line to AudioHardware. Instead it sends the message to the
> AudioPolicyService, which adds it to a vector that is processed by the
> AudioCommandThread's threadLoop. The AudioCommandThread sends the
> message to the AudioSystem, which in turn sends it to the AudioFlinger.
> The AudioFlinger then adds it to another list/queue which is processed
> by the relevant mixer thread, which finally sends it to the AudioHardware.
>
> These are a lot of intermediaries for the message to pass through. And
> when it finally gets to the AudioHardware you're still not done, because
> you might have to enable/disable the relevant devices, set the proper
> filters/volumes/etc, and in some cases turn on some amplifier (which
> incurs additional delays because you want to avoid click & pops).

That's interesting. The right place for this might be android-platform, but
that's interesting nevertheless. That's the kind of thing I sensed and that I
meant when I talked about "over-encapsulation".

> In my view there are too many layers in the audio chain. My main concern
> isn't latency though; it's the lack of centralized resource management
> for audio. *All* routing decisions, and *all* decisions about whether or
> not an application should be allowed to start/continue audio playback
> should IMO be centralized in some place at the lower levels, where all
> information about ongoing playback, available tracks, phone state,
> attached accessories, ringer mode etc is known. IIRC there used to be
> problems (and there may still be) where applications could hog all the
> playback resources so that when you got an incoming call there would be
> no ringtone. For an OS running primarily on mobile phones that's not
> exactly the optimal behavior.

Indeed, I can confirm that type of defect. In my app I even detect incoming
calls to stop recording when they occur. Otherwise, the whole system crashes on
several devices including the Milestone (Droid) running 2.1.

I thought that all the audio layers were mostly intended to protect the
telephony features, but they apparently don't.

--
Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
mic _
2010-12-28 14:43:57 UTC
Permalink
The latency isn't only the OS's fault though. A lot of it has to do with the
platform. I can only talk about Qualcomm platforms myself since they're the
ones I'm semi-familiar with (7227/7230), but I'm guessing they're quite
common among currently available Android phones. I know at least that for
those platforms the buffer size enforced by the hardware driver is something
like a few dozen ms for 16-bit stereo PCM, and no data is written to the
device until two buffers have been filled. So even if you manage to do the
routing faster you still have this additional latency of maybe 50 ms or more
to deal with. And lowering the buffer size can cause some serious stability
problems so it's not recommended. I don't know if their DSP has a
one-size-fits-all buffer size that it assumes all clients will use or what
the actual problem is.

/Michael

On Tue, Dec 28, 2010 at 3:22 PM, Olivier Guilyardi <***@samalyse.com>wrote:

> On 12/28/2010 11:12 AM, mic _ wrote:
> > I'll add some of my opinions, which are somewhat related to latency..
> >
> > When starting a new output stream, the AudioPolicyManager will force a
> > re-route of the audio. Of course that's up to each vendor to change, but
> > if you don't force a re-route you risk having some weird issues since
> > there are a *lot* of possible use cases you need to handle - not to
> > mention that the AudioFlinger can go behind the policy manager's back
> > and trigger some routing in AudioHardware on its own.
> >
> > To actually do the routing the AudioPolicyManager first picks a suitable
> > device to route the audio to based on the stream type, which accessories
> > you have attached etc. It then constructs a message (key/value pair)
> > that it sends to AudioHardware which is the class responsible for
> > actually enabling/disabling audio devices (among other things).
> >
> > Now the problem is that the AudioPolicyManager doesn't have a direct
> > line to AudioHardware. Instead it sends the message to the
> > AudioPolicyService, which adds it to a vector that is processed by the
> > AudioCommandThread's threadLoop. The AudioCommandThread sends the
> > message to the AudioSystem, which in turn sends it to the AudioFlinger.
> > The AudioFlinger then adds it to another list/queue which is processed
> > by the relevant mixer thread, which finally sends it to the
> AudioHardware.
> >
> > These are a lot of intermediaries for the message to pass through. And
> > when it finally gets to the AudioHardware you're still not done, because
> > you might have to enable/disable the relevant devices, set the proper
> > filters/volumes/etc, and in some cases turn on some amplifier (which
> > incurs additional delays because you want to avoid click & pops).
>
> That's interesting. The right place for this might be android-platform, but
> that's interesting nevertheless. That's the kind of thing I sensed and that
> I
> meant when I talked about "over-encapsulation".
>
> > In my view there are too many layers in the audio chain. My main concern
> > isn't latency though; it's the lack of centralized resource management
> > for audio. *All* routing decisions, and *all* decisions about whether or
> > not an application should be allowed to start/continue audio playback
> > should IMO be centralized in some place at the lower levels, where all
> > information about ongoing playback, available tracks, phone state,
> > attached accessories, ringer mode etc is known. IIRC there used to be
> > problems (and there may still be) where applications could hog all the
> > playback resources so that when you got an incoming call there would be
> > no ringtone. For an OS running primarily on mobile phones that's not
> > exactly the optimal behavior.
>
> Indeed, I can confirm that type of defect. In my app I even detect incoming
> calls to stop recording when they occur. Otherwise, the whole system
> crashes on
> several devices including the Milestone (Droid) running 2.1.
>
> I thought that all the audio layers were mostly intended to protect the
> telephony features, but they apparently don't.
>
> --
> Olivier
>
> --
> You received this message because you are subscribed to the Google Groups
> "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com.
> To unsubscribe from this group, send email to
> android-ndk+***@googlegroups.com<android-ndk%***@googlegroups.com>
> .
> For more options, visit this group at
> http://groups.google.com/group/android-ndk?hl=en.
>
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-28 15:11:34 UTC
Permalink
That shred some light, thank you. I never imagined that 50ms could be caused by
pure hardware issues on the current devices.

In addition to fixing the OS audio internals, there seem to be a real need to
ensure that proper hardware is provided, and that the audio hardware interface
developed by the manufacturer fits the job.

I guess that's where the CDD comes in, and the current one requires 10ms (warm)
hardware latency, instead of the 50ms that you mention. This could be reasonable
if a realtime audio thread in user-space code was an option. But I think that
such privileges will never be granted to apps, for security reasons.

So, an intermediary buffer, whatever that is, is needed. To avoid xruns, I think
that it must be at least double sized (?). That results in 20ms or more.

So even if the OS audio internals are greatly improved, the 10ms hardware
latency specified in the CDD will never be sufficient. I think that pure
hardware latency should be no more than 128 frames at 44Khz, that is 3ms.

Then the intermediary buffer could be something between 256 and 512 frames. That
is a maximum of 11.6ms. That would be acceptable I think, wouldn't it?

Olivier

On 12/28/2010 03:43 PM, mic _ wrote:
> The latency isn't only the OS's fault though. A lot of it has to do with
> the platform. I can only talk about Qualcomm platforms myself since
> they're the ones I'm semi-familiar with (7227/7230), but I'm guessing
> they're quite common among currently available Android phones. I know at
> least that for those platforms the buffer size enforced by the hardware
> driver is something like a few dozen ms for 16-bit stereo PCM, and no
> data is written to the device until two buffers have been filled. So
> even if you manage to do the routing faster you still have this
> additional latency of maybe 50 ms or more to deal with. And lowering the
> buffer size can cause some serious stability problems so it's not
> recommended. I don't know if their DSP has a one-size-fits-all buffer
> size that it assumes all clients will use or what the actual problem is.
>
> /Michael
>
> On Tue, Dec 28, 2010 at 3:22 PM, Olivier Guilyardi <***@samalyse.com
> <mailto:***@samalyse.com>> wrote:
>
> On 12/28/2010 11:12 AM, mic _ wrote:
> > I'll add some of my opinions, which are somewhat related to latency..
> >
> > When starting a new output stream, the AudioPolicyManager will force a
> > re-route of the audio. Of course that's up to each vendor to
> change, but
> > if you don't force a re-route you risk having some weird issues since
> > there are a *lot* of possible use cases you need to handle - not to
> > mention that the AudioFlinger can go behind the policy manager's back
> > and trigger some routing in AudioHardware on its own.
> >
> > To actually do the routing the AudioPolicyManager first picks a
> suitable
> > device to route the audio to based on the stream type, which
> accessories
> > you have attached etc. It then constructs a message (key/value pair)
> > that it sends to AudioHardware which is the class responsible for
> > actually enabling/disabling audio devices (among other things).
> >
> > Now the problem is that the AudioPolicyManager doesn't have a direct
> > line to AudioHardware. Instead it sends the message to the
> > AudioPolicyService, which adds it to a vector that is processed by the
> > AudioCommandThread's threadLoop. The AudioCommandThread sends the
> > message to the AudioSystem, which in turn sends it to the
> AudioFlinger.
> > The AudioFlinger then adds it to another list/queue which is processed
> > by the relevant mixer thread, which finally sends it to the
> AudioHardware.
> >
> > These are a lot of intermediaries for the message to pass through. And
> > when it finally gets to the AudioHardware you're still not done,
> because
> > you might have to enable/disable the relevant devices, set the proper
> > filters/volumes/etc, and in some cases turn on some amplifier (which
> > incurs additional delays because you want to avoid click & pops).
>
> That's interesting. The right place for this might be
> android-platform, but
> that's interesting nevertheless. That's the kind of thing I sensed
> and that I
> meant when I talked about "over-encapsulation".
>
> > In my view there are too many layers in the audio chain. My main
> concern
> > isn't latency though; it's the lack of centralized resource management
> > for audio. *All* routing decisions, and *all* decisions about
> whether or
> > not an application should be allowed to start/continue audio playback
> > should IMO be centralized in some place at the lower levels, where all
> > information about ongoing playback, available tracks, phone state,
> > attached accessories, ringer mode etc is known. IIRC there used to be
> > problems (and there may still be) where applications could hog all the
> > playback resources so that when you got an incoming call there
> would be
> > no ringtone. For an OS running primarily on mobile phones that's not
> > exactly the optimal behavior.
>
> Indeed, I can confirm that type of defect. In my app I even detect
> incoming
> calls to stop recording when they occur. Otherwise, the whole system
> crashes on
> several devices including the Milestone (Droid) running 2.1.
>
> I thought that all the audio layers were mostly intended to protect the
> telephony features, but they apparently don't.
>
> --
> Olivier
>
> --
> You received this message because you are subscribed to the Google
> Groups "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com
> <mailto:android-***@googlegroups.com>.
> To unsubscribe from this group, send email to
> android-ndk+***@googlegroups.com
> <mailto:android-ndk%***@googlegroups.com>.
> For more options, visit this group at
> http://groups.google.com/group/android-ndk?hl=en.
>
>
> --
> You received this message because you are subscribed to the Google
> Groups "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com.
> To unsubscribe from this group, send email to
> android-ndk+***@googlegroups.com.
> For more options, visit this group at
> http://groups.google.com/group/android-ndk?hl=en.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Ross Bencina
2011-01-01 05:09:22 UTC
Permalink
Just adding another data point: perhaps Glenn or Dave can comment on
whether this is still an issue in 2011, but I just found this post
(pasted below) from Dave Sparks in 2009 that suggests that the large
AudioFlinger mix buffer sizes are due to a 20ms scheduler interval. If
the audio mixer is stuttering under load It sounds to me like the
audio mixer is not running with sufficient priority -- it is after all
a hard real-time operation and it should be given sufficient priority
to complete by its deadline. For example, CoreAudio audio client
threads on iOS run in the real-time band.

