Created page with " == Audio Conference - Monday October 21st 2013 - Edinburgh == Attendees :- ----------------------------------------------------------------------- Mark Brown, Liam Gir..."
New page
== Audio Conference - Monday October 21st 2013 - Edinburgh ==
Attendees :-
-----------------------------------------------------------------------
Mark Brown, Liam Girdwood, Pierre Bossart, Patrick Lai, Lars-Peter Clausen,
Eric Laurent, David Henningsson, Takashi Iwai, Colin Guthrie, Tanu Kaskinen,
Arun Raghavan, Paul Handrigan, Daniel Mack, Dylan Reid, Sven Neuman,
Dmimtris Papastamos, Charles Keepax, Nariman Poushin, João Paulo Rechi Vita,
Alexander Patrakov, Micheal Wu, Lennart Poettering
Pulseaudio route management and UCM
-----------------------------------
There was interest for shared policy configuration file format across
different systems. No such sharing has been planned. Tanu's initial reaction
is that it's not feasible, but he is open to discussing it.
What to do if there are multiple paths from A to B? One will be chosen,
either arbitrarily or based on connection requirements or prioritization.
How to control effect offloading? Tanu's vague idea is that if some kind of
filtering is required/preferred, it's expressed as a requirement for a
connection, and the node implementation will then enable the details of
enabling the effect if such effect is available. Tanu doesn't know how well
that would work in practice - and probably unaware of many details that
would have to be taken into account.
UCM being used by some people but missing good documentation. Liam to write
documentation and use examples from Arun as "gold standard" example
configurations. More UCM configs to go online in alsa-lib repo.
UCM configs sometimes need updated for new kernel releases as Kcontrol strings
have changed. Suggestions made for additional UUID for each kcontrol to lessen UCM
config name changes (e.g. vendor-part-reg-shift-mask could be used as ASoC
kcontrol UUID). Naming changes also effect HDA kcontrols, but no HDA UCM configs
yet ? kcontrol UUID harder for DSP and HDA.
Arun reported UCM issue with combination of devices. e.g. Enable dev 1 & 2 then
disable dev 1 & 2. Maybe virtual device that combines devices ??
#include for UCM - should be easy to add. Simplifies config for family of similar
devices like Panda, PandaES etc
Device Names in UCM names. There are some standard names in UCM header but need
more. Liam to add more names.
Arbitration, UCM does not do arbitration but leaves that up to sound server.
Controls names need more standardisation. Kcontrol naming mainly based on HW
specified within data sheets. Difficult for userspace to cope with great degree
of naming variations and provide enough information for userspace to know
which volume is master etc.
Audio Delay and Timing granularity
----------------------------------
Some DMA drivers very bad at providing position information used to work out
buffer positions. Difficult for userspace to render audio and wake up correctly
with such HW unless we can provide mechanism for userspace to know the DMA position
accuracy. IOCTL suggested, info is in driver so easy to expose.
There already is the SNDRV_PCM_INFO_BATCH flag which is supposed to indicated that
the granularity is period-size-ish. It makes sense for applications
like pulse-audio to check whether this flag is set or not.
Hostless Audio
--------------
Hostless should now use kcontrol to join the graph links to enable or disable and
not hostless PCM device (old Android hack).
e.g. FM playback loopback to codec, Modem PCM handling.
i.e. CODEC <-> CODEC type. Need mixer to expose set config changes as link
config is static e.g. WB/NB configs.
Audio Offloading for Android
----------------------------
Avail in KK, supports > 1 compressed devices. No guarantee on no dropped samples
when changing compressed to PCM and vice versa (or enabling effects ????)
Not available yet when in use with video, no tech reason to stop this.
AV sync good. Some effects can be offloaded and some not.
Compressed API
--------------
Intel to upstream remaining fixes. AR - Liam, Vinod.
Stable updates with no users. 3.10 is stable but we can changes core API since
there are no users.
Android Audio Effects Framework Modifications
---------------------------------------------
Composite with HW and SW representations. Supports 3rd party GUI controls
but they must use the effect API and not bypass with private IOCTLs.
How do we do library, like tiny effects ? API exists in use by Factory,
can we extend to alsa ?
Nobody using compressed and effects outside of Android atm.
24 Bit HiRes Audio for Android
------------------------------
How do handle 24 bit audio at fast rates e.g. 192kHz. Being pushed by market.
Flac suggested as source media and not PCM.
We could use offload, on roadmap for Hires audio on android PCM,
maybe float instead of 24 bit. Need for 192kHz, Probably not use effects.
Effect Support in ALSA
----------------------
Several proposals for APIs.
Control device type, do we need new control or use existing CTL ? Per card or
per DSP / CODEC. ? Not to use control API for everything , could confuse some
userspace. Get/set could be any control type, not just binary ? Binary blobs
used to hide workings of effects.
Topology Query and edit - query the graph from HDA/DAPM -
used to establish connections and end points. per card based IOCTL. Could be new
alsamixer to allow users to control volumes.
Media controller API. How do we match to ALSA. Need non LGPL library for android.
Missing certain features, someone needs to write code. MC API is Video oriented,
library is just wrappers around IOCTLs, 1 man week of simple work to implement
non LGPL version.
Removal/creation - kernel can provide any type of information to user space key values.
Document new keys, generic structure. Channel mapping possible using tuples.
Need to support dynamic entity information. Handles static pipeline well,
and can also expose dynamic with some complexity (with more entities).
Abstract away sink+sources in userspace library, maybe too complex.
Trivial to use UUIDs for dynamic entities. Start with prototype and see what happens.
Action on Intel/Vinod ?
Topology 2 NODES and EFFECT IOCTLS could be replaced with media controller as MediaC
provides good match.