http://groups.google.com/group/android-framework/browse_thread/thread/c2d3d5e2343845ca
> --------- Forwarded message ----------
> From: Dave Sparks <***@android.com>
> Date: Mar 6 2009, 3:55 pm
> Subject: Audio Flinger and Mixer operations
> To: android-framework
> AudioFlinger mixes into a small buffer which determines the latency
> (along with the underlying hardware latency). The source buffers can
> be any size. The mix buffer size has to be larger than the scheduler
> interval (typically 20ms) or audio will stutter under load. That means
> best case latency is around 40ms for a ping-pong buffer audio driver.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-29 14:06:31 UTC
Permalink
On 12/26/2010 07:01 PM, Patrick Flanagan wrote:
> Regarding Java vs. Native, a few researchers at ETH Zurich were able
> to push latency in Java SE down to 0.5 ms. Paper here:
> http://diuf.unifr.ch/pai/wiki/doku.php?id=Publications&page=publication&kind=single&ID=16.
> They used RtAudio as an intermediary to talk to ALSA but only as a
> means to establish portability with Windows and MacOS. In a purely
> Linux environment, RtAudio wouldn't be necessary, according to the
> authors. This suggests to me a native solution isn't strictly
> necessary.

This isn't relevant. This project relies on direct access to ALSA, which means
exclusive hardware access. This can not happen on Android since there can be
multiple audio clients at any time (phone ring while listening to music, etc..).

So a sound server is needed, as AudioFlinger currently, which very likely
implies extra latency. But if the server and protocol are well designed and
running on proper hardware, low latency should be possible IMO.

--
Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Ross Bencina
2010-12-22 10:05:20 UTC
Permalink
On Dec 22, 8:31 am, Olivier Guilyardi <***@samalyse.com> wrote:
> 10ms sounds ok to me. You can't expect a professional zero-latency DAW anyway.

To me, if the current CDD said low-latency=10ms I would think "yeah
that's ok, but it isn't really low latency."

I don't expect a professional DAW. But I do expect things to keep up
with standard desktop audio APIs (e.g. CoreAudio, WASAPI) these, and
iOS, have significantly lower latency than 10ms, and you can bet that
QNX has pretty good latency too.

Setting the requirement at 45ms is so far from real low latency that
there is room for all levels of the system to fail to deliver (from
Dalvik, to NDK/OpenSL, to AudioFlinger, libaudio.so and friends, the
kernel etc). That's way too many subsystems to blame when things go
wrong. The best thing would be to provide documented latency
requirements at each level of the system. Kernel latency for example
will be affected by other drivers, not just audio. Seems like there
are already goals for Dalvik GC pause times. At the moment it seems
hard to get info about internal buffering bottlenecks inside
AudioFlinger and OpenSL implementations without diving into the
source.

Ross

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-22 23:02:08 UTC
Permalink
On 12/22/2010 11:05 AM, Ross Bencina wrote:
> On Dec 22, 8:31 am, Olivier Guilyardi <***@samalyse.com> wrote:
>> 10ms sounds ok to me. You can't expect a professional zero-latency DAW anyway.
>
> To me, if the current CDD said low-latency=10ms I would think "yeah
> that's ok, but it isn't really low latency."

Yeah, that's what I mean. A compromise, but written "10ms or less". Maybe that
some capable manufacturers would then provide true low latency. I don't known
what's under OpenSL, so I can't tell.

With the Java API, latency is apparently caused by extra buffers and IPC in what
stands between the app and the hardware. I'm not sure but I think that, as it
relates to telephony, sound is a critical part of the platform. Whereas you get
a rather direct access to the GPU, the audio subsystem seems over-encapsulated.

Plus, last time I checked, patches were not accepted for the audio stack
(critical again). So, commenting this CDD is quite the only thing that we can do
currently.

Apart maybe from releasing an Android variant where all audio is handled by
JACK. That would be something :)

--
Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Ross Bencina
2010-12-23 05:18:13 UTC
Permalink
On Dec 23, 10:02 am, Olivier Guilyardi <***@samalyse.com> wrote:
> With the Java API, latency is apparently caused by extra buffers and IPC in what
> stands between the app and the hardware.

I'm not sure native API audio reduces the IPC burden, although I am
yet to see an architecture diagram that explains how OpenSL is
supposed to be related to AudioFlinger. The most recent diagram I have
seen is this one, which is presumably pre-gingerbread:
http://www.kandroid.org/android_pdk/audio_sub_system.html

I don't know if there is a more official google link for that(Glenn ?)

If someone can post a link to an example of how to do simple low-
latency audio i/o with OpenSL that would be super helpful. Just an
audio pass-through or even playing a sawtooth wave like this year's
Google I/O "Advanced audio techniques" video would be enough.


>I'm not sure but I think that, as it
> relates to telephony, sound is a critical part of the platform.

I think for this reason google have left this detail to the platform
partners (so platform partners make telephony work well, and
everything we can do from Java or NDK sucks). Compare this to iOS
where media playback is a core device priority.

>Plus, last time I checked, patches were not accepted for the audio stack
>(critical again). So, commenting this CDD is quite the only thing that we can do
>currently.

Yes I read this too. Glenn can you please update us on the audio stack
roadmap and how we can get involved?



>Whereas you get
> a rather direct access to the GPU, the audio subsystem seems over-encapsulated.

Wow, what a can of worms that statement is. I agree, but I think it's
it's a mistake to think sound APIs should be like graphics APIs. Audio
hardware acceleration does not have the same status as graphics
acceleration in the marketplace -- otherwise we'd all be programming
in standardised audio dsp languages (equivalent of shader langauges)
and getting the hardware to compile the code on the fly and execute it
efficiently. That would be great, but that's not what we have. So, we
mainly write audio code for the CPU (if there are audio coprocessors,
they are specialised to certain tasks like codec acceleration). Based
on this, an audio API should look more like a direct frame buffer, a
way to get sound out of the device with low latency. That's what we're
asking for -- it's not an obscure request, it's what audio programmers
need to do what they do, and what they have already on other
platforms. Hardware codec acceleration can be handled by a library
that sits on the side, not something that must sit between the app and
the hardware.

I don't think it's so much a problem of over-encapsulation as it is a
poor choice of abstractions and layering. Audio programmers are used
to the audio equivalent of direct frame buffer access (or a good HAL
abstraction of it) as this is generally required for low latency. It
just doesn't make sense to put too much extra stuff in between the
hardware and the user app unless it is required for a particular
application. Things like OpenSL seem to me to be middle-ware -- good
for putting something together quickly, and they provide an
abstraction for HW vendors to support hardware acceleration of codecs
or effects. But like most middle-ware, it is not something you can
easily standardise for all use cases. Each middle-ware might be good
for it's target, but the platform needs a low-latency audio interface
to hardware so people can bypass middle-ware or use different middle-
ware.

One thing I've noticed, is that when application programmers start
learning how to do real-time audio programming on normal (non hard-
real-time) OSes there is a steep learning curve because they don't
understand what's required for real time code (no locks, no memory
allocation, no blocking apis, etc). I've been there, I've seen this on
many mailing lists on many platforms (ALSA, JACK, CoreAudio,
PortAudio, etc)... everyone goes through that stage. That isn't going
to change any time soon, because real time audio requires either a
real-time OS or code that avoids non-real-time features of a non-real-
time host OS. Perhaps it's kind of scary for google to offer an API
that actually allows proper real-time audio because it breaks with the
simplicity of what normal (non-audio) developers understand (cf
AudioTrack::write()). People like the idea of a read()/write() stream
interface for audio because it's easy to understand and elegant. But
in reality, a real-time audio data pump needs to do 2 things well: (1)
move the data to/from the hardware efficiently, (2) notify the user in
a timely manner to pump data by avoiding scheduling jitter. Read()/
write() interfaces don't directly answer requirement (2), at least on
non-real-time platforms they don't. The API implementation needs fine
grained control of the thread scheduling of the user proc that does
audio evaluation. Read()/write() interface give control of that to the
client and limit the mechanisms the API implementation can use. Oh,
and another problem with read()/write() is it isn't well suited to
full-duplex.

I haven't seen another low-latency API that uses read(),write() (rumor
has it that the Irix one worked well, but they had time constrained
threads I think) and I've seen people argue for read()/write()
interfaces because they simplify concurrency-related coding, but then
switch to async callbacks once they understand the issues. For Java,
where there are GC pauses and you need to buffer things anyway, I
think the read()/write() interface makes sense (unless the VM can
service async interrupts). For C++/NDK it makes no sense, and right
now, as I understand it, the audio flinger uses read()/write() at the
HardwareAudioInterface boundary so I'm hoping we get some clarity on
where that is heading and how it relates to OpenSL and real low-
latency.

Of course, you can sidestep all this if you defined "low-latency" as
45ms ;)

Sorry for another long post but I think Android is important enough
for this not to get f**kd up yet again...

Cheers

Ross.




--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Ross Bencina
2010-12-24 11:11:19 UTC
Permalink
To follow up on my questions:

After some more digging I discovered that the OpenSL ES interface's
Buffer Queue abstractions provides a callback notification interface.

An example of this is in the latest NDK at:
android-ndk-r5\samples\native-audio\jni\native-audio-jni.c
A lot of the questions I raised above are answered in the R5 NDK docs
at
android-ndk-r5/docs/opensles/index.html

there's an online copy of that page you can read at the link below
(google: that would be a useful doc to have online somewhere :-) :
http://mobilepearls.com/labs/native-android-api/opensles/index.html

As I think the CDD mentions too, it says:
>>>>
As OpenSL ES is a native C API, non-Dalvik application threads which
call OpenSL ES have no Dalvik-related overhead such as garbage
collection pauses. However, there is no additional performance benefit
to the use of OpenSL ES other than this. In particular, use of OpenSL
ES does not result in lower audio latency, higher scheduling priority,
etc. than what the platform generally provides. On the other hand, as
the Android platform and specific device implementations continue to
evolve, an OpenSL ES application can expect to benefit from any future
system performance improvements.
<<<<

So I guess the CDD latency guidelines Glenn posted earlier are to be
interpreted in the context of OpenSL buffer queues. Correct me if I'm
wrong (since I'm clutching at straws trying to understand what the CDD
latency numbers actually mean), but if the warm-start latency is 10ms,
and the continuous latency is 45ms, that implies to me that there is
an allowance for 35ms of scheduling jitter in the OpenSL buffer queue
notification callbacks. I don't know much about the Android kernel,
but on iOS, scheduling jitter for a real-time thread is around 200us..
so 35ms seems pretty high. Can someone who understands this better
explain what that 35ms covers please?

Glenn, can I suggest that the CDD reference concrete test cases in
both Java and the NDK that are supposed to exhibit the required
latency, that way we can understand exactly what is being guaranteed.
I am a little unclear whether the 45ms continuous figure includes a
margin for masking Dalvik GC pauses. It would make sense to have
separate requirements/constraints on the platform (Dalvik) and the C-
level and NDK capabilities.

In the end, a latency guarantee will require the vendor to make sure
that no driver will cause audio related native threads (in kernel or
userspace) to miss a deadline, that intermediate buffering can be
dimensioned to accommodate low latency, and that all inter-thread and
inter-process communication mechanisms are non-blocking (lock-free
FIFOS or similar mechanisms).

I am alarmed to see the following in the above cited document:
>>>>
But this point is asynchronous with respect to the application. Thus
you should use a mutex or other synchronization mechanism to control
access to any variables shared between the application and the
callback handler. In the example code, such as for buffer queues, we
have omitted this synchronization in the interest of simplicity.
However, proper mutual exclusion would be critical for any production
code.
<<<<

Employing mutexes will almost certainly cause priority inversion at
some stage and glitch audio rendering. The usual (and simplest) safe
technique to communicate with asynchronous audio callbacks are lock-
free fifo command queues. Try-locks are may also be an option in some
cases. I've already mentioned the learning curve involved in writing
real-time audio software on non-real time OSes.. clearly whoever wrote
that document hasn't traversed it.

Thanks

Ross.







--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-24 12:00:53 UTC
Permalink
On 12/24/2010 12:11 PM, Ross Bencina wrote:

> I am alarmed to see the following in the above cited document:

So am I!

> But this point is asynchronous with respect to the application. Thus
> you should use a mutex or other synchronization mechanism to control
> access to any variables shared between the application and the
> callback handler. In the example code, such as for buffer queues, we
> have omitted this synchronization in the interest of simplicity.
> However, proper mutual exclusion would be critical for any production
> code.
> <<<<

Yeah, of course, locking a mutex in an audio process callback. This is a newbie
audio development mistake.

> Employing mutexes will almost certainly cause priority inversion at
> some stage and glitch audio rendering. The usual (and simplest) safe
> technique to communicate with asynchronous audio callbacks are lock-
> free fifo command queues. Try-locks are may also be an option in some
> cases. I've already mentioned the learning curve involved in writing
> real-time audio software on non-real time OSes.. clearly whoever wrote
> that document hasn't traversed it.

I completely agree. This situation is not professional, it's 100% amateurism.

--
Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-24 12:07:33 UTC
Permalink
On 12/24/2010 01:00 PM, Olivier Guilyardi wrote:
> On 12/24/2010 12:11 PM, Ross Bencina wrote:
>
>> I am alarmed to see the following in the above cited document:
>
> So am I!
>
>> But this point is asynchronous with respect to the application. Thus
>> you should use a mutex or other synchronization mechanism to control
>> access to any variables shared between the application and the
>> callback handler. In the example code, such as for buffer queues, we
>> have omitted this synchronization in the interest of simplicity.
>> However, proper mutual exclusion would be critical for any production
>> code.
>> <<<<
>
> Yeah, of course, locking a mutex in an audio process callback. This is a newbie
> audio development mistake.
>
>> Employing mutexes will almost certainly cause priority inversion at
>> some stage and glitch audio rendering. The usual (and simplest) safe
>> technique to communicate with asynchronous audio callbacks are lock-
>> free fifo command queues. Try-locks are may also be an option in some
>> cases. I've already mentioned the learning curve involved in writing
>> real-time audio software on non-real time OSes.. clearly whoever wrote
>> that document hasn't traversed it.
>
> I completely agree. This situation is not professional, it's 100% amateurism.

Okay, I'm a bit harsh here, sorry. But really, all of this isn't serious.

--
Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Olivier Guilyardi
2010-12-24 11:53:31 UTC
Permalink
On 12/23/2010 06:18 AM, Ross Bencina wrote:
> On Dec 23, 10:02 am, Olivier Guilyardi <***@samalyse.com> wrote:
>
>> I'm not sure but I think that, as it
>> relates to telephony, sound is a critical part of the platform.
>
> I think for this reason google have left this detail to the platform
> partners (so platform partners make telephony work well, and
> everything we can do from Java or NDK sucks). Compare this to iOS
> where media playback is a core device priority.

Agreed. And aside of that, Apple has been making critical professional audio
hardware and software for decades. They clearly have the skills.

And, there are such skilled people amongst free audio software developers, but
for some reason it doesn't benefit to Android.

> I don't think it's so much a problem of over-encapsulation as it is a
> poor choice of abstractions and layering.

I agree.

> One thing I've noticed, is that when application programmers start
> learning how to do real-time audio programming on normal (non hard-
> real-time) OSes there is a steep learning curve because they don't
> understand what's required for real time code (no locks, no memory
> allocation, no blocking apis, etc). I've been there, I've seen this on
> many mailing lists on many platforms (ALSA, JACK, CoreAudio,
> PortAudio, etc)... everyone goes through that stage.

That's correct. I once submitted a JACK-related patch to the FFmpeg project. It
was accepted, but the FFmpeg devs, although really good at what they do, had
quite a lot of trouble understanding the requirements and semantics of realtime
audio, no memory allocations, lock-free ringbuffers, etc...

And what's confusing is that they are audio codecs experts, but that doesn't
make them application-level realtime audio devs.