Sending/Receiving Binary data to drivers
----------------------------------------
Use cases that go above 256/512 ? byte limit for binary controls.
Extend for larger ? Do something about it. Action for ????? Tiwai/Mark ????
MMAP for some parameters. Visualizer, peak meters. ? Multiple of 4k pages.
Can be used for DMA buffer positions on IPC IRQ based DSPs.
Avoid copying large amounts of data. trade off for DSP. Usable for DEBUG too.
Prototype and see what happens - Who ??
Expose API for sending large chunks of data e.g. 2MB. Nice for generic code to send
this data (no special tooling) Needs new API for large data chunks. What about write().
Patrick to prototype some code.
DSP resource management
-----------------------
Do we need pre-trigger to work out MIPS/MHz + buffer resources for DSP
resources. Tables used atm for most DSPs.
Something for policy or UCM to specify ? Could change for different FWs.
Peak and average loading gives difficulty.
Might not get feedback, latency in the pipe and need to agree on units of
measurements. Complex problem. Estimate worst case.
OMAP4 example uses DAPM, expose reource usage numbers, expose capability limits.
Liam to publish clock resource management logic for Haswell DSP. Can be used as
starting point for investigations into feasibility etc.
HDMI - Hotplug
--------------
Can cause changes to audio path. Do we need to tell userspace card has changed ?
Specific for ELD or generic problem ?
Suggestion to remove and then insert card not usable
Signal for monitor re-pluging, ELD not available right away.
How to sync ELD information, ELD time out signal ? DSPs can change capabilities,
ELDS can be garbage. PA maybe ask to quickly for ELD info, expose new jack for HDA
when ELD is known. Generic event to tell userspace that card has changed.
HDA/MIMI specific solution.
Generic API to talk between audio and DRM devices. Bespoke atm. Code lives in
the wrong place atm. Need to sync. Code is already upstream.
David to prototype a solution based on deferring the probe until ELD is ready.
Mic/LED hotkey
--------------
Few different scenarios - from slides. Multiple cards and multiple mics.
Which Mic will LED represent ? card or system or other.
We'll not use the LED class because it does not buy us anything: LED
class is writable for root only, and PA does not run as root.
Nobody seemed to have a firm convincing argument in any direction. As such,
keeping the current behaviour (reflect mute state of internal card) will be
kept for now.
Alexander suggested looking at what NetworkManager does with the wifi-off
indicator in the presence of multiple wireless devices with
(potentially) their own wifi-off switches.
Performance - Use case startup
------------------------------
Discussing optimisation from DPCM - users QC, TI and Intel. DSP can route to
multiple endpoints from a single PCM.
Startup call back -> PCM -> FE -> BE. Then hw_params(), followed by prepare()
and trigger(). Cold start latency 113ms since some slow ops on QDSP are serialised.
Solution to shift some prepare() work into DAPM. DAPM ops are tun in parrallel
and a workq.
Compressed offload
------------------
Video/Audio offload - pass audio timestamp. Metadata for QDSP stored at front of
compressed data by QC. Transport stream already supported by compressed API.
Timestamps may differ in DSP implementations. AV sync topic requested by V4Linux
for long time. Only discussed for PCM not offload, need TS for offload and support
of V4Linux guys.
Compressed offload transcoding. - How do we specify this transform. New API for
userspace to spec output of decoder. All sorts of routing combinations are possible
making implementation very difficult. Query the the BE connected to the FE, we may
be able to call existing PCM API's on the BE PCM. Patrick to prototype some
code.
Firmware based Kcontrols and Graphs
-----------------------------------
Patches have been developed to attach Kcontrol and Graph meta data to DSP FW.
This allows single DSP driver to be used with multiple FWs without modification.
Patch V1 reviewed last year before TI layoff and we should be ready for V2 before
end of year.
Consists of kernel patch and userspace tool to generate the control and graph
data. Current userspace tool is needing work to improve config generation.
There is no config file format at present. Paul and some Wolfson guys have
volunteered to assist in this effort. Code currently works on OMAP4 platforms.
Kernel code is here (top 5 patches) :-
https://git.kernel.org/cgit/linux/kernel/git/lrg/asoc.git/log/?h=topic/firmware
Userspace tool is here :-
git@gitorious.org:omap-audio/asoc-fw.git
Clean up ALSA code
------------------
Move plugins to their projects - move ffmpeg plugin to ffmpeg project, etc.
Drivers, requests to move to device probe model, away from alsa model.
Need to reg dev files at end of driver creation.
Input susbsys uses possible method that could be copied. sysfs files need sync
to avoid races for creation. Agreement.
Process, How do we work
-----------------------
Bug reports. Who looks at bugs. No takers.
New mailing lists. Use filters to hide emails.
Misc
----
Softvol, add SND_CTL for no softvol in ctl open. Use alsa conf file or ctl.
Arun encouraged to revive the email thread on the topic.
Possible bug where specially crafted javascript can set PA volume to 100%.
Defined pulseaudio behviour (Colin), related to flat volumes.
New ASoC drivers must now use generic DMA engine PCM driver.
Non atomic trigger requirments from DSP drivers. DSPs mostly sleep when performing
IPC and need a trigger call that can sleep. Patrick will investigate.
Propagate errors from BEs to FEs. Patrick to investigate. This will also include
reporting underruns too.
System restart for DSP FW crashes. This requires reloading all FW and trying to
recover all active streams. Patrick to investigate.
Lennart asked about the support of audio "mute" via systemd session management,
like input and drm drivers already do. Takashi took a look at input evdev code,
and actually it's revoking instead of muting. The advantage of revoking over
muting is that it's simpler and doesn't need the capability check, as far as he
read through discussion threads. But Takashi is not sure whether revoking
behavior is good for PA at all.