And I'm afraid that such skills seem to be missing in the Android teams.

> Of course, you can sidestep all this if you defined "low-latency" as
> 45ms ;)

Reading Glenn's answer about the difference between "warm" and "continuous"
latency, it seems that 35ms of these 45ms come from software layers, flinger and
the like.

> Sorry for another long post but I think Android is important enough
> for this not to get f**kd up yet again...

I agree, it's really time to improve this poor audio situation, but the more it
goes, the more I think there are critical design flaws in the OS.

Also, I looked at the OpenSL API and I clearly don't understand why so much work
is being put into bells and whistle such as reverb, etc.. Whereas the bare
minimum, reliable low latency pcm input/output, is not provided.

I'm sorry if I'm a bit harsh, but I've been working with Android audio APIs for
over a year now, and I feel like telling the truth.

That said, happy holidays to everyone!

--
Olivier

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Dan S
2011-01-15 14:10:55 UTC
Permalink
On Dec 22 2010, 10:05 am, Ross Bencina <***@audiomulch.com> wrote:
> On Dec 22, 8:31 am, Olivier Guilyardi <***@samalyse.com> wrote:
>
> > 10ms sounds ok to me. You can't expect a professional zero-latency DAW anyway.
>
> To me, if the current CDD said low-latency=10ms I would think "yeah
> that's ok, but it isn't really low latency."
>
> I don't expect a professional DAW. But I do expect things to keep up
> with standard desktop audio APIs (e.g. CoreAudio, WASAPI) these, and
> iOS, have significantly lower latency than 10ms, and you can bet that
> QNX has pretty good latency too.
>
> Setting the requirement at 45ms is so far from real low latency

I agree that 45ms is a long way from what I would ever have thought to
call "low latency"; but that aside, this discussion shows that there
are different criteria - some would be happy with 20ms, some with 2ms.

So the question is, why is android providing a binary "low-latency"
flag, why not a numerical "min-latency" value? That shouldn't be too
much trouble for manufacturers since they don't all have to measure
the system latency precisely, just give a rough value of the lowest
latency the system would be likely to provide on a bad day. That would
be something we could all then work with?

Best,
Dan

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
mic _
2011-01-15 14:16:01 UTC
Permalink
There already is such a value internally. Though it's based on the buffer
size used by the audio driver, so it doesn't take into account how long it
might take to turn on necessary amplifiers etc (i.e. how long before you
actually hear anything). I don't know if that value is exposed through any
public APIs.

/Michael

On Sat, Jan 15, 2011 at 3:10 PM, Dan S <***@gmail.com> wrote:

> On Dec 22 2010, 10:05 am, Ross Bencina <***@audiomulch.com> wrote:
> > On Dec 22, 8:31 am, Olivier Guilyardi <***@samalyse.com> wrote:
> >
> > > 10ms sounds ok to me. You can't expect a professional zero-latency DAW
> anyway.
> >
> > To me, if the current CDD said low-latency=10ms I would think "yeah
> > that's ok, but it isn't really low latency."
> >
> > I don't expect a professional DAW. But I do expect things to keep up
> > with standard desktop audio APIs (e.g. CoreAudio, WASAPI) these, and
> > iOS, have significantly lower latency than 10ms, and you can bet that
> > QNX has pretty good latency too.
> >
> > Setting the requirement at 45ms is so far from real low latency
>
> I agree that 45ms is a long way from what I would ever have thought to
> call "low latency"; but that aside, this discussion shows that there
> are different criteria - some would be happy with 20ms, some with 2ms.
>
> So the question is, why is android providing a binary "low-latency"
> flag, why not a numerical "min-latency" value? That shouldn't be too
> much trouble for manufacturers since they don't all have to measure
> the system latency precisely, just give a rough value of the lowest
> latency the system would be likely to provide on a bad day. That would
> be something we could all then work with?
>
> Best,
> Dan
>
> --
> You received this message because you are subscribed to the Google Groups
> "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com.
> To unsubscribe from this group, send email to
> android-ndk+***@googlegroups.com<android-ndk%***@googlegroups.com>
> .
> For more options, visit this group at
> http://groups.google.com/group/android-ndk?hl=en.
>
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
MrBuBBLs
2011-01-16 19:26:34 UTC
Permalink
I'm not sure but I think the buffer size DOES count for the "response
time" called latency.
Look at some existing tweakable ASIO drivers. When you tweak the
buffer size it directly changes the latency on those drivers.
e.g. http://www.asio4all.com

On 15 jan, 15:16, mic _ <***@gmail.com> wrote:
> There already is such a value internally. Though it's based on the buffer
> size used by the audio driver, so it doesn't take into account how long it
> might take to turn on necessary amplifiers etc (i.e. how long before you
> actually hear anything). I don't know if that value is exposed through any
> public APIs.
>
> /Michael
>
>
>
>
>
>
>
> On Sat, Jan 15, 2011 at 3:10 PM, Dan S <***@gmail.com> wrote:
> > On Dec 22 2010, 10:05 am, Ross Bencina <***@audiomulch.com> wrote:
> > > On Dec 22, 8:31 am, Olivier Guilyardi <***@samalyse.com> wrote:
>
> > > > 10ms sounds ok to me. You can't expect a professional zero-latency DAW
> > anyway.
>
> > > To me, if the current CDD said low-latency=10ms I would think "yeah
> > > that's ok, but it isn't really low latency."
>
> > > I don't expect a professional DAW. But I do expect things to keep up
> > > with standard desktop audio APIs (e.g. CoreAudio, WASAPI) these, and
> > > iOS, have significantly lower latency than 10ms, and you can bet that
> > > QNX has pretty good latency too.
>
> > > Setting the requirement at 45ms is so far from real low latency
>
> > I agree that 45ms is a long way from what I would ever have thought to
> > call "low latency"; but that aside, this discussion shows that there
> > are different criteria - some would be happy with 20ms, some with 2ms.
>
> > So the question is, why is android providing a binary "low-latency"
> > flag, why not a numerical "min-latency" value? That shouldn't be too
> > much trouble for manufacturers since they don't all have to measure
> > the system latency precisely, just give a rough value of the lowest
> > latency the system would be likely to provide on a bad day. That would
> > be something we could all then work with?
>
> > Best,
> > Dan
>
> > --
> > You received this message because you are subscribed to the Google Groups
> > "android-ndk" group.
> > To post to this group, send email to android-***@googlegroups.com.
> > To unsubscribe from this group, send email to
> > android-ndk+***@googlegroups.com<android-ndk%***@googlegr oups.com>
> > .
> > For more options, visit this group at
> >http://groups.google.com/group/android-ndk?hl=en.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Kevin Duffey
2011-01-17 05:24:09 UTC
Permalink
Wow..great thread!

I have a keen interest in some day understanding enough of all this to
actually make an audio app I had in mind. One of my favorite apps on any
platform is BeatMaker on iPhone.

At any rate, perhaps this will never get answered.. but I can't help to
ask... with the original iPhone being such a success, and out over a year
before any Android device (I'll keep in mind that google/android was
obviously being worked on probably not too long after iPhone came out),
knowing full well how good the games and audio capabilities were.. how could
anyone NOT see that iPhone was going to be a huge success for gamers and
eventually audio apps? I keep asking myself... how did probably a large
group of google and other engineers, brilliant all around for sure,
completely miss on one of the most critical requirements to make a
successful platform? I will say, I mean no disrespect to any of the android
team or those involved. With out a doubt, Android is huge, and going to be
far bigger in the coming years. I was blown away that the low latency audio
and in some respects, graphics too, were so far behind apple after almost
two years of being in the wild. I figured that 2.2 was going to be the one
that finally brought Android the low latency capabilities it needed to make
it take off for games and audio apps. I do realize that for just about
everything else, graphics/audio isn't needed to be super fast like it is for
games and audio apps... but with supposedly almost 50% of the apps on iPhone
being games and audio apps.. maybe even more now.. how could anyone have
kept putting such important features on the back burner? It seems
unfathomable to me that it would make any sort of good business sense to
almost cut out 1/2 of the leading competitions apps, out of Android.. touted
to be the king this year or next, and over the next few years will lead the
pack by a large margin.

So when I read that 2.3 added native low latency audio, fixed the input
issue that caused severe lag in games with multi-touch, and upped the
graphics a bit too..I thought for sure 2.3 is going to be the version that
finally brings Android to the front.. finally developers can port games,
write games and all those cool audio apps, for android.

So this thread is quite disheartening to me. It sounds like, for the most
part, drum machines, synth apps, and smoother playing games are not going to
be possible after all.

So the question that I have for any of you that might know.. will these
native apis help at all for games and audio apps? Or is it pointless to even
start on a multi-touch drum/MPC style audio app still? More so, even if it
might be possible.. is each and every different handset going to cause
problems? I am looking personally at the Motorola xoom tablet and Bionic
phone. I am thinking, dual core cpus, more memory.. it's gotta to be
smoking. I had a Pentium II years ago, 266Mhz, that could mix 64 audio
sounds together while working with 3D video. How can dual-core 1Ghz cpus not
even come close?

I sure hope that gamer developers can utilize these new apis some how to
make games more responsive than they have been. I look at a game like Angry
Birds, which is implemented very well on Android..runs solid on my Droid 1..
and those seem to be the type of games that do the best anyway, even on
iPhone. If that can be done now on Android, then the 2.3 enhancements
certainly have to improve upon that.. hopefully a magnitude more?



On Sun, Jan 16, 2011 at 11:26 AM, MrBuBBLs <***@gmail.com>wrote:

> I'm not sure but I think the buffer size DOES count for the "response
> time" called latency.
> Look at some existing tweakable ASIO drivers. When you tweak the
> buffer size it directly changes the latency on those drivers.
> e.g. http://www.asio4all.com
>
> On 15 jan, 15:16, mic _ <***@gmail.com> wrote:
> > There already is such a value internally. Though it's based on the buffer
> > size used by the audio driver, so it doesn't take into account how long
> it
> > might take to turn on necessary amplifiers etc (i.e. how long before you
> > actually hear anything). I don't know if that value is exposed through
> any
> > public APIs.
> >
> > /Michael
> >
> >
> >
> >
> >
> >
> >
> > On Sat, Jan 15, 2011 at 3:10 PM, Dan S <***@gmail.com> wrote:
> > > On Dec 22 2010, 10:05 am, Ross Bencina <***@audiomulch.com> wrote:
> > > > On Dec 22, 8:31 am, Olivier Guilyardi <***@samalyse.com> wrote:
> >
> > > > > 10ms sounds ok to me. You can't expect a professional zero-latency
> DAW
> > > anyway.
> >
> > > > To me, if the current CDD said low-latency=10ms I would think "yeah
> > > > that's ok, but it isn't really low latency."
> >
> > > > I don't expect a professional DAW. But I do expect things to keep up
> > > > with standard desktop audio APIs (e.g. CoreAudio, WASAPI) these, and
> > > > iOS, have significantly lower latency than 10ms, and you can bet that
> > > > QNX has pretty good latency too.
> >
> > > > Setting the requirement at 45ms is so far from real low latency
> >
> > > I agree that 45ms is a long way from what I would ever have thought to
> > > call "low latency"; but that aside, this discussion shows that there
> > > are different criteria - some would be happy with 20ms, some with 2ms.
> >
> > > So the question is, why is android providing a binary "low-latency"
> > > flag, why not a numerical "min-latency" value? That shouldn't be too
> > > much trouble for manufacturers since they don't all have to measure
> > > the system latency precisely, just give a rough value of the lowest
> > > latency the system would be likely to provide on a bad day. That would
> > > be something we could all then work with?
> >
> > > Best,
> > > Dan
> >
> > > --
> > > You received this message because you are subscribed to the Google
> Groups
> > > "android-ndk" group.
> > > To post to this group, send email to android-***@googlegroups.com.
> > > To unsubscribe from this group, send email to
> > > android-ndk+***@googlegroups.com<android-ndk%***@googlegroups.com>
> <android-ndk%***@googlegr oups.com>
> > > .
> > > For more options, visit this group at
> > >http://groups.google.com/group/android-ndk?hl=en.
>
> --
> You received this message because you are subscribed to the Google Groups
> "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com.
> To unsubscribe from this group, send email to
> android-ndk+***@googlegroups.com<android-ndk%***@googlegroups.com>
> .
> For more options, visit this group at
> http://groups.google.com/group/android-ndk?hl=en.
>
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Shijing Xian
2011-01-24 15:42:15 UTC
Permalink
Hello Guys,

I think we had pretty much discussion or complain on this 45ms "low-
latency", but still we may be lack of a good overview on the android
audio system and may forget or miss some piece of important
information.
I am recently doing a experiment in Nexus S using this native audio
implementation for steaming, and get some interesting numbers, I hope
they can be usefully and spur more discussions.

I made a loopback application, which gets data from the recording
device, puts the data to a ring buffer, and playout. All the data
processing(if any) is using a framesize of 10ms audio.

Setups for the native audio implementation:
Recording: simple buffer queue, two buffers with size of 10ms audio
for each.
Playback: buffer queue, two buffers with size of 10ms audio for each.
When the app is started, it will enqueue an empty 10ms buffer to the
recording queue, and 10ms data buffer to playout queue, to keep the
queues rolling. In the recording callback function, I copy the data to
the ring buffer and enqueue 10ms empty buffer; In the playout
callback, I pull 10ms data from ring buffer and enqueue it to the
playback (if no data, it will gets zero buffer, no waiting here).
It is pretty simple, isn't it? And my calculation for the delay is
something like this: recording delay + playout delay + ring buffer
delay
recording delay should be less than '10ms(buffering) + continuous
input latency(50ms)'
playback delay should be less than '10ms(buffering) + continuous
output latency(45ms)'
ring buffer delay is depending on the amount of recording data has
been put into the ring buffer, and the amount of data has been played.
It is scaling to the size of buffer in the ring, and the ideal case is
that we playback the data soon after it is recorded. This ring buffer
delay is quite depending on how the callbacks come.
So my expectation for the maximum delay is : 115 + ring buffer delay.
(Please correct me if I miss something in my model)
Surprisingly, I got a delay of 360ms for this loopback test, where the
ring buffer delay is about 100ms on average.
To clarify how I do the measurement, I use a latency instrument which
inputs a beep to the microphone while listening to the speaker, and it
reports the interval.

I understand the Nexus S device is not a low latency device, and it
may have a bigger continuous input/output latency, but 250ms in 10ms
buffering is definitely not acceptable to me. And the worst thing is
that I have no clue where is consuming the time.
Note, I can't tell how much the playback and recording is taking part
in this 250ms delay separately now, and I will investigate on this
later.

And another issue I want to address here is about the way how the
callbacks come. In my setup, I have two 10ms buffers for each
recording/playback queue, so I am expecting the callbacks to occur
every 10ms. But it is not true in my test, instead I usually get two
or three callbacks with the same timestamp every 20-30ms. The worst is
that at the beginning of the call I often get a stack of callbacks at
the same time, for example, 7 recording callbacks with the same
timestamp, then the next recording callback comes after 70ms. That may
be the reason why I have a 100ms ring buffer delay in my test, but I
have no idea what is causing it. Guys, any idea here?

To compare the performance, I run the same test but using AudioTrack
in Java. most of the setup is similar to the native audio
implementation, except that I use polling instead of callback.
Basically I run two aggressive threads to pull data from the recording
and put data for playback.
And the overall delay is 280ms.
Any comment here to explain why the Java implementation has less
latency?

While the similar application for iPhone is reporting a 60-70ms
overall delay.

Note, in the above test I am using 16000 as sampling rate for both
recording and playback.

I am not sure if any of you have already done a similar experiment
like what I did, any comments or suggestion will be appreciated. And I
do wonder if there any API for me to tweak the performance? Or any API
to get the hardware or os delay to get the accurate latency? (Though I
have already looked at the openSLES 1.0.1, but can't find any).

The only thing makes me happen is that the native audio implementation
does take less CPU usage, and the audio is more reliable than using
AudioTrack.


Best Regards,
/Shijing

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Robert Green
2011-01-25 19:31:39 UTC
Permalink
Would you be willing to rerun your tests at 44.1khz stereo? I found
it interesting that on my galaxy s, it used the same buffer size for
mono and stereo which made it so that stereo would have half the
latency of mono. I'm wondering if that's also the case on the nexus
s. It's still horrible latency, but half as horrible. :)

On Jan 24, 7:42 am, Shijing Xian <***@gmail.com> wrote:
> Hello Guys,
>
> I think we had pretty much discussion or complain on this 45ms "low-
> latency", but still we may be lack of a good overview on the android
> audio system and may forget or miss some piece of important
> information.
> I am recently doing a experiment in Nexus S using this native audio
> implementation for steaming, and get some interesting numbers, I hope
> they can be usefully and spur more discussions.
>
> I made a loopback application, which gets data from the recording
> device, puts the data to a ring buffer, and playout. All the data
> processing(if any) is using a framesize of 10ms audio.
>
> Setups for the native audio implementation:
> Recording: simple buffer queue, two buffers with size of 10ms audio
> for each.
> Playback:  buffer queue, two buffers with size of 10ms audio for each.
> When the app is started, it will enqueue an empty 10ms buffer to the
> recording queue, and 10ms data buffer to playout queue, to keep the
> queues rolling. In the recording callback function, I copy the data to
> the ring buffer and enqueue 10ms empty buffer; In the playout
> callback, I pull 10ms data from ring buffer and enqueue it to the
> playback (if no data, it will gets zero buffer, no waiting here).
> It is pretty simple, isn't it? And my calculation for the delay is
> something like this: recording delay + playout delay + ring buffer
> delay
> recording delay should be less than '10ms(buffering) + continuous
> input latency(50ms)'
> playback delay should be less than '10ms(buffering) + continuous
> output latency(45ms)'
> ring buffer delay is depending on the amount of recording data has
> been put into the ring buffer, and the amount of data has been played.
> It is scaling to the size of buffer in the ring, and the ideal case is
> that we playback the data soon after it is recorded. This ring buffer
> delay is quite depending on how the callbacks come.
> So my expectation for the maximum delay is : 115 + ring buffer delay.
> (Please correct me if I miss something in my model)
> Surprisingly, I got a delay of 360ms for this loopback test, where the
> ring buffer delay is about 100ms on average.
> To clarify how I do the measurement, I use a latency instrument which
> inputs a beep to the microphone while listening to the speaker, and it
> reports the interval.
>
> I understand the Nexus S device is not a low latency device, and it
> may have a bigger continuous input/output latency, but 250ms in 10ms
> buffering is definitely not acceptable to me. And the worst thing is
> that I have no clue where is consuming the time.
> Note, I can't tell how much the playback and recording is taking part
> in this 250ms delay separately now, and I will investigate on this
> later.
>
> And another issue I want to address here is about the way how the
> callbacks come. In my setup, I have two 10ms buffers for each
> recording/playback queue, so I am expecting the callbacks to occur
> every 10ms. But it is not true in my test, instead I usually get two
> or three callbacks with the same timestamp every 20-30ms. The worst is
> that at the beginning of the call I often get a stack of callbacks at
> the same time, for example, 7 recording callbacks with the same
> timestamp, then the next recording callback comes after 70ms. That may
> be the reason why I have a 100ms ring buffer delay in my test, but I
> have no idea what is causing it. Guys, any idea here?
>
> To compare the performance, I run the same test but using AudioTrack
> in Java. most of the setup is similar to the native audio
> implementation, except that I use polling instead of callback.
> Basically I run two aggressive threads to pull data from the recording
> and put data for playback.
> And the overall delay is 280ms.
> Any comment here to explain why the Java implementation has less
> latency?
>
> While the similar application for iPhone is reporting a 60-70ms
> overall delay.
>
> Note, in the above test I am using 16000 as sampling rate for both
> recording and playback.
>
> I am not sure if any of you have already done a similar experiment
> like what I did, any comments or suggestion will be appreciated. And I
> do wonder if there any API for me to tweak the performance? Or any API
> to get the hardware or os delay to get the accurate latency? (Though I
> have already looked at the openSLES 1.0.1, but can't find any).
>
> The only thing makes me happen is that the native audio implementation
> does take less CPU usage, and the audio is more reliable than using
> AudioTrack.
>
> Best Regards,
> /Shijing

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Shijing Xian
2011-01-27 11:54:42 UTC
Permalink
Thanks Robert, it is definitely interesting to know how much it
improves the latency by using the stereo audio. It may likely match
what Robert said, half of the latency of mono. But I am not sure if
44.1khz makes any difference, I tried the 44.1khz, 16khz, 8khz mono
recording/playback before, all the tests showed a similar latency,
unfortunately.
There are also some other reasons why I prefer 16khz to others, the
most important one is 16khz is usually supported by almost all Android
devices.
I will try the stereo recording and playback, and paste what I find
here later. In order to identify the problem, I will try stereo
recording and playback separately. So the combination will be:
stereo recording + mono playback
mono recording + stereo playback
stereo recording + stereo playback
And it will take a while.


On Jan 25, 8:31 pm, Robert Green <***@gmail.com> wrote:
> Would you be willing to rerun your tests at 44.1khz stereo?  I found
> it interesting that on my galaxy s, it used the same buffer size for
> mono and stereo which made it so that stereo would have half thelatencyof mono.  I'm wondering if that's also the case on the nexus
> s.  It's still horriblelatency, but half as horrible. :)
>
> On Jan 24, 7:42 am, Shijing Xian <***@gmail.com> wrote:
>
>
>
>
>
>
>
> > Hello Guys,
>
> > I think we had pretty much discussion or complain on this 45ms "low-
> >latency", but still we may be lack of a good overview on the android
> > audio system and may forget or miss some piece of important
> > information.
> > I am recently doing a experiment in Nexus S using this native audio
> > implementation for steaming, and get some interesting numbers, I hope
> > they can be usefully and spur more discussions.
>
> > I made a loopback application, which gets data from the recording
> > device, puts the data to a ring buffer, and playout. All the data
> > processing(if any) is using a framesize of 10ms audio.
>
> > Setups for the native audio implementation:
> > Recording: simple buffer queue, two buffers with size of 10ms audio
> > for each.
> > Playback:  buffer queue, two buffers with size of 10ms audio for each.
> > When the app is started, it will enqueue an empty 10ms buffer to the
> > recording queue, and 10ms data buffer to playout queue, to keep the
> > queues rolling. In the recording callback function, I copy the data to
> > the ring buffer and enqueue 10ms empty buffer; In the playout
> > callback, I pull 10ms data from ring buffer and enqueue it to the
> > playback (if no data, it will gets zero buffer, no waiting here).
> > It is pretty simple, isn't it? And my calculation for the delay is
> > something like this: recording delay + playout delay + ring buffer
> > delay
> > recording delay should be less than '10ms(buffering) + continuous
> > inputlatency(50ms)'
> > playback delay should be less than '10ms(buffering) + continuous
> > outputlatency(45ms)'
> > ring buffer delay is depending on the amount of recording data has
> > been put into the ring buffer, and the amount of data has been played.
> > It is scaling to the size of buffer in the ring, and the ideal case is
> > that we playback the data soon after it is recorded. This ring buffer
> > delay is quite depending on how the callbacks come.
> > So my expectation for the maximum delay is : 115 + ring buffer delay.
> > (Please correct me if I miss something in my model)
> > Surprisingly, I got a delay of 360ms for this loopback test, where the
> > ring buffer delay is about 100ms on average.
> > To clarify how I do the measurement, I use alatencyinstrument which
> > inputs a beep to the microphone while listening to the speaker, and it
> > reports the interval.
>
> > I understand the Nexus S device is not a lowlatencydevice, and it
> > may have a bigger continuous input/outputlatency, but 250ms in 10ms
> > buffering is definitely not acceptable to me. And the worst thing is
> > that I have no clue where is consuming the time.
> > Note, I can't tell how much the playback and recording is taking part
> > in this 250ms delay separately now, and I will investigate on this
> > later.
>
> > And another issue I want to address here is about the way how the
> > callbacks come. In my setup, I have two 10ms buffers for each
> > recording/playback queue, so I am expecting the callbacks to occur
> > every 10ms. But it is not true in my test, instead I usually get two
> > or three callbacks with the same timestamp every 20-30ms. The worst is
> > that at the beginning of the call I often get a stack of callbacks at
> > the same time, for example, 7 recording callbacks with the same
> > timestamp, then the next recording callback comes after 70ms. That may
> > be the reason why I have a 100ms ring buffer delay in my test, but I
> > have no idea what is causing it. Guys, any idea here?
>
> > To compare the performance, I run the same test but using AudioTrack
> > in Java. most of the setup is similar to the native audio
> > implementation, except that I use polling instead of callback.
> > Basically I run two aggressive threads to pull data from the recording
> > and put data for playback.
> > And the overall delay is 280ms.
> > Any comment here to explain why the Java implementation has less
> >latency?
>
> > While the similar application for iPhone is reporting a 60-70ms
> > overall delay.
>
> > Note, in the above test I am using 16000 as sampling rate for both
> > recording and playback.
>
> > I am not sure if any of you have already done a similar experiment
> > like what I did, any comments or suggestion will be appreciated. And I
> > do wonder if there any API for me to tweak the performance? Or any API
> > to get the hardware or os delay to get the accuratelatency? (Though I
> > have already looked at the openSLES 1.0.1, but can't find any).
>
> > The only thing makes me happen is that the native audio implementation
> > does take less CPU usage, and the audio is more reliable than using
> > AudioTrack.
>
> > Best Regards,
> > /Shijing

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
mic _
2011-01-27 12:34:19 UTC
Permalink
>> I found it interesting that on my galaxy s, it used the same buffer size
for mono and stereo

That's likely to be true for all Android phones using Qualcomm platforms.

Regarding stereo recording: some (maybe most?) current devices will just
record from a single mic and duplicate each sample to get a stereo signal.
Some might actually do stereo recording with two mics. But I guess it's
mainly the latency you're after here - not how the recording sounds.

/Michael

On Thu, Jan 27, 2011 at 12:54 PM, Shijing Xian <***@gmail.com>wrote:

> Thanks Robert, it is definitely interesting to know how much it
> improves the latency by using the stereo audio. It may likely match
> what Robert said, half of the latency of mono. But I am not sure if
> 44.1khz makes any difference, I tried the 44.1khz, 16khz, 8khz mono
> recording/playback before, all the tests showed a similar latency,
> unfortunately.
> There are also some other reasons why I prefer 16khz to others, the
> most important one is 16khz is usually supported by almost all Android
> devices.
> I will try the stereo recording and playback, and paste what I find
> here later. In order to identify the problem, I will try stereo
> recording and playback separately. So the combination will be:
> stereo recording + mono playback
> mono recording + stereo playback
> stereo recording + stereo playback
> And it will take a while.
>
>
> On Jan 25, 8:31 pm, Robert Green <***@gmail.com> wrote:
> > Would you be willing to rerun your tests at 44.1khz stereo? I found
> > it interesting that on my galaxy s, it used the same buffer size for
> > mono and stereo which made it so that stereo would have half thelatencyof
> mono. I'm wondering if that's also the case on the nexus
> > s. It's still horriblelatency, but half as horrible. :)
> >
> > On Jan 24, 7:42 am, Shijing Xian <***@gmail.com> wrote:
> >
> >
> >
> >
> >
> >
> >
> > > Hello Guys,
> >
> > > I think we had pretty much discussion or complain on this 45ms "low-
> > >latency", but still we may be lack of a good overview on the android
> > > audio system and may forget or miss some piece of important
> > > information.
> > > I am recently doing a experiment in Nexus S using this native audio
> > > implementation for steaming, and get some interesting numbers, I hope
> > > they can be usefully and spur more discussions.
> >
> > > I made a loopback application, which gets data from the recording
> > > device, puts the data to a ring buffer, and playout. All the data
> > > processing(if any) is using a framesize of 10ms audio.
> >
> > > Setups for the native audio implementation:
> > > Recording: simple buffer queue, two buffers with size of 10ms audio
> > > for each.
> > > Playback: buffer queue, two buffers with size of 10ms audio for each.
> > > When the app is started, it will enqueue an empty 10ms buffer to the
> > > recording queue, and 10ms data buffer to playout queue, to keep the
> > > queues rolling. In the recording callback function, I copy the data to
> > > the ring buffer and enqueue 10ms empty buffer; In the playout
> > > callback, I pull 10ms data from ring buffer and enqueue it to the
> > > playback (if no data, it will gets zero buffer, no waiting here).
> > > It is pretty simple, isn't it? And my calculation for the delay is
> > > something like this: recording delay + playout delay + ring buffer
> > > delay
> > > recording delay should be less than '10ms(buffering) + continuous
> > > inputlatency(50ms)'
> > > playback delay should be less than '10ms(buffering) + continuous
> > > outputlatency(45ms)'
> > > ring buffer delay is depending on the amount of recording data has
> > > been put into the ring buffer, and the amount of data has been played.
> > > It is scaling to the size of buffer in the ring, and the ideal case is
> > > that we playback the data soon after it is recorded. This ring buffer
> > > delay is quite depending on how the callbacks come.
> > > So my expectation for the maximum delay is : 115 + ring buffer delay.
> > > (Please correct me if I miss something in my model)
> > > Surprisingly, I got a delay of 360ms for this loopback test, where the
> > > ring buffer delay is about 100ms on average.
> > > To clarify how I do the measurement, I use alatencyinstrument which
> > > inputs a beep to the microphone while listening to the speaker, and it
> > > reports the interval.
> >
> > > I understand the Nexus S device is not a lowlatencydevice, and it
> > > may have a bigger continuous input/outputlatency, but 250ms in 10ms
> > > buffering is definitely not acceptable to me. And the worst thing is
> > > that I have no clue where is consuming the time.
> > > Note, I can't tell how much the playback and recording is taking part
> > > in this 250ms delay separately now, and I will investigate on this
> > > later.
> >
> > > And another issue I want to address here is about the way how the
> > > callbacks come. In my setup, I have two 10ms buffers for each
> > > recording/playback queue, so I am expecting the callbacks to occur
> > > every 10ms. But it is not true in my test, instead I usually get two
> > > or three callbacks with the same timestamp every 20-30ms. The worst is
> > > that at the beginning of the call I often get a stack of callbacks at
> > > the same time, for example, 7 recording callbacks with the same
> > > timestamp, then the next recording callback comes after 70ms. That may
> > > be the reason why I have a 100ms ring buffer delay in my test, but I
> > > have no idea what is causing it. Guys, any idea here?
> >
> > > To compare the performance, I run the same test but using AudioTrack
> > > in Java. most of the setup is similar to the native audio
> > > implementation, except that I use polling instead of callback.
> > > Basically I run two aggressive threads to pull data from the recording
> > > and put data for playback.
> > > And the overall delay is 280ms.
> > > Any comment here to explain why the Java implementation has less
> > >latency?
> >
> > > While the similar application for iPhone is reporting a 60-70ms
> > > overall delay.
> >
> > > Note, in the above test I am using 16000 as sampling rate for both
> > > recording and playback.
> >
> > > I am not sure if any of you have already done a similar experiment
> > > like what I did, any comments or suggestion will be appreciated. And I
> > > do wonder if there any API for me to tweak the performance? Or any API
> > > to get the hardware or os delay to get the accuratelatency? (Though I
> > > have already looked at the openSLES 1.0.1, but can't find any).
> >
> > > The only thing makes me happen is that the native audio implementation
> > > does take less CPU usage, and the audio is more reliable than using
> > > AudioTrack.
> >
> > > Best Regards,
> > > /Shijing
>
> --
> You received this message because you are subscribed to the Google Groups
> "android-ndk" group.
> To post to this group, send email to android-***@googlegroups.com.
> To unsubscribe from this group, send email to
> android-ndk+***@googlegroups.com<android-ndk%***@googlegroups.com>
> .
> For more options, visit this group at
> http://groups.google.com/group/android-ndk?hl=en.
>
>

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Shijing Xian
2011-01-27 15:39:05 UTC
Permalink
True, latency is my current main concern.

I am able to make the stereo playback working, but unfortunately have
an issue with the stereo recording. The audio from the stereo
recording doesnt sounds right, it sounds like the audio is played in a
wrong frequency, and there is lots of electronic noise.
So I printed out the first 10 data I got from the callback, it looks
quite strange to me:
01-27 16:24:09.710: in recording 406 406 532 532 133 133 -102 -102
01-27 16:24:09.710: in recording -193 145 335 63 -402 -131 -45 -284
01-27 16:24:09.710: in recording 822 822 441 441 70 70 54 54
01-27 16:24:09.710: in recording -80 -80 -176 -176 137 137 200 200
01-27 16:24:09.710: in recording -560 -560 -141 -141 265 265 281 281
01-27 16:24:09.710: in recording -227 -227 -284 -284 12 12 245 245
You can see that there are always more than 1 stack of data is not
duplicating each sample, and this happens every 5 callbacks. And all
the data doesnt look right, they are too big and varying too much.

Has any one ever tried stereo recording in Nexus S? I dont really want
to spend much time on investigating it here if the problem is in the
hardware.







01-27 16:24:09.066: DEBUG/---GIPS AD---(4982): in recording -567 -567
-328 -328 -178 -178 46 46



On Jan 27, 1:34 pm, mic _ <***@gmail.com> wrote:
> >> I found it interesting that on my galaxy s, it used the same buffer size
>
> for mono and stereo
>
> That's likely to be true for all Android phones using Qualcomm platforms.
>
> Regarding stereo recording: some (maybe most?) current devices will just
> record from a single mic and duplicate each sample to get a stereo signal.
> Some might actually do stereo recording with two mics. But I guess it's
> mainly the latency you're after here - not how the recording sounds.
>
> /Michael
>
> On Thu, Jan 27, 2011 at 12:54 PM, Shijing Xian <***@gmail.com>wrote:
>
>
>
>
>
>
>
> > Thanks Robert, it is definitely interesting to know how much it
> > improves the latency by using the stereo audio. It may likely match
> > what Robert said, half of the latency of mono. But I am not sure if
> > 44.1khz makes any difference, I tried the 44.1khz, 16khz, 8khz mono
> > recording/playback before, all the tests showed a similar latency,
> > unfortunately.
> > There are also some other reasons why I prefer 16khz to others, the
> > most important one is 16khz is usually supported by almost all Android
> > devices.
> > I will try the stereo recording and playback, and paste what I find
> > here later. In order to identify the problem, I will try stereo
> > recording and playback separately. So the combination will be:
> > stereo recording + mono playback
> > mono recording + stereo playback
> > stereo recording + stereo playback
> > And it will take a while.
>
> > On Jan 25, 8:31 pm, Robert Green <***@gmail.com> wrote:
> > > Would you be willing to rerun your tests at 44.1khz stereo?  I found
> > > it interesting that on my galaxy s, it used the same buffer size for
> > > mono and stereo which made it so that stereo would have half thelatencyof
> > mono.  I'm wondering if that's also the case on the nexus
> > > s.  It's still horriblelatency, but half as horrible. :)
>
> > > On Jan 24, 7:42 am, Shijing Xian <***@gmail.com> wrote:
>
> > > > Hello Guys,
>
> > > > I think we had pretty much discussion or complain on this 45ms "low-
> > > >latency", but still we may be lack of a good overview on the android
> > > > audio system and may forget or miss some piece of important
> > > > information.
> > > > I am recently doing a experiment in Nexus S using this native audio
> > > > implementation for steaming, and get some interesting numbers, I hope
> > > > they can be usefully and spur more discussions.
>
> > > > I made a loopback application, which gets data from the recording
> > > > device, puts the data to a ring buffer, and playout. All the data
> > > > processing(if any) is using a framesize of 10ms audio.
>
> > > > Setups for the native audio implementation:
> > > > Recording: simple buffer queue, two buffers with size of 10ms audio
> > > > for each.
> > > > Playback:  buffer queue, two buffers with size of 10ms audio for each.
> > > > When the app is started, it will enqueue an empty 10ms buffer to the
> > > > recording queue, and 10ms data buffer to playout queue, to keep the
> > > > queues rolling. In the recording callback function, I copy the data to
> > > > the ring buffer and enqueue 10ms empty buffer; In the playout
> > > > callback, I pull 10ms data from ring buffer and enqueue it to the
> > > > playback (if no data, it will gets zero buffer, no waiting here).
> > > > It is pretty simple, isn't it? And my calculation for the delay is
> > > > something like this: recording delay + playout delay + ring buffer
> > > > delay
> > > > recording delay should be less than '10ms(buffering) + continuous
> > > > inputlatency(50ms)'
> > > > playback delay should be less than '10ms(buffering) + continuous
> > > > outputlatency(45ms)'
> > > > ring buffer delay is depending on the amount of recording data has
> > > > been put into the ring buffer, and the amount of data has been played.
> > > > It is scaling to the size of buffer in the ring, and the ideal case is
> > > > that we playback the data soon after it is recorded. This ring buffer
> > > > delay is quite depending on how the callbacks come.
> > > > So my expectation for the maximum delay is : 115 + ring buffer delay.
> > > > (Please correct me if I miss something in my model)
> > > > Surprisingly, I got a delay of 360ms for this loopback test, where the
> > > > ring buffer delay is about 100ms on average.
> > > > To clarify how I do the measurement, I use alatencyinstrument which
> > > > inputs a beep to the microphone while listening to the speaker, and it
> > > > reports the interval.
>
> > > > I understand the Nexus S device is not a lowlatencydevice, and it
> > > > may have a bigger continuous input/outputlatency, but 250ms in 10ms
> > > > buffering is definitely not acceptable to me. And the worst thing is
> > > > that I have no clue where is consuming the time.
> > > > Note, I can't tell how much the playback and recording is taking part
> > > > in this 250ms delay separately now, and I will investigate on this
> > > > later.
>
> > > > And another issue I want to address here is about the way how the
> > > > callbacks come. In my setup, I have two 10ms buffers for each
> > > > recording/playback queue, so I am expecting the callbacks to occur
> > > > every 10ms. But it is not true in my test, instead I usually get two
> > > > or three callbacks with the same timestamp every 20-30ms. The worst is
> > > > that at the beginning of the call I often get a stack of callbacks at
> > > > the same time, for example, 7 recording callbacks with the same
> > > > timestamp, then the next recording callback comes after 70ms. That may
> > > > be the reason why I have a 100ms ring buffer delay in my test, but I
> > > > have no idea what is causing it. Guys, any idea here?
>
> > > > To compare the performance, I run the same test but using AudioTrack
> > > > in Java. most of the setup is similar to the native audio
> > > > implementation, except that I use polling instead of callback.
> > > > Basically I run two aggressive threads to pull data from the recording
> > > > and put data for playback.
> > > > And the overall delay is 280ms.
> > > > Any comment here to explain why the Java implementation has less
> > > >latency?
>
> > > > While the similar application for iPhone is reporting a 60-70ms
> > > > overall delay.
>
> > > > Note, in the above test I am using 16000 as sampling rate for both
> > > > recording and playback.
>
> > > > I am not sure if any of you have already done a similar experiment
> > > > like what I did, any comments or suggestion will be appreciated. And I
> > > > do wonder if there any API for me to tweak the performance? Or any API
> > > > to get the hardware or os delay to get the accuratelatency? (Though I
> > > > have already looked at the openSLES 1.0.1, but can't find any).
>
> > > > The only thing makes me happen is that the native audio implementation
> > > > does take less CPU usage, and the audio is more reliable than using
> > > > AudioTrack.
>
> > > > Best Regards,
> > > > /Shijing
>
> > --
> > You received this message because you are subscribed to the Google Groups
> > "android-ndk" group.
> > To post to this group, send email to android-***@googlegroups.com.
> > To unsubscribe from this group, send email to
> > android-ndk+***@googlegroups.com<android-ndk%***@googlegr oups.com>
> > .
> > For more options, visit this group at
> >http://groups.google.com/group/android-ndk?hl=en.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Shijing Xian
2011-02-02 10:56:11 UTC
Permalink
Just give an short update on my tests.

My previous tests results are not all right, mainly due to the noisy
instrument I used for measurement. In order to increase accuracy, I
have to add noise suppression to take away the noise.

The new measurement shows that
Using 44.1KHz does have about 60ms less latency than using 16KHz.
Stereo playback seems to help, but not much.
Stereo recording is not really my interest, because I only use 8K and
16KHz codec.

In general, the Native API has similar latency to the Java TrackAudio,
(maybe it is slightly better if I can trust the measurement).




On Jan 27, 4:39 pm, Shijing Xian <***@gmail.com> wrote:
> True,latencyis my current main concern.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Ross Bencina
2010-12-21 11:54:46 UTC
Permalink
On Dec 20, 8:30 pm, Robert Green <***@gmail.com> wrote:
> Low latency should be defined as an amount which allows for accurate
> and believable synchronization with video.

Robert, I disagree with this point but I think we agree in general.
Depending on which standards body you believe, acceptable A/V sync
latency is between 50ms and 140ms (the most recent standards that I
know of relate to Digital Broadcast TV). I believe frame rate is not
really relevant -- it's all about ear-eye perception and "lip sync"

As a long time professional audio developer I think it's an absolute
disgrace that Google are equating 45ms with "low latency." I strongly
disagree that the A/V sync requirement has anything to do with "low
latency" -- it is a requirement for sure, it should be a requirement
for _all_ Android devices, not those that claim "low latency". Heck,
every system I've programmed in the last 10 years could achieve 45ms
playback latency.

> Windows Audio (NOT DirectSound) has a latency lower than 45ms on my PC

WMME is legacy, even on Windows XP. The main latency is caused by the
"kernel mixer" presumably the equivalent of the AudioFlinger on
Android. Kernel mixing on Windows 7 is much lower latency.

> DirectSound often is around 20ms for most AC97 hardware (so I've
>read), but is still too high for real time audio manipulation.

Agreed. WASAPI on Windows Vista and Windows 7 achieves lower than 4ms.

> Core Audio (apple) is advertised at being able to be .2 to 1.5ms!

Yes. Well below 3ms on iOS too in my experience.

> ASIO drivers for audio production usually are below 11ms, typically
> hitting 5ms.

Agreed.


There are many applications that have more stringent latency
requirements than A/V synchronisation. Three use-cases I can think of
right now are:

1. Triggering sound from a user input. This could be in games where
the game reacts to gameplay, or a musical instrument where the user
"plays" the instrument and expects the instrument to respond
immediately.

2. Processing audio input into audio output "audio effects processors"
-- plugging your guitar in to your computer to use it as an effects
box for example.

3. Providing an input or output endpoint for an audio session where
the other end of the connection is somewhere else on the network
(Networked telepresence and musical performance come to mind).


>From memory, the upper limit of the perceptual threshold for fused
actuation->result for music performance is around 8ms. That means any
audio effect device capturing audio from a microphone and playing out
out a speaker needs to perform the whole round trip capture->process-
>playback in less than 8ms to be considered musically transparent (you
don't notice there's a computer involved). You will find these kind of
latencies on commercial electronic musical instruments where there is
a mechanical actuator and an analog audio output. Latencies for
digital audio effects units (where the input is an analog audio signal
and the output is also an analog signal) can (and should) be lower.
There was a paper published a few years ago that identified 25ms as
the upper threshold for round-trip audio transport for networked
musical performance (that's capture+network+playback latency).

"low latency" already has a well defined meaning in the audio software
world. On Linux for example, people have been working for years to get
single-direction latency below 2ms for audio applications (see for
example http://lowlatency.linuxaudio.org/). On MacOS and iOS these
kind of latencies are easily achievable. On Windows 7, < 4ms is also
easily achievable (on older versions of Windows it is achievable if
your code talks direct to the WDM layer).

In my experience, based on customer feedback, music and audio
processing software users now consider anything more than 5ms latency
to be "high latency".

To cite another data point: IEEE P802.1p QoS standard stipulates that
VOIP traffic be routed with less than 10ms latency. I don't personally
consider that low latency, but it gives you an idea of what the IEEE
think.

If Android was "just a phone OS" then I might be willing to let all
this slide, but the reality is that Android will be supporting a lot
more than Phones in the coming year and applications that meet or
surpass what is possible on the desktop should be enabled. And
besides, iOS can already deliver latencies below 2ms so the bar has
already been set for what "low latency" is.

If anyone is interested I can look up the references for this stuff.

And if anyone knows who to ping at Google to complain please let me
know.

Thanks!

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
shap
2011-04-18 07:05:29 UTC
Permalink
If Android was "just a phone OS" then I might be willing to let all

this slide...


Agree strongly that it isn't. But even if it were, the fact that competing
phone and tablet platforms are delivering dramatically better results is
cause for concern.

All that aside, failing to meet ASIO latency (which is more like 3ms for
firewire on a Core 2 Duo) on computationally comparable core hardware is
just plain *embarrassing*.

--
You received this message because you are subscribed to the Google Groups "android-ndk" group.
To post to this group, send email to android-***@googlegroups.com.
To unsubscribe from this group, send email to android-ndk+***@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/android-ndk?hl=en.
Loading...